Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

default · ~99.6k rows (showing the first 120k)

Split (2)

train · ~72.4k rows (showing the first 93.1k)

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
1709.07224	2	1709	2018-07-18T08:39:08	Local Communication Protocols for Learning Complex Swarm Behaviors with Deep Reinforcement Learning	[ "cs.MA", "cs.AI", "cs.LG", "eess.SY", "stat.ML" ]	Swarm systems constitute a challenging problem for reinforcement learning (RL) as the algorithm needs to learn decentralized control policies that can cope with limited local sensing and communication abilities of the agents. While it is often difficult to directly define the behavior of the agents, simple communication protocols can be defined more easily using prior knowledge about the given task. In this paper, we propose a number of simple communication protocols that can be exploited by deep reinforcement learning to find decentralized control policies in a multi-robot swarm environment. The protocols are based on histograms that encode the local neighborhood relations of the agents and can also transmit task-specific information, such as the shortest distance and direction to a desired target. In our framework, we use an adaptation of Trust Region Policy Optimization to learn complex collaborative tasks, such as formation building and building a communication link. We evaluate our findings in a simulated 2D-physics environment, and compare the implications of different communication protocols.	cs.MA	cs	Local Communication Protocols for Learning Complex Swarm Behaviors with Deep Reinforcement Learning Maximilian Hüttenrauch1, Adrian Šošić2, and Gerhard Neumann1 1 School of Computer Science, University of Lincoln, Lincoln, UK {mhuettenrauch,gneumann}@lincoln.ac.uk 2 Department of Electrical Engineering, Technische Universität Darmstadt, Darmstadt, Germany [email protected] Abstract. Swarm systems constitute a challenging problem for rein- forcement learning (RL) as the algorithm needs to learn decentralized control policies that can cope with limited local sensing and communica- tion abilities of the agents. While it is often difficult to directly define the behavior of the agents, simple communication protocols can be defined more easily using prior knowledge about the given task. In this paper, we propose a number of simple communication protocols that can be exploited by deep reinforcement learning to find decentralized control policies in a multi-robot swarm environment. The protocols are based on histograms that encode the local neighborhood relations of the agents and can also transmit task-specific information, such as the shortest dis- tance and direction to a desired target. In our framework, we use an adaptation of Trust Region Policy Optimization to learn complex collab- orative tasks, such as formation building and building a communication link. We evaluate our findings in a simulated 2D-physics environment, and compare the implications of different communication protocols. 1 Introduction Nature provides many examples where the performance of a collective of limited beings exceeds the capabilities of one individual. Ants transport prey of the size no single ant could carry, termites build nests of up to nine meters in height, and bees are able to regulate the temperature of a hive. Common to all these phenomena is the fact that each individual has only basic and local sensing of its environment and limited communication capabilities to its neighbors. Inspired by these biological processes, swarm robotics [4,1,5] tries to emulate such complex behavior with a collective of rather simple entities. Typically, these robots have limited movement and communication capabilities and can sense only a local neighborhood of their environment, such as distances and bearings to neighbored agents. Moreover, these agents have limited memory systems, such that the agents can only access a short horizon of their perception. As a consequence, the design of control policies that are capable of solving complex cooperative tasks becomes a non-trivial problem. 2 Hüttenrauch et al. In this paper, we want to learn swarm behavior using deep reinforcement learning [21,17,22,23,10] based on the locally sensed information of the agents such that the desired behavior can be defined by a reward function instead of hand-tuning controllers of the agents. Swarm systems constitute a challenging problem for reinforcement learning as the algorithm needs to learn decentralized control policies that can cope with limited local sensing and communication abilities of the agents. Most collective tasks require some form of active cooperation between the agents. For efficient cooperation, the agents need to implement basic communi- cation protocols such that they can transmit their local sensory information to neighbored agents. Using prior knowledge about the given task, simple commu- nication protocols can be defined much more easily than directly defining the behavior. In this paper, we propose and evaluate several communication proto- cols that can be exploited by deep reinforcement learning to find decentralized control policies in a multi robot swarm environment. Our communication protocols are based on local histograms that encode the neighborhood relation of an agent to other agents and can also transmit task- specific information such as the shortest distance and direction to a desired target. The histograms can deal with the varying number of neighbors that can be sensed by a single agent depending on its current neighborhood configura- tion. These protocols are used to generate high dimensional observations for the individual agents that is in turn exploited by deep reinforcement learning to ef- ficiently learn complex swarm behavior. In particular, we choose an adaptation of Trust Region Policy Optimization [21] to learn decentralized policies. In summary, our method addresses the emerging challenges of decentralized swarm control in the following way: 1. Homogeneity: explicit sharing of policy parameters between the agents 2. Partial observability: efficient processing of action-observation histories through windowing and parameter sharing 3. Communication: usage of histogram-based communication protocols over simple features To demonstrate our approach, we formulate two cooperative learning tasks in a simulated swarm environment. The environment is inspired by the Colias robot [2], a modular platform with two wheel motor-driven movement and various sensing systems. Paper Outline In Section 2, we review the concepts of Trust Region Policy Opti- mization and describe our problem domain. In Section 3, we show in detail how we tackle the challenges of modeling observations and the policy in the partially observable swarm context, and how to adapt Trust Region Policy Optimization to our setup. In Section 4, we present the model and parameters of our agents and introduce two tasks on which we evaluate our proposed observation models and policies. Local Communication Protocols 3 2 Background In this section, we provide a short summary of Trust Region Policy Optimization and formalize our learning problem domain. 2.1 Trust Region Policy Optimization achieved over a certain period of time. Trust Region Policy Optimization (TRPO) is an algorithm to optimize control policies in single-agent reinforcement learning problems [21]. These problems are formulated as Markov decision processes (MDP) which are compactly written as a tuple (cid:104)S,A, P, R, γ(cid:105). In an MDP, an agent chooses an action a ∈ A via some policy π(a s), based on its current state s ∈ S, and progresses to state ∈ S according to a transition function P (s(cid:48) s(cid:48) s, a). After each step, the agent is the expected cumulative reward E[(cid:80)∞ assigned a reward r = R(s, a), provided by a reward function R which judges the quality of its decision. The goal of the agent is to find a policy which maximizes k=t γk−tR(sk, ak)], discounted by factor γ, In TRPO, the policy is parametrized by a parameter vector θ containing weights and biases of a neural network. In the following, we denote this parame- terized policy as πθ. The reinforcement learning objective is expressed as finding a new policy that maximizes the expected advantage function of the current pol- , where A is an estimate of the advantage func- old(s). tion of the current policy πold which is defined as A(s, a) = Qπold(s, a)− V π Herein, state-action value function Qπold (s, a) is typically estimated by a single trajectory rollout while for the value function V πold (s) rather simple baselines are used that are fitted to the monte-carlo returns. The objective is to be maxi- mized subject to a fixed constraint on the Kullback-Leibler (KL) divergence of the policy before and after the parameter update, which ensures the updates to the new policy's parameters θ are bounded, in order to avoid divergence of the learning process. The overall optimization problem is summarized as icy, i.e., J TRPO = E(cid:104) πθ A(s, a) (cid:105) πθold (cid:20) πθ (cid:21) E θ maximize subject to E[DKL(πθoldπθ)] ≤ δ. A(s, a) πθold The problem is approximately solved using the conjugate gradient optimizer after linearizing the objective and quadratizing the constraint. 2.2 Problem Domain Building upon the theory of single-agent reinforcement learning, we can now formulate the problem domain for our swarm environments. Because of their limited sensory input, each agent can only obtain a local observation o from the vicinity of its environment. We formulate the swarm system as a swarm MDP (see [24] for a similar definition) which can be seen as a special case of a decentralized 4 Hüttenrauch et al. partially observed Markov decision process (Dec-POMDP) [20]. An agent in the swarm MDP is defined as a tuple A = (cid:104)S,O,A, O(cid:105), where, S is a set of local states, O is the space of local observations, and A is a set of local actions for each agent. The observation model O(os, i) defines the observation probabilities for agent i given the global state s. Note that the system is invariant to the order of the agents, i.e., given the same local state of two agents, the observation probabilities will be the same. The swarm MDP is then defined as (cid:104)N,E, A, P, R(cid:105), where N is the number of agents, E is the global environment state consisting of all local states S N of the agents and possibly of additional states of the environment, and P : S N ×S N ×AN → [0,∞) is the transition density function. Each agent maintains a truncated history hi t) of the current and past observations oi ∈ O and actions ai ∈ A of length η. All swarm agents are assumed to be identical and therefore use the same distributed policy π (now defined as π(a h)) which yields a sample for the action of each agent given its current history of actions and observations. The reward function R of the swarm MDP depends on the global state and, optionally, all actions of the swarm agents, i.e., R : S N ×AN → R. Instead of considering only one single agent, we consider multiple agents of the same type, which interact in the same environment. The global system state is in this case comprised of the local states of all agents and additional attributes of the environment. The global task of the agents is encoded in the reward function R(s, a), where we from now on write a to denote the joint action vector of the whole swarm. t−η+1, . . . , ai t−1, oi t = (ai t−η, oi 2.3 Related Work A common approach to program swarm robotic systems is by extracting rules from the observed behavior of their natural counterparts. Kube et al [13], for example, investigate the cooperative prey retrieval of ants to infer rules on how a swarm of robots can fulfill the task of cooperative box-pushing. Similar work can be found e.g. in [16], [12], [19]. However, extracting these rules can be tedious and the complexity of the tasks that we can solve via explicit programming is limited. More examples of rule based behavior are found in [5] where a group of swarming robots transports an object to a goal. Further comparable work can be found in [6] for aggregation, [18] for flocking, or [9] for foraging. In deep RL, currently, there are only few approaches tackling the multi-agent problem. One of these approaches can be found in [15], where the authors use a variation of the deep deterministic policy gradient algorithm [14] to learn a centralized Q-function for each policy, which, as a downside, leads to a linear increase in dimensionality of the joint observation and action spaces therefore scales poorly. Another algorithm, tackling the credit assignment problem, can be found in [8]. Here, a baseline of other agents' behavior is subtracted from a centralized critic to reason about the quality of a single agent's behavior. However, this approach is only possible in scenarios with discrete action spaces since it requires marginalization over the agents' action space. Finally, a different line of work concerning the learning of communication models between agents can be found in [7]. Local Communication Protocols 5 d (b) range histogram d (a) local agent config- uration φ bearing (c) togram his- φ (d) joint histogram Fig. 1: This Figure shows an illustration of the histogram-based observation model. Figure 1a shows an agent in the center of a circle whose neighborhood relations are to be captured by the histogram representation. The shaded green area is highlighted as a reference for Figures 1c and 1d. Figure 1b hereby shows the one dimensional histogram of agents over the neighborhood range d into four bins, whereas Figure 1c shows the histogram over the bearing angles φ into eight bins. Figure 1d finally shows the two dimensional joint histogram over range and bearing. 3 Multi-Agent Learning with Local Communication Protocols In this section, we introduce different communication protocols based on neigh- borhood histograms that can be used in combination to solve complex swarm behaviors. Our algorithm relies on deep neural network policies of special ar- chitecture that can exploit the structure of the high-dimensional observation histories. We present this network model and subsequently discuss small adap- tations we had to make to the TRPO algorithm in order to apply it to this cooperative multi-agent setting. 3.1 Communication Protocols Our communication protocols are based on histograms that can either encode neighborhood relations or distance relations to different points of interest. Neighborhood Histograms The individual agents can observe distance and bearing to neighbored agents if they communicate with this agent. We assume that the agents are constantly sending a signal, such that neighbored agents can localize the sources. The arising neighborhood configuration is an important source of information and can be used as observations of the individual agents. One of the arising difficulties in this case is to handle changing number of neighbors which would result in a variable length of the observation vector. Most policy representations, such as neural networks, expect a fixed input dimension. One possible solution to this problem is to allocate a fixed number of neighbor relations for each agent. If an agent experiences fewer neighborhood relations, 6 Hüttenrauch et al. standard values could be used such as a very high distance and 0 bearing. How- ever, such an approach comes with several drawbacks. First of all, the size of the resulting representation scales linearly with the number of agents in the system and so does the number of parameters to be learned. Second, the execution of the learned policy will be limited to scenarios with the exact same number of agents as present during training. Third, a fixed allocation of the neighbor relation in- evitably destroys the homogeneity of the swarm, since the agents are no longer treated interchangeably. In particular, using a fixed allocation rule requires that the agents must be able to discriminate between their neighbors, which might not even be possible in the first place. To solve these problems, we propose to use histograms over observed neigh- borhood relations, e.g., distances and bearing angles. Such a representation in- herently respects the agent homogeneity and naturally comes with a fixed di- mensionality. Hence, it is the canonical choice for the swarm setting. For our experiments, we consider two different types of representations: 1) concatenated one-dimensional histograms of distance and bearing and 2) multidimensional histograms. Both types are illustrated in Figure 1. The one-dimensional repre- sentation has the advantage of scalability, as it grows linearly with the number of features. The downside is that potential dependencies between the features are completely ignored. Shortest Path Partitions In many applications, it is important to transmit the location of a point of interest to neighbored agents that can currently not observe this point due to their limited sensing ability. We assume that an agent can observe bearing and distance to a point of interest if it is within its communication radius. The agent then transmits the observed distance to other agents. Agents that can not see the point of interest might in this case observe a message from another agent containing the distance to the point of interest. The distance of the sending agent is added to the received distance to obtain the distance to the point of interest if we would use the sending agent as a via point. Each agent might now compute several of such distances and transmits the minimum distance it has computed to indicate the length of the shortest path it has seen. The location of neighbored agents including their distance of the shortest path information is important knowledge for the policy, e.g. for navigating to the point of interest. Hence, we adapt the histogram representation. Each partition now contains the minimum received shortest path distance of an agent that is located in this position. 3.2 Weight Sharing for Policy Networks The policy maps sequences of past actions and observations to a new action. We use histories of a fixed length as input to our policy and a feed-forward deep neural network as architecture. To cope with such high input dimensionality, we Local Communication Protocols 7 propose a weight sharing approach. Each action-observation pair in an agent's history is first processed independently with a network using the same weights. After this initial reduction in dimensionality, the hidden states are concatenated in a subsequent layer and finally mapped to an output. The homogeneity of agents is achieved by using the same set of parameters for all policies. A diagram of the architecture is shown in Figure 2. Fig. 2: This diagram shows a model of our proposed policy with three hidden layers. The numbers inside the boxes denote the dimensionalities of the hid- den layers. The plus sign de- notes concatenation of vectors. 3.3 Adaptations to TRPO In order to apply TRPO to our multi-agent setup, some small changes to the original algorithm have to be made, similar to the formulation of [11]. First, since we assume homogeneous agents, we can have one set of parameters of the policy shared by all agents. Since the agents cannot rely on the global state, the advantage function is redefined as A(h, a). In order to estimate this function, each agent is assigned the same global reward r in each time step and all transitions are treated as if they were executed by a single agent. 4 Experimental Setup In this section, we briefly discuss the used model and state representation of a single agent. Subsequently, we describe our two experimental setups and the policy architecture used for the experiments. 4.1 Agent Model The local state of a single agent is modeled by its 2D position and orientation, i.e., si = [xi, yi, φi] ∈ S = {[x, y, φ] ∈ R3 : 0 ≤ x ≤ xmax, 0 ≤ y ≤ ymax, 0 ≤ φ ≤ 2π}. The robot can only control the speed of its wheels. Therefore, we apply a force to the left and right side of the agent, similarly to the wheels of the real robot. Our model of a single agent is inspired by the Colias robot 128at−η,ot−η+116128at−η+1,ot−η+216128at−1,ot16......16×η2actionl0,1l1,1l2,1l0,2l1,2l2,2l0,ηl1,ηl2,ηm0m1m28 Hüttenrauch et al. (a detailed description of the robot specifications can be found in [2]), but the underlying principles can be straightforwardly applied to other swarm settings with limited observations. Generally, our observation model is comprised of the sensor readings of the short and long range IR sensors (later denoted as 'sensor' in the evaluations). Furthermore, we augment this observation representation with the communication protocols developed in the following section. Our simulation is using a 2D physics engine (Box2D), allowing for correct physical interaction of the bodies of the agents. 4.2 Tasks The focus of our experiments is on tasks where agents need to collaborate to achieve a common goal. For this purpose, we designed the following two scenarios: Task 1: Building a Graph In the first task, the goal of the agents is to find and maintain a certain distance to each other. This kind of behavior is required, for example, in surveillance tasks, where a group of autonomous agents needs to maximize the coverage of a target area while maintaining their communication links. We formulate the task as a graph problem, where the agents (i.e. the nodes) try to maximize the number of active edges in the graph. Herein, an edge is considered active whenever the distance between the corresponding agent lies in certain range. The setting is visualized in Figure 3a. In our experiment, we provide a positive reward for each edge in a range between 10 cm and 16 cm, and further give negative feedback for distances smaller than 7 cm. Accordingly, the reward function is 1[0.1 m, 0.16 m](di m) − 5 1[0 m, 0.07 m](di m), (1) M(cid:88) M(cid:88) i=1 m>i M(cid:88) M(cid:88) i=1 m>i R(s, a) = (cid:112) where di the centers of agent i and agent m and (xi − xm)2 + (yi − ym)2 denotes the Euclidean distance between m = (cid:40) 1 if x ∈ [a, b], 0 else 1[a,b](x) = is an indicator function. Note that we omit the dependence of di state s to keep the notion simple. m on the system Task 2: Establishing a Communication Link The second task adds another layer of difficulty. While maintaining a network, the agents have to locate and connect two randomly placed points in the state space. A link is only established successfully if there are communicating agents connecting the two points. Figure 3b shows an example with an active link Local Communication Protocols 9 spanned by three agents between the two points. The task resembles the problem of establishing a connection between two nodes in a wireless ad hoc network [3,25]. In our experiments, the distance of the two points is chosen to be larger than 75 cm, requiring at least three agents to bridge the gap in between. The reward is determined by the length of the shortest distance between the two points dopt (i.e. a straight line) and the length of the shortest active link dsp spanned by the agents, (cid:40) dopt R(s, a) = dsp 0 if link is established otherwise. In this task, we use the shortest path partitions as communication protocol. Each agent communicates the shortest path it knows to both points of interests, resulting in two 2-D partitions that are used as observation input for a single time step. 4.3 Policy Architecture We decided for a policy model with three hidden layers. The first two layers process the observation-action pairs (ak−1, ok) of each timestep in a history in- dividually and map it into hidden layers of size 128 and 16. The output of the second layer is then concatenated to form the input of the third hidden layer which eventually maps to the two actions for the left and right motor. (a) edge task (b) link task Fig. 3: Illustration of the two cooperative tasks used in this paper. The green dots represent the agents, where the green ring segments located next to the agents indicate the short range IR front sensors. The outer green circles illustrate the maximum range in which distances / bearings to other agents can be observed, depending on the used observation model. (a) Edge task: The red rings show the penalty zones where the agents are punished, the outer green rings indicate the zones where legal edges are formed. (b) Link task: The red dots correspond to the two points that need to be connected by the agents. 10 Hüttenrauch et al. 5 Results We evaluate each task in a standardized environment of size 1 m × 1 m where we initialize ten agents randomly in the scene. Of special interest is how the amount of information provided to the agents affects the overall system per- formance. Herein, we have to keep in mind the general information-complexity trade-off, i.e., high-dimensional local observations generally provide more infor- mation about the global system state but, at the same time, result in a more complex learning task. Recall that the information content is mostly influenced by two factors: 1) the length of the history, and 2) the composition of the obser- vation. 5.1 Edge Task First, we evaluate how the history length η affects the system performance. Figure 4a shows an evaluation for η = {2, 4, 8} and a weight sharing policy using a two-dimensional histogram over distances and bearings. Interestingly, we observe that longer observation histories do not show an increase in the performance. Either the increase in information could not counter the effect of increased learning complexity, or a history length of η = 2 is already sufficient to solve the task. We use these findings and set the history length to η = 2 for the remainder of the experiments. Next, we analyze the impact of the observation model. Figure 4b shows the results of the learning process for different observation modalities. The first ob- servation is that, irrespective of the used mode, the agents are able to establish a certain number of edges. Naturally, a complete information of distances and bearing yields the best performance. However, the independent histogram rep- resentation yields comparable results to the two dimensional histogram. Again, this is due to the aforementioned complexity trade-off where a higher amount of information makes the learning process more difficult. 5.2 Link Task We evaluate the link task with raw sensor measurements, count based histograms over distance and bearing, and the more advanced shortest path histograms over distance and bearing. Based on the findings of the edge task we keep the history length at η = 2. Figure 4c shows the results of the learning process where each observation model was again tested and averaged over 8 trials. Since at least three agents are necessary to establish a link between the two points, the models without shortest path information struggle to reliably establish the connection. Their only chance is to spread as wide as possible and, thus, cover the area between both points. Again, it is interesting to see that independent histograms over counts seem to be favorable over the 2D histogram. However, both versions are surpassed by the 2D histogram over shortest paths which yields information about the current state of the whole network of agents, currently connected to each of the points. Local Communication Protocols 11 (a) Comparison of differ- ent history lengths η. (2D histogram) (b) Edge task: Compari- son of different observa- tion models. (η = 2) (c) Link task: Compari- son of different observa- tion models. (η = 2) Fig. 4: Learning curves for (a), (b) the edge task and (c) the link task. The curves show the mean values of the average undiscounted return of an episode (i.e. the sum of rewards of one episode, averaged over the number of episodes for one learning iteration) over the learning process plus / minus one standard deviation, computed from eight learning trials. Intuitively, the return in the edge task corresponds to the number of edges formed during an episode of length 500 steps. In the link task, it is a measure for the quality of the link. Legend: 2DSP: two dimensional histogram over shortest paths, 2D: two-dimensional histogram over distances and bearings, 1D: two independent histograms over distances and bearing, d: distance only histogram, b: bearing only histogram, sensor: no histogram. 6 Conclusions and Future Work In this paper, we demonstrated that histograms over simple local features can be an effective way for processing information in robot swarms. The central aspect of this new model is its ability to handle arbitrary system sizes without discrim- inating between agents, which makes it perfectly suitable to the swarm setting where all agents are identical and the number of agents in the neighborhood varies with time. We use these protocols and an adaptation of TRPO for the swarm setup to learn cooperative decentralized control policies for a number of challenging cooperative task. The evaluation of our approach showed that this histogram-based model leads the agents to reliably fulfill the tasks. Interesting future directions include, for example, the learning of an explicit communication protocol. Furthermore, we expect that assigning credit to agents taking useful actions should speedup our learning algorithm. Acknowledgments. The research leading to these results has received funding from EPSRC under grant agreement EP/R02572X/1 (National Center for Nu- clear Robotics). Calculations for this research were conducted on the Lichtenberg high performance computer of the TU Darmstadt. 0200400600800100002,0004,0006,0008,000TRPOIterationsAverageReturnη=2.0η=4.0η=8.00200400600800100002,0004,0006,0008,000TRPOIterationsAverageReturns2D1Ddbsensor020040060080010000100200300400TRPOIterationsAverageReturns2DSP2D1Dsensor12 Hüttenrauch et al. References 1. Alonso-Mora, J., Montijano, E., Schwager, M., Rus, D.: Distributed multi-robot formation control among obstacles: A geometric and optimization approach with consensus. In: Proceedings of the IEEE International Conference on Robotics and Automation. pp. 5356–5363 (2016) 2. Arvin, F., Murray, J., Zhang, C., Yue, S.: Colias: An autonomous micro robot for swarm robotic applications. International Journal of Advanced Robotic Systems 11(7), 113 (2014) 3. Basu, P., Redi, J.: Movement control algorithms for realization of fault-tolerant ad hoc robot networks. IEEE Network 18(4), 36–44 (2004) 4. Bayındır, L.: A review of swarm robotics tasks. Neurocomputing 172(C), 292 – 321 (2016) 5. Chen, J., Gauci, M., Gross, R.: A strategy for transporting tall objects with a swarm of miniature mobile robots. In: Proceedings of the IEEE International Conference on Robotics and Automation. pp. 863–869 (2013) 6. Correll, N., Martinoli, A.: Modeling and designing self-organized aggregation in a swarm of miniature robots. The International Journal of Robotics Research 30(5), 615–626 (2011) 7. Foerster, J., Assael, Y.M., de Freitas, N., Whiteson, S.: Learning to communicate with deep multi-agent reinforcement learning. In: Advances in Neural Information Processing Systems 29, pp. 2137–2145 (2016) 8. Foerster, J., Farquhar, G., Afouras, T., Nardelli, N., Whiteson, S.: Counterfactual multi-agent policy gradients. arXiv:1705.08926 (2017) 9. Goldberg, D., Mataric, M.J.: Robust behavior-based control for distributed multi- robot collection tasks (2000) 10. Gu, S., Lillicrap, T., Ghahramani, Z., Turner, R.E., Levine, S.: Q-prop: Sample- efficient policy gradient with an off-policy critic. In: Proceedings of the 5th Inter- national Conference on Learning Representations (2017) 11. Gupta, J.K., Egorov, M., Kochenderfer, M.: Cooperative multi-agent control using deep reinforcement learning. In: Proceedings of the Adaptive and Learning Agents Workshop (2017) 12. Hoff, N.R., Sagoff, A., Wood, R.J., Nagpal, R.: Two foraging algorithms for robot swarms using only local communication. In: Proceedings of the IEEE International Conference on Robotics and Biomimetics. pp. 123–130 (2010) 13. Kube, C., Bonabeau, E.: Cooperative transport by ants and robots. Robotics and Autonomous Systems 30(1), 85 – 101 (2000) 14. Lillicrap, T.P., Hunt, J.J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Sil- ver, D., Wierstra, D.: Continuous control with deep reinforcement learning. arXiv:1509.02971 (2015) 15. Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, P., Mordatch, I.: Multi-agent actor- critic for mixed cooperative-competitive environments. arXiv:1706.02275 (2017) 16. Martinoli, A., Easton, K., Agassounon, W.: Modeling swarm robotic systems: a case study in collaborative distributed manipulation. The International Journal of Robotics Research 23(4-5), 415–436 (2004) 17. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level control through deep reinforcement learning. Nature 518(7540), 529–533 (2015) 18. Moeslinger, C., Schmickl, T., Crailsheim, K.: Emergent flocking with low-end swarm robots, pp. 424–431 (2010) Local Communication Protocols 13 19. Nouyan, S., Gross, R., Bonani, M., Mondada, F., Dorigo, M.: Teamwork in self- organized robot colonies. IEEE Transactions on Evolutionary Computation 13(4), 695–711 (2009) 20. Oliehoek, F.A.: Decentralized POMDPs, pp. 471–503. Springer Berlin Heidelberg (2012) 21. Schulman, J., Levine, S., Moritz, P., Jordan, M., Abbeel, P.: Trust region policy optimization. In: Proceedings of the 32nd International Conference on Machine Learning. pp. 1889–1897 (2015) 22. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv:1707.06347 (2017) 23. Teh, Y.W., Bapst, V., Czarnecki, W.M., Quan, J., Kirkpatrick, J., Hadsell, R., Heess, N., Pascanu, R.: Distral: robust multitask reinforcement learning. arXiv:1707.04175 (2017) 24. ŠoŠić, A., KhudaBukhsh, W.R., Zoubir, A.M., Koeppl, H.: Inverse reinforcement learning in swarm systems. In: Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems. pp. 1413–1421 (2017) 25. Witkowski, U., El Habbal, M.A.M., Herbrechtsmeier, S., Tanoto, A., Penders, J., Alboul, L., Gazi, V.: Ad-hoc network communication infrastructure for multi-robot systems in disaster scenarios. In: Proceedings of the IARP/EURON Workshop on Robotics for Risky Interventions and Environmental Surveillance (2008)
1905.13191	2	1905	2019-08-13T13:06:20	Ridesharing with Driver Location Preferences	[ "cs.MA", "cs.GT" ]	We study revenue-optimal pricing and driver compensation in ridesharing platforms when drivers have heterogeneous preferences over locations. If a platform ignores drivers' location preferences, it may make inefficient trip dispatches; moreover, drivers may strategize so as to route towards their preferred locations. In a model with stationary and continuous demand and supply, we present a mechanism that incentivizes drivers to both (i) report their location preferences truthfully and (ii) always provide service. In settings with unconstrained driver supply or symmetric demand patterns, our mechanism achieves (full-information) first-best revenue. Under supply constraints and unbalanced demand, we show via simulation that our mechanism improves over existing mechanisms and has performance close to the first-best.	cs.MA	cs	Ridesharing with Driver Location Preferences∗ Duncan Rheingans-Yoo1 , Scott Duke Kominers2 , Hongyao Ma3 and David C. Parkes3 1Harvard College 2Harvard Business School and Department of Economics 3Harvard School of Engineering and Applied Sciences {d rheingansyoo@college, kominers@fas, hma@seas, parkes@eecs}.harvard.edu 9 1 0 2 g u A 3 1 ] A M . s c [ 2 v 1 9 1 3 1 . 5 0 9 1 : v i X r a Abstract We study revenue-optimal pricing and driver com- pensation in ridesharing platforms when drivers have heterogeneous preferences over locations. If a platform ignores drivers' location preferences, it may make inefficient trip dispatches; moreover, drivers may strategize so as to route towards their preferred locations. In a model with stationary and continuous demand and supply, we present a mech- anism that incentivizes drivers to both (i) report their location preferences truthfully and (ii) always provide service. In settings with unconstrained driver supply or symmetric demand patterns, our mechanism achieves the full-information, first-best revenue. Under supply constraints and unbalanced demand, we show via simulation that our mecha- nism improves over existing mechanisms and has performance close to the first-best. 1 Introduction Uber connected its first rider to a driver in the summer of 2009,1 and since then, ridesharing platforms have dramat- ically changed the way people get around in urban areas. Ridesharing platforms allow a wide array of people to be- come drivers and -- in contrast to traditional taxi systems -- use dynamic "surge pricing" at times when demand exceeds supply. Properly designed, dynamic pricing improves sys- tem efficiency [Castillo et al., 2017], increases driver sup- ply [Chen and Sheldon, 2015], and makes the system reliable for riders [Hall et al., 2015]. A growing literature studies how to structure prices for rid- ers and compensation for drivers so as to optimally account for variation in supply and demand [Banerjee et al., 2015; Bimpikis et al., 2016; Castillo et al., 2017; Ma et al., 2019]. However, existing models leave aside driver heterogeneity. In practice, some drivers may prefer to drive in the city and oth- ers in the suburbs, and many prefer to end their days close to home. A matching system that treats drivers as homogeneous makes inefficient dispatches, with drivers preferring to fulfill each other's dispatches instead of their own. ∗The full version of this paper is available at arXiv:1905.13191. 1https://www.uber.com/newsroom/history/, visited 02/25/2019. The problem goes beyond simple efficiency loss. A core feature of ridesharing platforms is that drivers retain the flex- ibility to choose when and where to provide service. Every ride the platform proposes needs to be accepted voluntarily, forming an optimal response for the driver [Ma et al., 2019]. This incentive alignment simplifies participation for drivers and also makes behavior more predictable. Without account- ing for heterogeneity, a platform cannot fully understand a driver's preferences or achieve full incentive alignment. Indeed, platforms have experimented with methods to in- corporate driver heterogeneity. As of Summer 2019, Uber al- lows drivers to indicate -- twice a day -- that they would like to take trips in the direction of a particular location.2 However, mechanisms that account for driver preferences can also have unintended consequences if not designed properly. By saying "I want to drive South," a driver biases her dispatches in a way that could in principle promote more profitable trips.3 In this paper, we introduce the study of driver location pref- erence in a mechanism design framework. In Section 2, we adapt a model originally conceived by Bimpikis et al. [2016] to an economy where drivers prefer a particular location. In Section 3, we present the Preference-Attentive Rideshar- ing Mechanism (PARM), which elicits driver preferences and sets a revenue-optimal pricing policy. We show that PARM is incentive-compatible, and that it achieves the full- information, first-best revenue when supply is unconstrained or when demand is symmetric. In Section 4, we study set- tings with constrained supply and asymmetric demand, using simulations to compare the revenue and welfare performance of PARM to existing ridesharing mechanisms. We show that PARM achieves close to first-best revenue and typically out- performs even the best case for preference-oblivious pricing (where strategic behavior hurts efficiency). Proofs not pre- sented in the text are deferred to Appendix A of the full paper. 1.1 Related Work Existing research on pricing and dispatching in ridesharing platforms does not account for driver heterogeneity. We build on work of Bimpikis et al. [2016], who show that under a continuum model, and with stationary demand and 2https://help.uber.com/partners/article/set-a-driver-destination? nodeId=f3df375b-5bd4-4460-a5e9-afd84ba439b9, visited 2/25/19. 3https://therideshareguy.com/uber-drops-destination-filters- back-to-2-trips-per-day/, visited 2/25/19. unlimited supply, a ridesharing platform's revenue is maxi- mized when the demand pattern across different locations is balanced. They show via simulation that in comparison to set- ting a uniform price for all locations, pricing trips differently depending on trip origins improves revenue. Relative to the Bimpikis et al. [2016] model, we allow limited driver supply; moreover, each driver in our model has a preferred location. We thus introduce a reporting phase, in which drivers report their preferred locations. We then modify the matching and pricing formulations in order to align incentives. Ma et al. [2019] study the incentive alignment of drivers in the presence of spatial imbalance and temporal variation of supply and demand. Castillo et al. [2017] show that dy- namic pricing mitigates inefficient "wild goose chase" phe- nomena for platforms that employ myopic dispatching strate- gies. Modeling a shared vehicle system as a continuous-time Markov chain, Banerjee et al. [2017] establish approxima- tion guarantees for a static, state-independent pricing pol- icy. Ostrovsky and Schwarz [2019] study the economy of self-driving cars, focusing on car-pooling and market equilib- rium. Queuing-theoretic approaches have also been adopted: Banerjee et al. [2015] show the robustness of dynamic pric- ing; Af`eche et al. [2018] study the impact of driver autonomy and platform control; and Besbes et al. [2018] analyze the relationship between capacity and performance. (cid:80) There are various empirical studies, analyzing the impact of dynamic pricing [Hall et al., 2015; Chen and Sheldon, 2015], the labor market for drivers [Hall and Krueger, 2016; Hall et al., 2017], consumer surplus [Cohen et al., 2016], the value of flexible work [Chen et al., 2017], the gender earnings gap [Cook et al., 2018], and the commission vs. medallion lease-based compensation models [Angrist et al., 2017]. 2 Model We consider a discrete time, infinite horizon model of a ridesharing network with discrete locations, L = {1, . . . , n}. Following the baseline model of Bimpikis et al. [2016], we assume unit distances, i.e., it takes one period of time to travel in between any pair of locations. At the beginning of each time period, for each location i ∈ L, there is a continuous mass θi ≥ 0 of riders requesting trips from i. The fraction of riders at i with destination j ∈ L is given by αij ∈ [0, 1] (thus j∈L αij = 1). We assume that the components of rider de- mand θ = {θi}i∈L and α = {αij}i,j∈L are stationary and do not change over time. Riders' willingness to pay for trips are i.i.d. random variables with CDF F . Thus, for any i, j ∈ L, the number of trips demanded from i to j at price pij ≥ 0 would be θiαij(1 − F (pij)). (Riders who are unwilling to pay the stated prices for their rides leave the market.) Each driver has a preferred location τ ∈ L. Drivers receive I ≥ 0 additional utility whenever they start a period in their preferred locations (irrespective of whether they have a rider); this preferred location is private information and represents a driver's type. For each location τ ∈ L, the total mass of avail- able drivers of type τ is given by s(τ ) ≥ 0. Drivers have a discount factor of δ ∈ (0, 1), and an outside option that deliv- t=0 δtI = I/(1 − δ) < w, 4Throughout the paper, we consider δ to be very close to 1 -- this ers utility w ≥ 0.4 We assume(cid:80)∞ meaning that the utility from being in one's favorite location at all times does not outweigh the outside opportunity. A ridesharing mechanism elicits drivers' preferred loca- tions, matches drivers and riders to trips, sets riders' trip prices and drivers' compensation, and (potentially) imposes drivers' penalties for strategic behavior. Before the beginning of the first time period, the mechanism elicits the preferred locations from potential drivers. At the beginning of each period, a driver whose previous trip ended at location i chooses whether or not to provide ser- vice at location i. If a driver provides service, the mecha- nism may dispatch that driver to (i) pick up some rider with trip origin i, (ii) relocate to some location, or (iii) stay in the same location. If a rider going from i to j is picked up by some driver, then the rider pays the platform the trip price pij ≥ 0. If a driver of reported type τ is dispatched from i to ij ≥ 0, regardless of if her dis- j, the platform pays them c(τ ) patch was to pick up a rider or relocate.5 Drivers who choose not to provide service in a period can relocate to any loca- tion j in the network, are not compensated by the mechanism in this period, and may be charged a penalty Pj.6 Denote p (cid:44) {pij}i,j∈L, c (cid:44) {c(τ ) ij }i,j,τ∈L and P (cid:44) {Pj}j∈L. Based on rider demand (θ, α) and the reported supply of drivers of each type, a mechanism determines rider and driver flow, trip prices p, driver compensation c, and driver penalties P . Drivers decide whether or not to participate, consider- ing pricing, penalties, and their outside options. Given entry decisions by drivers, and decisions made by drivers since en- try, the platform then dispatches drivers to trips and processes payments and penalties accordingly in each period. 2.1 Steady-State Equilibrium In this section we will analyze a steady-state equilibrium while ignoring penalties. This will be helpful because it es- tablishes that under truthful reporting of types, drivers will always follow the proposed dispatches. We will separately handle incentives to report truthful types, considering the ef- fect of penalties on aligning these incentives. It will turn out that drivers are only charged penalties for the first time they deviate, not for future deviations. i i (cid:80) i∈L x(τ ) At the beginning of each period, let x(τ ) be the number of drivers of reported type τ at location i, and let x(τ ) (cid:44) denote the total number of drivers of reported type ij }i,j,τ∈L, ij ≥ 0 is the number of riders from i to j assigned ij ≥ 0 be the mass of drivers is natural, since an annual interest rate of 4% implies an exponential discount factor of 0.9999992 over the course of ten minutes. τ on the platform. Let the trip flow be f (cid:44) {f (τ ) where f (τ ) to drivers of type τ. Let y(τ ) 5It bears mentioning that c(τ ) ij need not be a fixed proportion of pij. In fact, Bimpikis et al. [2016] find that for certain types of net- works, making driver compensation a fixed proportion of trip price drastically reduces platform revenue. 6Drivers only choose whether to provide service at a location and cannot decline dispatches based on the trip destination. This is consistent with current ridesharing platforms, which hide trip desti- nations because of concern that drivers will cherry pick rides. i ij ≤ x(τ ) i }i,τ∈L and y (cid:44) {y(τ ) ij + y(τ ) ij ≤ θiαij(1 − F (pij)) and for all i, j ∈ L. j∈L f (τ ) τ∈L f (τ ) For a trip with origin i and destination j, of type τ at i who are dispatched to relocate to j without a rider, and set x (cid:44) {x(τ ) ij }i,j,τ∈L. No driver or rider can be matched multiple times in the same pe- riod, so assuming drivers always provide service, we have for all i ∈ L and all τ ∈ L, and (cid:80) (cid:80) rider demand exceeds driver supply (i.e., if (cid:80) timization, we can assume without loss that (cid:80) if the total τ∈L f (τ ) ij < θiαij(1−F (pij))), the mechanism may increase the trip price pij and achieve higher revenue. Therefore for revenue op- ij = θiαij(1 − F (pij)). When x(τ ) i > 0, meaning that some drivers with reported type τ are at location i, the probability that a given driver of reported type τ is dispatched to destina- tion j is (f (τ ) . Assuming a driver of type τ has truthfully reported her type and will provide service in all pe- riods, her lifetime expected utility for starting from location i is of the form ij + y(τ ) ij )/x(τ ) τ∈L f (τ ) i π(τ ) i = (c(τ ) ij + δπ(τ ) j ) f (τ ) ij + y(τ ) ij x(τ ) i + I · 1{i = τ}, (1) where 1{·} is the indicator function. The first term in (1) is the expected compensation and future utility a driver gets when dispatched to one of the n possible destinations. The second term corresponds to the idiosyncratic utility drivers get from starting trips in their favorite locations. Definition 1 (Steady-State Equilibrium). A steady-state equilibrium under pricing policy (p, c) is a tuple (f, x, y) s.t.: (C1) (Driver best-response) Drivers providing service always i > 0 ⇒ maximizes their payoff, i.e. ∀i, τ ∈ L, x(τ ) ∀k ∈ L, π(τ ) i ≥ I · 1{i = τ} + δπ(τ ) k . (C2) (Flow balance) For all locations i ∈ L and driver types (cid:88) j∈L i =(cid:80) (C3) (Market-clearing)(cid:80) τ ∈ L, x(τ ) j∈L f (τ ) ji + y(τ ) ji . τ∈L f (τ ) ij = θiαij(1 − F (pij)). (C4) (Individually rational driver entry) Participating drivers get at least their outside option w; with excess supply of drivers with type τ, all participating type-τ drivers get exactly their outside option w. ij , x(τ ) ij , y(τ ) i.e., ∀i, j, τ ∈ L, f (τ ) straints are satisfied, i.e., ∀τ ∈ L,(cid:80) (C5) (Feasibility) Rider and driver flows are non-negative, i ≥ 0; the supply con- i ≤ s(τ ). The full information first best revenue (FB) is the highest revenue a mechanism can achieve in stationary-state equilib- rium, if the mechanism has full knowledge of driver types (therefore does not need to determine dispatching and com- pensation in order to incentivize truthful reporting of types): i∈L x(τ ) pij · f (τ ) ij − c(τ ) ij (f (τ ) ij + y(τ ) ij ) (2) (cid:88) (cid:88) (cid:88) i∈L j∈L τ∈L max p,c s.t. (f, x, y) is a steady-state equilibrium under (p, c). The design problem is to compute rider prices p, driver compensation c, and driver penalties P to optimize plat- form revenue in the steady state equilibrium, in a way that drivers will truthfully report their location preferences and will choose to always provide service. 3 The Preference-Attentive Ridesharing Mechanism (PARM) We now introduce our Preference-Attentive Ridesharing Mechanism (PARM) and show that this mechanism (i) truth- fully elicits drivers' location preferences, (ii) incentivizes drivers to provide service, and (iii) achieves first-best revenue when supply is unconstrained or when demand is symmetric. 3.1 Alternate Form of the Optimization The optimization problem (2) need not be convex, and more- over, even when an optimal solution can be found, it may not incentivize drivers to report their types truthfully. Denoting W (cid:44) w(1 − δ), we present an alternative problem (3), which guarantees that any optimal solution can be converted into an optimal solution for (2) using compensation scheme (4) -- while preserving the objective. Specifically, we consider: max p,f,x,y s.t. x(τ ) i = (cid:16) (cid:88) i,j,τ∈L x(τ ) τ f (τ ) ij pij x(τ ) i + I (cid:17)− W (cid:88) (cid:88) (cid:88) i,τ∈L ji , ∀i ∈ L, ∀τ ∈ L y(τ ) (cid:88) j∈L ij = θiαij(1 − F (pij)), ∀i, j ∈ L f (τ ) i ≤ s(τ ), ∀τ ∈ L x(τ ) f (τ ) ji + τ∈L j∈L (3) i −(cid:88) j∈L ij = x(τ ) y(τ ) ij , ∀i ∈ L, ∀τ ∈ L f (τ ) j∈L f (τ ) ij , y(τ ) ij , x(τ ) i ≥ 0, ∀i, j ∈ L, ∀τ ∈ L. (cid:88) (cid:88) (cid:88) i∈L τ∈L Our approach is analogous to a similar move by Bimpikis et al. [2016] -- assuming that F is distributed U[0, 1], the solution space is convex, and the optimization problem is quadratic. We also go a step further by accounting for driver heterogeneity and the possibility of zero demand at a location, the latter by paying drivers for relocation dispatches. Consider the following compensation scheme: ij =W− I · 1{i = τ},∀i, j, τ ∈ L. c(τ ) (4) Lemma 1. Consider an optimal solution (p, f, x, y) to prob- lem (3), and let c be the compensation scheme (4). Then: (i) (p, c) is an optimal solution to optimization problem (2) with steady-state equilibrium (f, x, y); and (ii) The expected lifetime utility (payment and location value) of a truthful driver is exactly w starting from ev- ery location, i.e. π(τ ) i = w for all i, τ ∈ L. Briefly, feasible solutions to (3) satisfy conditions (C2), (C3) and (C5). Moreover, with W > I, the compensation c as in (4) is non-negative. Given (4), drivers receive utility W in expectation per period (so w over their lifetimes); this implies (C1) and (C4). Furthermore, the solution is optimal, since no compensation scheme can lower the total payment to drivers while fulfilling the same rider trip flow f. 3.2 Constructing PARM Definition 2. Given rider demand (θ, α), the Preference- Attentive Ridesharing Mechanism (PARM): 1. Elicits the location preferences from drivers. 2. Solves (3) with an additional constraint τ ≥ x(τ ) x(τ ) , ∀i ∈ L, ∀τ ∈ L i (5) for dispatching and pricing, and determine driver compen- sation by (4). 3. If a driver with reported type-τ did not provide service and relocated to location i (cid:54)= τ ∈ L, the platform treats her as a type-i driver from then on. If this is the first deviation for this driver, the driver pays penalty Pτ (cid:44) max{maxk∈L{P k→τ}, 0} for P k→τ as solved for in the following linear system: W + δ (cid:88) j f (τ ) ij + y(τ ) ij πk→τ j x(τ ) i  + πk→τ i =1{i (cid:54)= τ} (cid:88) 1{i = τ}(δw− P k→τ ) + 1{i = k}I, ∀i, k ∈ L; (6) i /x(τ ), ∀k ∈ L. x(τ ) πk→τ i w = i i The system (6) has n2 + n linear equations and n2 + n and n of the P k→τ ). Intuitively, unknowns (n2 of the πk→τ πk→τ is the expected utility of a driver of type k pretend- i ing to be of type τ and providing service everywhere except τ, where she instead relocates to k. By construction, P k→τ is the minimum penalty needed to equalize driver earnings between this deviation and truth telling plus always provid- ing service. We take the maximum over such penalties so no driver can benefit from pretending to be of type τ and employ this strategy. 7 If a driver declines to provide service but relocates to her reported preferred location, she is charged no penalty. A driver might have a legitimate (idiosyncratic) reason for not being able to provide service in a period, but if she relocates to a location she did not report as preferred, that is taken as an indication that her original report was not truthful. We now prove, under the assumption that drivers always provide service and as a result are never charged any penalty, that imposing (5) is sufficient to guarantee truthful reporting. Theorem 1. Assuming all drivers always provide service, it is a dominant strategy for drivers to report their location pref- erences truthfully under PARM. Proof. Observe that by being truthful, each driver gets utility W = w(1 − δ) per period -- getting paid W − I at preferred 7 The penalty Pτ is set to zero if P k→τ < 0 for all k ∈ L. This case arises if this deviation is itself bad for drivers of all types, in which case the only way to make the deviating drivers' utility equal to w is to pay those drivers. locations, and W at every other location. As a result, π(τ ) i = w for all i, τ ∈ L. Suppose an infinitesimal driver of type k ∈ L reports she is of type τ (cid:54)= k. At all i (cid:54)= k, τ she gains utility W per period. At k, she makes W + I because the platform, treating her as a type-τ driver, is still paying her W . At τ, she is paid W − I and does not get the extra utility I. With δ → 1, misreporting τ in place of k leads to an in- crease in the expected payoff in static steady-state equilib- rium if and only if in equilibrium, the driver with reported type τ spends more time in location k than in location τ. Considering the location of a driver with reported type τ as a Markov chain, then {x(τ ) i /x(τ )}i∈L is the stationary distri- bution. (5) then guarantees that a driver with reported type τ spends a plurality of her time at location τ, therefore no driver benefits from misreporting her type if all drivers always pro- vide service. We now consider drivers who may strategically decline to provide service and show such deviations are not useful under PARM, which updates its belief about a driver's type after deviations and imposes a penalty on the first such deviation. Theorem 2. Under PARM, it is an ex post Nash equilibrium for drivers to report their types truthfully and to always pro- vide service. Briefly, Theorem 1 and the following Lemma 2 imply that (i) a profitable misreport must be paired with post-reporting deviation(s), and (ii) the most profitable deviation must be the driver providing service everywhere except her reported preferred location. The penalties ensure the driver does not get a utility higher than w from this deviation (or any other), so there does not exist a profitable deviation. Drivers are never charged any penalty under the equilib- rium outcome, but the threat of a penalty is necessary to en- sure truthful reporting. In certain special economies, a misre- porting driver might spend many periods at her true preferred location before being sent to her reported preferred location. Without penalties, she may simply decline service and relo- cate back to her actual preferred location, thereby sacrificing one period of income for the possibility of many periods of extra idiosyncratic utility. See Appendix B of the full paper for an example and discussions. Lemma 2. Consider a driver of true type k ∈ L and re- ported type τ ∈ L, and assume that the rest of the drivers always provide service. If τ = k (truthful), always providing service is a best response. If τ (cid:54)= k, one of the following is a best-response: (i) always providing service, or (ii) providing service at every location except τ, where the driver instead drives to k. We now outline the proof of Lemma 2. We first show that a truthful driver earns W at every location, and it is always op- timal for her to provide service.An untruthful driver gets the least utility when at her reported preferred location τ, so relo- cating to τ is worse than providing service; in any period, she should either provide service or relocate to a location i (cid:54)= τ. In fact, because she is charged the same penalty for relocating to any location i (cid:54)= τ, her optimal relocation is her true pre- ferred location k. Intuitively, if she relocates to i, she will be at i and treated as type i from then on, which is sub-optimal for her unless i = k. Given that the optimal relocation is then her true preferred location -- after which she will make W ev- ery period -- the only location where she might profitably not provide service is τ, her reported preferred location and the only place she currently makes less than W in-period. 3.3 Cases with First-Best Revenue and No Penalty Although the IC constraint (5) may reduce revenue, we can characterize some settings where imposing IC does not lead to a revenue loss: when supply is unconstrained, or when rider demand is symmetric, PARM achieves the full- information first-best revenue. Furthermore, no penalty is necessary to ensure incentive compatibility. Theorem 3. Suppose s(τ ) = ∞ for all τ ∈ L. Then PARM achieves full-information first-best revenue, and no penalty is necessary to ensure incentive compatibility. Briefly, the IC constraint (5) does not bind because drivers cost less to the platform when at their preferred locations. If there are more drivers with reported type τ at location i than at τ, the platform can improve its revenue by replacing the type-τ drivers with type-i drivers (there are unlimited type-i drivers). Incentive compatibility holds without penalties be- cause each driver visits her reported preferred location before visiting any other location too many times. Thus, she cannot profitably use a misreport-plus-deviation to sacrifice one pe- riod of income for many periods of idiosyncratic utility (as described following Theorem 2). Note that the preceding ar- gument makes no assumption on the demand pattern and re- quires only the availability of supply. Definition 3. Rider demand (θ, α) is symmetric if ∀i, j, k, l ∈ L we have θi = θk and αij = αkl. Theorem 4. Suppose that rider demand is symmetric. Then we can construct a solution to optimization problem (3) with incentive compatibility constraint (5) such that PARM achieves full-information first-best revenue, and no penalty is necessary to ensure incentive compatibility. To understand Theorem 4, we prove two additional lemmas. Lemma 3. With symmetric demand, any optimal solution to (3) satisfies f (τ ) τ τ for all i, τ ∈ L. τ τ + y(τ ) ii ≤ f (τ ) ii + y(τ ) Intuitively, type-τ drivers cost less when at location τ, so it is optimal for the marginal ride they give at location τ to have a lower price than at other locations. If demand is symmetric, this means drivers with reported type-τ provide more rides at location τ than at any other location. Lemma 4. If the demand pattern is symmetric, we can con- struct an optimal solution to (3) such that for all i, j ∈ L and all τ ∈ L, f (τ ) ji and y(τ ) ij = f (τ ) ij = y(τ ) ji = 0. Intuitively, there is no need for drivers to relocate when demand is fully symmetric. Moreover, given any optimal so- lution to (3), we can construct an alternative optimal solution, where the flow of drivers of each type can be decomposed as cycles with length 2, i.e., f (τ ) ij = f (τ ) We can now sketch the proof of Theorem 4. With symmet- ric demand, Lemma 3 implies driver flow for within-location . ji trips satisfies the IC constraint (5). For all inter-location trips, Lemma 4 lets us focus only on bilateral driver flow between pairs of locations i and j. Type-τ drivers cost less at τ, so they will naturally fill more rides between τ and j than be- tween i and j -- and this holds for all j, so type-τ drivers fill more rides in and out of τ than i. Combining the two cases, drivers of type τ do not spend more time at another location i (cid:54)= τ than they do at τ, so imposing the IC constraint (5) is without loss of revenue. As in Theorem 3, incentive compat- ibility holds without penalties because each driver will visit her reported preferred location before visiting any other too many times. Thus, she cannot profitably use a misreport-plus- deviation to sacrifice one period of income for many periods of idiosyncratic utility (as described following Theorem 2). 4 Simulation Results In this section, we use simulations to analyze the revenue and social welfare under PARM for settings outside the cases cov- ered by Theorems 3 and 4 -- i.e., settings with limited supply and unbalanced demand. Social welfare is defined as the total rider value plus drivers' utilities from being in their preferred locations, mi- nus the total opportunity costs incurred by drivers. We compare PARM with the full-information first-best, and also a Preference-Oblivious Ridesharing Mechanism (PORM) which sets prices as in Bimpikis et al. [2016] without con- sidering drivers' location preferences, while assuming that drivers always follow dispatches. In Section 4.2, we also study the equilibrium outcome under PORM, allowing driver autonomy. For ease of illustration, we consider two locations L = {0, 1} throughout the analysis. 4.1 Varying Demand Patterns Suppose that there are an equal number of drivers favoring each location: s(0) = s(1) = 100. Drivers have outside option w =40, discount factor δ = 0.99, and gain utility I = 0.2W = 0.2w(1−δ) per period from being in their preferred locations. Each rider has value independently drawn ∼ U[0, 1]. Varying Total Demand. We first assume an unbalanced trip flow α00 = α10 = 0.25 and α01 = α11 = 0.75 (i.e., three quarters of riders from each location would like to go to lo- cation 1). Fixing the total demand at location 1 at θ1 = 1000, and varying θ0 from 0 to 1000, the revenue and welfare under PARM and benchmarks are as in Figure 1. Although PARM only necessarily achieves first-best revenue when θ0 = 1000 (symmetric demand), we see that PARM achieves the first- best and outperforms PORM unless θ0 is very small, such that demand from the two locations is highly asymmetric. When θ0 (cid:28) θ1, almost all rides originate and terminate at location 1, thus the first-best and PORM dispatch most drivers of both types to provide service at location 1. Fig- ure 2 illustrates the rider trip flows fulfilled by drivers of each type under different mechanisms, when θ0 = 50. To satisfy PARM's incentive compatibility (IC) constraint, how- ever, drivers of type 0 must spend a plurality of their time at location 0. Therefore, PARM completes fewer trips at loca- tion 1, dispatches more type 0 drivers to fulfill (the less prof- itable) between-location trips, and asks many type 0 drivers to (a) Revenue (b) Welfare (a) Revenue (b) Welfare Figure 1: Revenue and welfare varying demand θ0 at location 0. Figure 4: Equilibrium revenue and welfare varying I/W . 3.3 65.4 6.2 6.2 1.4 84.0 15.7 0 1 15.7 35.0 0 1 18.8 + 16.2 0 7.3 7.3 1 (a) Type 0: FB (b) Type 0: PARM (c) Type 0: PORM 0.0 100.0 0.0 100.0 1.4 84.0 0 0.0 0.0 1 0 0.0 0.0 1 0 7.3 7.3 1 (d) Type 1: FB (e) Type 1: PARM (f) Type 1: PORM Figure 2: Total rider trips fulfilled by drivers of each type, with (αi0, αi1) = (0.25, 0.75), θ = (50, 1000), and s = (100, 100). (a) Revenue (b) Welfare Figure 3: Revenue and welfare varying αi0 for i = 0, 1. relocate back to 1 once they arrive at location 0 (the numbers after the "+" sign represent driver relocation flow), resulting in lower revenue and social welfare. Varying Imbalance in Demand. Fixing θ0 = θ1 = 1000 and varying αi0 for i = 0, 1 (i.e., changing the proportion of rides with destination 0), the revenue and welfare achieved by different mechanisms are shown in Figure 3. Similar to Figure 1, PARM achieves first-best revenue and outperforms PORM for a wide range of αi0 (although demand is only symmetric when αi0 = 0.5). For similar reasons as in the above scenario, we see a decline of revenue and welfare under PARM when demand becomes highly unbalanced -- in this case, when αi0 approaches 0 or 1 and almost all riders have the same destination. 4.2 PORM in Equilibrium In this section, we analyze a scenario for which we are able to compute the equilibrium outcome given the pricing under PORM, and under the setting where drivers are given the flex- ibility to decide how to drive. Consider two locations L = 80 0 120 100 100 75 100 + 25 1 0 1 0 1 (a) Type 1: PARM Figure 5: Rider trips fulfilled by type-1 drivers, with s = (0, 200), θ = (1000, 1000), α00 = α11 = 1, α01 = α10 = 0, and I = 0.2W . (c) Type 1: PORM Eq. (b) Type 1: PORM {0, 1} and drivers of type 1 only: s(0) = 0, s(1) = 200. All trips start and end in the same location, i.e., α00 = α11 = 1. Being oblivious to drivers' preferences, PORM sets the same trip price for the two locations and expects the spatial distri- bution of drivers to be proportional to the distribution of de- mand. In equilibrium, however, more drivers decide to drive in location 1 (the preferred location), such that in each period drivers in 1 are dispatched with probability less than 1 and achieve the same expected utility as drivers in 0. Varying Location Preference I. In Figure 4, we fix de- mand θ0 = θ1 = 1000 and plot revenue and welfare as I, the idiosyncratic driver utility, varies from 0 to W . As I in- creases, welfare and revenue under PARM coincide with the first-best and increase as expected. However, revenue under PORM (assuming driver compliance) remains constant since the mechanism is oblivious to drivers' preferences. We also see a decrease in welfare and revenue achieved in equilib- rium under PORM, since more drivers decide to supply in location 1, instead of in location 0 as dispatched, resulting in unfulfilled rides in 0 and idle drivers in 1. Beyond I = 0.5W , revenue and welfare remain constant, since all drivers are al- ready supplying location 1. Figure 5 illustrates rider trip flow fulfilled by the type 1 drivers when I = 0.2W . PARM assigns more drivers to lo- cation 1 than location 0, but PORM does not. However, in equilibrium more drivers end up at location 1 anyway, lead- ing to 25 units of drivers idling at location 1. Varying Demand Ratio θ0/θ1. In Figure 6, we fix θ1 = 1000, I/W = 0.2, and vary θ0 from 0 to 2000. We see that PARM revenue coincides with the first-best and significantly exceeds the revenue of PORM. The revenue and welfare of the equilibrium outcome under PORM is much lower, how- ever, because drivers over-supply the preferred location 1, leaving rider trips in 0 unfulfilled. It is curious that with highly unbalanced demand, an increase in θ0 initially leads to reduced equilibrium revenue and welfare -- this is because with higher demand at location 0, PORM sets a higher price at location 1 and accepts fewer location 1 trips in order to com- (a) Revenue (b) Welfare Figure 6: Equilibrium revenue and welfare varying θ1/θ0 (a) Revenue (b) Welfare Figure 9: Equilibrium revenue and welfare varying s1 plete more trips in 0. The drivers, however, are only willing to drive in 0 when θ0 is high enough that the low probability of getting a ride in 1 offsets the extra utility I. Varying Demand Ratio θ1/θ0. In Figure 7, we set I/W = 0.2, θ0 = 1000 and vary θ1 from 0 to 2000. We see a sim- ilar trend as in Figure 1, with PARM doing worse than even equilibrium PORM for very small values of θ1. Figure 8 il- lustrates driver flow for θ1 = 100 = 0.1θ0 -- PARM employs drivers to idle at location 1 in order to satisfy IC, and this is very costly. It is worth noting that as θ1 increases, the social welfare achieved by the equilibrium outcome under PORM in fact does not increase, due to the increased amount of idle drivers at location 1. (a) Revenue (b) Welfare Figure 7: Equilibrium revenue and welfare varying θ1/θ0. 181.8 18.2 177.3 18.2 + 4.5 100.0 50.0 + 50.0 0 1 0 1 0 1 (a) Type 1: PORM (b) Type 1: PORM Eq. (c) Type 1: PARM Figure 8: Rider trip and idle driver flows, with θ = (1000, 100), α = [(1.0, 0.0), (0.0, 1.0)], s = (0, 200), and I = 0.2W . Varying Driver Supply s1. We now examine the effect of varying the supply of type 1 drivers (while still keeping the supply of type 0 drivers at 0). In Figure 9, we set I = 0.2W , θ0 = θ1 = 1000, and vary s1 from 0 to 1000. Revenue and welfare under PARM coincide with first-best and outperform PORM. All the mechanisms improve in profit and welfare as supply increases, but PARM is better able to use the ad- ditional drivers. Under PORM in equilibrium, drivers again over-supply the preferred location 1, causing rides at location 0 to get dropped. Eventually, there are so many drivers that they can fill all the demand, even with drivers idling at loca- tion 1. At this point, equilibrium PORM revenue coincides with PORM revenue, though the welfare is still lower. 5 Discussion We have proposed the Preference-Attentive Ridesharing Mechanism (PARM) for pricing and dispatch in the pres- ence of driver location preferences. It is an equilibrium under PARM for drivers to report their preferred locations truthfully and always provide service. PARM achieves first-best rev- enue in settings with unconstrained driver supply or symmet- ric rider demand, and we show via simulations that even out- side those scenarios, PARM achieves close to first-best wel- fare and revenue and outperforms a mechanism that is obliv- ious to location preferences. Our analysis suggests that incorporating drivers' location preferences is compatible with other aspects of ridesharing pricing and marketplace design -- even though drivers could in principle game the system by expressing preferences for locations associated with more highly compensated rides. There are two key elements to our approach that both seem likely to provide practical insight beyond the specific frame- work and mechanism considered here: First, we recognize that respecting drivers' location preferences creates value, which can at least partially substitute for cash compensa- tion. Then, we incentivize truthful location preference reve- lation through a variation on a revealed preference approach. PARM uses drivers' deviations from proposed dispatches to learn about their preference types -- a driver who chooses to drive to i instead of her assigned location is inferred to prefer location i and subsequently faces the compensation profile of other drivers with that preference. For this approach to work, it is important that preferences do not change frequently over the course of the day. Otherwise, it would be much harder to enforce incentive compatibility by tracking endogenous re- sponses to dispatch assignments. Acknowledgments We appreciate the helpful comments of Peter Frazier, Jonathan Hall, Hamid Nazerzadeh, Michael Stewart, and the Lab for Economic Design. Rheingans-Yoo gratefully ac- knowledges the support of a Harvard Center of Mathemati- cal Sciences and Applications Economic Design Fellowship, a Harvard College Program for Research and Science and En- gineering Fellowship, and the Harvard College Research Pro- gram. Rheingans-Yoo also appreciates the academic inspira- tion and personal support of Ross Rheingans-Yoo. Komin- ers gratefully acknowledges the support of National Science Foundation grant SES-1459912, as well as the CMSA's Ng Fund and Mathematics in Economics Research Fund. Ma gratefully acknowledges the support of a Siebel scholarship. [Ma et al., 2019] Hongyao Ma, Fei Fang, and David C. Parkes. Spatio-temporal pricing for ridesharing platforms. In Proceedings of the 2019 ACM Conference on Eco- nomics and Computation. ACM, 2019. [Ostrovsky and Schwarz, 2019] Michael Ostrovsky and Michael Schwarz. Carpooling and the economics of the 2019 ACM self-driving cars. Conference on Economics and Computation. ACM, 2019. In Proceedings of References [Af`eche et al., 2018] Philipp Af`eche, Zhe Liu, and Costis Maglaras. Ride-hailing networks with strategic drivers: The impact of platform control capabilities on perfor- mance. Working Paper, 2018. [Angrist et al., 2017] Joshua D. Angrist, Sydnee Caldwell, and Jonathan V. Hall. Uber vs. taxi: A driver's eye view. NBER Working Paper No. 23891, 2017. [Banerjee et al., 2015] Siddhartha Banerjee, Ramesh Johari, and Carlos Riquelme. Pricing in ride-sharing platforms: A queueing-theoretic approach. In Proceedings of the Six- teenth ACM Conference on Economics and Computation, pages 639 -- 639. ACM, 2015. [Banerjee et al., 2017] Siddhartha Banerjee, Daniel Freund, and Thodoris Lykouris. Pricing and optimization in shared In Pro- vehicle systems: An approximation framework. ceedings of the 2017 ACM Conference on Economics and Computation, pages 517 -- 517. ACM, 2017. [Besbes et al., 2018] Omar Besbes, Francisco Castro, and Ilan Lobel. Spatial capacity planning. Working Paper, 2018. [Bimpikis et al., 2016] Kostas Bimpikis, Ozan Candogan, and Daniela Saban. Spatial pricing in ride-sharing net- works. Working Paper, 2016. [Castillo et al., 2017] Juan Camilo Castillo, Dan Knoepfle, and E. Glen Weyl. Surge pricing solves the wild goose In Proceedings of the 2017 ACM Conference chase. on Economics and Computation, pages 241 -- 242. ACM, 2017. [Chen and Sheldon, 2015] M. Keith Chen and Michael Shel- don. Dynamic pricing in a labor market: Surge pricing and the supply of Uber driver-partners. Working Paper, 2015. [Chen et al., 2017] M. Keith Chen, Judith A. Chevalier, Pe- ter E. Rossi, and Emily Oehlsen. The value of flexible work: Evidence from Uber drivers. NBER Working Paper No. 23296, 2017. [Cohen et al., 2016] Peter Cohen, Robert Hahn, Jonathan Hall, Steven Levitt, and Robert Metcalfe. Using big data to estimate consumer surplus: The case of Uber. NBER Working Paper No. 22627, 2016. [Cook et al., 2018] Cody Cook, Rebecca Diamond, Jonathan Hall, John List, and Paul Oyer. The gender earnings gap in the gig economy: Evidence from over a million rideshare drivers. Working Paper, 2018. [Hall and Krueger, 2016] Jonathan V. Hall and Alan B. Krueger. An analysis of the labor market for Uber's driver- partners in the United States. NBER Working Paper No. 22843, 2016. [Hall et al., 2015] Jonathan V. Hall, Cory Kendrick, and Chris Nosko. The effects of Uber's surge pricing: A case study. Working Paper, 2015. [Hall et al., 2017] Jonathan V. Hall, John J. Horton, and Daniel T. Knoepfle. Labor market equilibration: Evidence from Uber. Working Paper, 2017. Appendix We provide in Appendix A the proofs that are omitted from the body of the paper. We present in Appendix B an example of an incentive issue when no penalty is imposed. A Proofs A.1 Proof of Lemma 1 Lemma 1. Consider an optimal solution (p, f, x, y) to prob- lem (3), and let c be the compensation scheme (4). Then: (i) (p, c) is an optimal solution to optimization problem (2) with steady-state equilibrium (f, x, y); and (ii) The expected lifetime utility (payment and location value) of a truthful driver is exactly w starting from ev- ery location, i.e. π(τ ) i = w for all i, τ ∈ L. Proof. The first term of the objective of (3) is the same as the income portion of the objective of (2). The constraints on this optimization are exactly the equilibrium conditions (C2), (C3) and (C5). Thus, it suffices to devise a compensation scheme where driver income everywhere is exactly equal to outside option, i.e. π(τ ) i = w (this will necessarily satisfy (C1) and (C4)). This will be revenue-optimal because by (C4) drivers cannot be making less than w. Consider compensation scheme (4): ij =W− I · 1{i = τ} c(τ ) The second term means that any idiosyncratic utility a driver gets is extracted by the platform, so any dispatched driver makes exactly W in that period. Thus, π(τ ) i = w = W 1−β exactly when probability of dispatch at every location is 1, i.e. x(τ ) i = f (τ ) ij + y(τ ) ij , ∀i ∈ L, ∀τ ∈ L (cid:88) j∈L This is equivalent to the second-to-last constraint of (3), so i = w, ∀i ∈ L, so any optimal solution to (3) will have π(τ ) (C1) and (C4) are satisfied. Thus an optimal solution to (3) corresponds to an optimal solution to (2) under compensation scheme (4). A.2 Proof of Theorem 2 Theorem 2. Under PARM, it is an ex post Nash equilibrium for drivers to report their types truthfully and to always pro- vide service. Proof. By Theorem 1, if a driver is going to always provide service, she cannot profitably misreport. By Lemma 2, if a driver reports her type truthfully, it is not profitable for her to strategically decline to provide service. So for a misreport to be profitable, it must be paired with post-reporting devi- ation (strategically declining to provide service). Lemma 2 characterizes what this deviation must be: providing service everywhere except her reported preferred location. There she instead drives to her true preferred location. Thus to show incentive compatibility without penalty, it suffices to show this strategy is not more profitable than truth- fully reporting and always providing service. Recall how we calculate penalties:  f (τ ) ij + y(τ ) ij xi πk→τ j W + δ (cid:88) j πk→τ i ,∀k. πk→τ i =1{i (cid:54)= τ} (cid:88) i i(cid:80) x(τ ) j x(τ ) j w = + I · 1{i = k} + 1{i = τ}(δw − P k→τ ),∀i, k; Where Pτ (cid:44) max{maxk∈{P k→τ}, 0}. If we had set penal- ties for switching types from k to τ to be P k→τ , a driver of type k pretending to be type τ following the strategy from Lemma 2 would get utility as defined in (6). The first term is the compensation from providing service at locations that are not τ. The second term is idiosyncratic utility from being at her favorite location. The third term is the utility from de- clining to provide service at τ and driving to k, after which she pays the penalty, the platform updates her type, and she makes w afterwards. Then with random initialization, by (6) the driver's expected utility is exactly w. Because we set Pτ = maxk P k→τ , The equality in (6) is an inequality, but still the driver's expected utility is no more than w, which is what she would make by reporting truthfully and always pro- viding service. Therefore, it is a best response for each driver to report truthfully and always provide service. A.3 Proof of Lemma 2 Lemma 2. Consider a driver of true type k ∈ L and re- ported type τ ∈ L, and assume that the rest of the drivers always provide service. If τ = k (truthful), always providing service is a best response. If τ (cid:54)= k, one of the following is a best-response: (i) always providing service, or (ii) providing service at every location except τ, where the driver instead drives to k. Proof. Suppose we have a driver of type τ who has reported she was of type j. If j = τ, then the driver makes W every period she provides service. If j (cid:54)= τ, the driver makes W +I every period she provides service at τ (extra utility but not paid less), makes W − I every period they provide service at j (paid less but not extra utility), and makes W every period elsewhere. First, we will show that the driver will never choose to not provide service and drive to τ. This move will not change how the platform treats the driver in the future. So in either case (j = τ and j (cid:54)= τ) the driver has lost W in income and then relocated to the location where she makes weakly least in the network. This is not profitable, so the driver will not choose to not provide service and drive to τ. Next, we will show that if a driver chooses to not provide service, she will drive to her true preferred location. Sup- pose this driver of type τ declines to provide service. As es- tablished in the previous paragraph, she will not relocate to her previously reported preferred location, so she may pay a penalty, but the penalty is the same no matter where she relo- cates to. she has the choice of where to drive in the network and will choose the location i that maximizes her expected lifetime earnings given that she will be treated as type-i in the future. If location i is such a best-response choice, then for every future deviation, the symmetry of the situation im- plies that i will be a best-response choice then too. So we can assume that for all future deviations, the driver drives to i. In the case that i = τ, the driver expects to make W = w(1 − δ) in every period after arriving at location τ. Now suppose that i (cid:54)= τ. The driver gets utility W +I every period she is at location τ, and W −I every period she is at i, and W everywhere else. With δ → 1, this means choice of i (cid:54)= τ can only be better than τ if she spends more time at τ than i before her next deviation, at which point she faces the same choice. However, we know that x(i) τ , which by the Ergodic Theorem implies that a driver treated as type-i will spend on average at least as much time at location i than location τ. This implies that if the deviant driver starts at location i, the expected number of times she visits location τ before return- ing to i is at most 1 (otherwise she would spend more time on , where t is the number of time periods and N (i) j (t) is the num- ber of periods in which the driver is in location j. So there is no time in the future by which point the driver expects to have been at τ more than i if she follows platform instruc- tions. The driver can deviate from platform instructions, but as established previously, without loss of best response she will drive to location i, which puts in the same position as before. average at τ than i). So for any t, E(cid:104) (cid:105) ≤ E(cid:104) i ≥ x(i) τ (t) N (i) N (i) (cid:105) (t) Finally, we will show that a driver will choose to provide service everywhere if j = τ and will choose to provide ser- vice everywhere except possibly j if j (cid:54)= τ. We have already established that a driver's best relocation is her true preferred location but that a driver will not decline to provide service and drive to her previously reported preferred location. It follows immediately that truthful drivers (those whose pre- viously reported preferred location is their true preferred lo- cation) will always provide service. So we assume j (cid:54)= τ. Because a driver's post-deviation relocation is her preferred relocation, she will make W in every period thereafter, after paying the platform a penalty. Before deviation, the driver makes W + I at τ, W − I at j, and W everywhere else. So the only location she might not want to provide service at is j. Everywhere else her earnings are the same as post-deviation, and the only time it will be less is when at j. So the driver will always choose to provide service at locations other than j. i A.4 Proof of Theorem 3 Theorem 3. Suppose s(τ ) = ∞ for all τ ∈ L. Then PARM achieves full-information first-best revenue, and no penalty is necessary to ensure incentive compatibility. Proof. We show full-information first-best revenue by show- ing the IC constraint does not bind. Consider an optimal so- lution to (3) without IC constraint (5). The flow constraint j allows us to decompose the flow of type-τ drivers into cy- cles with various mass. Suppose one of these cycles does not go through location τ. Then it is optimal to replace that driver flow with drivers of type-j, for location j in the cy- cle. (this can necessarily happen because there is no supply constraint). The new drivers do exactly what the old drivers did, so the demand met is exactly the same and the flow con- straints are still satisfied. However, they end up in their pre- ferred location strictly more, so the last term of the objective strictly increases. So the previous solution was not optimal. So all the cycles of type-τ driver flow go through location τ. ,∀τ, j because all flow of τ drivers This means x(τ ) through j also goes through τ. So any optimal solution to (3) naturally satisfies the IC constraint (5). So imposing the IC constraint does not lead to an objective loss. τ ≥ x(τ ) We show no penalty is necessary to achieve incentive com- patibility by showing that the strategy described in Lemma 2 is not profitable. Consider δ → 1 and suppose a driver of type τ reports he is of type i. Every time he visits location τ, he will visit location i before returning to τ. Otherwise, there would be some non-negative flow in a cycle through τ but not through i, which the platform could more optimally fill with reported type-τ drivers. So the driver will always eventually visit i and will never visit location τ more than once before doing so. The driver makes W +I at location τ, then W every period before giving a ride to i, then 0 at location i, before re- locating to τ and being thereafter treated as τ, making W in every subsequent period. So compared to truthful behavior, the driver makes a maximum of I extra at τ and loses a min- imum of W ≥ at location i before revealing his true type and making W every period thereafter. With δ → 1, this strategy is not more profitable than truthful reporting and always pro- viding service. So even without a penalty, PARM is incentive compatible. A.5 Proof of Lemma 3 Lemma 3. With symmetric demand, any optimal solution to (3) satisfies f (τ ) τ τ for all i, τ ∈ L. τ τ + y(τ ) ii ≤ f (τ ) ii + y(τ ) ii = y(τ ) Proof. Consider an arbitrary i, τ. First notice that in any optimal solution, y(τ ) τ τ = 0 because these are the drivers are not fulfilling rides and staying in the same location, which adds cost to the system without fulfilling demand or helping to satisfy any of the constraints. Next, if τ τ > 0 for j (cid:54)= τ, then it is more optimal to f (τ ) ii > 0 and f (j) switch drivers of type j going τ → τ with drivers of type τ going i → i. This fulfills the same demand, but with less cost because you have drivers of type τ at location τ more. This violates the optimality of the original solution, so we can τ τ = 0 ∀j (cid:54)= τ and that ∃j (cid:54)= τ assume that f (τ ) s.t. f (j) ii = 0. So we find ourselves in one of two cases: ii > 0 ⇒ f (j) τ τ > 0 ⇒ f (τ ) ii = 0: If this is the case, then f (τ ) ii ≤ f (τ ) τ because Case 1. f (τ ) τ τ ≥ 0. f (τ ) Case 2. f (τ ) f (τ ) ii > f (τ ) τ τ . Then(cid:80) ii > 0: Then f (j) ii > (cid:80) j f (j) τ τ = 0 ∀j (cid:54)= τ. Suppose τ τ ⇒ pii < pτ τ be- j f (j) cause the demand pattern is symmetric. I will show it is more optimal for the platform to raise pii, f (τ ) τ τ and lower pτ τ , f (τ ) ii by infinitesimal amounts. Making substitutions and differen- tiating, we get the derivative of the first term of the objective with respect to pii is given as follows: OBJ1 = 1 − 2 pii θiαii d dpii We know that θiαii = θtαtt and that pii < pτ τ . So d dpii OBJ1 > d dpτ τ OBJ1 τ τ ∀i, τ. ii ≤ f (τ ) So the platform can make a marginal increase in the first term of the objective by raising pii and lowering pτ τ by infinitesi- mal amounts, shifting an infinitesimal amount of f (τ ) to f (τ ) τ τ ii to make the market clear. The same number of drivers are in the system, so it does not change the second term of the objective. And it increases the third term in the objective because we just shifted drivers of type τ to only be at lo- cation τ. So this shift brings us to a more optimal solution, which is a contradiction. So without the IC constraint im- posed, f (τ ) A.6 Proof of Lemma 4 Lemma 4. If the demand pattern is symmetric, we can con- struct an optimal solution to (3) such that for all i, j ∈ L and all τ ∈ L, f (τ ) Proof. Consider an optimal solution to the platform's opti- mization problem. Now set pij = pji and f (τ ) ij + f (τ ) ji ), y(τ ) ji ). The same number of drivers of each type is used, so the supply constraint is still satis- fied. Flow into and out of a location before was the same, so it will be the same now too, so the flow constraint is sat- isfied. And the symmetry of the demand pattern implies that θiαij = θjαji and so if the market cleared before, setting prices and flows to be the same will still clear the market. ji and y(τ ) ij = f (τ ) ij + y(τ ) ij = y(τ ) ij = 1 ji = 0. ij = 1 2 (f (τ ) 2 (y(τ ) So all we need to show is that revenue is weakly better when pij = pji, holding number of drivers used (and thus pij + pji) constant. Let pij + pji = q. Then isolating the part of the objective that changes under this switch yields: θiαijpij(1 − pij) + θjαjipji(1 − pji) =θiαijpij(1 − pij) + θiαij(q − pij)(1 − (q − pij)) (cid:16) ∂ ∂pij =θiαij 0 =2q − 4pij pij = q 2 (cid:17) 1 − 2pij − 1 + 2q − 2pij So holding pij + pji constant, equal prices is optimal. So any solution to the optimization problem can be made without loss of optimality into a solution with f (τ ) ij = ji ∀i, j, τ. Having drivers flowing back and forth without y(τ ) giving rides is sub-optimal, so this implies y(τ ) ji = 0. ij = f (τ ) ij = y(τ ) , y(τ ) ji A.7 Proof of Theorem 4 Theorem 4. Suppose that rider demand is symmetric. Then we can construct a solution to optimization problem (3) with incentive compatibility constraint (5) such that PARM achieves full-information first-best revenue, and no penalty is necessary to ensure incentive compatibility. τ j(cid:54)=i f (τ ) ji > f (τ ) ij >(cid:80) Proof. We will show full-information first-best revenue by showing the IC constraint does not bind. By Lemma 3, ii ≤ the symmetry of the demand pattern means that f (τ ) τ τ ∀i, τ. By Lemma 4, we can restrict our attention to so- f (τ ) lutions where all flow of drivers is bilateral and there are no i > x(τ ) floating drivers. Suppose in an optimal solution x(τ ) . τ j , which in turn implies ij = f (τ ) τ j = f (τ ) jτ . This implies(cid:80) j(cid:54)=τ f (τ ) that there exists a location j s.t. f (τ ) We find ourselves in one of two cases: 1. Exists k (cid:54)= τ s.t.f (k) τ j = f (τ ) jτ > 0. Then we can switch some type-τ drivers going i → τ → i with some type k drivers going τ → j → τ. Formally, we make f (k) ij , f (τ ) smaller by and make f (τ ) τ j , f (τ ) jτ , f (k) larger by . All the same demand is filled, the same number of drivers of each type are used, and we've swapped drivers in a way that preserves the flow constraint, so all the constraints are satisfied. We've just strictly increased x(τ ) τ while weakly increasing xkk (in- crease in the case that k = i), which decreases our cost. This improvement is a contradiction with the optimality of the original solution. τ j , f (k) ij , f (k) 2. DNE k (cid:54)= τ s.t. f (k) k f (k) ji > (cid:80) τ j = (cid:80) ij = jτ . The symmetry of the demand pattern then implies that pij = pji < pτ j = pjτ , and so it is an objective improvement to raise pij, pji, f (τ ) ij , f (τ ) ji by . This violates the optimality of the original solution. jτ by and lower pτ j, pjτ , f (τ ) jτ > 0. Then (cid:80) τ j = f (τ ) τ j , f (τ ) jτ , f (τ ) (cid:80) k f (k) k f (k) k f (k) ji ji So there exists an optimal bilateral-flow solution, and this so- lution must have x(τ ) τ ∀i, τ. i ≤ x(τ ) We will show no penalty is necessary to achieve incen- tive compatibility by showing that the deviation described in Lemma 2 is not profitable for the solution constructed in Lemma 4. Consider δ → 1. We know that for all i, j, τ that iτ ≥ f (τ ) f (τ ) ij , so a driver of reported type τ is more likely to be sent to τ than j no matter what location she is at. So, with random initialization of drivers, a driver of type j pretend- ing to be type τ is more likely to be sent to τ before than j than the other way around. Even if sent to j first, she is more likely than not to visit τ before visiting j again. When the driver visits j, she declines to provide service, giving up W in income and then relocating to τ, making W in every sub- sequent period. When the driver visits τ, she gets I ≤ W in extra idiosyncratic utility. However, the probability she visits i (and lose W ) before τ is at least 1 2. So from the beginning of the game to the end of her first period in location τ or i, she loses more in expectation than she gains. Even if she is sent to τ first, the probability she visits i (losing W and then revealing her true preference) before returning to τ and mak- ing another I is less than 1 2, so she is still making less than if she had truthfully reported and always provided service. B Incentive Problem Without Penalty Figure 10 illustrates driver flow for θ1 = θ0 = 100, α = [(1.0, 0.0), (0.0, 1.0)], s = (200, 5), I = 0.2W . Importantly, there is no demand between locations, so once a driver is as- signed to a location, they will continue giving rides there and never visit another location, unless they decline to provide service in a period and relocate. Then a type 1 driver has a useful deviation. She reports that she is type 0, then if she is sent to location 1, she provides service in every period. Thus she prefers location 1 but is being paid as if she does not, and is never sent to location 0 (where she would be paid less) because there is no demand between locations. If this happens, she gets utility w + I 1−δ = 48 over her lifetime. If instead she is sent to location 0, she does not provide ser- vice the first period and relocates to location 1, after which she provides service and the platform treats her as type 1. If this happens, she gets utility w − W = 39.4 over her life- time. So her expected lifetime utility from this deviation is 59 · 39.4 + 25 59 · 48 ≈ 43.16, which is greater than her util- 34 ity from reporting truthfully and providing service, which is w = 40. So this deviation is useful. The key thing going on here is that although more type 0 drivers are at location 0 than 1, an individual type 0 driver might spend their entire lifetime at location 1, which incentivizes type 1 drivers to misreport, taking the risk of one lost period of income in order to get a lifetime of idiosyncratic utility. In this case, the markov chain describing the movement of a type 0 driver is discon- nected, but this issue also occurs when the markov chain is only very weakly connected and the driver is balancing one period of income versus many periods of idiosyncratic util- ity. The penalty is calculated such that if a driver does the deviation described here, the extra loss when they decline to provide service makes the deviation not useful. And in our proof of Theorem 2, we show that this deviation is the best of all non-truthful strategies, so the penalty ensures incentive compatibility. 34 0 25 1 0 0 5 1 (a) Type 0: PARM (b) Type 1: PARM Figure 10: Rider trip and idle driver flows, with θ = (100, 100), α = [(1.0, 0.0), (0.0, 1.0)], s = (200, 5), and I = 0.2W .
1902.09687	1	1902	2019-02-26T01:15:47	Market-Based Model in CR-WSN: A Q-Probabilistic Multi-agent Learning Approach	[ "cs.MA", "cs.GT" ]	The ever-increasingly urban populations and their material demands have brought unprecedented burdens to cities. Smart cities leverage emerging technologies like the Internet of Things (IoT), Cognitive Radio Wireless Sensor Network (CR-WSN) to provide better QoE and QoS for all citizens. However, resource scarcity is an important challenge in CR-WSN. Generally, this problem is handled by auction theory or game theory. To make CR-WSN nodes intelligent and more autonomous in resource allocation, we propose a multi-agent reinforcement learning (MARL) algorithm to learn the optimal resource allocation strategy in the oligopoly market model. Firstly, we model a multi-agent scenario, in which the primary users (PUs) is the sellers and the secondary users (SUs) is the buyers. Then, we propose the Q-probabilistic multiagent learning (QPML) and apply it to allocate resources in the market. In the multi-agent interactive learning process, the PUs and SUs learn strategies to maximize their benefits and improve spectrum utilization. Experimental results show the efficiency of our QPML approach, which can also converge quickly.	cs.MA	cs	1 Market-Based Model in CR-IoT: A Q- Probabilistic Multi-agent Learning Approach Dan Wang, Wei Zhang, Bin Song*, Xiaojiang Du, IEEE, and Mohsen Guizani Abstract -- The ever-increasingly urban population and its material demands have brought an unprecedented burden to cities. Smart cities leverage emerging technologies like CR-IoT to provide better QoE and QoS for all citizens. However, resource allocation is an important problem in CR-IoT. Currently, this problem is handled by auction theory or game theory. To make CR-IoT nodes smarter and more autonomously allocate resources, we propose a multiagent reinforcement learning (MARL) algorithm to learn the optimal resource allocation strategies in the oligopolistic market model. Firstly, we model a multi-agent scenario with two primary users (PUs) as sellers and multiple secondary users (SUs) as buyers, which is a Bertrand market model. Then, we propose the Q-probabilistic multiagent learning (QPML) and apply it for allocating resource in a market. In the situation of multiagent interaction, the PUs and SUs learn strategies to maximize their revenue. Experimental results prove our multiagent learning approach performs well and convergence also well. Index Terms -- CR-IoT, MARL, resource allocation, market model I. INTRODUCTION addressing urban W ITH the increasing world's population of cities, there has been estimated to increase to over 65 percent by 2050 of the total population [1]. Such rapid population growth and material demands are bringing an unprecedented burden to cities, such as urban infrastructure and social fabric. Therefore, for accelerating urbanization and creating a livable environment for worldwide urban residents, the smart city is proposed [2]. A smart city is one which leverages Information Communications Technology (ICT) and emerging technologies such as the Internet of Things (IoT), cloud computing and big data mining to provide the best quality of life for all citizens while minimizing the consumption of energy and resources. resource-strapped, The IoT is regarded as one of the most critical enabling technologies in the construction of the smart city. With the implementation of smart cities, a large number of various devices have been deployed everywhere [3]. According to the forecasts [4], more than 50 billion devices will be connected to This work has been supported by the National Natural Science Foundation of China (Nos.61772387), the Fundamental Research Funds of Ministry of Education and China Mobile（MCM20170202）, and also supported by the ISN State Key Laboratory. D. Wang and B. Song (corresponding author) are with the State Key Laboratory of Integrated Services Networks, Xidian University, 710071, China (e-mail: [email protected], [email protected]) the IoT by the year 2020, which will provide a seamless connection between things and networks in smart cities [5]. Accessing so many devices in the IoT is posing many new and unprecedented opportunities and challenges, such as the ever- growing data and ever-increasing demands for spectrum resources, as well as the rationality of management and use of spectrum resources. If not addressed the above problems, these will become the bottleneck of the future IoT and smart city, especially the usage problem of spectrum resources [6]. Consequently, to solve these problems, many types of research have been done in industry and academia. In [6], the CR-IoT is proposed, which considers the emerging cognitive radio (CR) technology and cognitive radio network (CRN) in the IoT. CR enable IoT nodes to intelligently and dynamically manage themselves, effectively utilize spectrum communications resource, avoid resource waste, and improve spectrum efficiency. However, there are still existing some problems in real-time resource allocation in CR-IoT, for example, a centralized resource allocation mechanism leads to high network cost and low resource utilization. From another point of view, resources are cost-effective and resource utilization should be maximized at the lowest cost (that is, resources are no longer wasted). A fully distributed and autonomous resource allocation policy should be considered as a feasible solution [7]. Therefore, a sophisticated resource allocation mechanism in CR-IoT is expected, which in addition to being capable of improving resource utilization, can have smarter learning and decision-making capabilities to allocate resource by interacting with the environment. Up to now, game theory and auction theory are two prevalent tools for modeling and analyzing interactive decision-making processes. However, these approaches not very "smart" in decision-making, which lacks autonomy and self-adaptation. Fortunately, the era of artificial intelligence (AI) is coming [8]. The AI technologies pave the way for resource allocation in CR-IoT, like reinforcement learning (RL) method [9]. In this paper, we focus our emphasis on the CR-IoT scenario and propose a resource allocation W. Zhang is with the Science and Technology on Information Systems Engineering Laboratory, National University of Defense Technology, China (email: [email protected]) X. Du is with Dept. of Computer and Information Sciences, Temple University, Philadelphia PA 19122, USA (email: [email protected]) M. Guizani is with Dept. of Electrical and Computer Engineering, University of Idaho, Moscow, ID 83844, USA (email: [email protected]) 2 mechanism from a relatively novel perspective - the market economy. Generally, the market is a natural multi-agent system. As illustrated in Fig. 1, we focus on a CR-IoT scenario, which is considered as an oligopolistic market from an economic perspective. In this paper, from the perspective of the cost and benefit of the market economy, we present a market economy model of resource allocation. We consider one scenario that there are two primary users (PUs) and multiple secondary users (SUs) in the Bertrand market. To make CR-IoT nodes intelligently and autonomously allocate resources, we propose a multiagent reinforcement learning (MARL) algorithm to learn the optimal resource allocation strategies in the market model. To motivate fair allocation and use of resources with a fully distributed autonomous manner in CR-IoT. Our contributions in this work are as follows. (1) We have described the price decision problem of the Bertrand market as a distributed multi-agent dynamic resource allocation problem; (2) We have modeled the two PUs and multiple SUs scenario as an oligopolistic market, from the perspective economic theory; (3) We have analyzed the behaviors of PUs and SUs in the multi-agent system; (4) We have proposed the Q-probabilistic multi-agent learning (QPML) and applied it for allocating resources in the Bertrand market. The remainder of this paper is organized as follows. Section 2 briefly introduces the related work of resource allocation in CR-IoT. Following that, we expound a multi-agent market model design in section 3. Then we propose the Q-probabilistic multi-agent learning and applied it for allocating resource in the Bertrand market in section 4. Section 5 indicates that the performance of our proposed algorithm is better verified by experiments. And section 6 concludes the paper. II. RELATED WORK Resource allocation in CR-IoT is challenging, and a lot of works has been done in academia in recent years. In the following sections, we first review some of the recent kinds of literature that address resource allocation issues in different ways. Then, we discuss the latest research on applying artificial intelligence (AI) algorithms to solve the resource allocation problem. A. Game Theory It analyzes Game theory is a powerful mathematical tool for learning interactive decision-making processes. the interaction strategies among players, and each player gains its own maximum benefit by game [10]. Nowadays, game theory has been widely recognized as a popular tool in many fields such as economics, engineering, science and computer science. Also, leveraging game theory for CRN resource allocation has become a hot research topic. An introduction to the application of game theory in CRN is provided in [11], which details the most fundamental concepts of game theory and explains how to use these concepts in cognitive tasks. Specifically, this paper Fig. 1. A CR-IoT scenario of two primary users (PUs) and multiple secondary users (SUs). This scenario is considered as an oligopolistic market. discusses four kinds of game theory methods in spectrum sharing, namely non-cooperative games, cooperative games, stochastic games, economic games. A detailed overview of cooperative game and non-cooperative game model applications in CR spectrum allocation is presented in [12]. In a similar literature, a non-cooperative spectrum sharing games in CRN has been discussed in [13], such as repeated game, potential game and supermodular game. A resource allocation based on game theory for distributed implementation has been presented in [14], mainly considering the potential games and coexistence problems of two wireless networks operating on the same spectrum. In addition, in [12], a spectrum trading was also discussed. It formulates the game problem between the primary user and multiple secondary users, with the theory of evolutionary game and non-cooperative game. In addition, game theory has been used to address resource allocation issues in various scenarios of the IoT, like device-to-device (D2D) communication scenarios [15]. B. Auction Theory Auction theory and its application are an important branch of economics and now be used to solve various engineering problems [16]. During the auction, the buyers submit the bids and sellers submit the sale. The actual item or service can be regarded as the auction commodity. Then trade according to the preferences of buyers and sellers. The auction can be designed for different types of radio resource allocation issues [17], such as a reverse auction, double auction and combinatorial auction all can be used for a different scenario. Joint spectrum allocation and relay allocation are modeled as hierarchical auctions in [18]. Moreover, the two auction mechanisms are proposed to improve spectrum utilization and increase the use of spatial diversity in the CRNs. In addition, auction theory can be combined with game theory and applied to dynamic resource allocation problems. In [19], the author presented an auction- based approach that uses game theory mechanism in a non- cooperative environment and repeats the game with incomplete information to determine bidders in the auction, thereby making resource provider to obtain maximum benefits and users obtain the resources they need. Both of the above methods are derived from the perspective 3 of economics and can be used to solve the problem of dynamic resource allocation in cognitive wireless networks. Game theory tends to study the behavioral strategies of agents to maximize the benefits. However, the auction theory is more commercial and more dependent on market factors. The auction is usually designed to address resource allocation issues. From the perspective of economics, as the number of agents continues to increase, the ability to rationally allocate resources using these two methods will continue to decline due to inadequate adaptability and self-organization capabilities. C. Reinforcement Learning Recently, the emerging technology reinforcement learning also has decision-making capability, which is similar to the two approaches above. RL is a dynamic online learning method, which mainly includes single-agent reinforcement learning (SARL) and multi-agent reinforcement learning (MARL) [20]. Researchers widely use the SARL to solve resource allocation problems in CRN, such as spectrum sensing, spectrum management and spectrum resource allocation [21]-[23]. A multi-layer cooperative mechanism has been proposed in [24], which uses the popular method RL, Q-learning. In [25], a market-based resource allocation mechanism is proposed that uses demand functions to give agents an allocation preference. In addition, a probabilistic reinforcement learning has been proposed in [26]. From the perspective of economics, the monopolistic market model is modelled to solve the resource allocation problem of one PU and multiple SUs in CR. However, although the RL algorithm has good performance in CRN, numerous communication scenarios are multi-agent systems. Therefore, the method of SARL is not enough to meet the demand. Therefore, the multi-agent approach is expected to solve the resource allocation problem of multi-agent scenarios. The [27] proposed a MARL method to common-pool resource appropriation to predict spatial and temporal resource dynamics allocation. Although there are three common methods for resource allocation in CRN, the self-adaptation and autonomy of RL- based methods still need to be improved. In the CR-IoT network, the intelligent control and distributed methods should be presented in sophisticated scenarios. the resource allocation problem to solve III. MODEL DESIGN AND USERS' BEHAVIORS ANALYSIS Different from the aforementioned works, we consider the scenario where two adjacent PUs have spectrum resources in a market. Meanwhile, there are many SUs bid to meet their own resource needs. In this section, we first briefly introduce the market model and then discuss the behaviors of PUs and SUs in the multi-agent system (MAS) model. A. The Oligopolistic market Model for CR-IoT Generally, the market is a sophisticated multi-agent environment, which is a mixture of monopolistic and oligopolistic markets. With the emergence of the CR-IoT networks, the capabilities of agents and network nodes have improved and became smarter. The users of CR-IoT are more Fig. 2. The interactions learning among PUs and SUs. likely to voluntarily send requests to get resources instead of waiting for resource allocation. Hence, we study the phenomenon of multi-agent interaction in a market economy model, in which the network nodes have the capability of making decisions about resources. The market environment is a natural multi-agent system with a variety of interesting agents. When it comes to market, the competition and allocation of resources is a universal phenomenon. Based on the economic aspects of resource allocation market, the prices of resources are important, which decide whether to exchange resources. Inspired by economic theory and market model approach, we consider the allocation problem of the CR-IoT resource as a market problem, where PUs are regarded as sellers and SUs are regarded as buyers. Under the market, the PUs specify prices for their spectrum resources, and SUs also give bids to fulfil their own needs. The PUs and SUs actively update prices and bids according to behaviors and environmental changes, while learning the best resource allocation strategy. SUs bid on the resource and PUs decide whether to sell resources according to the bids of SUs. The scenario is illustrated in Fig.2, in this market, feedbacks are generated continuously between PUs and SUs during the interaction process. During the learning process, the PUs update prices based on feedback from SUs, and the SUs also update the bids according to the prices of the resources designed by PUs. The prices of all PUs will reach an equilibrium, which will balance the market revenue. The motivation of this paper is to propose a multi-agent reinforcement learning method to establish a model and allocate resources. Next, in this section, we will discuss the oligopolistic model involving PUs and SUs. B. Problem Definition in Market Model As mentioned previously, the Bertrand market model can be used to balance market resources and prices. When the CR-IoT environment is regarded as markets, the network nodes can make price decisions, making CR-IoT systems analogous to the real market. Here, we describe this price decision problem in the Bertrand market as a distribute multi-agent dynamic resource allocation problem (DMDRAP). In DMDRAP, the agents represent network users (PUs and SUs). All agents can be considered as two types of agents where 4 relationship between all PUs are one type of homogeneous agents, and the other SUs are another type of homogeneous agents. Therefore, there is a competitive the same categories. Furthermore, all agents make price decisions autonomously and adaptively. In PUs, each agent's decision depends not only on its own state but also on neighbor agent's states and policies. In general, the states and behaviors of the agents must be considered in the reinforcement learning. However, we considered the stateless multi-agent Q-learning in our model, since the user's previous states have almost no effects on DMDRAP. Each resource allocation does not depend on the previous state. The DMDRAP is described as follows. In this problem, the PUs and SUs can trade resources under the Bertrand market. The SUs are buyers in this market and they decide to pay the price of the PUs. The main goal of DMDRAP is to develop decision strategies for each agent to maximize the benefits and achieve an overall price balance of the entire Bertrand market. Mathematically, we consider this market model has N PUs (P = {𝑝1, 𝑝2, … , 𝑝𝑁}), all PUs submit their prices and they have C subchannels. For simplicity, we focus on the scenario of the two PUs, which have 2 × C subchannels. And meanwhile, there are M SUs (S = {𝑠1, 𝑠2, … , 𝑠𝑀}) are submitting their bids and competing for the spectrum resources in the 2 × C subchannels. Then, we introduce the behaviors of the PUs and SUs. C. Behaviors of SUs and PUs 1) The behaviors of SUs We consider that each SU as an agent, while other users are considered to be external environment. Each agent learns and makes price decisions from feedback by interacting with the environment and maximize the efficiency of each SU. Therefore, the behaviors of the agents should be analyzed and an appropriate multi-agent reinforcement learning method should be proposed. And the details of these contents are as follows. As mentioned previously, the SUs are buyers, which give bids to compete for the spectrum resources. In general, each SU can compete for all subchannel resources of PUs, but each subchannel can only be allocated to one SU, otherwise, channel conflict will occur. Initially, the SU agents obtain the resource status of PUs'. There is a competitive relationship among SUs since each agent cannot communicate. During the interaction and learning of PUs and SUs, the demands of each SU is defined as a set of probabilities in all subchannels resources, as D = {𝑑1, 𝑑2, … 𝑑𝑐, … , 𝑑2𝑐}, (𝑑𝑖ϵ[0,1], ∑ 𝑑𝑖 = 1 ) . Originally, the demand for the subchannel spectrum resources of each SU is random. In a forward transaction market, the buyers (SUs) bid for subchannels from sellers (PUs). There is a dynamic process, the SUs generally updates their strategies based on the prices of the PUs design. The bids of SUs are decided to their demands for resources because demands represent how much they want to pay. Hence, according to the demands of SUs, then bids can be obtained by some policy function with reinforcement learning, as 𝐵𝑖𝑑𝑡 = π(D). 𝑖 However, the demands of agents are always dynamic, so we consider that the demands are static for a period of time. During this period, bidders can sample and estimate their demands by interacting with the PUs, and the trial and error method is effective. In other words, within a certain period of time, we can measure the difference between the actual demand values and the estimated values of the users to evaluate the performance of our estimates. 𝑠 = {𝑎1 𝑠,𝑡, 𝑎2 𝑠,𝑡, … , 𝑎𝐶 𝑠,𝑡}, (𝑎𝑖 𝑠,𝑡 … , 𝑎2𝐶 We denote a vector 𝐴𝑡 𝑠,𝑡 ∈ 𝑠,𝑡 {0,1}) , which is a set of actions of SUs, and each element 𝑎𝑖 is the SU's choices of channels at time 𝑡. Generally, the SUs' strategy has three aspects, namely, choosing the PU, selecting subchannel resource and bidding on the resources. Bids are obtained by an interaction between agents and the environment, which are considered as the sample, as 𝐵𝑖𝑑𝑡 = 𝐴𝑡 𝑠 ∙ D. (1) In the above formula, the actions is a set of discrete values, so the bids here is a set of discrete values based on demands. Based on the above interaction, SUs' revenue can be defined as follows: 𝑆𝑈 = { 𝑅𝑡 𝑈𝑡 − 𝑏𝑖, 𝑖𝑓 𝑏𝑖 𝑖𝑠 max 𝑏𝑖𝑑 𝑎𝑛𝑑 𝑏𝑖 > 𝑝𝑖 −𝑏𝑖, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 （2） where 𝑈𝑡 is the marginal cost of the resources, for simply experiment, we consider 𝑈𝑡 as a constant and 𝑏𝑖 is the bid of SU i at time t. It should be noted that the channel resource quality of each PU is the same, and the number of subchannels owned by PU is the same. The two conditions in the above formula indicate that when the SU's bid is the highest of the same user's bid and is higher than the price of the resource, the channel resource will be purchased, so the reward value is positive. Conversely, when the price is lower than the price, the reward is negative and no resources are not obtained. 2) The behaviors of PUs is complicated because Because the CR-IoT nodes have the capability of intelligent behaviors, which can perceive the environments and make a smart decision about the prices of resources. Learning in a multi-agent environment the environment changes effectively as other PUs learn. However, in two agents' Bertrand market environment, an agent PU learn to design the prices for the C subchannels, while the other agent PU and the SUs are both regarded as a static environment. Hence, the above learning process of the agent PU is affected by the action of another PUs. In a multi-agent system of two PUs, each PU pursues its own target and choose actions independently, while there is no communication between PUs. In addition, the learning framework between PUs is a general- sum game framework. Obviously, in the market, the PU as a seller, its ideal goal is to maximize revenue in a multi-agent system. To achieve this goal in the market, PUs design their own resource prices. Implementation enables the network nodes to allocate resources intelligently. the environment, what PU learn from it is turned into their own behavior, that is, the designed prices of resources. This learning process is actually the process of dealing with SUs' needs and bids. interaction between PU and the In Now we analyze behaviors of PUs, the action space, and the 5 𝑝,𝑡, 𝑎2 𝑝,𝑡 , 𝑎𝐶+2 𝑝,𝑡, … , 𝑎𝐶 𝑝,𝑡 , … , 𝑎2𝐶 𝑝 = {𝑎1 𝑝 = {𝑎𝐶+1 reward function for learning price decision policy. Since there 𝑝,𝑡} are two PUs, so the PU's action is 𝐴1,𝑡 𝑝,𝑡} . and the another PU's action is 𝐴2,𝑡 Each PU's action is the same, that is, their action is to design the price based on the interaction with the SUs and to select SUs to allocate the resources. There is no communication between PUs, but they know each other how to take action in Bertrand market. The agent receives feedback from the external environment (the SUs) and its neighbor (another PU). Thus, other agents' decision policies affect the agent's price decision policy. The PUs design a set of prices, represented as P = 1,𝑡 ∈ [0,1]) is the {𝑃𝑡 2,𝑡 ∈ [0,1]) is first PU's price, and 𝑃𝑡 the another PU's price. The behavior of PUs is to allocate channel resources according to the bid of SUs, and the users with high bids obtain resources. We assume that there is a standard that each SU obtains only one channel resource. To motivate agents to achieve the goal of balancing market prices, supplies and demands, the PUs are using the same reward functions to learn. Each PU receives a positive reward in learning episode. This reward is given by the function: 1,𝑡, 𝑝2 2 = {𝑝1 1,𝑡, … , 𝑝𝐶 2,𝑡, 𝑝2 1,𝑡}, (𝑝𝑖 2,𝑡, … , 𝑝𝐶 2,𝑡}, (𝑝𝑖 1, 𝑃𝑡 2} , where 𝑃𝑡 1 = {𝑝1 𝑃𝑈 = ∑ R𝑡 𝑃𝑈𝑖) (𝑟𝑡 𝑖∈[1,2] . (3) 𝑃𝑈𝑖 is the reward of agent PU and R𝑡 𝑃𝑈 is the sum of where 𝑟𝑡 two agents. The reward 𝑟𝑡 𝑃𝑈𝑖 is defined as follows: 𝑃𝑈𝑖 = max{𝐵𝑖𝑑𝑡 − P − 𝐶𝑜𝑠𝑡 , 0}. 𝑟𝑡 (4) In Bertrand market, we assume that marginal cost is constant. Hence, when ignoring cost (𝐶𝑜𝑠𝑡 = 0), we simplify the formula to calculate the maximum reward of PU. Based on the Bertrand market model above, the revenue of the whole multiagent 𝑃𝑈. Therefore, the system to be maximized, that is, maximize R𝑡 policy of one PU is adjusted based on interaction and feedback with another PU and the SUs. When the number of users seeking to purchase is too small and the supply of resources exceeds demand, the price of the resource should be appropriately reduced. On the contrary, when there are too many users and demand exceeds supply, the price should be appropriately increased to make the entire market equilibrium and curb the development of monopoly industries. In a dealer market, there is a problem for the market seller (PUs) and buyers (SUs) learn how to deals their bids and judge whether the prices are acceptable. Specifically, this issue is a question of pricing policy decisions. Reinforcement learning is a method to solve decision problems. Therefore, in the next section, we will propose the details of the Q-probabilistic multi- agent algorithm to solve the price decision problem. IV. Q-PROBABILISTIC MULTIAGENT LEARNING METHOD We study the phenomenon of price decision problem in the CR-IoT market model of agents' interaction. We use the Bertrand market model, which is an interesting and dynamic multi-agent system. Our model is based on the framework of a general-sum stochastic game in which each agent's payoff also depends on other agents' policy. We are interested in the equilibrium price in the Bertrand market model because we want to design learning agents in resource allocation problem that can learn allocation policies, adaptively. In a single-agent system, reinforcement learning algorithms, such as Q-learning, is usually used to address a problem like a resource allocation in a monopolistic market. Also, in a multi- agent system, the multiagent reinforcement learning (MARL) algorithms have been developed to address the problem of resource allocation in CR-IoT. To address DMDRAP, we proposed a multiagent reinforcement learning algorithm, called Q-probabilistic multiagent learning (QPML). The QPML can solve the problem of resource allocation in an oligopolistic market. We consider a probabilistic policy to control the actions of PUs and SUs. In single-agent reinforcement learning, the Q-learning algorithm is the off-policy approach, which is very popular. Its target is to update Q-value, which is defined as the follows: Q(s, a) = (1 − α)Q(s, a) + α[r + β max 𝑏 𝑄(𝑠′, 𝑏)]. (5) where α is the learning rate, β is the discount factor and r is the reward. The 𝑄(𝑠′, 𝑏) is a future state-action pair. Generally, α determines whether convergence, convergence speed, and accuracy of expectation, which is a trade-off value. In the learning process, the state-action pair that leads to higher reward will be reinforced. In multi-agent Q-learning, we not only maximize our own Q- values but also consider joint optimization since the one agent's Q-value also depend on other agents. In a multi-agent system, the rules for updating the Q-values are as follows: 𝑄1(s, 𝑎1, 𝑎2) = (1 − α)Q(s, 𝑎1, 𝑎2) + α[𝑟1 + β𝜋1(𝑠′)𝑄1(𝑠′)𝜋1(𝑠′)]. (6) 𝑄2(s, 𝑎1, 𝑎2) = (1 − α)Q(s, 𝑎1, 𝑎2) + α[𝑟2 + β𝜋1(𝑠′)𝑄2(𝑠′)𝜋2(𝑠′)]. (7) where the α is the learning rate and β is the discount factor, 𝑟1 is the reward of agent 1, 𝑟2 is the reward of agent 2. Here, (𝜋1, 𝜋2) is the strategy function of 𝑃𝑈1 and 𝑃𝑈2 respectively, and the (𝑄1(𝑠′), 𝑄2(𝑠′)) is a biomatrix game policy. Therefore, each agent maintains two Q tables that the one is its own Q table and the other is another agent's Q table. Furthermore, each PU can observe the rewards and previous actions of other PUs during the learning period. In this paper, a scenario with 2C subchannels is considered where there are two PUs and multiple SUs. Each PU has 2𝐶 − 1 states and actions. Also, each PU learns two Q tables, each of them has (2𝐶 − 1)2 elements. In such a large state and action space, we consider using stateless Q-probabilistic multi-agent algorithm. As mentioned previously, we considered the stateless multi- agent Q-learning in our model, since each resource allocation does not depend on the previous states. Algorithm 1 describes ALGORITHM 1 ALGORITHM 2 6 Loop t= 0, Initialization: All action of multi-agent Random initialization Q-value Algorithm 1: Q-probabilistic multiagent learning (QPML) Algorithm begin end update Q-value function, respectively Calculate rewards Foreach action let t=t+1 end 𝑆 is other agents' action, P is a probabilistic form of Q-value 𝑆) Initialization: t= 0, 𝐴𝑡 Random initialization: 𝑄𝑡+1(𝐴𝑡 Algorithm 2: QPML for SU learning procedure begin end for each period 𝑡 End Calculate P as (11) Choose action 𝐴𝑡 Calculate 𝐵𝑖𝑑𝑡 as (1) Calculate rewards as (2) 𝑟𝑆 ←the reward for action 𝐴𝑡 foreach action 𝐴𝑡 Calculate KL divergence as (12) 𝑠, according to P 𝑠 do end Update Q-value function as (10) 𝑆 of other agents where D is the true demand distribution of SUs and D is the estimated distribution. Then K(𝐷‖𝐷) represents the relative entropy of D to D, that is, the similarity estimation when D is fitted to D. During the learning process, we present the details learning methods for SUs and PUs, respectively. A. Learning procedure of SUs Based on the behaviors of the SUs and the QPML algorithm depicted above, the details of the learning method of SUs will be introduced as follows. In the Bertrand market, each SU bids to compete for the spectrum resources of two PUs. When the SU selects one subchannel, the action is 1, Otherwise 0. Each SU has a selection probability, and this value is determined by the Boltzmann distribution in the QPML algorithm. In the market, SUs are enough active buyers who trade of high liquidity with PUs. Each SU submits its bid, which aims to obtain as many resources as possible. Then, PUs receive the SUs' bids and determine whether to accept the incoming bids and resource asks. The market PU can accept a bid from SUs when bid exceeds the PU's price. Then, the resource allocation can be executed right after the moment that a bid and an ask are matched in a trading. Otherwise, the ask of SU is rejected and the trade is over. In the multi-agent system of the market, the environment is made up of SUs and PUs, and the learning process of SUs and PUs is closed-loop. The SU's reward comes from PU, and PUs receives SU's feedback during the learning process, which is an interactive process. Algorithm 2 shows the price decision-making process of SUs, which repeats at each epoch. B. Learning procedure of SUs Formally, in our multi-agent market model, the PU's price not only depends on their own behaviors but also depends on another agent's actions. When designing two learning agents, we must consider what information is available to agents, such as preference, price vector and another agent's action. We assume an agent PU can observe the actions of other agents (whether resources are allocated). the QPML policy update process, in which we adopt the stateless Q-learning in the learning process of PUs, as follows: 1 (𝐴1,𝑡 𝑄𝑡+1 𝑝 , 𝐴2,𝑡 𝑝 ) = 𝑄𝑡 1(𝐴1,𝑡 𝑝 , 𝐴2,𝑡 𝑝 ) + α[r𝑃 1 − 𝑄𝑡 1(𝐴1,𝑡 𝑝 , 𝐴2,𝑡 𝑝 )]. 𝑄𝑡+1 2 (𝐴1,𝑡 𝑝 , 𝐴2,𝑡 𝑝 ) = 𝑄𝑡 2(𝐴1,𝑡 𝑝 , 𝐴2,𝑡 𝑝 ) + α[r𝑃 2 − 𝑄𝑡 2(𝐴1,𝑡 𝑝 , 𝐴2,𝑡 𝑝 )]. (8) (9) 𝑝 is a set of actions of PU1 and 𝐴2,𝑡 1 is the reward of PU1 and r𝑃 𝑝 where 𝐴1,𝑡 is a set of actions 2 is the reward of of PU2. The r𝑃 2 are the Q value of PU1 and PU2. Furthermore, the 𝑄𝑡 PU2. For two PUs, the action space is 2 × A2, here, we assume 𝑝 = A. Similar to that of the PUs', the stateless 𝐴1,𝑡 Q-learning is also adopted in SUs learning, the following rule is used to update Q-value function: 𝑝 = 𝐴2,𝑡 1 and 𝑄𝑡 𝑄𝑡+1(𝐴𝑡 𝑆) = 𝑄𝑡 (𝐴𝑡 𝑆) + α[𝑟𝑆 − 𝑄𝑡 (𝐴𝑡 𝑆)]. (10) 𝑆 donates a set of action of SUs, that is the bids of the where 𝐴𝑡 SUs. Here, the action (bids) in Q-learning is stochastic and the Q value is not probabilistic, to assure that all actions will be converted to a probabilistic form. We consider Boltzmann distribution to normalize the Q-value, as P = 𝑡 𝑄𝑖𝑗,𝑘 (𝑎) ⁄ 𝜏 𝑒 𝑡 𝑄𝑖𝑗,𝑘 (𝑎) ⁄ 𝜏 ∑ 𝑒 𝑖 (11) 𝑡 (𝑎) is the Q-value for action a when SU i select channel here 𝑄𝑖𝑗 k of the PU j (𝑗 ∈ (1,2)) at time t. The τ is a temperature parameter, which controls the fluctuation of this Boltzmann distribution. Hence, we use P as the estimated demand D to control the actions of agents. In addition, we use KL divergence to measure the difference between the demand D and estimate value D. The KL divergence is a method used to describe the difference between the probability, D and estimate value D as follows: K(𝐷‖𝐷) = ∑ 𝐷𝑙𝑜𝑔 𝑛 𝑖=1 𝐷 𝐷 (12) ALGORITHM 3 t= 0, 𝑝 , 𝐴2,𝑡 Initialization: For each period 𝑡 𝑝 is 𝑃𝑈2's action 𝑝 ), 𝑄𝑡+1 1 (𝐴1,𝑡 𝑝 is 𝑃𝑈1's action, 𝐴2,𝑡 All 𝐴1,𝑡 Random initialization: 𝑄𝑡+1 Algorithm 3: QPML for PU learning procedure begin 2 ←the reward for action 𝐴1,𝑡 1 ←the reward for action 𝐴1,𝑡 ≥ 𝑀2 𝑀 update 𝑃𝑈1 Q-value function as (13) update 𝑃𝑈2 Q-value function as (14) Obtain the reward 𝑟𝑡 r𝑃 r𝑃 Input bids of SUs Calculate payoff 𝑝1,𝑡, 𝑝2,𝑡 𝑝 , 𝐴2,𝑡 𝑝 , 𝐴2,𝑡 𝑃𝑈𝑖 as (4) then else 𝑀1 𝑀 if 𝑀1 𝑀 < 𝑀2 𝑀 , when update 𝑃𝑈1 Q-value function as (15), update 𝑃𝑈2 Q-value function as (16) end end end 2 (𝐴1,𝑡 𝑝 , 𝐴2,𝑡 𝑝 ) 𝑝 of 𝑃𝑈1 𝑝 of 𝑃𝑈2 Initially, the prices of the resources are generated randomly, which is in [0,1]. We assume that the quantity and quality of resources owned by two PUs are the same for each time period. During a learning process, the prices of PUs' resources are constantly changing. That is, the PUs update price strategies to change the prices during the learning process. Specifically, the price is adjusted according to the difference between the maximum bid of each subchannel and actual price. When 𝐵𝑖𝑑𝑡 − P ≥ 0, the PU allocate resource to SUs, otherwise, the trade is failed. When the trade fails, the PU should consider adjusting their own designed prices. Each PU has two update rules, which are to increase prices and lower prices. ≥ 𝑀2 𝑀 𝑀1 𝑀 − . When 2, and when 𝑝1,𝑡 > 0, denote as 𝑝1,𝑡 As mentioned previously, there are M SUs, we assume 𝑀1 SUs obtain resource of 𝑃𝑈1 and 𝑀2 SUs obtain resource of 𝑃𝑈2, 1 and 𝑝2,𝑡 = where 𝑀1 + 𝑀2 ≤ 𝑀. We denote 𝑝1,𝑡 = 𝐵𝑖𝑑𝑡 − 𝑃𝑡 + , otherwise denote 𝐵𝑖𝑑𝑡 − 𝑃𝑡 as 𝑝1,𝑡 , the transaction success rate of 𝑃𝑈1on the market is greater than 𝑃𝑈2. In the next round of trading, 𝑃𝑈1 will continue to raise its prices in order to obtain more benefits. In contrast, 𝑃𝑈2 will lower its prices to get more buyers. However, 𝑃𝑈1 may also continue to suppress prices, which is a malicious competition in the market. In this paper, we don't consider the malicious competition. Then, 𝑃𝑈1 update Q-value function as the following rule: 𝑄𝑡+1 1 (𝑝1,𝑡 + , 𝑝2,𝑡 − ) = 𝑄𝑡 1(𝑝1,𝑡 + , 𝑝2,𝑡 − ) + α[r𝑃 1 − 𝑄𝑡 1(𝑝1,𝑡 + , 𝑝2,𝑡 − )]. (13) 7 𝑀1 𝑀 𝑀2 𝑀 < , 𝑃𝑈1 will lower its prices and 𝑃𝑈2 will raise its When prices. The 𝑃𝑈1 update Q-value function as the following rule: 𝑄𝑡+1 + ) + α[−r𝑃 + ) = 𝑄𝑡 1 (𝑝1,𝑡 − , 𝑝2,𝑡 1 − 𝑄𝑡 − , 𝑝2,𝑡 − , 𝑝2,𝑡 1(𝑝1,𝑡 1(𝑝1,𝑡 + )]. (15) The 𝑃𝑈2 update Q-value function as the following rule: 𝑄𝑡+1 + ) + α[r𝑃 + ) = 𝑄𝑡 2 (𝑝1,𝑡 − , 𝑝2,𝑡 − , 𝑝2,𝑡 − , 𝑝2,𝑡 2 − 𝑄𝑡 + )]. 2(𝑝1,𝑡 2(𝑝1,𝑡 (16) Algorithm 3 shows the learning process of 𝑃𝑈1and 𝑃𝑈2, which repeats at each epoch. V. EXPERIMENTAL EVALUATION In this section, we present experimental to evaluate our proposed Q-probabilistic multi-agent learning method. Our experiments are based on a Window 10 operating system (Intel(R) Core™ i5-7200U CPU @ 2.50GHz 2.71GHz). In Bertrand market model, two PUs interact with many SUs. From a market perspective of the Bertrand model, in order to achieve a balance of the resource market, our agent PUs have the same quality resources. Initially, the prices of the two PUs' resources are random (prices are higher than costs), and then during the learning process, the two PUs change the price decisions based on the multi-agent environment. The Bertrand model can better describe strategies choice because quantity decisions are generated immediately by price choices. The goal of our agents' learning is to determine prices and maximize personal interests in the short term, but the long-term goal is to balance the Bertrand market. interaction with In our experiments, we explored agents' behaviors strategies learning separately and tracked the benefits of each SU and PU We have tested the MAQL approach on multiple CR-IoT network models with different numbers of SUs and different numbers of channels, all of which show similar results, but there are also significant differences. Here, we provide detailed results of the users' revenues with the three cases in Table 1, the experimental results as shown in Fig. 3, Fig. 4 and Fig. 5. Next, we analyze the different effects of the experiments in three cases. Case 1: In this case, we have 10 SUs and 2 PUs. We set the three different channel resources for each PU: (a) 600 channels, THE THREE CASES INDIFFERENT DYNASTIC ENVIRONMENTS TABLE 1 CASE CASE1 (M=10) CASE2 (M=20) CASE3 (M=50) NETWORK ENVIRONMENTS 𝐶1 = 600 𝐶2 = 600 𝐶1 = 600 𝐶2 = 600 𝐶1 = 600 𝐶2 = 600 𝐶1 = 300 𝐶2 = 300 𝐶1 = 300 𝐶2 = 300 𝐶1 = 300 𝐶2 = 300 𝐶1 = 100 𝐶2 = 100 𝐶1 = 100 𝐶2 = 100 𝐶1 = 100 𝐶2 = 100 The 𝑃𝑈2 update Q-value function as the following rule: The M represents the numbers of SUs, 𝐶1 represents the numbers of 𝑃𝑈1's 2 (𝑝1,𝑡 + , 𝑝2,𝑡 − ) = 𝑄𝑡 2(𝑝1,𝑡 + , 𝑝2,𝑡 − ) + α[−r𝑃 2 − 𝑄𝑡 2(𝑝1,𝑡 + , 𝑝2,𝑡 − )]. 𝑄𝑡+1 (14) channels, the 𝐶2 represents the numbers of 𝑃𝑈2's channels 8 (a) the numbers of PUs' channels are 𝐶1 = 600 , 𝐶2 = 600 (b) the numbers of PUs' channels are 𝐶1 = 300 , 𝐶2 = 300 (c) the numbers of PUs' channels are 𝐶1 = 100 , 𝐶2 = 100 Fig. 3. Case 1.The average rewards for PUs , the efficiency of SUs' bids, and the average revenue convergence curve in the market (the number of SUs is N=10). (a) the numbers of PUs' channels are 𝐶1 = 600 , 𝐶2 = 600 (b) the numbers of PUs' channels are 𝐶1 = 300 , 𝐶2 = 300 (c) the numbers of PUs' channels are 𝐶1 = 100 , 𝐶2 = 100 Fig. 4. Case 2. The average rewards for PUs, the efficiency of SUs' bids, and the average revenue convergence curve in the market (the number of SUs is N=20). (b) 300 channels, (c) 100 channels. As shown in Fig. 3 (a), (b), (a) the numbers of PUs' channels are 𝐶1 = 600 , 𝐶2 = 600 (b) the numbers of PUs' channels are 𝐶1 = 300 , 𝐶2 = 300 (c) the numbers of PUs' channels are 𝐶1 = 100 , 𝐶2 = 100 9 Fig. 5. Case 3. The average rewards for PUs, the efficiency of SUs' bids, and the average revenue convergence curve in the market (the number of SUs is N=50). (c), these figures show the average rewards of 𝑃𝑈1 and 𝑃𝑈2 , also show the efficiency of SUs' bids during the learning process. It can be clearly seen that at first PUs' average rewards are very small, and as the numbers of trade increases, their rewards gradually improve. We carefully analyze the behavior of PU and SU in the four sets of images in Fig. 3 (a). The trend of other group diagrams is similar. As shown in the first and second figure of Fig. 3 (a), In the PUs' learning process, each time the SUs bid is received, and then the profit is calculated based on the market price of the design. As the price changes from one iteration to the next, each PU's revenues in the process is also constantly changing. But the total average revenues will converge to the value. This simulation result indicates that the optimal price design of resource can be learned in a Bertrand market model. The third image of Fig. 3 (a) depicts a long-term efficiency of SUs' bidding in multi-agent environments. SUs mainly learn to make bids based on whether or not the feedback information of the resources is obtained, and the pricing behavior of each SU is different. We averaged the rewards for SUs in the curve and analyzed the overall benefits of pricing. As can be seen from the figure, the overall benefit closes to be a value when the iteration is displayed about 600 iterations on the horizontal axis. The number of iterations on the horizontal axis on the chart represents one run of the trade. The last graph in Fig. 3 (a) is the average revenue of all users in the market. The red line， the blue line and the green line represent the revenues of SUs, 𝑃𝑈1 and 𝑃𝑈2 , respectively. From this figure, as the number of iterations was increased, the TABLE 2 THE CONVERGENCE VALUE CASE C PU1 PU2 SUs CASE1 (M=10) CASE2 (M=20) CASE3 (M=50) 100 1.00 ± 0.01 1.06 ± 0.01 0.90 ± 0.00 300 0.64 ± 0.01 0.68 ± 0.00 0.62 ± 0.01 600 0.42 ± 0.01 0.44 ± 0.01 0.41 ± 0.00 100 1.14 ± 0.00 1.24 ± 0.01 0.94 ± 0.01 300 0.89 ± 0.01 0.94 ± 0.01 0.82 ± 0.00 600 0.62 ± 0.01 0.66 ± 0.01 0.6 ± 0.00 100 1.25 ± 0.00 1.36 ± 0.01 0.97 ± 0.00 300 1.10 ± 0.00 1.18 ± 0.01 0.93 ± 0.01 600 0.90 ± 0.01 0.96 ± 0.01 0.83 ± 0.00 M represents the numbers of SUs, C represents the numbers of a PU. Fig. 6. The KL divergence of SUs' demands. three curves gradually converge. The convergence is within 600 iterations. Here, one iteration means one complete strategy updates for all users. After 600 iterations, the three curves closer 10 to their own optimal value, respectively. It shows that our algorithm converges. Fig. 3 (b) and (c) also depicts the users' rewards and the average revenue of users in the market. The behaviors of SUs and PUs are similar to that described previously, but their rewards, market revenues, and convergence times are all different. In Fig.4 and Fig.5 of the Case 2 and Case 3, respectively, the behaviors of SUs and PUs are similar for the case 1. We summarize the results of the final convergence values in the three cases in Table 2, and then analyzed the behaviors of users in the market under different conditions. The conclusions we observed are as follows: (1) We observe that the number of users is the same, the smaller the number of channels, and the greater the average market revenue and the user's personal reward. For example, in case 1, the number of SUs is 10, when 𝑃𝑈1 and 𝑃𝑈2 all have 100 channels, their average reward and SU's bidding efficiency are the best. It can be seen from this that the smaller the number of resources in the market, the greater the market gain. And resources are allocated reasonably to SUs, no waste. If the amount of resources is too large, it will become cheaper and the overall market income will decrease. (2) In three cases, our algorithm does not reduce its performance as the number of users increases. For example, when the number of channels is 600, the convergence iterations of the algorithm in the three cases are about 500 iterations. This proves that the fullness of market users will not affect our resource allocation. (3) As the number of SUs increases, the average reward for PUs and SUs increases gradually, indicating that our algorithm can allocate resources and maximize market efficiency in the high-volume user market. The more users, the better the performance of the algorithm. In addition, in order to verify the validity of the ability of the SUs to estimate demand after interacting with the dynamic environment, we propose to use KL divergence to calculate the similarity between the probability distribution of demand D and estimated value D . As illustrated in Fig. 6, it plots the KL divergence curve between the original resources demand distribution of the SUs and the estimated resources distribution of agents' online learning. The curve fluctuates dynamically between the value of 1.7 and 1.9. It can be seen that the KL divergence of the actual demand and the estimated demand is small, so their similarity is greater. Therefore, it is effective to estimate the demand function using the Boltzmann distribution normalized Q value. In the proposed multi-agent learning method, both PUs know each other behaviors and benefits, so they can determine the price that can bring the most profit for themselves, that is, each user will adjust their own prices. These prices adapt to the price of other users on the market. It can be indicated by the above experimental results that the proposed method is effective in the multiple users' scenario of CR-IoT, and the implementation of the method suppresses the monopolistic behavior of the merchants in the resource market and maintains the market balance. VI. CONCLUSION AND FUTURE WORK This paper has proposed the use of the Q-probabilistic multiagent learning to solve dynamic resource allocation problem in a sophisticated market model. From the perspective of economic theory, we have tackled a distribute multi-agent dynamic resource allocation problem. Each learning agent PU can observe the behavior of other PUs, which design their own resource price in Bertrand market model. Also, the SUs compete for resources by bidding. The empirical results of this paper plausibly suggest that the interactive learning between multi-agent SUs and PUs is a promising method for balancing market price, optimizing resource allocation and suppressing monopolistic behavior in resource market. In this paper, we have only designed a multi-agent system with two PUs and multiple SUs in CR-IoT, but can be extended to n-PUs scenario in future work. REFERENCES [1] "Harnessing the Smart City opportunity.", KPMG, 2017, vailable online: http://www.un.org/en/index. [2] " Where's Wally? In Search of Citizen Perspectives on the Smart City.", Sustainability, Vol. 8, pp. 207. [3] X. He, et al. "QoE-Driven Big Data Architecture for Smart City." IEEE Communications Magazine, vol. 56, no. 2, Feb. 2018, pp. 88-93. [4] S. H. Shah, and I. Yaqoob. "A Survey: Internet of Things (IOT) Technologies, Applications and Challenges.", Smart Energy Grid Engineering IEEE, Aug. 2016, pp. 381-385. I. A. T. Hashem, V. Chang, N. B. Anuar, et al. "The role of big data in smart city." International Journal of Information Management, vol. 36, no. 5, Oct. 2016, pp. 748-758. [5] [6] P. Rawat, K. D. Singh, and J. M. Bonnin. "Cognitive radio for M2M and Internet of Things: A survey." Computer Communications, vol. 94, Nov. 2016, pp. 1-29. [7] Bae, B., Park, J. and Lee, S. "A Free Market Economy Model for Resource Management in Wireless Sensor Networks" Wireless Sensor Network, Vol.7, No.6, June 2015, pp. 76-82. [8] Dan Wang, Dong Chen, Bin Song, et al. "From "IoT" to "5G I-IoT": The Next Generation IoT-Based Intelligent Algorithms and 5G Technologies." IEEE Communications Magazine, 2018, Accept. [9] X. Wang, X. Li, and V. C. M. Leung. "Artificial Intelligence-Based Techniques for Emerging Heterogeneous Network: State of the Arts, Opportunities, and Challenges." IEEE Access, vol. 3, Aug. 2015, pp. 1379-1391. [10] A. Mackenzie and L. Dasilva. "Game theory for wireless engineers." Synthesis Lectures on Communications, vol. 1, 2006, pp. 1-86. [11] B. Wang, Y. Wu, and K. J. R. Liu. "Game theory for cognitive radio networks: An overview." Computer Networks, vol. 54, no. 14, Oct. 2010, pp. 2537-2561. [12] A. A. Anghuwo, et al. "Spectrum Allocation Based on Game Theory in Cognitive Radio Networks." Journal of Networks, vol. 8, no. 3, Mar. 2013, pp. 1-9. [13] Kim, Sungwook, et al. "Game Theory for Cognitive Radio Networks." Game Theory Applications in Network Design (2014). [14] Re, E. Del, et al. "A power allocation strategy using Game Theory in Cognitive Radio networks." ICST International Conference on Game Theory for Networks IEEE Press, Jun. 2009, pp.117-123. [15] Kaur, Veerpal, and S. Thangjam. "A stochastic geometry analysis of RF energy harvesting based D2D communication in downlink cellular networks." India International Conference on Information Processing IEEE, 2017, pp. 1-5. [16] S. Parsons, J. A. Rodriguez-Aguilar, and M. Klein. "Auctions and bidding:A guide for computer scientists. " Acm Computing Surveys, vol. 43, no. 2, Jan. 2011, pp. 1-59. [17] Y. Zhang, et al. "Auction-based resource allocation in cognitive radio systems." Communications Magazine IEEE, vol. 50, no. 11, Nov. 2012, pp. 08-120. 11 [18] X. Wang, et al. "Auction-based resource allocation for cooperative cognitive radio networks." Computer Communications, vol. 97, Jan,2017, pp. 40-51. [19] A. Nezarat, and G. Dastghaibifard. "Efficient Nash equilibrium resource allocation based on game theory mechanism in cloud computing by using auction." International Conference on Next Generation Computing Technologies IEEE, Jan. 2016, pp. e0138424. [20] L. Busoniu, R. Babuska, and B. De Schutter. "A Comprehensive Survey of Multiagent Reinforcement Learning." IEEE Transactions on Systems Man & Cybernetics Part C, vol. 38, no. 2, Mar. 2008, pp. 156-172. [21] Yau, K. L., et al., "Application of Reinforcement Learning in Cognitive Radio Networks: Models and Algorithms." The Scientific World Journal, Jun. 2014. [22] Ling, Mee Hong, et al. "Application of reinforcement learning for security enhancement in cognitive radio networks." Applied Soft Computing vol. 37, 2015, pp. 809-829. [23] B. F. Lo, and I. F. Akyildiz. "Reinforcement learning-based cooperative sensing in cognitive radio ad hoc networks." IEEE, International Symposium on Personal Indoor and Mobile Radio Communications IEEE, Dec. 2010, pp.2244-2249. [24] J. Peng, et al. "Multi-relay Cooperative Mechanism with Q-Learning in Cognitive Radio Multimedia Sensor Networks." IEEE, International Conference on Trust, Security and Privacy in Computing and Communications IEEE, Nov. 2011, pp. 1624-1629. [25] Yue Zhang, Bin Song, Xiaojiang Du and Mohsen Guizani. "Monopolistic Models for Dynamic Spectrum Access in Cognitive Radio Networks: A Probabilistic Reinforcement Learning Approach." IEEE Access, 2017. [26] Zhang, Y., et al. "Market Model for Resource Allocation in Emerging Sensor Networks with Reinforcement Learning." Sensors, vol. 16, Dec. 2016, pp. 2021. [27] Perolat, J., Leibo, Joel Z., et al. "Vinicius; A multi-agent reinforcement learning model of common-pool resource appropriation", 2017, arXiv170706600P.
1608.04851	1	1608	2016-08-17T04:36:27	Essentials of an Integrated Crowd Management Support System Based on Collective Artificial Intelligence	[ "cs.MA" ]	The simulation of the dynamical behavior of pedestrians and crowds in spatial structures is a consolidated research and application context that still presents challenges for researchers in different fields and disciplines. Despite currently available commercial systems for this kind of simulation are growingly employed by designers and planners for the evaluation of alternative solutions, this class of systems is generally not integrated with existing monitoring and control infrastructures, usually employed by crowd managers and field operators for security reasons. This paper introduces the essentials and the related computational frame- work of an Integrated Crowd Management Support System based on a Collective Artificial Intelligence approach encompassing (i) interfaces from and to monitored and controlled environments (respectively, sen- sors and actuators), (ii) a set of software tools supporting the analysis of pedestrians and crowd phenomena taking place in the environment to feed a (iii) faster than real-time simulation of the plausible evolution of the current situation in order to support forms of inference provid- ing decision support to crowd managers, potentially directly controlling elements of the environment (e.g. blocking turnstiles, escalators), com- municating orders to operators on the field or trying to influence the pedestrians by means of dynamic signage or audible messages.	cs.MA	cs	Essentials of an Integrated Crowd Management Support System Based on Collective Artificial Intelligence Giuseppe Vizzari1 and Stefania Bandini1,2 1 Complex Systems and Artificial Intelligence Research Centre, {giuseppe.vizzari, stefania.bandini}@disco.unimib.it 2 Researche Center for Advanced Science and Technology, University of Milano-Bicocca The University of Tokyo Abstract. The simulation of the dynamical behavior of pedestrians and crowds in spatial structures is a consolidated research and application context that still presents challenges for researchers in different fields and disciplines. Despite currently available commercial systems for this kind of simulation are growingly employed by designers and planners for the evaluation of alternative solutions, this class of systems is generally not integrated with existing monitoring and control infrastructures, usually employed by crowd managers and field operators for security reasons. This paper introduces the essentials and the related computational frame- work of an Integrated Crowd Management Support System based on a Collective Artificial Intelligence approach encompassing (i) interfaces from and to monitored and controlled environments (respectively, sen- sors and actuators), (ii) a set of software tools supporting the analysis of pedestrians and crowd phenomena taking place in the environment to feed a (iii) faster than real-time simulation of the plausible evolution of the current situation in order to support forms of inference provid- ing decision support to crowd managers, potentially directly controlling elements of the environment (e.g. blocking turnstiles, escalators), com- municating orders to operators on the field or trying to influence the pedestrians by means of dynamic signage or audible messages. Keywords: pedestrians and crowd simulation, monitoring and control, crowd management 1 Introduction The simulation of the movement of pedestrians and crowds in spatial structures is a consolidated research and application context that still presents challenges for researchers in different fields and disciplines: both the automated analysis and the synthesis of pedestrian and crowd behaviour, as well as attempts to in- tegrate these complementary and activities [26], present open challenges as well as significant opportunities in a smart environment perspective [20]. Although 6 1 0 2 g u A 7 1 ] A M . s c [ 1 v 1 5 8 4 0 . 8 0 6 1 : v i X r a the currently available commercial tools are used on a day-to-day basis by de- signers and planners, there is still room for innovations in models, to improve their effectiveness in modeling pedestrians and crowd phenomena, their expres- siveness (i.e. simplifying the modeling activity or introducing the possibility of representing phenomena that were still not considered by existing approaches) and efficiency. Moreover, and this is the particular focus of this paper, they are generally not integrated with existing monitoring and control infrastructures, that are employed by crowd managers and field operators. As testified by an analysis of a recent disaster related to a crowding situ- ation [13], a more systemic perspective on the overall work flow leading from planning and preparation, to monitoring and control of a crowded event, can help preventing these kinds of situations. This idea of a more thorough and comprehensive computational support to crowd management is also discussed in [31]. This paper introduces the notion of an Integrated Crowd Management Support System, a system encompassing both interfaces from and to a moni- tored and controlled environment (respectively sensors and actuators) as well as a set of software tools supporting the analysis of phenomena taking place in the environment, a faster than real-time simulation of the plausible evolution of the current situation in order to support forms of inference providing deci- sion support to crowd managers, potentially directly controlling elements of the environment (e.g. blocking turnstiles, escalators), communicating orders to op- erators on the field or trying to influence the pedestrians by means of dynamic signage or audible messages. The following Section provides a brief discussion of the relevant state of the art in pedestrians and crowd automated analysis and synthesis, then some early works providing relevant examples of prototypes of systems partly implementing a crowd control center will be presented and dis- cussed in order to sketch a general architecture. Conclusions and future works end the paper. 2 Pedestrians and Crowd Analysis and Synthesis The research on pedestrian dynamics is basically growing on two lines. On the analysis side, literature is producing methods for an automatic extraction of pedestrian trajectories (e.g. [5,6]), charactertization of pedestrian flows (e.g. [15]) automatic recognition of pedestrian groups [22], recently gaining importance due to differences in trajectories, walking speeds and space utilization [2]. The synthesis side -- where the contributions of this work are concentrated -- has been even more prolific, starting from preliminary studies and assumptions provided by [14] or [11] and leading to quite complex, yet not usually validated, models exploring components like panic [7] or other emotional variables. Most of the literature has been focussed on the reproduction of the physics of the system, so on the lowest level, where a significant knowledge on the fun- damental diagram achieved with different set of experiments and in different environment settings (see, e.g, [32,33]) allows a robust validation of the models. Literature of this level can be classified regarding the scope of the modeling approach. Macroscopic models describe the earliest approach to pedestrian mod- eling, based on analogies between behavior of dense crowds and kinetic gas [14] or fluids [11], but essentially abstracting the concept of individual. A microscopic approach is instead focused on modeling the individual behavior, effectively im- proving the simulations precision also in low density situations. The microscopic approach is as well categorized in two classes describing the representation of space and movement: continuous models simulate the dynamics by means of a force-based approach, which finds its basis on the well-known social force model by [12]. These models design pedestrians as particles moved by virtual forces, that drive them towards their destination and let them avoid obstacles or other pedestrians. Latest models in this class are the centrifugal force model by [9] and the stride length adaptation model by [21]. Other examples consider also groups of pedestrians by means of attractive forces among persons inside the group [19]. The usage of a discrete environment is mostly employed by the cellular au- tomata (CA) based models, and describes a less precise approach in the re- production of individuals trajectories that, on the other side, is significantly more efficient and still able to reproduce realistic aggregated data. This class derives from vehicular modeling and some models are direct adaptations of traf- fic ones, describing the dynamics with ad hoc rules (e.g. [3,4]). Other models employs the well-know floor field approach from [8], where a static floor field drives pedestrians towards a destination and a dynamic floor field is used to generate a lane formation effect in bi-directional flow. [23] is an extension of the floor field model, introducing the anticipation floor field used to manage cross- ing trajectories and encourage the lane formation. [16] discussed methods to deal with different speeds, in addition to the usage of a finer grid discretization that decreases the error in the reproduction of the environment, but significantly impacts on the efficiency of the model. An alternative approach to represent dif- ferent speeds in a discrete space is given by [1]. [27] is another extension of the floor-field model, where groups of pedestrians are also considered. 3 Towards an Integrated Crowd Management Support System The first work proposing the idea of employing faster than real-time pedestrian simulation to support decision making activities of crowd managers was proposed in [10]: the authors were essentially considering the idea of feeding a CA-based simulator with data from computer vision algorithms characterizing a given evac- uation situation. The simulation of a short-term future could detect plausible congested positions and, therefore, actions influencing pedestrian movement (i.e. activation of sound and/or optical signals) could be enacted to prevent them. As for a more recent work by the same research group [24], the authors also de- scribe a hardware (FPGA) implementation of the system, for achieving increased performance of both the computer vision and CA simulation systems. Fig. 1. A high level architecture of an Integrated Crowd Management Support System. Another relevant prototypical implementation of this kind of system was carried out in the context of the Hermes project [29,28]: once again, also this work basically considered evacuation situations, in which the intention of pedestrians is generally to vacate the area in the shortest possible time, which generally means following the shortest path. Both the above mentioned works present separate blocks for the interpreta- tion of inputs to the system (mostly computer vision techniques for analyzing video footage and perform pedestrian detection, counting, tracking) and the short-term simulation of the plausible evolution of the system. Both works sim- ply suggest, but do not actually describe in details, the presence of a module actually coordinating these modules: triggering the start of a simulation accord- ing to the passage of time (i.e. repeatedly perform a 10 minute simulation every minute) or due to the detection of a certain kind of event (e.g. an increased incoming flow from a certain entrance), interpreting simulation results (e.g. to detect and characterize congested areas), correlating these results and available information on the spatial structure of the environment (e.g. regions and pas- sages connecting them) to support decision making activities by users, that are generally supposed to be "in -- the -- loop" (i.e. actually taking final decisions on Faster than Real-Time SimulatorCrowd Analysis toolsCorrelation and crowd control logic Sensors (turnstile counters, video cameras, RFID devices, etc.)Actuators (doors, signs, audible signals, turnstiles, escalators, etc.)Crowd manager user interfaceIntegrated CrowdManagement Support SystemMonitored and controlled environmentactions to be enacted). The overall resulting architecture can be described by the schema depicted in Figure 1: an Integrated Crowd Management Support System is essentially responsible to support the control of an area subject to crowding conditions calling for a careful planning of the operations, monitoring of actual crowd dynamics, decision support to the crowd manager (and poten- tially supporting a smooth communication with field operators) and potential enactment of actions (e.g. different forms of information provisioning to pedes- trians or actual changes in the environmental conditions) to influence overall crowd dynamics. A different and additional line of work, described in [18] and [17], and devel- oped in the context of a Spanish research project, does not employ pedestrian simulation systems, but it rather relies on consolidated results in the area of route optimization to provide guidance in smart cities, large public events and evacuation situations. This work is in earlier stages compared to the previously described ones, but it is interesting to mention both to suggest that predictions on system evolution can be carried out also by more coarse grained models of the environment, and also to stress the fact that a growing number of additional sensors can be considered to complement and enrich the information acquired by means of computer vision tools. These works also suggest that different Crowd Management Support Systems could be fruitfully connected, both to leverage the monitoring capabilities and be able to foresee changes in the local situations whenever the monitored areas are connected and reachable from one another, and more generally to support a more comprehensive and systemic view of the overall environment, reducing at the same time the computational costs related to the simulation of a potentially very high number of pedestrians. Finally, this line of work suggests that, besides supporting decisions of ex- perts, either designers or crowd managers, this kind of system could be used to provide a distributed awareness of the state of the overall (large scale) environ- ment to actually trigger collaborative forms of diffused, bottom-up behaviours that, at the same time, lead to individually and collectively beneficial choices: by being informed about the state of a transportation network and informing (either explicitly or implicitly, through a behaviour implicit form of communication) the overall system, pedestrians could avoid situations of congestion (causing delays or even simply unpleasant travels), making a reasonable individual choice that could also play a role in limiting the growth of the already existing congestion. Although research on this line of work in the transportation area is still quite open [25], the idea of investigating a collective intelligence scheme [30] in which different forms of AI techniques are employed is appealing and promising. 4 Conclusions and Future Developments This paper has introduced the notion of an Integrated Crowd Management Sup- port System, an comprehensive set of tools supporting the both analysis of crowd dynamics and its simulation (faster than real-time) for supporting short term predictions on the evolution of the monitored environment. A control logic mod- ule is responsible both to trigger the simulation, to analyze its results and supply the crowd manager indications on potential issues and potential mitigation ac- tions. Whereas the state of the art presents abundant and significant results both in the analysis of crowd dynamics and in their simulation (even respect- ing the faster than real-time speed requirement) more work has to be carried out to define a proper control logic for a Crowd Management Support System: this module should encompass both the ability to analyze and correlate the flow of information coming from sensors, to trigger simulation and other forms of predictions, but also to support decisions on the most appropriate mitigation actions to prevent or solve potential issues. Acknowledgements The authors would like to thank Davide Bogni, Guido Bonazza, Luca Crociani, Andrea Gorrini and Sara Manzoni for the fruitful discussions that contributed to the conceptualization of the Crowd Management Support System concept and architecture. References 1. Bandini, S., Crociani, L., Vizzari, G.: Heterogeneous Pedestrian Walking Speed in Discrete Simulation Models. In: Chraibi, M., Boltes, M., Schadschneider, A., Seyfried, A. (eds.) Traffic and Granular Flow '13, pp. 273 -- 279. Springer Interna- tional Publishing (2015) 2. Bandini, S., Gorrini, A., Vizzari, G.: Towards an integrated approach to crowd analysis and crowd synthesis: A case study and first results. Pattern Recognition Letters 44, 16 -- 29 (2014), http://dx.doi.org/10.1016/j.patrec.2013.10.003 3. Blue, V., Adler, J.: Modeling Four-Directional Pedestrian Flows. Transportation Research Record: Journal of the Transportation Research Board 1710(-1), 20 -- 27 (Jan 2000) 4. Blue, V.J., Adler, J.L.: Cellular Automata Microsimulation of Bi-Directional Pedestrian Flows. Transportation Research Record 1678, 135 -- 141 (1999) 5. Boltes, M., Zhang, J., Seyfried, A., Steffen, B.: T-junction: Experiments, trajectory collection, and analysis. In: Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on. pp. 158 -- 165 (Nov 2011) 6. Boltes, M., Seyfried, A.: Collecting pedestrian trajectories. Neurocomputing 100, 127 -- 133 (2013), http://dx.doi.org/10.1016/j.neucom.2012.01.036 7. Bosse, T., Hoogendoorn, M., Klein, M.C.A., Treur, J., van der Wal, C.N., van Wissen, A.: Modelling collective decision making in groups and crowds: Integrat- ing social contagion and interacting emotions, beliefs and intentions. Autonomous Agents and Multi-Agent Systems 27(1), 52 -- 84 (2013) 8. Burstedde, C., Klauck, K., Schadschneider, A., Zittartz, J.: Simulation of pedes- trian dynamics using a two-dimensional cellular automaton. Physica A: Statistical Mechanics and its Applications 295(3 - 4), 507 -- 525 (2001) 9. Chraibi, M., Seyfried, A., Schadschneider, A.: Generalized centrifugal-force model for pedestrian dynamics. Phys. Rev. E 82(4), 46111 (2010) 10. Georgoudas, I.G., Sirakoulis, G.C., Andreadis, I.: An anticipative crowd manage- ment system preventing clogging in exits during pedestrian evacuation processes. IEEE Systems Journal 5(1), 129 -- 141 (2011), http://dx.doi.org/10.1109/JSYST. 2010.2090400 11. Helbing, D.: A fluid dynamic model for the movement of pedestrians. arXiv preprint cond-mat/9805213 (1998) 12. Helbing, D., Molnar, P.: Social force model for pedestrian dynamics. Physical re- view E (1995) 13. Helbing, D., Mukerji, P.: Crowd disasters as systemic failures: analysis of the love parade disaster. EPJ Data Science 1(1), 1 -- 40 (2012), http://dx.doi.org/ 10.1140/epjds7 14. Henderson, L.F.: The Statistics of Crowd Fluids. Nature 229(5284), 381 -- 383 (Feb 1971) 15. Khan, S.D., Bandini, S., Basalamah, S.M., Vizzari, G.: Analyzing crowd behav- ior in naturalistic conditions: Identifying sources and sinks and characterizing main flows. Neurocomputing 177, 543 -- 563 (2016), http://dx.doi.org/10.1016/ j.neucom.2015.11.049 16. Kirchner, A., Klupfel, H., Nishinari, K., Schadschneider, A., Schreckenberg, M.: Discretization effects and the influence of walking speed in cellular automata mod- els for pedestrian dynamics. Journal of Statistical Mechanics: Theory and Experi- ment 2004(10), P10011 (2004) 17. Lujak, M., Giordani, S., Ossowski, S.: Distributed safety optimization in evacuation of large smart spaces by marin lujak, stefano giordani and sascha ossowski. In: Proceedings ot the Ninth Workshop on Agents in Traffic and Transportation (ATT 2016) (2016) 18. Lujak, M., Ossowski, S.: Intelligent People Flow Coordination in Smart Spaces, pp. 34 -- 49. Springer International Publishing, Cham (2016), http://dx.doi.org/ 10.1007/978-3-319-33509-4_3 19. Qiu, F., Hu, X.: Modeling group structures in pedestrian crowd simulation. Simu- lation Modelling Practice and Theory 18(2), 190 -- 205 (2010) 20. Sassi, A., Borean, C., Giannantonio, R., Mamei, M., Mana, D., Zambonelli, F.: Crowd steering in public spaces: Approaches and strategies. In: 14th IEEE In- ternational Conference on Ubiquitous Computing and Communications, IUCC 2015, Liverpool, United Kingdom, October 26-28, 2015. pp. 2098 -- 2105 (2015), http://dx.doi.org/10.1109/CIT/IUCC/DASC/PICOM.2015.312 21. von Sivers, I., Koster, G.: Dynamic Stride Length Adaptation According to Utility And Personal Space p. 30 (Jan 2014) 22. Solera, F., Calderara, S., Cucchiara, R.: Socially constrained structural learning for groups detection in crowd. IEEE Trans. Pattern Anal. Mach. Intell. 38(5), 995 -- 1008 (2016), http://dx.doi.org/10.1109/TPAMI.2015.2470658 23. Suma, Y., Yanagisawa, D., Nishinari, K.: Anticipation effect in pedestrian dynam- ics: Modeling and experiments. Physica A: Statistical Mechanics and its Applica- tions 391(1-2), 248 -- 263 (Jan 2012) 24. Tsiftsis, A., Georgoudas, I.G., Sirakoulis, G.C.: Real data evaluation of a crowd supervising system for stadium evacuation and its hardware implementation. IEEE Systems Journal 10(2), 649 -- 660 (2016), http://dx.doi.org/10.1109/JSYST. 2014.2370455 25. Varga, L.: On intention-propagation-based prediction in autonomously self- adapting navigation. Scalable Computing: Practice and Experience 16(3), 221 -- 232 (2015), http://www.scpe.org/index.php/scpe/article/view/1098 26. Vizzari, G., Bandini, S.: Studying pedestrian and crowd dynamics through in- tegrated analysis and synthesis. IEEE Intelligent Systems 28(5), 56 -- 60 (2013), http://dx.doi.org/10.1109/MIS.2013.135 27. Vizzari, G., Manenti, L., Crociani, L.: Adaptive Pedestrian Behaviour for the Preservation of Group Cohesion. Complex Adaptive Systems Modeling 1(7) (2013) 28. Wagoum, A.U.K., Chraibi, M., Mehlich, J., Seyfried, A., Schadschneider, A.: Ef- ficient and validated simulation of crowds for an evacuation assistant. Journal of Visualization and Computer Animation 23(1), 3 -- 15 (2012), http://dx.doi.org/ 10.1002/cav.1420 29. Wagoum, A.U.K., Steffen, B., Seyfried, A., Chraibi, M.: Parallel real time compu- tation of large scale pedestrian evacuations. Advances in Engineering Software 60, 98 -- 103 (2013), http://dx.doi.org/10.1016/j.advengsoft.2012.10.001 30. Weld, D.S., Mausam, Lin, C.H., Bragg, J.: Handbook of Collective Intelligence, chap. Artificial Intelligence and Collective Intelligence. MIT Press (2015) 31. Wijermans, N., Conrado, C., van Steen, M., Martella, C., Li, J.: A land- scape of crowd-management support: An integrative approach. Safety Science 86, 142 -- 164 (2016), http://www.sciencedirect.com/science/article/pii/ S0925753516300030 32. Zhang, J., Klingsch, W., Rupprecht, T., Schadschneider, A., Seyfried, A.: Empirical study of turning and merging of pedestrian streams in T-junction. ArXiv e-prints p. 8 (2011) 33. Zhang, J., Klingsch, W., Schadschneider, A., Seyfried, A.: Ordering in bidirec- tional pedestrian flows and its influence on the fundamental diagram. Journal of Statistical Mechanics: Theory and Experiment (02), 9 (2011)
1912.05676	1	1912	2019-12-11T22:39:51	Biases for Emergent Communication in Multi-agent Reinforcement Learning	[ "cs.MA", "cs.CL", "cs.LG" ]	We study the problem of emergent communication, in which language arises because speakers and listeners must communicate information in order to solve tasks. In temporally extended reinforcement learning domains, it has proved hard to learn such communication without centralized training of agents, due in part to a difficult joint exploration problem. We introduce inductive biases for positive signalling and positive listening, which ease this problem. In a simple one-step environment, we demonstrate how these biases ease the learning problem. We also apply our methods to a more extended environment, showing that agents with these inductive biases achieve better performance, and analyse the resulting communication protocols.	cs.MA	cs	Biases for Emergent Communication in Multi-agent Reinforcement Learning Yoram Bachrach DeepMind London, UK [email protected] Tom Eccles DeepMind London, UK [email protected] Guy Lever DeepMind London, UK [email protected] Thore Graepel DeepMind London, UK [email protected] Angeliki Lazaridou DeepMind London, UK [email protected] Abstract We study the problem of emergent communication, in which language arises because speakers and listeners must communicate information in order to solve tasks. In temporally extended reinforcement learning domains, it has proved hard to learn such communication without centralized training of agents, due in part to a difficult joint exploration problem. We introduce inductive biases for positive signalling and positive listening, which ease this problem. In a simple one-step environment, we demonstrate how these biases ease the learning problem. We also apply our methods to a more extended environment, showing that agents with these inductive biases achieve better performance, and analyse the resulting communication protocols. Introduction 1 Environments where multiple learning agents interact can model important real-world problems, ranging from multi-robot or autonomous vehicle control to societal social dilemmas [4, 27, 23]. Further, such systems leverage implicit natural curricula, and can serve as building blocks in the route for constructing aritficial general intelligence [22, 1]. Multi-agent games provide longstanding grand-challenges for AI [17], with important recent successes such as learning a cooperative and competitive multi-player first-person video game to human level [14]. An important unsolved problem in multi-agent reinforcement learning (MARL) is communication between independent agents. In many domains, agents can benefit from sharing information about their beliefs, preferences and intents with their peers, allowing them to coordinate joint plans or jointly optimize objectives. A natural question that arises when agents inhibiting the same environment are given a communication channel without an agreed protocol of communication is that of emergent communication [34, 37, 8]: how would the agents learn a "language" over the joint channel, allowing them to maximize their utility? The most naturalistic model for emergent communication in MARL is that used in Reinforced Inter-Agent Learning (RIAL) [8] where agents optimize a message policy via reinforcement from the environment's reward signal. Unfortunately, straightforward implementations perform poorly [8], 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada. driving recent research to focus on differentiable communication models [8, 30, 35, 12], even though these models are less generally applicable or realistic. RIAL offers the advantage of having decentralized training and execution; similarly to human communication, each agent treats others as a part of its environment, without the need to have access to other agents' internal parameters or to back-propagate gradients "through" parameters of others. Further, agents communicate with discrete symbols, providing symbolic scaffolding for extending to natural language. We build on these advantages, while facilitating joint exploration and learning via communication-specific inductive biases. We tackle emergent communication through the lens of Paul Grice [10, 31], and capitalize on the dual view of communication in which interaction takes place between a speaker, whose goal is to be informative and relevant (adhering to the equivalent Gricean maxims), and a listener, who receives a piece of information and assumes that their speaker is cooperative (providing informative and relevant information). Our methodology is inspired by the recent work of Lowe et al. [25], who proposed a set of comprehensive measures of emergent communication along two axis of positive signalling and positive listening, aiming at identifying real cases of communication from pathological ones. Our contribution: we formulate losses which encourage positive signaling and positive listening, which are used as auxiliary speaker and listener losses, respectively, and are appended to the RIAL communication framework. We design measures in the spirit of Lowe et al. [25] but rather than using these as an introspection tool, we use them as an optimization objective for emergent communication. We design two sets of experiments that help us clearly isolate the real contribution of communication in task success. In a one-step environment based on summing MNIST digits, we show that the biases we use facilitate the emergence of communication, and analyze how they change the learning problem. In a gridworld environment based on search of a treasure, we show that the biases we use make communication appear more consistently, and we interpret the resulting protocol. 1.1 Related Work Differentiable communication was considered for discrete messages [8, 30] and continuous mes- sages [35, 12], by allowing gradients to flow through the communication channel. This improves performance, but effectively models multiple agents as a single entity. In contrast we assume agents are independent learners, making the communication channel non-differentiable. Earlier work on emergent communication focused on cooperative "embodied" agents, showing how communication helps accomplish a common goal [8, 30, 6], or investigating communication in mixed cooperative- competitive environments [26, 3, 15], studying properties of the emergent protocols [21, 18, 25]. Previous research has investigated independent reinforcement learners in cooperative settings [36], with more recent work focusing on canonical RL algorithms. One version of decentralized Q-learning converges to optimal policies in deterministic tabular environments without additional communi- cation [19], but does not trivially extend to stochastic environments or function approximation. Centralized critics [26, 9] improve stability by allowing agents to use information from other agents during training, but these violate our assumptions of independence, and may not scale well. 2 Setting We apply multi-agent reinforcement learning (MARL) in partially-observable Markov games (i.e. partially-observable stochastic games) [33, 24, 11], in environments where agents have a joint communication channel. In every state, agents take actions given partial observations of the true world state, including messages sent on a shared channel, and each agent obtains an individual reward. Through their individual experiences interacting with one another and the environment, agents learn to broadcast appropriate messages, interpret messages received from peers and act accordingly. Formally, we consider an N-player partially observable Markov game G [33, 24] defined on a finite state set S, with action sets (A1, . . . ,AN ) and message sets (M1, . . . ,MN ). An observation function O : S × {1, . . . , N} → Rd defines each agent's d-dimensional restricted view of the true t = O(St, i), and the state space. On each timestep t, each agent i receives as an observation oi t−1 sent in the previous state for all j (cid:54)= i. Each agent i then select an environment action messages mj t ) ∈ (A1, . . . ,AN ) t ∈ Ai and a message action mi ai the state changes based on a transition function T : S × A1 × ··· × AN → ∆(S); this is a stochastic transition, and we denote the set of discrete probability distributions over S as ∆(S). Every t ∈ Mi. Given the joint action (a1 t , . . . , aN 2 t , . . . , mN t ), and ot = (o1 t , . . . , aN t ), mt = (m1 t , . . . oN t ). We write m¯i,t for (m1 EπA,πM ,T (cid:2)(cid:80)∞ t, and M¯i for (M1, . . . ,MN ), excluding Mi. t : S × A1 × ··· × AN → R for player i. We use the notation t ), t , . . . , mN agent gets an individual reward ri at = (a1 excluding mi In our fully cooperative setting, each agent receives the same reward at each timestep, ri t = ∀i, j ≤ N, which we denote by rt. Each agent maintains an action and a message policy rj t t), and which can in from which actions and messages are sampled, ai general be functions of their entire trajectory of experience xi t). t−1, mt−1, oi 1, ..., ai These policies are optimized to maximize discounted cumulative joint reward J(πA, πM ) := {π1 M}. Although the objective J(πA, πM ) is a joint objective, our model is that of decentralized learning and execution, where every agent has its own experience in the environment, and independently optimizes the objective J with respect to its own action and message policies πi M ; there is no communication between agents other than using the actions and message channel in the environment. Applying independent reinforcement learning to cooperative Markov games results in a problem for each agent which is non-stationary and non-Markov, and presents difficult joint exploration and coordination problems [2, 5, 20, 28]. (cid:3) (which is discounted by γ < 1 to ensure convergence), where πA := t=1 γt−1rt A }, πM := {π1 t) and mi t ∼ πi t := (m0, oi M (·xi 1, ai t ∼ πi A(·xi M , ..., πN A and πi A, ..., πN 3 Shaping Losses for Facilitating Communication One difficulty in emergent communication is getting the communication channel to help with the task at all. There is an equilibrium where the speaker produces random symbols, and the listener's policy is independent of the communication. This might seem like an unstable equilibrium: if one agent uses the communication channel, however weakly, the other will have some learning signal. However, this is not the case in some tasks. If the task without communication has a single, deterministic optimal policy, then messages from policies sufficiently close to the uniform message policy should be ignored by the listener. Furthermore, any entropy costs imposed on the communication channel, which are often crucial for exploration, exacerbate the problem, as they produce a positive pressure for the speaker's policy to be close to random. Empirically, we often see agents fail to use the communication channel at all; but when agents start to use the channel meaningfully, they are then able to find at least a locally optimal solution to the communication problem. We propose two shaping losses for communication to alleviate these problems. The first is for positive signalling [25]: encouraging the speaker to produce diverse messages in different situations. The second is for positive listening [25]: encouraging the listener to act differently for different messages. In each case, the goal is for one agent to learn to ascribe some meaning to the communication, even while the other does not, which eases the exploration problem for the other agent. We note that most policies which maximize these biases do not lead to high reward. Much information about an agent's state is unhelpful to the task at hand, so with a limited communication channel positive signalling is not sufficient to have useful communication. For positive listening, the situation is even worse -- most ways of conditioning actions on messages are actively unhelpful to the task, particularly when the speaker has not developed a good protocol. These losses should therefore not be expected to lead directly to good communication. Rather, they are intended to ensure that the agents begin to use the communication channel at all -- after this, MARL can find a useful protocol. 3.1 Bias for positive signalling The first inductive bias we use promotes positive signalling, incentivizing the speaker to produce different messages in different situations. We add a loss term which is minimized by message policies that have high mutual information with the speaker's trajectory. This encourages the speaker to produce messages uniformly at random overall, but non-randomly when conditioned on the speaker's trajectory. We denote by πi M the average message policy for agent i over all trajectories, weighted by how often they are visited under the current action policies for all agents. The mutual information of agent i's message mi I(mi t is: t with their trajectory xi txi t) − H(mi t, xi t) t) = H(mi (cid:88) m∈Mi M (mxi πi t) log(πi M (mxi t)) (1) (2) = − (cid:88) m∈Mi πi M (m) log(πi M (m)) + E xi t 3 txi t, xi We estimate this mutual information from a batch of rollouts of the agent policy. We calculate H(mi t) exactly for each timestep from the agent's policy. To estimate H(πm), we estimate πm as the average message policy in the batch of experience. Intuitively, we would like to maximize I(mi t), so that the speaker's message depends maximally on their current trajectory. However, adding this objective as a loss for gradient descent leads to poor solutions. We hypothesize that this is due to properties of the loss landscape for this loss. Policies which maximize mutual information are t, but uniformly random unconditional on xi deterministic for any particular trajectory xi t. At such policies, the gradient of the term H(πi M (·xi t)) is infinite. Further, for any c < log(2) the space of policies which have entropy at most c is disconnected, in that there is no continuous path in policy space between some policies in this set. To overcome these problems, we instead use a loss which is minimized for a high value for H(πi and a target value for H(πi Msi). The loss we use is: M ) Lps(πi (3) for some target entropy Htarget, which is a hyperparameter. This loss has finite gradients around its minima, and for suitable choices of Htarget the space of policies which minimizes this loss is connected. In practice, we found low sensitivity to Htarget, and typically use a value of around log(A)/2, which is half the maximum possible entropy. M ) − (H(mi M , si) = −E(cid:0)λH(πi txt) − Htarget)2(cid:1), t for 1 ≤ t ≤ T . Algorithm 1 Calculation of positive signalling loss 1: πM i ← 0. 2: Lps ← 0. 3: Target conditional entropy Htarget. 4: Weighting λ for conditional entropy. 5: for b=1; b ≤ B; b++ do # Batch of rollouts. t for 1 ≤ t ≤ T . 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: end for 22: Lps ← Lps + T × B × H. Observations oi Actions ai Other agent messages mt,¯i from M¯i for 1 ≤ t ≤ T . Initial hidden state hi 0. Action set Ai, message set Mi, observation space Oi, hidden state space H i. Message policy πi Hidden state update rule hi : Oi × Ai × H i × M¯i (cid:55)→ H i. for t = 1; t ≤ T; t++ do t−1, hi t−1, mt−1,¯i). t/(T × B). t(m)). t(m) log(pi t − Htarget)2. t ← hi(oi t, ai hi t ← πi t, hi M (oi pi πM ← πM + πi Hi m∈Mi pi Lps ← Lps + λ(Hi M : Oi × Ai × H i × M¯i (cid:55)→ Mi. 21: H =(cid:80) m∈Mi π(m) log(π(m)). t ←(cid:80) t−1, mt−1,¯i). end for 3.2 Bias for positive listening The second bias promotes positive listening: encouraging the listener to condition their actions on the communication channel. This gives the speaker some signal to learn to communicate, as its messages have an effect on the listener's policy and thus on the speaker's reward. The way we encourage positive listening is akin to the causal influence of communication, or CIC [15, 25]. In [15], this was used as a bias for the speaker, to produce influential messages, and in [25] as a measure of whether communication is taking place. We use a similar measure as a loss for the listener to be influenced by messages. In [15, 25], CIC was defined over one timestep as the mutual information between the speaker's message and the listener's action. We extend this to multiple timesteps using the mutual information between all of the speaker's previous messages on a single listener action -- using this as an objective encourages the listener to pay attention to all the speaker's messages, rather than just the most recet. For a listener trajectory xt = (o1, a1, m1, . . . , ot−1, at−1, mt−1, ot), we define 4 x(cid:48) t = (o1, a1, . . . , at−1, ot) (this is the trajectory xt, with the messages removed). We define the multiple timestep CIC as: CIC(xt) = H(atx(cid:48) = DKL t) − H(atxt). (cid:0)(atxt)(atx(cid:48) t)(cid:1) (4) (5) (6) (8) We estimate this multiple timestep CIC by learning the distribution πi t). We do this by per- forming a rollout of the agent's policy network, with the actual observations and actions in the trajectory, and zero inputs in the place of the messages. We fit the resulting function πi t) to predict πi A(·xt), using a cross-entropy loss between these distributions: A(·x(cid:48) A(·x(cid:48) t)), A(ax(cid:48) A(axt) log(πi πi A(axt) term. For a given policy πi A(·x(cid:48) A, this loss is where we backpropagate only through the πi t) = E(πi minimized in expectation when πi A is trained to be an approximation of the listener's policy unconditioned on the messages it has received. The multi-timestep CIC can then be estimated by the KL divergence between the message-conditioned policy and the unconditioned policy: t)). Thus πi A(·x(cid:48) Lce(xt) = − (cid:88) a∈Ai CIC(xt) ≈ DKL(πi A(·xt)πi A(·x(cid:48) t)). (7) For training positive listening we use a different divergence between these two distributions, which we empirically find achieves more stable training. We use the L1 norm between the two distributions: Lpl (xt) = − (cid:88) (cid:12)(cid:12)πi a∈Ai A(axt) − πi A(ax(cid:48) t)(cid:12)(cid:12). t for 1 ≤ t ≤ T . A : Oi × Ai × H i × M¯i (cid:55)→ Ai. t for 1 ≤ t ≤ T . 0 = h(cid:48) 0. Algorithm 2 Calculation of positive listening losses 1: Observations oi 2: Actions ai 3: Other agent messages mt,¯i from M¯i for 1 ≤ t ≤ T . 4: Initial hidden state hi 5: Action set Ai, observation space Oi, hidden state space H i. 6: Action policy πi 7: Hidden state update rule hi : Oi × Ai × H i × M¯i (cid:55)→ H i. 8: Lce ← 0. 9: Lpl ← 0. 10: for i = 1; t ≤ T; t++ do t−1, hi 11: t−1, mt−1,¯i). 12: t−1, h(cid:48) t−1, 0). 13: t−1, h(cid:48) t, 0). 14: a∈A stop_gradient(pi 15: a∈A pi 16: 17: end for Lce ← Lce +(cid:80) Lpl ← Lpl +(cid:80) t(a)) log(pi t(a) − stop_gradient(pi t ← hi(oi hi t ← πi pi A(oi t ← hi(oi h(cid:48) t ← πi pi A(oi t, ai t, hi t, ai t, ai t(a)). t(a)). t−1, mt−1,¯i). 4 Empirical Analysis We consider two environments. The first is a simple one-step environment, where agents must sum MNIST digits by communicating their value. This environment has the advantage of being very amenable to analysis, as we can readily quantify how valuable the communication channel currently is to each agent. In this environment, we provide evidence for our hypotheses about how the biases we introduce in Section 3 ease the learning of communication protocols. The second environment is a new multi-step MARL environment which we name Treasure Hunt. It is designed to have a clear 5 performance ceiling for agents which do not utilise a communication channel. In this environment, we show that the biases enable agents to learn to communicate in a multi-step reinforcement learning environment. We also analyze the resulting protocols, finding interpretable protocols that allow us to intervene in the environment and observe the effect on listener behaviour. The full details of the Treasure Hunt environment, together with the hyperparameters used in our agents, can be found in the supplementary material. 4.1 Summing MNIST digits In this task, depicted in Figure 1, the speaker and listener agents each observe a different MNIST digit (as an image), and must determine the sum of the digits. The speaker observes an MNIST digit, and selects one of 20 possible messages. The listener receives this message, observes an independent MNIST digit, and must produce one of 19 possible actions. If this action matches the sum of the digits, both agents get a fixed reward of 1, otherwise, both receive no reward. The agents used in this environment consist of a convolutional neural net, followed by an multi-layer perceptron and a linear layer to produce policy logits. For the listener, we concatenate the message sent to the output of the convnet as a one-hot vector. The agents are trained independently with REINFORCE. Figure 1: Summing MNIST environment. In this example, both agents would get reward 1 if a = 9. The purpose of this environment is to test whether and how the biases we propose ease the learning task. To do this, we quantify how useful the communication channel is to the speaker and to the listener. We periodically calculate the rewards for the following policies: 1. The optimal listener policy πlc, given the current speaker and the labels of the listener's MNIST digits. 2. The optimal listener policy πlnc, given the labels of the listener's MNIST digits and no communication channel. 3. The optimal speaker policy πsc, given the current listener and the labels of the speaker's MNIST digits. 4. The uniform speaker policy πsu, given the current listener. We calculate these quantities by running over many batches of MNIST digits, and calculating the optimal policies explicitly. The reward the listener can gain from using the communication channel is Pl(πs) = R(πlc, πs) − R(πlnc , πs), so this is a proxy for the strength of the learning signal for the listener to use the channel. Similarly, Ps(πs) = R(πl, πsc ) − R(πl, πsu ) is how much reward the speaker can gain from using the communication channel, and so is a proxy for the strength of the learning signal for the speaker. The results (Figure 2) support the hypothesis that the bias for positive signalling eases the learning problem for the listener, and the bias for positive listening eases the learning problem for the speaker. When neither agent has any inductive bias, we see both Pl and Ps stay low throughout training, and the final reward of 0.1 is exactly what can be achieved in this environment with no communication. When we add a bias for positive signalling or positive listening, we see the communication channel used in most runs (Table 1), leading to greater reward, and Ps and Pl both increase. Importantly, when we add our inductive bias for positive listening, we see Ps increase initially, followed by Pl. This is consistent with the hypothesis that the positive listening bias bias produces a stronger learning signal for the speaker; then once the speaker has begun to learn to communicate meaningfully, the listener also has a strong learning signal. When we add the bias for positive signalling the reverse is true -- Pl increases before Ps. This again fits the hypothesis that speaker's bias produces a stronger learning signal for the speaker. We also ran experiments with the speaker getting an extra reward for positive listening, as in [15]. However, we did not see any gain from this over the no-bias baseline; in our setup, it seems the 6 Figure 2: (a) Both biases lead to more reward. (b,c, d) Listener and speaker power in various settings. Listener power increases first with positive signalling, and speaker power increases first with positive listening. speaker agent was unable to gain any influence over the listener. We think that there is a natural reason this bias would not help in this environment; for a fixed listener, the speaker policy which optimizes positive listening has no relation to the speaker's input. Thus this bias does not force the speaker to produce different messages for different inputs, and so does not increase the learning signal for the listener. Proportion of good runs Final reward of good runs Biases No bias Social influence Positive listening Positive signalling Both 0.00 0.00 0.88 0.98 1.00 CI 0.00-0.07 0.00-0.07 0.76-0.94 0.90-1.00 0.93-1.0 N/A N/A 0.92 ± 0.03 0.99 ± 0.00 0.98 ± 0.01 Table 1: Both biases lead to consistent discovery of useful communication. We define a good run to be one with final average reward greater than 0.2. Averages are over 50 runs for each setting. 4.2 Treasure Hunt We propose a new cooperative RL environment called Treasure Hunt, where agents explore several tunnels to find treasure 1. When successful, both agents receive a reward of 1. The agents have a limited field of view; one agent is able to efficiently find the treasure, but can never reach it, while the other can reach the treasure but must perform costly exploration to find it. In the optimal solution, the agent which can see the treasure finds it and communicates the position to the agent which can reach it. Agents communicate by sending one of five discrete symbols on each timestep. The precise generation rules for the environment can be found in the supplementary material. The agents used in this environment are Advantage Actor-Critic methods [29] with the V-trace correction [7]. The agent architecture employs a single convolutional layer, followed by a multi-layer perceptron. The message from the other agent is concatenated to the output of the MLP, and fed into an LSTM. The network's action policy, message policy and value function heads are linear layers. 1Videos for this environment can be found at https://youtu.be/eueK8WPkBYs and https://youtu. be/HJbVwh10jYk. 7 Figure 3: Treasure hunt environment. Figure 4: Positive signalling and listening biases leads to more reward. Proportion good Final reward (good runs) Biases No bias Positive signalling Positive listening Both 0.28 0.84 0.64 0.94 CI 0.18-0.42 0.71-0.92 0.50-0.76 0.84-0.98 Final reward 12.45 ± 0.48 14.22 ± 0.36 13.94 ± 0.44 15.14 ± 0.33 14.67 ± 0.29 14.69 ± 0.18 14.95 ± 0.20 15.41 ± 0.14 36.1 ± 3.3 41.3 ± 7.9 Table 2: Proportion and average reward of good runs. Values are means over 50 runs with 95% confidence intervals, calculated using Wilson approximation in the case of Bernoulli variables. Run Mean time (unmodified) Mean visit time (modified) Median Best 100.6 ± 14.7 85.4 ± 14.1 Table 3: Visit time to tunnel, with and without modified messages. Values are means over 100 episodes with 95% confidence intervals. Our training follows the independent multiagent reinforcement learning paradigm: each agent is trained independently using their own experience of states and actions. We use RMSProp [13] to adjust the weights of the agent's neural network. We co-train two agents, each in a consistent role (finder or collector) across episodes. The results are shown in Table 2. We find that biases for positive signalling and positive listening both lead to increased reward, and adding either bias to the agents leads to more consistent discovery of useful communication protocols; we define these as runs which get reward greater than 13, the maximum final reward in 50 runs with no communication. With or without biases, the agents still frequently only discover local optima - for example, protocols where the agent which can find treasure reports on the status of only one tunnel, leaving the other to search the remaining tunnels. This demonstrates a limitation of these methods; positive signalling and listening biases are useful for finding some helpful communication protocol, but they do not completely solve the joint exploration problem in emergent communication. However, among runs which achieve some communication, we see greater reward on average among runs with both biases, corresponding to reaching better local optima for the communication protocol on average. We also ran experiments with the speaker getting an extra reward for influencing the listener, as in [15]. Here, we used the centralized model in [15], where the listener calculates the social influence of the speaker's messages, and the speaker gets an intrinsic reward for increasing this influence. We did not see a significant improvement in task reward, as compared to communication with no additional bias. We analyze the communication protocols for two runs, which correspond to the two videos linked in1. One is a typical solution among runs where communication emerges; we picked this run by taking the median final reward out of all runs with both positive signalling and positive listening biases enabled. Qualitatively, the behaviour is simple -- the finder finds the rightmost tunnel, and then reports whether there is treasure in that tunnel for the remainder of the episode. The other run we analyze is the one with the greatest final reward; this has more complicated communication behaviour. To analyze these runs, we rolled out 100 episodes using the final policies from each. 8 First, we relate the finder's communication protocol to the actual location of the treasure on this frame; in both runs, we see that these are well correlated. In the median run, we see that one symbol relates strongly to the presence of treasure; when this symbol is sent, the treasure is in the rightmost tunnel around 75% of the time. In the best run, where multiple tunnels appear to be reported on by the finder, the protocol is more complicated, with various symbols correlating with one or more tunnels. Details of the correlations between tunnels and symbols can be found in the supplementary material. Next, we intervene in the environment to demonstrate that these communication protocols have the expected effect on the collector. For each of these pairs of agents, we produce a version of the environment where the message channel is overridden, starting after 100 frames. We override the channel with a constant message, using the symbol which most strongly indicates a particular tunnel. We then measure how long the collector takes to reach a square 3 from the bottom, where the agent is just out of view of the treasure. In Table 3, we compare this to the baseline where we do not override the communication channel. In both cases, the collector reaches the tunnel significantly faster than in the baseline, indicating that the finder's consistent communication is being acted on as expected. 5 Conclusion We introduced two new shaping losses to encourage the emergence of communication in decentralized learning; one on the speaker's side for positive signalling, and one on the listener's side for positive listening. In a simple environment, we showed that these losses have the intended effect of easing the learning problem for the other agent, and so increase the consistency with which agents learn useful communication protocols. In a temporally extended environment, we again showed that these losses increase the consistency with which agents learn to communicate. Several questions remain open for future research. Firstly, we investigate only fully co-operative environments; does this approach can help in environments which are neither fully cooperative nor fully competitive? In such settings, both positive signalling and positive listening can be harmful to an agent, as it becomes more easily exploited via the communication channel. However, since the losses we use mainly serve to ensure the communication channel starts to be used, this may be as large a problem as it initially seems. Secondly, the environments investigated here have difficult communication problems, but are otherwise simple; can these methods be extended to improve performance in decentralized agents in large-scale multi-agent domains? There are a few dimensions along which these experiments could be scaled -- to more complex observations and action spaces, but also to environments with more than two players, and to larger communication channels. References [1] Trapit Bansal, Jakub Pachocki, Szymon Sidor, Ilya Sutskever, and Igor Mordatch. Emergent complexity via multi-agent competition. In 6th International Conference on Learning Repre- sentations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, 2018. [2] Daniel S. Bernstein, Shlomo Zilberstein, and Neil Immerman. The complexity of decentralized control of Markov Decision Processes. In UAI '00: Proceedings of the 16th Conference in Uncertainty in Artificial Intelligence, Stanford University, Stanford, California, USA, June 30 - July 3, 2000, pages 32 -- 37, 2000. [3] Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent communication through negotiation. arXiv preprint arXiv:1804.03980, 2018. [4] Yongcan Cao, Wenwu Yu, Wei Ren, and Guanrong Chen. An overview of recent progress in the study of distributed multi-agent coordination. IEEE Trans. Industrial Informatics, 9(1):427 -- 438, 2013. [5] Caroline Claus and Craig Boutilier. The dynamics of reinforcement learning in cooperative mul- tiagent systems. In Proceedings of the Fifteenth National Conference on Artificial Intelligence and Tenth Innovative Applications of Artificial Intelligence Conference, AAAI 98, IAAI 98, July 26-30, 1998, Madison, Wisconsin, USA., pages 746 -- 752, 1998. [6] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Michael Rab- bat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. arXiv preprint arXiv:1810.11187, 2018. 9 [7] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu, T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures. ArXiv e-prints, February 2018. [8] Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pages 2137 -- 2145, 2016. [9] Jakob N. Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 2974 -- 2982, 2018. [10] H Paul Grice. Utterer's meaning, sentence-meaning, and word-meaning. Language, and Artificial Intelligence, pages 49 -- 66. Springer, 1968. In Philosophy, [11] Eric A. Hansen, Daniel S. Bernstein, and Shlomo Zilberstein. Dynamic programming for partially observable stochastic games. In Proceedings of the 19th National Conference on Artifical Intelligence, AAAI'04, pages 709 -- 715. AAAI Press, 2004. [12] Matthew John Hausknecht. Cooperation and communication in multiagent deep reinforcement learning. PhD thesis, University of Texas at Austin, Austin, USA, 2016. [13] G. Hinton, N. Srivastava, , and K. Swersky. Lecture 6a: Overview of mini -- batch gradient descent. Coursera, 2012. [14] Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio García Castañeda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in first-person multiplayer games with population-based deep reinforcement learning. CoRR, abs/1807.01281, 2018. [15] Natasha Jaques, Angeliki Lazaridou, Edward Hughes, Caglar Gulcehre, Pedro A Ortega, DJ Strouse, Joel Z Leibo, and Nando de Freitas. Intrinsic social motivation via causal influence in multi-agent rl. International Conference of Machine Learning, 2019. [16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. [17] Hiroaki Kitano, Minoru Asada, Yasuo Kuniyoshi, Itsuki Noda, and Eiichi Osawa. Robocup: The robot world cup initiative. In Agents, pages 340 -- 347, 1997. [18] Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. Natural language does not emerge'naturally'in multi-agent dialog. arXiv preprint arXiv:1706.08502, 2017. [19] Martin Lauer and Martin A. Riedmiller. An algorithm for distributed reinforcement learning in cooperative multi-agent systems. In Proceedings of the Seventeenth International Conference on Machine Learning (ICML 2000), Stanford University, Stanford, CA, USA, June 29 - July 2, 2000, pages 535 -- 542, 2000. [20] Guillaume J. Laurent, Laëtitia Matignon, and Nadine Le Fort-Piat. The world of Indepen- dent learners is not Markovian. International Journal of Knowledge-Based and Intelligent Engineering Systems, 15(1):55 -- 64, March 2011. [21] Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic communication from referential games with symbolic and pixel input. arXiv preprint arXiv:1804.03984, 2018. [22] Joel Z. Leibo, Edward Hughes, Marc Lanctot, and Thore Graepel. Autocurricula and the emergence of innovation from social interaction: A manifesto for multi-agent intelligence research. CoRR, abs/1903.00742, 2019. 10 [23] Joel Z. Leibo, Vinícius Flores Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. Multi-agent reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, AAMAS 2017, São Paulo, Brazil, May 8-12, 2017, pages 464 -- 473, 2017. [24] Michael L. Littman. Markov games as a framework for multi-agent reinforcement learning. In In Proceedings of the Eleventh International Conference on Machine Learning, pages 157 -- 163. Morgan Kaufmann, 1994. [25] Ryan Lowe, Jakob Foerster, Y-Lan Boureau, Joelle Pineau, and Yann Dauphin. On the pitfalls of measuring emergent communication. arXiv preprint arXiv:1903.05168, 2019. [26] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi- agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pages 6379 -- 6390, 2017. [27] Laëtitia Matignon, Laurent Jeanpierre, and Abdel-Illah Mouaddib. Coordinated multi-robot exploration under communication constraints using decentralized markov decision processes. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, July 22-26, 2012, Toronto, Ontario, Canada., 2012. [28] Laëtitia Matignon, Guillaume J. Laurent, and Nadine Le Fort-Piat. Independent reinforcement learners in cooperative markov games: a survey regarding coordination problems. Knowledge Eng. Review, 27(1):1 -- 31, 2012. [29] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lilli- crap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928 -- 1937, 2016. [30] Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi- agent populations. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. [31] Stephen Neale. Paul grice and the philosophy of language. Linguistics and philosophy, 15(5):509 -- 559, 1992. [32] Malcolm Reynolds, Gabriel Barth-Maron, Frederic Besse, Diego de Las Casas, Andreas Fidjeland, Tim Green, Andria Puigdomenech, Sébastien Racanière, Jack Rae, and Fabio Viola. Open sourcing Sonnet - a new library for constructing neural networks. https: //deepmind.com/blog/open-sourcing-sonnet/, 2017. [33] L. S. Shapley. Stochastic games. Proceedings of the National Academy of Sciences of the United States of America, 39(10):1095 -- 1100, 1953. [34] Luc Steels. Evolving grounded communication for robots. Trends in cognitive sciences, 7(7):308 -- 312, 2003. [35] Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropagation. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 2244 -- 2252, 2016. [36] Ming Tan. Multi-agent reinforcement learning: Independent versus cooperative agents. In Ma- chine Learning, Proceedings of the Tenth International Conference, University of Massachusetts, Amherst, MA, USA, June 27-29, 1993, pages 330 -- 337, 1993. [37] Kyle Wagner, James A Reggia, Juan Uriagereka, and Gerald S Wilkinson. Progress in the simulation of emergent communication and language. Adaptive Behavior, 11(1):37 -- 69, 2003. 11 A Environment details To generate a map for the treasure hunt environment, we: 1. Create a rectangle of grey pixels with height 18 and width 24. 2. Draw a black tunnel on the second row up, including all but the leftmost and rightmost pixels. 3. Draw a black tunnel on the second row down, including all but the leftmost and rightmost pixels. 4. Pick 4 starting positions on the top horizontal tunnel for vertical tunnels. These are randomly selected among sets of positions which are all at least 3 pixels apart (so no tunnel is visible from another). Draw black tunnels from these positions to 2 pixels above the bottom tunnel. 5. Place the yellow treasure at the bottom of a random tunnel. 6. Place one agent uniformly at random in the top tunnel, and one uniformly at random in the bottom tunnel. The episode length is 500 timesteps. The agents have 5 actions, corresponding to the 4 directions and a no-op action. They can move in the black tunnels and onto the treasure, but not onto the grey walls. The agent observation is a 5 × 5 square, centered on the agent. When an agent moves onto the treasure, both agents receive reward 1, and the treasure respawns at the bottom of a random tunnel. The RGB values of the colors of the pixels in the observations are: • Blue self: (0, 0, 255). • Red partner agent: (255, 0, 0). • Grey walls: (128, 128, 128). • Black tunnels: (0, 0, 0). • Yellow treasure: (255, 255, 0). B Treasure hunt communication protocols In this section, we give more details on the communication protocols in Treasure Hunt. First, we give full details of the correlations between messages and treasure location in the two runs discussed in the main text. These are the median and the best runs in terms of final reward, in the setting where we use both positive signalling and positive listening biases. We generated 100 episodes using the final policies for the 2 agents, and recorded the treasure locations and symbols on each timestep. Each cell shows the probability of each treasure location, given that the speaker transmits a particular symbol. In Table 4, we see that symbol 2 is particularly meaningful, fairly reliably indicating the presence of treasure in the final tunnel. In Table 5, we see that three symbols appear to be used; 0 and 2 correlate with righthand tunnels, and 3 with lefthand ones. P(T = 1S) 0.12 ± 0.01 0.36 ± 0.02 0.06 ± 0.01 0.32 ± 0.01 0.27 ± 0.01 P(T = 2S) 0.08 ± 0.01 0.35 ± 0.02 0.05 ± 0.00 0.35 ± 0.02 0.28 ± 0.02 P(T = 4S) S 0.46 ± 0.03 0 0.07 ± 0.01 1 0.75 ± 0.02 2 0.09 ± 0.01 3 0.20 ± 0.01 4 Table 4: Message tunnel correlations for median Treasure Hunt run. P(T = 3S) 0.35 ± 0.03 0.22 ± 0.01 0.14 ± 0.01 0.24 ± 0.01 0.24 ± 0.01 12 S 0 1 2 3 4 P(T = 1S) 0.08 ± 0.01 0.15 ± 0.01 0.12 ± 0.01 0.56 ± 0.03 0.23 ± 0.02 P(T = 2S) 0.12 ± 0.01 0.31 ± 0.02 0.15 ± 0.01 0.19 ± 0.02 0.32 ± 0.02 P(T = 3S) 0.32 ± 0.02 0.31 ± 0.02 0.29 ± 0.02 0.16 ± 0.02 0.27 ± 0.02 P(T = 4S) 0.48 ± 0.02 0.22 ± 0.01 0.44 ± 0.02 0.09 ± 0.02 0.18 ± 0.02 Table 5: Message tunnel correlations for best Treasure Hunt run. We also show the correlations between messages and listener actions, in Tables 6 and 7. In both cases, these are as expected from the correlations between tunnels and messages; we see the listener move more in the direction which the messages correlate with. P(RightS) 0.46 ± 0.01 0.05 ± 0.00 0.32 ± 0.01 0.04 ± 0.00 0.03 ± 0.00 P(DownS) 0.34 ± 0.01 0.49 ± 0.00 0.17 ± 0.01 0.09 ± 0.00 0.73 ± 0.00 P(N oopS) 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 P(Lef tS) 0.05 ± 0.00 0.26 ± 0.00 0.03 ± 0.00 0.09 ± 0.00 0.09 ± 0.00 P(U pS) 0.14 ± 0.01 0.20 ± 0.00 0.48 ± 0.01 0.78 ± 0.00 0.14 ± 0.00 Table 6: Message action correlations for median Treasure Hunt run. S 0 1 2 3 4 S 0 1 2 3 4 P(N oopS) 0.01 ± 0.00 0.01 ± 0.00 0.01 ± 0.00 0.01 ± 0.00 0.00 ± 0.00 P(U pS) 0.49 ± 0.01 0.37 ± 0.01 0.25 ± 0.01 0.47 ± 0.01 0.26 ± 0.01 P(RightS) 0.28 ± 0.01 0.07 ± 0.00 0.23 ± 0.01 0.04 ± 0.01 0.07 ± 0.00 P(DownS) 0.11 ± 0.00 0.38 ± 0.01 0.42 ± 0.01 0.27 ± 0.01 0.52 ± 0.01 P(Lef tS) 0.10 ± 0.01 0.17 ± 0.00 0.09 ± 0.00 0.21 ± 0.01 0.14 ± 0.00 Table 7: Message action correlations for best Treasure Hunt run. C Network details and hyperparameters C.1 MNIST sums experiments In the MNIST sums environment, the agent architecture used was from an existing MNIST classifier; we did not optimize this, as the goal was to investigate the effect of communication biases rather than to achieve optimal performance. This architecture is: • A convolutional neural network, with 2 layers, which have 32 and 64 channels respectively. Both layers have kernel size 5, stride 1, and rectified linear unit (ReLU) activations. We use max pooling, with stride and kernel 2. • One hidden linear layer with 1024 neurons, with ReLU activations. We used the Adam optimizer [16], with a learning rate of 0.0003 and parameters β1 = 0.9, β2 = 0.999, = 10−8. These layers were implemented using the Sonnet library [32]. The agents were trained using REINFORCE; the total loss for each agent consists of the REINFORCE loss, entropy regularization, and the biases for positive signalling (Lps) and positive listening, (Lpl and Lce). For the listener agent, the message is concatenated to the flattened output of the convolutional net before the hidden linear layer. The final hyperparameters used for the 5 settings were: To select the hyperparameters for the no bias setting, we performed a joint sweep over action and message entropy bonuses, consider a range of values from 0.0 to 0.3 for each. No values were found which improved over the no communication policy; the final values reported here are those which worked best in the other settings. 13 Hyperparameter Batch size Action policy entropy bonus Message policy entropy bonus Target message entropy Htarget Weight of Lpl Weight of Lps Weight of Lce λ for Lps No bias 32 0.03 0.03 N/A N/A N/A N/A N/A PL PS 32 32 0.03 0.03 0.03 0.0 1.0 1.0 N/A 0.01 N/A 0.1 N/A 0.001 N/A 3.0 Both 32 0.03 0.0 N/A 0.01 0.1 0.001 3.0 SI 32 0.01 0.01 N/A N/A N/A 0.001 N/A Table 8: Hyperparameters for final MNIST experiments. In the positive listening setting, we performed sweeps: • Over the weight of Lpl, using values in (0.01, 0.03, 0.1, 0.3). • Over the entropy costs for messages and actions, using values (0.01, 0.03, 0.1). • Over the weight of Lce, using values of (0.0, 0.001, 0.01, 0.1); aside from 0, which unsur- prisingly produced worse results, there was no significant difference in the results of these runs. In the positive signalling setting, we performed sweeps: • Jointly over the weight of Lps and λ, using values in (0.01, 0.03, 0.1) for Lps, and (0.01, 0.03, 0.1) for the product λLps. In the setting with both biases, we ran no additional sweeps, simply combining the hyperparameters from the best runs with positive signalling and positive listening. In the social influence setting, we performed sweeps: • Over the weight of the social influence reward, using values (0.01, 0.03, 0.1, 0.3, 1.0, 10.0). • Over the entropy costs for messages and actions, using values (0.01, 0.03, 0.1). For all hyperparameter sweeps, we ran 5 runs, and picked the setting with the highest average final reward. For the final sets of hyperparameters, we then ran 50 runs. C.2 Treasure Hunt experiments In our experiments, we use 32 parallel environment copies for the Asynchronous Advantage Actor- Critic algorithm [29] with the V-trace correction [7]. The total loss for each agent consists of the A3C loss, including entropy regularization, and the biases for positive signalling (Lps) and positive listening, (Lpl and Lce). The two agents have the same architecture, which consists of: • A single convolutional layer, using 6 channels, kernel size of 1 and stride of 1. • A multi-layer perceptron with 2 hidden layers of size 64. • An LSTM, with hidden size 128. • Linear layers mapping to policy logits for the action and message policies, and to the baseline value function. The message from the other agent is concatenated to the flattened output of the convolutional net before the hidden linear layer. We used the RMSProp optimizer [13] for gradient descent, with a initial learning rate of 0.001, exponentially annealed by a factor of 0.99 every million steps. The other parameters are η = 0.99, and = 10−6. The final hyperparameters used for the 5 settings were: In the no communication setting, we performed sweeps: • Over the entropy costs for actions, using values (0.001, 0.003, 0.006, 0.01). 14 Parameter Batch size No comms No bias 16 16 PS 16 PL 16 0.006 Action entropy regularization Message entropy regularization Target message entropy Htarget 0.0 N/A N/A 0.01 0.001 N/A N/A Table 9: Hyperparameters for final Treasure Hunt experiments. 0.006 N/A N/A N/A N/A N/A N/A N/A 0.006 0.0 0.8 0.003 N/A 0.001 3.0 N/A 0.006 0.0 N/A N/A N/A N/A N/A N/A Weight of Lpl Weight of Lce Weight of Lps λ for Lps Weight of Lsi Both 16 0.006 0.0 0.8 0.003 0.01 N/A 3.0 N/A SI 16 0.006 0.0 N/A N/A 0.01 N/A N/A 0.01 • Over the sizes of the MLP layers, using values (32, 64, 128). • Over the sizes of the LSTM hidden size, using values (64, 128). We then fixed these parameters for the other settings. In the positive listening setting, we performed sweeps: • Over the weight of Lpl, using values in (0.001, 0.003, 0.01). • Over the weight of Lce, using values of (0.0, 0.001, 0.01). In the positive signalling setting, we performed sweeps: • Jointly over the weight of Lps and λ, using values in (0.001, 0.003, 0.01) for Lps, and (0.001, 0.003, 0.01) for the product λLps. • Over the target entropy Ht, using values in (0.4, 0.8, 1.6). In the setting with both biases, we ran no additional sweeps, simply combining the hyperparameters from the best runs with positive signalling and positive listening. In the social influence setting, we performed a sweep over the weighting of the reward for social influence to the speaker, using values in (0.01, 0.03, 0.1, 0.3, 1.0). For all hyperparameter sweeps, we ran 5 runs, and picked the setting which exceeded the no- communication baseline most frequently, terminating runs early if the result was clear. For the final sets of hyperparameters, we then ran 50 runs. D Multi-step CIC ablation In the positive listening bias for the listener, we use the multiple timestep CIC. Recall that for a listener trajectory xt = (o1, a1, m1, . . . , ot−1, at−1, mt−1, ot), we define x(cid:48) t = (o1, a1, . . . , at−1, ot) (this is the trajectory xt, with the messages removed). We define the multiple timestep CIC as: t) − H(atxt). CIC(xt) = H(atx(cid:48) (9) We could also choose to use the single timestep CIC, defined in [15, 25] as the mutual information between the speaker and the listener's actions. Defining x(cid:48)(cid:48) t = (o1, a1, m1, . . . , at−1, ot) -- which is the trajectory with only the final message removed -- this would be: t ) − H(atxt). CIC(st) = H(atx(cid:48)(cid:48) (10) A(·x(cid:48) Similarly to in the multiple timestep CIC, we estimate the single timestep CIC by learning the t). We do this by performing a rollout of the agent's policy network, with the distribution πi actual observations and actions in the trajectory, including all message except the final one, which A(·xt), using a cross-entropy loss is zeroed out. We fit the resulting function πi between these distributions. The single-timestep CIC can then be estimated by: t ) to predict πi A(·x(cid:48)(cid:48) CIC(xt) ≈ DKL(πi A(·xt)πi A(·x(cid:48)(cid:48) t )). (11) 15 Figure 5: One step CIC does not lead to better results than no speaker-side bias. As with the multi-timestep CIC, we used the L1 loss for Lpl (xt) = − (cid:88) (cid:12)(cid:12)πi t)(cid:12)(cid:12). A(axt) − πi A(ax(cid:48) a∈Ai (12) Using the single timestep CIC, with or without positive signalling, we did not find improvements over a baseline with no speaker-side bias -- see Figure 5. E Statistical methodology All confidence intervals shown are 95% confidence intervals. For confidence intervals of Bernoulli variables -- the proportion of runs with reward above a certain threshhold -- we use the Wilson approximation. For graphs depicting average performance over multiple runs, we first take the mean reward per run in 100 time windows over training. The interval shown is the 95% confidence interval for this mean. 16
1903.04726	1	1903	2019-03-12T04:40:53	Self-triggered distributed k-order coverage control	[ "cs.MA", "math.OC" ]	A k-order coverage control problem is studied where a network of agents must deploy over a desired area. The objective is to deploy all the agents in a decentralized manner such that a certain coverage performance metric of the network is maximized. Unlike many prior works that consider multi-agent deployment, we explicitly consider applications where more than one agent may be required to service an event that randomly occurs anywhere in the domain. The proposed method ensures the distributed agents autonomously cover the area while simultaneously relaxing the requirement of constant communication among the agents. In order to achieve the stated goals, a self-triggered coordination method is developed that both determines how agents should move without having to continuously acquire information from other agents, as well as exactly when to communicate and acquire new information. Through analysis, the proposed strategy is shown to provide asymptotic convergence similar to that of continuous or periodic methods. Simulation results demonstrate that the proposed method can reduce the number of messages exchanged as well as the amount of communication power necessary to accomplish the deployment task.	cs.MA	cs	Self-triggered distributed k-order coverage control Daniel Tabatabai Mohanad Ajina Cameron Nowzari 9 1 0 2 r a M 2 1 ] A M . s c [ 1 v 6 2 7 4 0 . 3 0 9 1 : v i X r a Abstract -- A k-order coverage control problem is studied where a network of agents must deploy over a desired area. The objective is to deploy all the agents in a decentralized manner such that a certain coverage performance metric of the network is maximized. Unlike many prior works that consider multi- agent deployment, we explicitly consider applications where more than one agent may be required to service an event that randomly occurs anywhere in the domain. The proposed method ensures the distributed agents autonomously cover the area while simultaneously relaxing the requirement of constant communication among the agents. In order to achieve the stated goals, a self-triggered coordination method is developed that both determines how agents should move without having to continuously acquire information from other agents, as well as exactly when to communicate and acquire new information. Through analysis, the proposed strategy is shown to provide asymptotic convergence similar to that of continuous or periodic methods. Simulation results demonstrate that the proposed method can reduce the number of messages exchanged as well as the amount of communication power necessary to accomplish the deployment task. I. INTRODUCTION This paper studies a multi-agent coordination problem where a network of agents perform a deployment task to statically position themselves over a desired area. For exam- ple, a mobile sensor network where it is required to deploy sensors to positions that will maximize total coverage of the desired sensing environment. This is commonly referred to as coverage control. Specific applications include topics such as environmental monitoring [1], [2], survelliance [3], data collection [4], [5], and search and rescue [6]. More specifically, we consider a generalization of the coverage problem that extends to scenarios where more than one agent may be required to overlap a region in the coverage area. This is referred to as k-order coverage control [7] -- [9], where k > 1 agents must overlap coverage of the same point q. Our contributions focus on the development of coordination strategies that will reduce the amount of communication necessary between agents while performing the deployment task. This is accomplished by the design of a self-triggered algorithm where agents autonomously decide when they require information from other agents in the network. With respect to coverage control, the majority of previous research has focused on scenerios where an individual agent is capable of servicing events that occur in the agent's respective region of responsibility without the assistance of other agents. As an example, consider a monitoring application where a wirless sensor network must monitor the The authors and Com- puter Engineering, George Mason University, Fairfax, VA 22030, USA, {dtabatab,majina,cnowzari}@gmu.edu are with the Department of Electrical environment. If a random event occurs in the vicinity of a particular sensor then that sensor has the ability to measure and capture the event independent of other sensors in the network. However, various applications exists where agents do not possess the capability to capture or respond to events independently. These applications require multiple agents to work collectively in order to service events. One example of this type of application is that of Time Difference of Arrival (TDOA) localization [10] -- [12] where the requirement is that three or more sensors that are located at different positions must measure the same event. Another example is emergency response vehicles where two or more vehicles may be required to respond to a particular event, such as a fire or burglary. In other scenarios, two or more agents may not be necessary for event handling, but the application may require redundant agents to overlap areas for fail-safe purposes. Literature review: The topic of multi-agent coverage control has been studied by a number of authors in the past including the seminal work [13] where coverage control based on agents moving to the centroids of a Voronoi partition was introduced. In [14], the authors consider the coverage control problem where each node is constrained to have m neighboring nodes. The authors use an approach based on vector potential fields where each node acts as a repelling force in order to maximize coverage and acts as an attracting force in order to satisfy the m neighbor constraint. In [15], the authors consider heterogeneous and non-point source nodes as well as non-convex enviroments. In [16], the authors study the problem in the context of using sensor measurements to estimate regions of importance in the mission space thus driving nodes to concentrate in these areas. Common to all the above mentioned works is the fact that they study the coverage control problem in terms of a first-order coverage problem where each agent is solely responsible for covering a sub-region of the mission space. As previously mentioned, the interest of our work is the generalized k-order coverage control problem where multiple agents overlap coverage of sub-regions in the mission space. The k-order coverage control problem was studied in [17] -- [19] where a method using higher-order Voronoi parti- tions was proposed. The authors present a method for deploy- ing agents over a bounded area when more than one agent must have overlapping coverage of the same point. However, to realize the proposed contol law in [19], it is assumed that continuous communication between agents is achievable. For many real-world systems, continuous communication is not feasible and periodic solutions can be resource inefficient and may not be neccessary. As alternatives to continuous and periodic solutions, self-triggered and event-triggered approaches have been proposed in the literature to handle similiar problems in networked systems [8], [20] -- [23]. For self- and event-triggered solutions, the exact time at which agents perform actions, e.g. wirelessly communicate or up- date a control signal, is autonomously decided by the agents rather than occurring at periodic time intervals. In [24], [25], the concepts of self-triggered control was applied to the case of first-order optimal deployment. In our current work, we extend the self-triggered centroid algorithm presented in [24] by considering the higher-order coverage control problem studied in [17] -- [19] and develop a self- triggered coordination strategy to relax the synchronous, periodic communication requirement while guaranteeing that each agent moves such that it does not contribute negatively to the task. Statement of contributions: The main contribution of this work is the development of a distributed self-triggered control strategy that deploys a set of agents to static locations in a convex area in order to achieve k-order optimal coverage. Our solution relaxes the need for continuous or periodic communication among agents as is done in prior works [19]. More specifically, our algorithm is comprised of two major sub-components. The first being an update decision policy where each agent decides when to acquire new information from neighboring agents through a wireless communication network. The decision to comunicate is based on the level of uncertainty each agent has accumulated over time. This uncertainty is due to not having up-to-date information that results from the lack of communication with other agents. We extend the notion of uncertain spatial partitioning [26] -- [28] used for optimal deployment in [24] by the use of k- order guaranteed and dual-guaranteed Voronoi partitions. The second major sub-component is a motion control law that determines how agents should move given possibly outdated information about the location of other agents in the network. Each agent determines a motion plan that is guaranteed to contribute positively to the higher-order deployment task. Organization: Section II outlines some important notions from computational geometry. Section III formally presents the problem statement. Section IV formulates the concepts of k-order guaranteed and k-order dual-guaranteed Voronoi partitions. Section V presents the algorithm design. In section VI convergence analysis of the algorithm is discussed. Section VII presents simulation results and section VIII assimilates the conclusions. II. PRELIMINARIES Let R≥0 and Z≥0 be the set of non-negative real, integer values respectively. With the Euclidean norm defined by k · k A. Basic geometric notions We denote by [p, q] ⊂ Rd the closed segment with extreme points p and q ∈ Rd. Let φ : Rd → R≥0 be a bounded measurable function that we term density. For MS = ZS φ(q)dq, CS = qφ(q)dq. 1 MS ZS S ⊂ Rd, the mass and center of mass of S with respect to φ are Let s1, s2, ..., sn be n subsets of S and {s1, s2, ..., sn} be a partition of S then mass and center of mass with respect to φ and the n partitions, MS = Msi, nXi=1 CS = Pn Pn i=1 MsiCsi i=1 Msi The circumcenter ccs of a bounded set S ⊂ Rd is the center of a closed ball of minimum radius that contains S. The circumradius crs of S is the radius of this ball. The diameter of S is diam(S) = maxp,q∈S kp − qk. Given v ∈ Rd \ {0}, let unit(v) be the unit vector in the direction of v. Given a convex set S ⊂ Rd and p ∈ Rd, let prS(p) denote the orthogonal projection of p onto S, i.e., prS(p) is the point in S closest to p. The to-ball-boundary map tbb : (Rd × R≥0)2 → Rd takes (p, δ, q, r) to (p + δ unit(q − p) prB(q,r) if kp − prB(q,r)(p)k ≥ δ, if kp − prB(q,r)(p)k ≤ δ. Figure 1 illustrates the action of tbb. p δ tbb(p, δ, q, r) p ≤ δ tbb(p, δ, q, r) r q r q Fig. 1: Graphical representation of the action of tbb when (a) kp − prB(q,r)(p)k > δ and (b) kp − prB(q,r)(p)k ≤ δ. We denote by B(p, r) the closed ball centered at p ∈ S with radius r and by Hpo = {q ∈ Rd kq − pk ≤ kq − ok} the closed halfspace determined by p, o ∈ Rd that contains p. B. 1-order Voronoi partitions The methods developed in this work rely heavily on the concept of Voronoi partitioning [29]. In the following sub- sections a brief discussion of 1-order Voronoi partitions is presented. Let S be a convex polygon in R2 and P = (p1, . . . , pn) be the location of n agents. A partition of S is a collection of n polygons K = {K1, . . . , Kn} with disjoint interiors whose union is S. The Voronoi partition V(P ) = {V1, . . . , Vn} of S generated by the points P = (p1, . . . , pn) is Vi = {q ∈ S kq − pik ≤ kq − pjk , ∀j 6= i}. Intuitively, the Voronoi cell Vi represents all the points that are closer to the agent at position pi than to any of the other agents in the network. When the Voronoi regions Vi and Vj are adjacent (i.e., they share an edge), pi is called a (Voronoi) neighbor of pj (and vice versa). P = (p1, . . . , pn) is a centroidal Voronoi configuration if it satisfies that pi = CVi , for all i ∈ {1, . . . , n}. III. PROBLEM STATEMENT A. k-order Voronoi partitions 2 Intuitively, a k-order Voronoi cell represents all the points that are closer to k agents located at positions {pi1, pi2 , ..., pik } = PI than to any of the other agents in the network. A k-order Voronoi partition would then be a collection of all the k-order Voronoi cells. In the first-order case, the space is partitioned into n cells such that each agent is closest to every point in their cell than any of the other agents. However, in the case of a second-order partition the space is partitioned into n(n−1) cells and there are many cells that can be empty because certain agents may not share overlapping responsibility for any points in the space. The difference between the first-order and second-order case can be seen in figure 2. In figure 2a an example of a first-order partition is illustrated for five agents. Note that each agent is enclosed in their respective cells and there is exactly five cells, one per agent. Figure 2b illustrates the second-order partition for the same five agents. Note that in figure 2b the number of cells are greater than the number of agents. Also note that some cells contain multiple agents while other cells do not contain any agents at all. Furthermore, some agent combinations are not associated to a cell at all, e.g. V(1, 5) and V(3, 4). This is due to the fact that these agents do not share points in the space that are mutually closer to them combined than to any of the other agents. A more formal definition of the k-order Voronoi partition follows. Let S ⊂ R2 be a convex polygon in a 2-dimensional space. Let A = {1, . . . , n} be a finite set of integers representing the agents in a n-agent network. Let P = {p1, . . . , pn} ∈ S be the set of positions of the agents A in the domain S. Let I = (i1, . . . , ik) ∈ I be a k-tuple element of the set I where I = {(i1, . . . , ik) ∈ Ak i1 < · · · < ik} is the set of k-tuples in Ak that do not repeat, for example (i1, i2, i3) ∈ I , but (i3, i2, I1) /∈ I. Let PI = {pi1, . . . pik } ⊂ P be the subset of agent positions corresponding to the agents (i1, . . . , ik) ∈ I. The collection of all elements I ∈ I that include a particular agent i is denoted by I i where I i = {I ∈ I ∀ 1 ≤ α ≤ k, i = Iα} A k-order partition of S is a collection of m polygons R = {R1, . . . , Rn} with disjoint interiors whose union is S and where an element in R is associated with k-agents. The k-order Voronoi partition of a convex polygon S can be defined as follows. Given a set of agents A with positions P. For k < n, let PI ⊂ P with PI = k. Then the k-order Voronoi region associated with agents I = (i1, . . . , ik) with generating sites PI = (pi1 , . . . , pik ) is defined as VI = {q ∈ S kq−pk ≤ kq−p′k, ∀ p ∈ PI, ∀ p′ ∈ P \PI}. For example, if k = 2 then PI = {pi, pj} and the second- order Voronoi cell for agents (i, j) ∈ I becomes, Vij = {q ∈ S kq − pik ≤ kq − p′k, kq − pjk ≤ kq − p′k, ∀ p′ ∈ P \ {pi, pj}} For every point q in VI , the distance from q to any agent position in PI is less than or at most equal to the distance from q to all other agent positions not in PI . For k = 2, the second-order Voronoi partition with I = (i, j) and PI = {pi, pj} would mean that the two agents i and j are closer to or at most as close to all the points in Vij than any of the other agents A \ {i, j}. An alternative interpretation would be that the agents i and j are considered responsible for the region defined by Vij . Combining all k-order Voronoi regions in S, the k-order Voronoi partition of the environment S becomes V(P ) = . The environment S can be considered as the union (cid:8)VI(cid:9)I∈I of all k-order Voronoi cells S =SI∈I VI . Figure 2 presents an example of the difference between a first-order (2a) and second-order (2b) Voronoi partition for five agents. For any agent i with position pi ∈ P, there can be multiple sets PI ⊂ P that contain pi meaning that an agent i located at position pi can be responsible for multiple k-order Voronoi cells. The collection of k-order Voronoi cells associated with agent i is given by V i = {VI }I∈I i. All k-order cells associated with agent i can be combined to form a single region of S that agent i is responsible for and this cell is referred to as the dominant region of agent i. The dominant region for agent i is be defined by Wi = [I∈I i VI . The dominant cell Wi represents the region of S that agent i is responsible for covering. Note that the first-order cell Vi and the k-order cell VI are not equivalent, but both are convex. However, the dominant cell Wi may not be convex. The k-order neighbors of agent i is denoted by Ni. For a k- order Voronoi partition, P = (p1, . . . , pn) is a centroidal k- order Voronoi configuration if it satisfies pi = CWi , for all i ∈ {1, . . . , n}. Next, optimal deployment for k-order Voronoi partitioning is discussed. B. Objective for higher-order coverage The interest is in applications where k > 1 agents are required to service an event occuring at a random point q ∈ S. This is in contrast to the 1-order problem where for any point q ∈ S only one agent is responsible. In order to optimally deploy agents throughout the mission space, an objective function for the higher order deployment problem must be defined. For the 1-order case, from [13], the objective function in terms of Voronoi partitions is defined as H(P ) = kq − pik2φ(q)dq (1) nXi=1ZVi a1 a3 a2 a4 a5 a1 V(1,4) V(1,3) V(1,2) a4 V(2,4) a2 V(4,5) a3 V(2,3) V(3,5) a5 V(2,5) (a) 1-order Voronoi diagram (b) 2-order Voronoi diagram Fig. 2: Example of 1-order Voronoi diagram (left) and 2-order Voronoi diagram (right). In the 1-order case, cells can be represented by the agent that covers the cell. In the case of 2-order partitions, two agents share coverage over a cell. The cells are represented by two agents that cover the particular partition. Note that not all agents share cells. In particular, agents (1, 5) and agents (3, 4) do not share cells, i.e. V(1,5) = ∅ and V(3,4) = ∅ The objective here is to minimize the distance from agent i's position pi to all points q ∈ Vi. Taking advantage of the parallel axis theorem, H(P ) may be expressed as, H(P ) = nXi=1 JVi,CVi + nXi=1 MVikpi − CVi k2 (2) where JVi,CVi is the polar moment of inertia of the 1-order Voronoi cell Vi centered at the centroid CVi . Taking the partial derivative of (2) with respect to pi and evaluating at zero will produce the minimum H at position p∗ i for agent i. The partial derivative of (2) with respect to pi is given by, ∂H ∂pi = MVi(pi − CVi ) This demonstrates that the objective function H in the 1- order case is minimal when pi is located at the centroid CVi . For Voronoi partitions of the k-order, a similar approach to the order-1 partition can be followed. In [19] an objective function for higher-order coverage control with a general per- formance measure was introduced and a detailed derivation with performance measured defined by Euclidean distance for k = 2 was presented. For completeness, the objective function is restated for arbitrary k. The objective function in terms of a k-order partition of S is defined as, H(P, R) = 1 kXI∈IZRI f (q, p1, . . . , pk)φ(q)dq. (3) Where f (q, ·) is the performance measure given by, f (q, p1, . . . , pk) = kq − pik2. kXi=1 The objective function in terms of k-order Voronoi partitions is defined as, H(P ) = = 1 kXI∈IZVI kXI∈IZVI(cid:16) kXi=1 1 f (q, p1, . . . , pk)φ(q)dq (4) kq − pik2(cid:17)φ(q)dq. Unlike like the first-order Voronoi objective function where the integration occurred over each cell and there was a cell for each agent, the k-order case does not have a one-to- one relationship between cells and agents. The performance measure is based on the distance k-agents are from each point q in VI . However, by manipulation, the objective function can be written in terms of the contribution of each agent separately. By distributing the integral, H(P ) = kq − pi1 k2φ(q)dq + . . . . . . +ZVI kq − pik k2φ(q)dqi and summing over all cells for each agent, H(P ) = kq − pi1 k2φ(q)dq + . . . 1 kXI∈IhZVI 1 kXI∈IZVI . . . + 1 kXI∈IZVI kq − pik k2φ(q)dq, the higher-order objective function can be expressed in terms of the polar moment of inertia, H(P ) =XI∈IhJV,CV + + 1 k 1 k MVI kpi1 − CVI k2 + . . . MVI kpik − CVI k2i, (5) From (5), it can be seen that the value of H depends on the distance from an agent to the centroid of a given cell. Clearly an agent cannot be located at the centroid of all the cells it is responsible for. To solve for the optimal location for agents to be located, the function H is described in matrix form as follows, H(P ) = 1⊤(JVI ,CVI )1 + . . . + + 1 k(cid:0)pi1 k(cid:0)pik 1 1 − CVi1(cid:1)⊤ 1 − CVik(cid:1)⊤ MVi1(cid:0)pi1 MVik(cid:0)pik 1 − CVi1(cid:1) 1 − CVik(cid:1). Where 1 is a vector of ones, JVI ,CVI is a diagonal matrix, is a vector of cell centroids associated with agent i, CVi and MVi is a diagonal matrix with elements on the diagonal represent the mass of the respective cell. Now the the optimal position p∗ i for agent i can be solved by, 1(cid:1)−1(cid:0)1⊤MVi CV i CVi(cid:1) j = CWi . p∗ i =(cid:0)1⊤MVi = PV i PV i j=1 MV i j j=1 MV i j CWi is the centroid of the dominant cell Wi. As mentioned in the previous section, the cell Wi is the dominant cell of agent i, which is the union of all the k-order Voronoi cells associated with agent i. The objective function H is minimal when pi is located at the centroid CWi of the dominant cell Wi. This leads to the following lemma. Lemma III.1 Given P ∈ Sn and a k-order partition R of S, H(P, V(P )) ≤ H(P, R), i.e., the optimal partition is the k-order Voronoi partition. For P ′ ∈ S with kp′ i − CWi k ≤ kpi − CWi k, i ∈ {1, . . . , n}, H(P ′, R) ≤ H(P, R), i.e., the optimal positions of agents are the centroids. As discussed in [19], for continuous control and commu- nication the gradient descent control law is given by ui = −k(pi − CWi ) for gain k > 0. However, implementing this in continuous time assumes that agents have exact position information about their neighbors at all times. Instead, we next discuss how to relax this requirement without resorting to a synchronous, periodic implementation. C. Communication between agents We assume agent i has access to its own position pi(t) at all times t ∈ Z≥0, but must communicate with neighbors j ∈ Ni to obtain their positions pj . Similar to [24], a request- response communication model is used where agent i is able to request position pj from agent j and agent j immediately responds with this information. We assume that packet loss does not occur and that round-trip latency is negligible such that agent i can request and receive information in- stantaneously i.e. the action of requesting and responding information occurs within the same timestamp. More specifically, let {ti ℓ}ℓ∈Z≥0 ⊂ Z≥0 be the sequence of times at which agent i requests information from some neighbor j ∈ Ni. Then, agent i only has access to the position of agent j at these times, e.g., at timestep t agent i has access to {pj(t′)}t′∈{ti ℓ≤t}. ℓti D. Agent state representation If an agent does not request information on every timestep then that agent does not have access to the current position of other agents. Therefore, agent i maintains state information pertaining to the most recent known position of agent j in addition to information that is able to model the evolution of uncertainty over time that exists with respect to agent j's current position. Given that agent i has acquired position pj(tℓ) from agent j at timestep tℓ, let τ > tℓ be the amount of time that has elapsed since agent i has communicated with agent j. Then the position pj(t) where t = tℓ + τ will be unknown to agent i at time t. However, if the maximum speed vmax for agent j is known then agent i can determine the set of all possible positions where agent j could have traveled to in time duration τ . The set of possible positions for agent j can be represented by a closed ball with center at pj(tℓ) and radius rj = vmaxτj . To maintain state, each agent stores pj(tℓ) and rj in memory for every agent in the network. The data storage for agent i is then defined by, Di =(cid:0)(pi 1, ri 1), . . . , (pi n, ri n)(cid:1) ∈ (S × R≥0)n (6) where ri i = 0 for all time since it is assumed that agent i always has access to it's own position pi i at every timestep. There exists two methods for which the contents of the data structure Di may be updated. The first is a time evolution update where all values ri increase in magnitude based j on the the time duration τ . The second update method, referred to as the information/position update, corresponds to the acquisition of a new position value pi j via means of communication with agent j. When a position update occurs for pi j = 0. This is due to the fact that the exact position of agent j is known at the instance in time that pi j has been received and stored in memory by agent i. In addition, two explicit methods for agent i to extract information from Di. The first is the map loc : (S ×R≥0)n → Sn that allows agent i to extract position information (p1 i )) from Di. The second extraction map π : (S × R≥0)n → (S × R≥0)m, where m ≤ n, allows agent i to extract a subset π(Di) ⊂ Di from data storage. j is reset i.e. ri j, the value ri i , . . . , pn E. Agent dynamics Considering the set A of agents moving in a convex polygon S with positions P = (p1, . . . , pn). We consider discrete-time, single-integrator dynamics pi(t + 1) = pi(t) + ui(t)∆t, (7) where ∆t > 0 denotes the length of time of one timestep, and ui(t) denotes the input at timestep t with kui(t)k ≤ vmax for each agent i ∈ A. The interest is in optimally deploying these agents in the domain S such that k agents overlap responsibility for every point q ∈ S. Equipped with a communication model, a state data model, and agent dynamics the formal problem may now be presented by the following, Problem III.2 Given a set A = {1, . . . , n} of agents mov- ing in a convex polygon S ⊂ R2 with dynamics (7), maximum speed vmax > 0, spatial density φ : S → R, and only depending on information local to agent i, find a distributed communication and control strategy such that pi → CWi . Based on the data that each agent stores in memory, the exact computation of the k-order Voronoi cell cannot necessarily be achieved at each timestep. Next we address the issue of space partitioning with uncertainty for general cases of k-order. IV. SPACE PARTITION WITH UNCERTAIN INFORMATION If agent i does not have access to the exact location pj of agent j, then the uncertain position of agent j with respect to agent i can be represented to be within a set of points Dj ∈ S. This set Dj represents all the possible points where agent j is guaranteed to be located relative to agent i. The consequence of this representation is that agent i cannot compute it's dominant region exactly. However, because the position of agent j is guaranteed to be constrained to the set Dj , it is possible for agent i to compute regions in S that pertain to a) the points that are certain to be part of its dominant cell, b) the points that are certain not to be part of agent i's dominant cell, and c) the region where it is uncertain if the points belong to agent i's dominant cell or not. The region of points that are certain to be part of agent i's dominant cell is referred to as the k- order guaranteed dominant cell of agent i. The region of points that are certain to not be a part of agent i's dominant cell is referred to as the k-order dual-guaranteed dominant cell. Similar to the case of certain sites, we construct the guaranteed and dual-guaranteed dominant cell of an agent i by means of the k-order guaranteed and dual-guaranteed Voronoi cells. The the k-order guaranteed Voronoi partition is described next. A. k-order guaranteed Voronoi partitions To assist in the exposition that follows, the first-order guaranteed Voronoi cell is briefly mentioned. The first-order guaranteed Voronoi cell for agent i is given by, gVi =nq ∈ S (cid:12)(cid:12)(cid:12) max x∈Di kq − xk ≤ min y∈Dj kq − yk, ∀j 6= io. The cell gVi contains the points of S that are guaranteed to be closer to pi than to any other pj , with i 6= j. The uncertain regions Di and Dj are considered to be closed balls B(pi, ri,) and B(pj, rj ,) centered at pi and pj with radius ri and rj , respectively. The set D = {D1, . . . , Dn} is the collection of uncertain regions for n agents. Similar to the discussion of k-order Voronoi partitions of certain sites where I = (i1, . . . , ik) ∈ I and PI ⊂ P, a subset of D is defined by DI = {D1, . . . , Dk} Given DI and with DJ = D \ DI , the k-order guaranteed Voronoi cell associated with I agents is defined by, gVI =nq ∈ S (cid:12)(cid:12)(cid:12) max x∈Di kq − xk ≤ min y∈Dj kq − yk ∀Di ∈ DI , ∀Dj ∈ DJo The k-order guaranteed Voronoi cell represents the points that are guaranteed to be closer to the k-agents in I with positions in PI than to the agents with positions in PJ . For example, with k = 2 the I-agents becomes I = {Di, Dj} and has positions PI = {pi, pj} such that the second-order guaranteed Voronoi cell associated with agents i and j is given by, gVij =(cid:26)q ∈ S (cid:12)(cid:12)(cid:12) max x∈Di max x∈Dj kq − xk ≤ min y∈DJ kq − xk ≤ min y∈DJ kq − yk, kq − yk(cid:27) with DJ = D \ {Di, Dj}. For agent i, dominant region can be defined by, the guaranteed gWi = [I∈I i gVI The cell gWi represents the region that agent i is guaranteed to be responsible for covering. B. k-order dual-guaranteed Voronoi partitions In [24], the concept of dual-guaranteed Voronoi partitions was presented. Here we extend this concept to the case of k- order dual-guaranteed Voronoi partitions. Again using DI ⊂ D and DJ = D \ DI , the k-order dual-guaranteed Voronoi cell for agents in I is defined by, dgVI =nq ∈ S (cid:12)(cid:12)(cid:12) min x∈Di kq − xk ≤ max y∈Dj kq − yk, ∀Di ∈ DI , ∀Dj ∈ DJo The region outside of the cell dgVI represents the points that are guaranteed to be closer to agents in J than to the agents in I. The dual-guaranteed dominant region associated with agent i is given by, dgWi = [I∈I i dgVI The region outside of cell dgWi represents the points that agent i is guaranteed not to be responsible for covering. Next, a solution that includes both the design of a motion control law and a communication strategy for the above stated problem is presented. V. SELF-TRIGGERED HIGHER ORDER COVERAGE OPTIMIZATION Given the problem described in Section III, one possi- ble approach would be for agent i to periodically acquire position information from other agents. This would occur (a) r = 0 (b) r = 1 (c) r = 2 Fig. 3: The guaranteed k-order Voronoi cells (k = 2) for a single agent represented by the black asterisk located close to the center of each diagram. Together, the diagrams illustrate the difference between cells when the radii changes, r = 0 (a), r = 1 (b), r = 2 (c). (a) r = 0 (b) r = 1 (c) r = 2 Fig. 4: The dual-guaranteed k-order Voronoi cells (k = 2) for a single agent represented by the black asterisk located close to the center of each diagram. Together, the diagrams illustrate the difference between cells when the radii changes, r = 0 (a), r = 1 (b), r = 2 (c). at each time step where agent i would 1) acquire new information, 2) Compute it's dominant cell Wi, 3) compute the centroid CWi , 4) move towards CWi at vmax, 5) repeat. However, similar to continuous communication, a periodic method requires frequent communication and a potentially unnecessary computational burden. The method proposed in the following section attempt to alleviate the communication and computational burden by a two part approach. The first component is a motion control law to determine how agents move when the information they possess is not up to date with respect to the most recent time step. The second component is an information update policy that allows each agent to decide when information from other agents should be acquired. A. Motion control If agent i has access to the exact positions of other agents then agent i is capable of computing the exact dominant cell Wi. Consequently, agent i can compute the centroid CWi . Once CWi has been computed, agent i may simply move towards it. When agent i does not communicate with the other agents in the network, the exact location of the other agents will be unknown to agent i. Since the exact locations of other agents is unknown, each agent must rely on the data that it does possess as a means for deciding how to move. The data that an agent does possess at any given time step includes the most recent position update that it has received from the other agents and the time that has elapsed since the last update. Informally, the motion control law is described by the following. At each time step each agent uses the information that it has stored to compute it's k-order guaranteed and dual-guaranteed Voronoi cells. Next, each agent computes it's guaranteed and dual-guaranteed dominant cells. Once the agent has computed these cells, the agent then computes the centroid for the guaranteed dominant cell gWi and begins moving toward it. The motion control law assumes that each agent has ac- cess to the value of the density φ over it's k-order guaranteed dominant cell. The motion control law as describe above does not necessarily guarantee that agents will move closer to the centroids of their dominant cells without applying additional constraints on agent movement. As in [24], the following lemma applies. Lemma V.1 Given p 6= q, q∗ ∈ R2, let p′ ∈ [p, q] such that kp′ − qk ≥ kq∗ − qk, then kp′ − q∗k ≤ kp − q∗k. Following lemma V.1, if p = pi is the position of agent i that is moving toward p′ in the direction of the computed goal q = CgWi then the distance to q∗ = CWi decreases while kp′ − CgWi k ≥ kCVi − CgWi k (8) holds. Since CW i is unknown to agent i the right hand side of 8 cannot be computed. However the value kCWi −CgWik can be bounded such that kp′ − CgWik ≥ bndi (9) where bndi is given by bndi = bnd(gWi, dgWi) = 2crdgWi(cid:18)1 − MgWi MdgWi(cid:19) (10) Therefore, agent i moves towards CgWi as much as possible in one timestep while maintaining the condition in (9). The motion control law is formally defined in table (1). For every consecutive time step that an agent goes without receiving updated information, (10) increases making the condition of (9) less likely to be achievable. Leading to the condition where the agent can no longer move in a manner that does not increase the distance to CWi . Therefore, a decision mechanism that governs when an agents will acquire new information is required and is discussed next. Algorithm 1 : motion control law Agent i ∈ {1, . . . , n} performs: 1: set D = Di 2: compute L = gWi(D) and U = dgWi(D) 3: compute q = CL and r = bnd(L, U ) 4: set d = vmax∆t 5: set p′ 6: move to p′ i j = ri 7: set ri j + d j = ( pi 8: set Di i = (p′ 9: set Di j, min {ri i, 0) i = tbb(pi, d, q, r) j, diam(S)} ) B. Update decision policy The second major aspect of the self-triggered deploy- ment strategy provides a decision mechanism that deter- mines when an agent must perform an information update via communication with other agents. Updates to position information will be necessary for an agent to reduce the level of uncertainty that it has accumulated since the last time an update occurred. As previously mentioned, as time elapses without receiving position information from other agents, the true location of CWi will be unknown and the set of possible locations for CWi will continue to increase in size. Based on the motion control law presented in the previous section, agent i will rely on moving towards CgWi so long as condition (9) holds. If it becomes infeasible for agent i to move due to condition (9) not being satisfied, then agent i must perform an information update at that moment in time in order to maintain condition (9). Therefore, the update decision policy can be describe as follows. For every timestep, each agent computes their k-order guaranteed and dual guaranteed dominant cells, as well as computing the bound (10). Then each agent decides whether or not to perform a position data update. Agent i will decide to perform the update when the bound (10) becomes greater than or equal to kpi − CgWi k. It is possible that the points pi and CgWi may become close to one another i.e. kpi − CgWi k < ε for ε > 0.In this case, the bound (10) may not be able to become small enough such that a position update is not required. To handle this condition, the value of kpi − CgWik is clamped at ε so that a minimum amount of time will pass before and update will occur. The update policy is described formally in table 2. Algorithm 2 : one-step-ahead update policy Agent i ∈ {1, . . . , n} performs: 1: set D = Di 2: compute L = gWi(D) and U = dgWi(D) 3: compute q = CL and r = bnd(L, U ) 4: if r ≥ max {kq − pik, ε} then 5: 6: end if reset Di by performing a position update C. The k-order self-triggered centroid algorithm A self-triggered deployment strategy can be formulated by combining the motion control law defined in Table 1 and the update decision policy from Table 2. First, it is noted that combining the two algorithms from Table 1 and Table 2 without modification would provide an event-triggered deployment strategy. The event-triggered strategy would be performed on each timestep where agent i runs the update decision policy followed by running the motion control law. This requires agent i to compute L, U , CL, and r from Table 2 on every timestep. However, agent i is in possession of all the information necessary to predict its motion trajectory up to the time in the future where r ≥ max {kq − pik, ε} occurs. The self-triggered algorithm is presented in Table 4. In addition, note that a trivial update mechanism would provide each agent with up-to-date locations for all other agents in the network i.e. using all information stored in Di. However, this is costly from a communications point of view. Instead, a localized algorithm is proposed that limits the number of agents that agent i must acquire information from. To compute gWi and dgWi, agent i must have knowledge of only a subset of agent positions. The subset of agents used by agent i can be found by first defining Ai(q) = {j ∈ A kpj − qk < kpi − qk, j 6= i} where Ai(q) ≥ k. Based on this definition we can redefine the cell Wi by Wi = {q ∈ S (Ai(q)) ≤ k − 1} To locally compute Wi at the specific time when step 4: is executed, the Dominant cell computation is used. This is borrowed from [30] and presented in Algorithm 3 set out ← true set ρ ← ρ + γ set Ni(ρ) ← {j kpj − pik < ρ} for all {q ∈ S kq − pik = ρ/2} do Algorithm 3 : Dominant cell computation 1: initialize ρ = 0 2: repeat 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: until out = true 14: compute Wi from Ni(ρ) set Ai(q) ← {j ∈ Ni(ρ) kpj − qk < kpi − qk, j 6= i} if Ai(q) < k then set out ← f alse break end for end if The Dominant cell computation is based on agent i gradually increasing its communication radius until all the information required to construct its exact k-order Voronoi cell has been obtained. Combining Algorithms 1- 3 leads to the complete k-order self-triggered centroid algorithm described in Algorithm 5. reset Di by performing a position update initialize tsleep = 0 while r < max {kq − pik, ε} do Algorithm 4 : multiple-steps-ahead update policy Agent i ∈ {1, . . . , n} performs: 1: set D = Di 2: compute L = gWi(D) and U = dgWi(D) 3: compute q = CL and r = bnd(L, U ) 4: if r ≥ max {kq − pik, ε} then 5: 6: else 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: end if set tsleep = tsleep + 1 set d = vmax∆t set p′ move to p′ i j = ri set ri j + d j = ( pi set Di i = (p′ set Di set D = Di compute L = gWi(D) and U = dgWi(D) compute q = CL and r = bnd(L, U ) end while wait for tsleep timesteps repeat j, min {ri i, 0) i = tbb(pi, d, q, r) j, diam(S)} ) Algorithm 5 : k-order self-triggered centroid algorithm Initialization 1: execute Dominant cell computation At time step ℓ ∈ Z≥0, agent i ∈ {1, . . . , n} performs: reset Di by running Dominant cell computation set D = π(Di) compute L = gWi(D) and U = dgWi(D) compute q = CL and r = bnd(L, U ) 1: set D = π(Di) 2: compute L = gWi(D) and U = dgWi(D) 3: compute q = CL and r = bnd(L, U ) 4: if r ≥ max {kq − pik, ε} then 5: 6: 7: 8: 9: end if 10: set d = vmax∆t 11: set p′ 12: move to p′ i 13: set ri j = ri j + d j = ( pi 14: set Di i = (p′ 15: set Di j, min {ri i, 0) i = tbb(pi, d, q, r) j, diam(S)} ) the number of agents in the network, the area of the task space, and the initial agent configuration. This presents chal- lenges when attempting to analyze the convergent properties in a similar fashion to that of the continuous-time continuous- update policy. Instead, our analysis assumes that agents move according to the motion control law given in Table (1) while considering information updates that occur randomly in time. We show that regardless of how agents share information, trajectories governed by the motion control law, in particular the constraints laid out by tbb, will at least converge to a positively invariant set and that if the update decision policy is followed, the network will converge to the centroidal configuration. To achieve this, a set-valued map T is defined that describes the evolution of the network state represented by the data storage of all agents. Then by applying the LaSalle Invariance Principal for set-valued maps, it is shown that all trajectories generated under the state evolution map T provide values of the performance function H that are monotonically non-increasing. It is also shown that there exist under T a weakly positively invariant set that is specif- ically contained in the trajectories that follow the update decision policy of Table (2). Finally, we deduced that this set coincides with the centroidal network configuration in task space. This is proposed formally by the following: Proposition VI.1 For ε ∈ [0 diam(S)], the agent position evolving under the self-triggered deployment algorithm from any initial network configuration in S converges to the k- order Voronoi centroidal configuration. VI. CONVERGENCE ANALYSIS A detailed analysis is provided in this section to demon- strate that agents following the motion and information update strategies presented thus far will generate a network configuration such that all agents converge to their centroidal positions. The asynchronous timing of information exchange that occurs during the network evolution is dependent on the Di = (cid:0)(pi To proceed, we formally define D = (D1, . . . , Dn) ∈ (S × R≥0)n2 as the sate of an n agent network where 1, ri 1), . . . , (pi n)(cid:1) is the state of agent i. We n, ri define M : (S × R≥0)n2 as the map that updates both the motion pi i and uncertainty evolution ri j in D. Recall that the magnitude of ri j increases over time when information updates do not occur. We define → (S × R≥0)n2 (a) Initial configuration (b) Trajectories (c) Final configuration Fig. 5: Initial configuration (a), self-triggered centroid algorithm trajectories (b), and final configuration (c) for 20 agents running the k-order → (S × R≥0)n2 fu : (S × R≥0)n2 as a mapping of the network state into itself and it describes the information update evolution when the update decision policy from Table (2) is followed. Note that the self-triggered algorithm can be described as the composition fst = M ◦ fu = M(fu(D)). Let U : (S × R≥0)n2 be the set-valued map that represents any possible information-update evolution. For D′ ∈ U(D), the ith component of D′ is described by, ⇒ (S × R≥0)n2 Di =((cid:0)(pi (cid:0)(pi 1), . . . , (pi 1, ri 1, 0), . . . , (pi n, ri n)(cid:1), no update n, 0)(cid:1), update occurred Note that fu(D) ∈ (S × R≥0)n2 but U(D) ⊂ (S × R≥0)n2 further, fu(D) ∈ U(D) is one outcome in U(D). is an element of the domain, is a subset of the domain and Given the definition of M and U , the full state evolution is defined by the set-valued map T : (S × R≥0)n2 ⇒ (S × R≥0)n2 where T = U ◦ M. Since U is closed and M is continuous, the evolution map T is closed. For a trajectory γ = {D(tℓ)}t∈Z>0 generated by the self-triggered algorithm and γ′ = {D′(tℓ)}t∈Z>0 given by D′(tℓ) = fu(D(tℓ)) then, D′(tℓ+1) = T (D′(tℓ)) (11) Let loc : (S × R≥0)n → Sn be a map that extracts the n)) from D such that H(loc(D)) = 1, . . . , pn positions P = (p1 H(P ). Lemma VI.2 H : (S × R≥0)n2 increasing along the trajectories of T . → R is monotonically non- Proof: Let D ∈ (S × R≥0)n2 and D′ ∈ T (D). Let P = loc(D) and P ′ = loc(D′) = loc(M(D))). To demonstrate that H(P ′) ≤ H(P ), first the k-order partition V(P ) is fixed. Then for each i ∈ A, if the condition i − CgWi k ≤ bnd(Di) is true then p′ kp′ i = pi. This is due to the fact that agent i strictly follows the definition of tbb. If instead kp′ i − CgWi k > bnd(Di) then it is true that kp′ i − CWi k < kpi − CWi k by lemma V.1 and (10). For both cases, H(P ′, V(p)) ≤ H(P, V(P )) and furthermore, from lemma III.1, H(P ′, V(P ′)) ≤ H(P ′, V(P )) Lemma VI.3 Let γ′ be a trajectory of (11). Then the ω- limit set Ω(γ′) ⊂ (S × R≥0)n2 with Ω(γ′) 6= ∅ belongs to H−1(c) for some constant c ∈ R≥0 and is weakly positively invariant. Let γ′ be a trajectory of (11). Proof: Let γ′ be a trajectory of (11). First, note that γ′ being bounded implies Ω(γ′) 6= ∅ and for D′ ∈ Ω(γ′) there exists a converging sub-sequence {D′(tℓm m ∈ Z≥0} of γ′ such that D′(tℓm) → D′ as m → ∞. In addition, the sequence {D′(tℓm+1) m ∈ Z≥0} is also bounded and has a converging sub-sequence where for bD′ the sequence D′(tℓm+1 → bD′ for m → ∞. Since by definition bD′ ∈ Ω(γ′) and T is closed, this implies Ω(γ′) is weakly positive invariant. Since γ is bounded and H is non-increasing along γ for all of (S×R≥0)n2 , the sequence H◦γ = {H(γ(l)) l ∈ Z≥0} is decreasing and bounded from below and therefore convergent. Since for any z ∈ Ω(γ) there is a converging subsequence γ(ℓm) in Ω(γ) that converges to z and since H is continuous, H(γ(ℓm)) → H(z) = c as m → ∞ where c ∈ R is a constant. Proof of Proposition VI.1 Let γ = {D(tℓ)}t∈Z≥0 be an evolution of the self-triggered centroid algorithm. Define γ′ = {D(tℓ)}t∈Z≥0 by D′(tℓ) = fu(D(tℓ)). Note that loc(D(tℓ)) = loc(D′(tℓ)). Since γ′ is a trajectory of T , lemma VI.3 guarantees that Ω(γ′) is weakly positively invariant and belongs to H−1(c) for some c ∈ R. Next, it is shown that Ω(γ′) ⊂ {D ∈ (S × R≥0)n2 i ∈ A, kpi i − CgWi k ≤ bndi} (12) We reason by contradiction. Assume there exists D ∈ Ωγ for which there is i ∈ A such that kpi i − CgWik > bndi. By lemma III.1, V.1 and the constraint given by (8), any possible evolution from D under T will strictly decrease H. This is in contradiction with the fact that Ω(γ′) is weakly positively invariant for T . (a) Initial configuration (b) Trajectories (c) Final configuration Fig. 6: self-triggered centroid algorithm Initial configuration (a), trajectories (b), and final configuration (c) for 5 agents running the k-order benchmark st: = 5 2500 2000 1500 1000 500 benchmark st: = 5 benchmark st: = 5 1500 1000 t n u o c g s m 500 10 9 8 7 6 5 4 3 2 1 0 0 5 10 15 time 20 25 30 0 0 5 10 15 time 20 25 30 0 0 5 10 15 time 20 25 30 (a) Performance (b) Message count (c) Power Fig. 7: Performance (a), messages communicated between agents (b), and power (c) versus time. It is also noted that for each i the inequality bndi < i − CgWi k, ε} is satisfied at D′(tℓ)), for all ℓ ∈ Z≥0 max {kpi an by continuity, this holds for Ω(γ′) as well. That is, bndi < max {kpi i − CgWi k, ε} (13) i = p′ for all i ∈ A and all D ∈ Ω(γ′). Now it is shown that Ω(γ′) ⊂ {D ∈ (S × R≥0)n2 i = CWi }. Consider i ∈ A, pi i gets updated information then bndi = 0 and consequently from (12), pi i = CgWi = CWi and the result follows. If 1) > eD ∈ Ω(γ′). Since Ω(γ′) is weakly positively invariant, there exists eD1 ∈ Ω(γ′) ∩ T (eD). Note that (12) implies that loc(eD1) = loc(eD) We consider two cases depending on whether agents have received information in eD1. If agent agent i does not get updated information then bnd(eDi bnd(eD1) and gWi(eD1) ⊂ gWi(eD). Again using the fact eD2 ∈ Ω(γ′) ∩ T (eD1) Reasoning repeatedly in this manner, the only case that needs to be discarded is when agent i never receives updated information. In this case kpi i − CgWi k → 0 while bndi monotonically increases towards diam(S). For sufficiently large ℓ, kpi i − CgWik < ε. Then (13) implies bndi < ε, which contradicts the fact that bndi tends towards diam(S). that Ω(γ′) is a weakly positively invariant set, there exist VII. SIMULATIONS In this section, simulation results for the self-triggered deployment algorithm are presented. Simulations were per- formed with n = 5 agents moving in a 50m× 50m area. The timestep was set to ∆t = 0.1s and all agents were given the same maximum velocity of vmax = 1m/s. Multiple simulation iterations were performed by selecting different values of ε and generating random initial positions for agents on each iteration. Twenty iterations were carried out for each value of ε. The values selected for ε were ε = {0.5, 1.0, 2.5, 5.0}. To quantify the performance of the self-triggered method, the objective function H, the total transmission power, and the total number of messages transmitted were computed on every timestep. As in [24], the power output model for agent i is given by Pi = 10 log10" Xj∈A\{i} β 100.1Pi→j+αkpi−pj k# where α > 0 and β > 0 are parameters that are dependent on the wireless medium and Pi→j is the power received from agent i at agent j in decibel-milliwatts. Simulation results were evaluated against a benchmark case that rep- resents a centroidal continuous information update method 10 9 8 7 6 5 4 3 2 1 0 benchmark = 0.5 st: st: = 1.0 = 2.5 st: st: = 5.0 benchmark st: = 0.5 = 1.0 st: = 2.5 st: st: = 5.0 1600 1400 1200 1000 800 600 400 200 t n u o c g s m benchmark = 0.5 st: = 1.0 st: = 2.5 st: st: = 5.0 1800 1600 1400 1200 1000 800 600 400 200 0 5 10 15 time 20 25 30 0 0 5 10 15 time 20 25 30 0 0 5 10 15 time 20 25 30 (a) Performance (b) Message count (c) Power Fig. 8: Performance (a), messages communicated between agents (b), and power (c) versus time. Average over 20 random initial configurations for different values of ε. 1.006 1.005 1.004 1.003 1.002 1.001 % 1 0.999 0.998 0.997 0.996 0 1 2 3 4 5 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 % 0 1 2 3 4 5 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 % 0 1 2 3 4 5 (a) Performance (b) Message count (c) Power Fig. 9: Convergence of H (a), total network message count (b), and total network power (c) averaged over 20 random initial configurations for each value of ε = (0, 0.5, 1, 2.5, 5) where ε = 0 corresponds to the benchmark case of continuous communication. where agents move toward their dominant cell centroid and positions are updated on every timestep ∆t = 0.1s. Figures 6 and 7 display the results for the execution of a single simulation instance. Figures 6 provides illustration of the initial configuration (6a), the trajectories traveled (6b), and the final configuration (6c) of all agents following the self-triggered deployment strategy. Figure (7) shows a comparison against the benchmark case of the convergence of H (7a), the total message count (7b), and the communi- cation power (7c) at each timestep. The results from figure 7 demonstrate how the self-triggered strategy can reduce both the total amount of communication and the power required to perform the deployment task. This is accomplished while still being capable of achieving convergence performance similar to that of a continuous or periodic communication strategy. Figures 8 and 9 further illustrate this point by presenting results for combined values of ε where twenty random initial configurations for each ε are averaged together. In figure 9, the value ε = 0 corresponds to the benchmark case. These figures illustrate how varying ε affects the overall performance. It can be seen that the total message count and communication power decreases when the value of ε increases, while the the convergence rate of H degrades. However, the convergence degradation of H can be con- sidered minimal when compared to the reduction in both message count and power. For the largest value ε = 5, the convergence of H degrades by less than one-percent, while message count and communication power see a decrease of more than eighty-percent. VIII. CONCLUSIONS This paper presented a k-order self-triggered centroid algorithm for optimal deployment of k- order coverage control scenarios. The presented strategy combined an information update policy with a motion control law. The information update policy provided a method to determine when each agent should communicate with other agents in the network. Agents communicate in order to update their data storage. The decision to communicate is based on whether an agent can continue to contribute positively to the deployment objective. The motion control law provided a method for agents to move when the locations [17] B. Jiang, Z. Sun, and B. D. Anderson, "Higher order voronoi based mobile coverage control," in American Control Conference (ACC), 2015. IEEE, 2015, pp. 1457 -- 1462. [18] B. Jiang, Z. Sun, B. D. O. Anderson, and C. Lageman, "Higher to localization," CoRR, vol. abs/1703.02424, 2017. [Online]. Available: http://arxiv.org/abs/1703.02424 control with order mobile coverage application [19] B. Jiang, Z. Sun, B. D. Anderson, and C. Lageman, "Higher order mobile coverage control with applications to clustering of discrete sets," Automatica, vol. 102, pp. 27 -- 33, 2019. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0005109818306356 [20] W. Heemels, K. H. Johansson, and P. Tabuada, "An introduction to event-triggered and self-triggered control," in Decision and Control (CDC), 2012 IEEE 51st Annual Conference on. IEEE, 2012, pp. 3270 -- 3285. [21] D. V. Dimarogonas, E. Frazzoli, and K. H. Johansson, "Distributed self-triggered control for multi-agent systems," in Decision and Con- trol (CDC), 2010 49th IEEE Conference on. IEEE, 2010, pp. 6716 -- 6721. [22] D. V. Dimarogonas and K. H. Johansson, "Event-triggered control for multi-agent systems," in Decision and Control, 2009 held jointly with the 2009 28th Chinese Control Conference. CDC/CCC 2009. Proceedings of the 48th IEEE Conference on. IEEE, 2009, pp. 7131 -- 7136. [23] M. Mazo and P. Tabuada, "On event-triggered and self-triggered control over sensor/actuator networks," in Decision and Control, 2008. CDC 2008. 47th IEEE Conference on. IEEE, 2008, pp. 435 -- 440. [24] C. Nowzari and J. Cort´es, "Self-triggered coordination of robotic networks for optimal deployment," Automatica, vol. 48, no. 6, pp. 1077 -- 1087, 2012. [25] C. Nowzari, J. Cort´es, and G. J. Pappas, "Team-triggered coordination of robotic networks for optimal deployment," Chicago, IL, Jul. 2015, pp. 5744 -- 5751. [26] W. Evans and J. Sember, "Guaranteed voronoi diagrams of uncertain sites," in 20th Canadian Conference on Computational Geometry, 2008, pp. 207 -- 210. [27] M. Jooyandeh, A. Mohades, and M. Mirzakhah, "Uncertain voronoi diagram," Information processing letters, vol. 109, no. 13, pp. 709 -- 712, 2009. [28] R. Cheng, X. Xie, M. L. Yiu, J. Chen, and L. Sun, "Uv-diagram: A voronoi diagram for uncertain data," in Data Engineering (ICDE), 2010 IEEE 26th International Conference on. IEEE, 2010, pp. 796 -- 807. [29] M. Senechal, "Spatial tessellations: Concepts and applications of voronoi diagrams," Science, vol. 260, no. 5111, pp. 1170 -- 1173, 1993. [30] F. Li, J. Luo, S. Xin, W. Wang, and Y. He, "Autonomous deployment for load balancing k-surface coverage in sensor networks," vol. 14, no. 1, pp. 279 -- 293, 2015. of other agents is uncertain due to the lack of communica- tion. Through analysis, the proposed strategy was shown to provide guaranteed asymptotic convergence. The results have shown convergence similar to that of continuous and periodic position update methods. Simulation results were able to demonstrate the potential benefits of the proposed method by illustrating the ability of the k-order self-triggered centroid algorithm to not only reduce the amount of communication necessary to achieve the deployment goal, but also reducing the power consumed from communication. REFERENCES [1] T. B. Curtin, J. G. Bellingham, J. Catipovic, and D. Webb, "Au- tonomous oceanographic sampling networks," Oceanography, vol. 6, no. 3, pp. 86 -- 94, 1993. [2] Q. Lu, Q.-L. Han, B. Zhang, D. Liu, and S. Liu, "Cooperative control of mobile sensor networks for environmental monitoring: An event-triggered finite-time control scheme," IEEE transactions on cybernetics, vol. 47, no. 12, pp. 4134 -- 4147, 2017. [3] J. R. Peters, S. J. Wang, and F. Bullo, "Coverage control with anytime updates for persistent surveillance missions," in American Control Conference (ACC), 2017. IEEE, 2017, pp. 265 -- 270. [4] P. E. Rybski, N. P. Papanikolopoulos, S. A. Stoeter, D. G. Krantz, K. B. Yesin, M. Gini, R. Voyles, D. F. Hougen, B. Nelson, and M. D. Erickson, "Enlisting rangers and scouts for reconnaissance and surveillance," IEEE Robotics & Automation Magazine, vol. 7, no. 4, pp. 14 -- 24, 2000. [5] M. Zhong and C. G. Cassandras, "Distributed coverage control and data collection with mobile sensor networks," IEEE Transactions on Automatic Control, vol. 56, no. 10, pp. 2445 -- 2455, 2011. [6] A. Macwan, G. Nejat, and B. Benhabib, "Optimal deployment of robotic teams for autonomous wilderness search and rescue," in Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ International Conference on. IEEE, 2011, pp. 4544 -- 4549. [7] A. Gallais and J. Carle, "An adaptive localized algorithm for multiple sensor area coverage," in 21st International Conference on Advanced Information Networking and Applications (AINA '07), May 2007, pp. 525 -- 532. [8] J. Wang, S. Medidi, and M. Medidi, "Energy-efficient k-coverage for wireless sensor networks with variable sensing radii," in Global Telecommunications Conference, 2009. GLOBECOM 2009. IEEE. IEEE, 2009, pp. 1 -- 6. [9] J. Yu, S. Wan, X. Cheng, and D. Yu, "Coverage contribution area based k -coverage for wireless sensor networks," IEEE Transactions on Vehicular Technology, vol. 66, no. 9, pp. 8510 -- 8523, Sep. 2017. [10] F. Gustafsson and F. Gunnarsson, "Positioning using time-difference of arrival measurements," in Acoustics, Speech, and Signal Processing, 2003. Proceedings.(ICASSP'03). 2003 IEEE International Conference on, vol. 6. IEEE, 2003, pp. VI -- 553. [11] W. A. Gardner and C.-K. Chen, "Signal-selective time-difference-of- arrival estimation for passive location of man-made signal sources in highly corruptive environments. theory and method," IEEE Transactions on signal processing, vol. 40, no. 5, pp. 1168 -- 1184, 1992. i. [12] G. Mellen, M. Pachter, and J. Raquet, "Closed-form solution for determining emitter location using time difference of arrival mea- surements," IEEE Transactions on Aerospace and Electronic Systems, vol. 39, no. 3, pp. 1056 -- 1058, 2003. [13] J. Cortes, S. Martinez, T. Karatas, and F. Bullo, "Coverage control for mobile sensing networks," IEEE Transactions on robotics and Automation, vol. 20, no. 2, pp. 243 -- 255, 2004. [14] S. Poduri and G. S. Sukhatme, "Constrained coverage for mobile sensor networks," in Robotics and Automation, 2004. Proceedings. ICRA'04. 2004 IEEE International Conference on, vol. 1. IEEE, 2004, pp. 165 -- 171. [15] L. C. Pimenta, V. Kumar, R. C. Mesquita, and G. A. Pereira, "Sensing and coverage for a network of heterogeneous robots," in Decision and Control, 2008. CDC 2008. 47th IEEE Conference on. IEEE, 2008, pp. 3947 -- 3952. [16] M. Schwager, J.-J. Slotine, and D. Rus, "Decentralized, adaptive con- trol for coverage with networked robots," in Robotics and Automation, 2007 IEEE International Conference on. IEEE, 2007, pp. 3289 -- 3294.
1901.09660	1	1901	2018-12-09T06:34:33	Swarm Intelligent Algorithm For Re-entrant Hybrid Flow shop Scheduling Problems	[ "cs.MA" ]	In order to solve Re-entrant Hybrid Flowshop (RHFS) scheduling problems and establish simulations and processing models, this paper uses Wolf Pack Algorithm (WPA) as global optimization. For local assignment, it takes minimum remaining time rule. Scouting behaviors of wolf are changed in former optimization by means of levy flight, extending searching ranges and increasing rapidity of convergence. When it comes to local extremum of WPA, dynamic regenerating individuals with high similarity adds diversity. Hanming distance is used to judge individual similarity for increased quality of individuals, enhanced search performance of the algorithm in solution space and promoted evolutionary vitality.A painting workshop in a bus manufacture enterprise owns typical features of re-entrant hybrid flowshop. Regarding it as the algorithm applied target, this paper focus on resolving this problem with LDWPA (Dynamic wolf pack algorithm based on Levy Flight). Results show that LDWPA can solve re-entrant hybrid flowshop scheduling problems effectively.	cs.MA	cs	Swarm Intelligent Algorithm For Re-entrant Hybrid Flow shop Scheduling Problems Zhonghua Han Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, Liaoning, China Department of Digital Factory, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, Liaoning, China Email: [email protected] Xutian Tian* Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China Email: [email protected] Corresponding author Xiaoting Dong Faculty of Electrical Engineering, Sichuan College of Architectural Technology, Deyang, Sichuan ,China Email: [email protected] Fanyi Xie Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China Email: [email protected] Biographical notes: Zhonghua Han received his PhD degree at Shenyang Institute of Automation, China in 2014. He is currently a Professor with the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. His main research includes production and operation management, of automation system in enterprise, and the engineering application research of production scheduling method. technology integrated XuTian Tian is a master candidate from the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. He currently does his study under the supervising of Zhonghua Han.His main research area is production scheduling. Xiaoting Dong is a postgraduate student from the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. Her main research area is production scheduling. Fanyi Xie is a master candidate from the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. She currently does her study under the supervising of Zhonghua Han.Her main research area is production scheduling. Abstract: In order to solve Re-entrant Hybrid Flowshop (RHFS) scheduling problems and establish simulations and processing models, this paper uses Wolf Pack Algorithm (WPA) as global optimization. For local assignment, it takes minimum remaining time rule. Scouting behaviors of wolf are changed in former optimization by means of levy flight, extending searching ranges and increasing rapidity of convergence. When it comes to local extremum of WPA, dynamic regenerating individuals with high similarity adds diversity. Hanming individual similarity for distance to increased quality of individuals, enhanced search performance of the algorithm in solution space and promoted evolutionary vitality.A painting workshop in a bus manufacture enterprise owns typical features of re-entrant hybrid flowshop. Regarding it as the algorithm applied target, this paper focus on resolving this problem with LDWPA (Dynamic wolf pack algorithm based on Levy Flight). Results show that LDWPA can solve re-entrant problems effectively. scheduling flowshop is used judge hybrid Key words: Re-entrant Hybrid flow shop; simulations and processing models; Hanming distance; levy flight ; Swarm Intelligent Algorithm 1 Introduction Kumar(1993)first proposed the RHFS line(Li and Wang, 2010) as the third type of scheduling problem that distinguishes it from Flowshop and Jobshop. Reentrant manufacturing system have all jobs of reentrant working procedure featuring re-entering the waiting matching area. It leads to instability scheduling problems of reentrant workshop and obstruct, making scheduling problems of reentrant manufacturing workshop more complicated than general issue. Many tasks, various processes, multi work stations, and different processing time of jobs in the same working procedure are features of hybrid flowshop problems (HFSP) (Han et al., 2016; Karimi et al., 2011). If reentrant manufacturing lines appears in the RHFS, more complex production process, sharply increased production loads, unbalanced instability added and equipment loading and other questions greatly result in HFS difficulty. RHFS scheduling problem (Ying et al., 2015)is a typical kind of NP-hard problem. RHFS scheduling problems can be found in semi-conduct manufacturing, bus production and steel smelting. Multi chip packages of semiconductor packaging production line, multiple and multi-sticking crape masking process of a painting workshop have obvious characters of RHFS. bonding chip line Many scholars have made varying degrees of progress in research on RHFS scheduling. Liu et al.(2011) and Zhu and Chen(2018) use particle swarm with fast convergence and strong global searching ability of genetic algorithm for an exchanged result, designing a genetic particle swarm optimization for reentrant manufacturing dispatch algorithm. (Fan et al., 2012) study aero-engine reassembly line in a workshop after multi decomposition. With the stochastic matrix coding method, crossover method and variation method, an optimization scheduling method based on genetic algorithm is proposed to find the optimal method of assembly workshop scheduling. (Lin et al., 2016) aimed to meeting the characteristics of the dynamic reach ability and re-entry of jobs in mold heat treatment, weighted tardiness and energy consumption index are established as optimization objectives, having in mould heat treatment been in rolling schedule for heuristic algorithm. As for obstacles of two classes of RFFS, (Sangsawang et al., 2015) investigates GA based on fuzzy logic controller and PSO about Cauchy distribution to solve the problem. tempering process According to relevant references in recent years, existing works(Collart and Verschueren, 2014;Lin D et al., 2012) have supplied optimization algorithm to Re-entrant Flowshop scheduling problems and achieve good results. Algorithm is more accurate for optimization after improvement algorithm and combined with various algorithms, especially the traditional meta heuristic algorithm and improved algorithm which includes (Re-entrant flexible Flowshop, (RFFS) scheduling problems, there are few references about RHFS scheduling problems. This paper investigates LDWPA (The Dynamic wolf pack algorithm based on Levy Flight), and show LDWPA has faster speed and higher accuracy compared with other existing algorithm 2 Describe and mold about RHFS dispatch problems jobs processed As for RHFS scheduling problems, it can be described as in working procedures. Jobs are produced according to the specified sequence of operations in the process flow. One working procedure of all contains at least one work station. There are differences about process time of jobs. At least one process with re-entrant features work station must be supplied jobs traversing all process are re-inserted into the queue to be processed. Jobs with the least remaining processing time are chosen in free work station. Scheduling problems are about work station distribution, process, start time and completion time. Process must follow the rules below: 1 Every workstation can manufacture only one job's process flow .After to job at a time. 2 Every procedure of one job are processed in one workstation. 3 The whole procedure of job process are not allow to be interrupted. 4 Transit time of jobs is neglected during processes. nrm During the scheduling process of RHFS, in the number non-reentrant manufacturing process, each workstation consists of a set of the . In the number same work stations rm re-entrant manufacturing phase, each workstation composes a set of the same work stations ; the job iJ needs to 1 ~ M  re-enter this phase times, while different job M M nrm rm irts 1 ~ nrm M nrm has its own re-entrant times. In this problem, each job in accordance with the specified process flow goes through various workstations for processing. Each job will be processed for times. Flow process chart is nrm rm rts i shown in Figure 1.   Figure 1 the chart of process sequence Non- reentrant working procedure 1 M1 M2 …… M1 Non- reentrant working procedure nrm M1 M2 …… Reentrant working procedure1 Reentrant working procedure rm M1 M2 …… M1 M2 …… Mnrm Mnrm+1 Mnrm+rm 2.1 Mathematic simulations and processing model description 2.1.1 Parameter setting In this section, there are many mathematical variables of the re-entrant hybrid flow shop for establishing its mathematics planning models. n means the largest number of processing jobs iJ means the i th job m means procedures. the whole number of working {1,....., } n  . i jOP means j process. jM means the largest work station of jOP working procedure. ,j kWS means the kth work station of jOP working procedure. qFL means the process flow of jobs, it is processed working procedure collection of jobs according to the process flow. rm means the re-entrant processes number in production line. nrm means numbers in production line. non-re-entrant processes the irts means iJ job experience times of reentrant working procedure. iom means the whole number of processed jobs in production procedure qFL . Cumulative the whole number of counting about re-entrant working procedure, l means the sequence number of being processed working in procedure iom m the . qFL l  {1,....., om }i process flow. S l , , i j k means iJ job, the jOP working procedure in the process flow of qFL , the start process time. C l , i , j k means iJ job, the jOP working procedure in the process flow of qFL , the finished process time. i j kWT , l , iJ job, the jOP working procedure in the process flow of qFL , the process time. ,j kWS means the serial number of job in the kth station of jOP working procedure. 2.1.2 Assumed variables and basic constrained relationships In this section, many mathematical formulas are established with the variable set up in the through which constraint previous section, in relations scheduling optimization the scheduling procedure the process included, are and features of reflecting of optimization in the Re-entrant Hybrid flow shop. workshop.（5）indicates that the whole number of At ti , , i j k    0      1   i , j k of J doesn t process on ' WS t i Workpiece the work station O P j during the number working . procedure Workpiece the work station during the numbe proce process on of WS r wor t i OP j king dure , j k J i The variable ,j kWS represents if the jobs are in the qFL process. In the l th processed of jOP working procedure is about the process situation of ,j kWS job. C l , , i j k  S l , , i j k  WT l , , i j k , i  {1,....., } n , l  {1,....., } om , j  {1,....., } m (1) C l , , i j k S l , , i j k , i  {1,....., } n , l  {1,....., } om , j  {1,....., } m (2) omi  ti  1 At ti , , i j k  om i , i  {1,....., } n , l om i n , j k omn i   ti  1  1 i At ti , , i j k (3) (4)    nrm rm rts i (5) om i Formula (1) indicates relationships among job's start process time, process time and finished process time in its processing flow with the re-entrant process. Formula (2) indicates limited relationships of the same job between continuous process start and finished process time. Formula (3) indicates every job should experience their whole process flow. Formula (4) indicates that it is necessary to accumulate number of repetitively processed jobs when counting the number of jobs processed re-entrant in a workstation in jobs in qFL equals to the whole number of iJ in non-reentrant and reentrant working procedures. 3 WPA designs The Dynamic wolf pack algorithm based on Levy Flight (LDWPA) after utilizing WPA(Yang et al., 2007;Zhou et al., 2013) is in this paper, which includes two improvements. 1 Scouting behaviors based on Levy Flight: scouting behaviors are about position changes from present situation to new one. The search ranges of this kind of local random walking is focused in such small scope. Basically, the Levy Flight is a random walk combines of long step and small step. The Levy Flight has a larger search ranges and ability compared with random walking because of directed long step. Founding wolf position is improved according to levy flight, the wolf rushed to its own searching area of low-probability with a long step in solution space. It also extends searching range for a development of wolf optimization. 2 The species dynamic mechanism based on hanming distance: After a certain number of iterations, WPA is shows the phenomenon of evolutionary stagnation. The reason is that with the increased iterations, more similar and even the same genes of individuals and useless communication between individuals attribute to this result. Therefore, a dynamic updating population method based on stagnation evolution iterations is introduced. If the optimum value of stagnated revolution iterations exceeds threshold, the hanming distance can be used to judge the similarity among the individuals with the highest similarity of best one and individual with low similarity and large differences, making a diversity, highlighting local extremum and individuals left new abandon and (cid:160) keeping vitality of algorithm. 3.1 Population dynamic renewal mechanism to In order The basic concept and characteristics of Hanming distance are charted in 3.1.1 firstly. In 3.1.2 is about real examples of LDWPA to reflect objectively the distance of long code word based on Hanming distance. judge differences among individuals in population on the condition that the iterative process falls to hysteresis, the method of using Hanming distance the difference among individuals for less difference one is introduced in 3.1.3. Besides, it can be used to increase new individuals with more differences, enhance diversity and maintain evolutionary vitality. evolution during judge to 3.1.1 Hanming distance Hanming distance (Harada et al., 2017; Atallah and Duket 2011; Haider et al., 2015) is a basic concept in information theory, describing the distance of two long code words. D x y ( , )  n  x k k  1 x  (x , x , ..., x ) 1 2 n , y k , y  (y , 1 2y , ..., y ) n (6) , {0,1} kx  represents  XOR operation, D x y means the whole number of two code words in the same position. It reflects the differences between two code words and evidences for the similarity. ky  {0,1} ( , ) , and improve the time efficiency. Through the similarity D between wolves, that is, the ratio of the same gene in the two individuals to the total number of genes in the individual. Then compare with the threshold to find similar individuals. Firstly, individual is ranked by fitness. The similarity D between the leader wolf and other individuals in wolf pack can be calculated. If D is greater than the threshold Rt , two individuals are similar individuals which leads to a temporary subpopulation 2 StPop is built in the rest individuals by repeating the above operations for better individuals till the end. In every temporary subpopulation, many similar individuals can be ranked according to adjustment degree. Partial individuals can be reserved based on Kr ratio of better individual as its revolution adaptation out of others. StPop . 1 d , i ti 0      1      a , i t i a , i t i       ' a , i t ' a , i t     i i     The variate ,i tid means upper line of two different individuals npX and 'npX . The responding gene section    reflects situation of the upper line. If the situation   and a '   ,i ti ,i tia  is same, i tid  ; if not, 1 , i tid  . 0 , N D omn i   ti  1  1 i D , i ti (7) 3.1.2 The individual similar judgment based on hanming distance ND in (7) means the number of two individual with the same gene. In WPA, the individual is the sequence of real numbers and matrix. The individual similarity can be found from the individual distribution of population. In view of the fact that similarity judgment method is in the evolve algorithm, in order to reduce the complexity of the operation D  D N N Gene (8) D is the similarity of individual and the proportion between DN of two individuals with the same gene section and the whole number of GeneN . If it surpasses the threshold Rt , two individuals is similar. 3.1.3 Dynamic renewable similar individual The individual with a high similarity can be found after every generation when it comes to a certain iterations. Reserving superb individuals, eliminating similar individuals, then generate new individuals with large differences from the population and replace individuals with high similarity in order to preserve good distribution of population in solution space and strengthen the global searching ability. The purpose of using starting generation StartGen mainly consider that the population still has good evolutionary vitality at the early stage of evolution. When it comes to a certain number of iterations, the vitality will decrease and then the operation of dynamic updating population is introduced which reduce computation of the algorithm during the entire evolution process. it et in flight(Palyulin improvement of natural random phenomenon, such as Brownian movement. Levy al., 2014)conforms to random searching routines of the levy distribution. It is a walking way combined with short distance searching and occasionally long distance walking. Levy fight is adopt to upgrade many population optimization. Many scholars (Ibrahim, 2016; Ehsan et al., 2013) utilize informational interaction among individuals and searching for the optimal solution in the solution space. In a conclusion, there is a possibility of population individual advancing to former small searching range, expanding the range. It has achieved satisfactory results increasing population diversity and also improved the rapidity and veracity of algorithm. Therefore, the use of Swarm Intelligent Algorithm based on Levy flight makes it easier to jump out of the local extremum which can effectively enhance the algorithm's optimization ability. in The new method of competitive wolf's position based on levy flight scouting behaviors. ( ) t x i  ( , ) stepa levy u v  (9)   ( 1) t x i  1...... n 3.2 Scouting behaviors based on levy flight i This section combines the dynamic renewal populations based on Hanming distance mentioned above and the Scouting behavior depended on Levy flight. By creating the constraint relation of different variables and listing the changing variables in the evolution process. This part describes all steps of the LDWPA proposed in this paper in detail and makes the logical relations more clearly based on the flow chart of the algorithm. French mathematician Levy put forward a probability distribution called levy distribution in 1930s, which leads to a large amount of scholars' researches. Till now, scholars can prove that foods searching routines of many animals and insects (albatross, bee and fruit fly) are the same with levy distribution. It also explains many ( )t ix means the t th position of competitive wolf； is multiplication; step is step element. to Levy flight basically is a random step which levy distribution. It has been conforms achieved because of levy distribution, which leads to the frequent usage of Mantegna algorithm. Mantegna algorithm about levy flight is as followed. the complicated levy  ( ) ~ u   u 1 v  (10)  u           (1  )sin(   [1   ]  2  2   (1 2 1  ) )        , ~ (0, u N  u 2 ) (11) 2  1,v ~ N( v   v   1.5 ) (12) A larger searching range and improvement of WPA can been achieved by advancing to a small possibility range of levy flight. 3.3 LDWPA operations the Step1 The parameter of original algorithm and Np wolves form initial population, initializing positions of every wolf. The maximum number of iteration is Genmax , the is q , each number of scouting wolf scouting wolf hunt prey in h directions, the is SCmax and . The searching step maximum SCmax  searches 15 stepa and move step are stepb . There are bw wolves will be eliminated in each iteration. Step2 q wolves are used for the leader wolf by hunting. The cwp k wolf's scouting behaviors searches for better positions based on levy flight and (9). Step3 Other wolves which fail to implement scouting behaviors followed the best wolf as the leader wolf. It's position is changed through   x kd x kd   ( rand stepb x ld   x kd ) : rand is a random number between (0,1). ldx is d dimensional vector of the leader wolf. Step4 Regenerating population according to the wolf distribution principle, removing the worst bwwolf and creating newbwwolves by initializing position through X kbw  x min  ( rand x  x max min ) limitation and lower limitation of value. The leader wolf randomly moves to find the food, notifies the other wolves around the prey by howling, and the other wolves surround the leader wolf. For this behavior, a random number kr generated in [0,1] is generated. If kr is smaller than (a preset threshold), the wolf kx does not move. If kr is greater than  .The wolf kx surrounds the prey with the leader wolf. The position new of k wolf is 'k x  x k  rand ra  ( ) ra t  ra min  ( x max  x min  ) exp( ln( ra min / ra maxt max )  t ) Step5 The and best individual revolution limitation can be judged or not. If it meets the condition, it will be finish; if it does not meet the condition, it will continue. Step6 If the population revolution is none of newer best individual, it is necessary to stop evolve iteration. stopgen stopgen  . 1  Step7 The evolvement reaches StartGen iteration at the beginning of new population similar individual or not. If gen StartGen , which indicates the calculating of iteration and turn to Step 2; Otherwise, it starts operational formula of dynamic update similar individuals. it does not reach that  SI that similarity SI the Individuals among Step8 Calculating are individuals. greater than the threshold Rt are made into a temporary subpopulation, Subpopulation of subpopulation temporary is ltPop .Each member Nt in ltPop contains several subpopulations stPop. The number of each stPop is stnp . temporary subpopulation these maxx and minx represents the upper Step9 Assuming the individual count variable of a is inp to better . temporary new population According individual reserve percentage Kr . Preserving better individuals of subpopulation are . That is, preserving individuals Kr with fitness values that match evolutionary trends. The best individual is put into a new group newPop and in p ＝ Kr each  stnp temporary stnp .  individual number of newPop reaches to N p and set a new individual count variable tnp  . 0 Step10 If the Step11 If the new individual number tnp bigger or the same with  N p  in p  and then to the step 13; if not, go on to the Step 12. above newPop individual of Step12 New individuals came into being in this way to judge the similarity between the new individual and individual of newPop . This new individual will be abandoned as long as the similarity between the new individual and the threshold Rt . This new individual will be reserved as long as the similarity between the individual of newPop lower or the same with the threshold Rt . This individual is added to newPop ， tnp = tnp +1, and go on to Step 11. Regarding Pop as a new group Pop with  gen q is 5, searching direction his 4, searching step is the first step before Step 2. individual new and gen  1 Step13 stepa is 0.6  x max  x min and moving step stepb is 0.3. t is the current iteration number, m a x r a maxt is the maximum number of iterations, =400 is the maximum surrounding and ra steps. =0.5 is minimum surrounding steps. Similarity threshold Rt is 0.6. Better individual reserve ratio K r is 0.3. m in Figure 2 The chart of algorithm sequence Start Initial population Y Gen>maxt raining generations Genmax N Scouting behaviors based on the levy flight Competition for the leader wolf Move to the the leader wolf . Beleaguering behaviors Exit and output the best results Judge of meeting terminational conditions Y N N New optimization or not Y Stopgen=Stopgen+1 Stopgen=0 N Stopgen>stopgenmax Y According to the SI between individuals, ItPop of each temporary subpopulation is formed, includingNttemporary subpopulationstPop.ThetNzequals to0. inpis0. N tNt>= Nt Y Based on the reserved ratio Kr and temporary subpopulationstPop, the best individual of temporary subpopulationstPopcan be found and reserved it in a populationnewPop. tNt=tNt+1，inp=inp+KrstPopm Increasing the individual number of populationnewPopbyNP,tnp=0 Tnp=Np-inp N Y A new individualwas createthroughtheindividual built method of initial population N The similarity between individuals innewPop over or equals to Rt Y Reserve the new individual in population newPop Abandon the new individual tnp=tnp+1 Pop=newPop,gen= gen+1 4 Simulation experiments In this section, The LDWPA is used to optimize benchmarks of different scales in this section. The optimization results are compared with these obtained of GA, WPA and WOA simulation examples in order to prove the effectiveness and superiority of the LDWPA algorithm. 4.1 Algorithm analysis To verify effectiveness of LDWPA algorithm mentioned in the former part, the flexible flow shop examples are used to compare performance of the algorithm under the real facts that researches on re-entrant hybrid flow shop scheduling optimization problem is still in the primary stage, and lack of examples. Adopt the benchmark published by (Carlier and Néron, 2000), this section includes three hard instances: j15c5d2, three easy j15c5d4 and instances: j15c5a4, j15c5a5, j15c5b1. j15c5d3, Comparing the LDWPA algorithm with the NEH algorithm mentioned in (Ribas et al., 2010) research and the WPA algorithm, the best lower 'LB is known as the best optimal bound scheduling result. The effectiveness of the algorithm is evaluated by the following indexes: deviation d refers to the deviation between maxC and 'LB . Among in d refers to: d  (( C max  LB ') / LB  ') 100% Based on CPU 2.5GHz and 4G internal memory, MATLAB environment is the prerequisite of GA, WOA, WPA, LDWPA algorithms in this paper. examples are compared with those obtained by WOA, WPA and LDWPA selected in this paper. This proves the validity and superiority of the algorithm in dealing with various scale data. Table 1 the results of instance in three algorithm Figure 3 The Iterative Curve Diagram of Four kinds of algorithms to solve the Scheduling Problems of instance of J15c5d3 in the easy In table 1, it can be concluded from the table that three algorithms have obtained best optimization results instances. However, the advantages of LDWPA algorithm is more obvious in hard problems. The optimization effects of WPA and NEH are nearly the same as for the easy instances. In the difficult instances, such as J15c5d3, the effects of WPA optimization algorithm are worse than NEH because that WPA algorithm is easy to the local extremum and lack of evolutionary vitality. Both the optimal solutions and average optimal solutions from LDWPA are better than those from the contrast algorithm NEH, which shows that the LDWPA algorithm mentioned this paper greatly improves the overall optimization effects of the WPA algorithm, effectiveness and superiority of LDWPA algorithm. in In this paper, two small scale examples and five large scale examples mentioned in (Sun et al., 2017) research are used for simulation test. By using GA algorithm mentioned in Sun Y's paper, the optimization results of different scale till the small stable Figure 3 is a proper revolution curve graph from optimization algorithm data of GA, WOA, WPA and LDWPA. With the development of exercise in figure 3, the fitness value of four groups is smoothly situation. Optimized curve convergence rate is slow in GA algorithm. The optimization is inferior to other three methods in limited evolution frequency. The optimized curve convergence rate with WOA algorithm is rapidly and stop within 20th generation, leading to the local extremum and bad evolution situation. The bad fitness reflects the fast convergence rate、local extremum and bad optimization. The optimized WPA algorithm between 10th generation and sharp decreased 20th generation is superior to whales optimization algorithm in optimization speed. It also exists the weak local extremum capacity and fragile vitality. The worse first generation optimization fitness through the LDWPA algorithm is the same with the initial generation among 20th generation. Thus, the optimized curve by WPA is nearly reach the fitness. The optimization speed is faster than WOA and WPA featuring fast speed. It is concluded that the Levy flight optimizes the local searching and fosters the algorithm optimization. With the dynamic new population, algorithm among 10th generation and 40th generation keeps evolution after its local extremum and jump from the local extremum for a better fitness value 86. The dynamic upgraded population greatly improves WPA of being in local extremum. Figure 4 The chart of four type algorithms Figure 4 is the Box-plot about fitness value after 20 times same date by means of optimizing GA, WOA, WPA, LDWPA. The abscissa is the algorithm of example date and the ordinate is about fitness values. The box's tendency is down in the process of optimization from the Box-plot above, testifying the whole result is up in the comparable four algorithms. The minimum of improved algorithm is smaller than other three algorithms, resulting to the higher possibility of qualitative results in the improved algorithm. 1st quartile and maximum are equal to 3rd quartile of WPA and minimum, which is also smaller than the range of median of GA、3rd quartile and fitness value of algorithm. It can be concluded that the overall results' quality through the LDWPA algorithm is superior to other two algorithms. The unusual number 90 is related to fitness value of initial population. The worst result appears when the fitness value of initial population is 20. From the partial solution examples in table 2, the fitness value of every algorithm is similar for the small range examples. The LDWPA is still better to get results, whose average is also less than that of other algorithms. That is to say, the LDWPA has good optimization ability in small range of date. It is necessary to notice that the former WPA , with the feature of being into local extremum, exists phenomena of lacking evolve vitality and relying on optimal value of initial population. From the chart, GA algorithm has a feature of jumping out of local extremum. But results of GA algorithm is better than those of WPA with enough generations. With population dynamic mechanism and the levy flight in neighbor searching, the LDWPA's results is superior to GA algorithm in medium and big date, which proves the efficiency and stability of this algorithm in the paper. Table 2 The results of examples in different algorithms 4.2 Experimental analysis on LDWPA results and RHFS dispatching problems This section uses the LDWPA algorithm, which has proved its superiority in the former section to optimize the operation section of multiple masking procedures in the painting shop with the characteristics of re-entrant hybrid flow shop in the actual bus manufacturing enterprises. By establishing different evaluation indexes, the superiority of LDWPA algorithm to solve the re-entrant shop scheduling optimization problem can be proved. hybrid flow Figure 5 Deployment Diagram of Work Station in Multi-Sticking Crape Masking Procedure Some automobile manufacturing enterprises provide scheduling date about the practical information of multi-sticking crape process masking working coating workshops. This figure 5 is about the distribution of multi-sticking crape masking working procedure. procedures in The result is about the scheduling problems crape masking working in 15 automobiles' bodies. The of multi-sticking procedures working procedures of  O P O P O P 3 1 2 , ,  mean stickingcolor ribbon、spraying and baking paint in multi-sticking crape masking working procedures. These three working procedures contain parallel work stations of 2,3,4. painting housefor spraying as well as painting house for baking paint are treated as work stations in thescheduling procedure.  1 J J , , J J 3 , 4 , J J 5 , 6 2 , J , J 8 7 J , J 10 , 9 J 11 12 J , , J 13 , J 14 , J 15 r  in all working procedure epresent jobs of the fifteen types of car body to be processed. Fifteen cars belong to fifteen types manufactured by its order. X in the list means automobile process time of multi-sticking crape masking working procedures. The manufacture time is different because of different types of buses. There is no people in the process of baking paint in painting house. Therefore, the baking paint time in different painting house about the same type of automobiles is no different and working time difference is related to types. Table 3 Time of multi-sticking crape masking working procedures（min） times factory. In the working multi-sticking crape masking working procedures of spraying workshop in automobile manufacturing The technological persons disperse color ribbon into different colors according to the difficulty of color ribbon models and spray various color ribbon patterns on bus external parts. Spraying one kind of color after sticking color ribbon(crape masking) 、 spraying and baking paint and the whole pattern come into being after several Therefore, multi-sticking crape masking working procedure is a typical re-entrant stage, including many working procedures and work stations. This working procedure has obvious features of hybrid flowshop because that various abilities of technological persons attribute to different process times. Aimed at scheduling problems of multi-sticking working procedures in spraying workshop, this paper optimizes it with LDWPA and resolve scheduling problems in RHFS. crape masking spraying. of Working Procedure Work station Time of multi-sticking crape masking working procedure J1 J2 J3 J4 J5 J6 J7 J8 J9 J10 J11 J12 J13 J14 J15 OP1 1stC OLO RRI OP2 BBO N 2nd COL ORR IBB ON 3rdC OLO OP3 OP1 OP2 OP3 OP1 RRI OP2 BBO N OP3 2 5 1 WS1,1 WS1,2 WS2,1 WS2,2 WS2,3 3 WS3,1 WS3,3 4 WS1,1 WS1,2 WS2,1 WS2,2 WS2,3 6 WS3,1 WS3,3 7 WS1,1 WS1,2 WS2,1 WS2,2 WS2,3 9 WS3,1 WS3,3 8 FL1 FL2 12 15 FL3 15 FL4 12 FL5 FL6 FL7 FL8 FL9 FL10 FL11 FL12 FL13 FL14 FL15 15 15 15 20 12 18 15 15 15 20 15 15 15 20 18 25 15 18 22 20 25 12 15 15 18 20 18 18 15 20 20 20 20 25 20 25 18 25 25 20 25 18 25 18 20 25 15 18 20 18 20 10 12 15 15 15 18 10 15 15 18 20 12 10 15 10 15 18 12 18 15 20 20 15 15 20 15 15 20 20 20 20 22 25 15 10 15 15 20 20 18 20 20 22 25 15 12 15 15 20 20 15 15 18 15 20 15 12 18 15 18 20 10 12 10 15 15 18 18 18 20 15 20 15 12 15 18 15 18 12 10 15 10 15 18 12 20 16 15 20 12 15 15 18 15 18 12 10 12 15 10 18 The GA, WPA and LDWPA algorithms are used to compare the different evaluation indexes. The makespan maxC is used as the fitness value function of the overall optimization algorithm during the optimization process. At the same time, many evaluation indexes related to the practical application of the production line are established. These include the total load balance costTLB, the total workstation free time TWT and the total equipment utilization rate FUR. Except for the total equipment utilization FUR, the other evaluation index value is smaller and better. The evaluation indicators are described as following: WT j           TLB  M j n om       1 1 1 j l i WT l , , i j k  t i , , i j k At         M j j Mm  om      1 1 j l          1  k WT l , , i j k  At i t , , i j k            WT j 2          (13) (14) jWT in formula (13) means the average processing time of jM work stations of jOp working procedure. By summing up the Total load balance TLB. total processing time at each station ,j kWS of the procedure jOp and the average processing time jWT of the work stations and the square of the average processing time jWT .The load balance cost TLB is established in formula (14). j The total load balance cost TLB is the sum of the cost of load balance of all working procedures during the whole process. the Total equipment utilization ratio FUR FUR  M j m   1 k     i n      1  om lWTi     ,   1 1 l j  t i At , i , j k , j k   min             l , i , j k  t i At , i , j k         (15) , j k  S m   1 j  M j        1 k max  C l , i FUR in formula (15) is the total equipment utilization ratio of all stations in the flexible flow shop . It is also the ratio of the all effective processing time of work stations and the span of the work station working time. This time span is a period from the first processing task to the final processing task. the beginning of Table 4 The results of evaluation indexes in three algorithms Optimization performances of each optimization algorithm is analyzed from the data in Table 4. The average of maximum completion time maxC from LDWPA algorithm is better than that of GA and WPA algorithm. The three algorithms are similar to each other in terms of furniture utilization ratio FUR , but the optimization effects of LDWPA is still slightly higher than that of the other two algorithms. The optimization effect of WPA algorithm for total load balance TLB and free time of total work station TWT is obviously worse than that of GA algorithm. However, the optimization effect of LDWPA algorithm on every evaluation index is better than that of GA algorithm, which shows that the optimization ability and the quality of solution obtained by LDWPA algorithm are greatly improved compared with WPA algorithm. LDWPA can optimize every index, which shows that LDWPA can reduce the free time of work station better, arrange the processing task more reasonably, make the equipment load more balanced, and have better optimization performance. Figure 6 shows the gantt of the re-entrant hybrid flow shop based on the actual data during production process. The work 7J , traversing the process route, passes through the work station W S 1 2 , W S W S 2 1 3 2 , W S W S W S , , 1 2 2 1  3 3 , and the work 2J firstly it is produced in station enters when the Re-entrant working 11WS procedure 1OP .The job is produced in the work station 12W S when it first comes to re-entrant working procedure 1OP . Besides, the two stations in the working procedure 1OP is different. Both 2J and the work of them indicates that the scheduling of work production tasks is constantly changing during the ongoing production procedure, reflecting that Re-entrant Hybrid flow shop scheduling problem is a dynamic problem. Special attention is paid to entering the work 7J firstly the second working procedure 2OP of the re-entrant working procedure part at the same time. According to the local assignment rule specified in this paper, the work with the smaller remaining has a higher priority to enter the station for processing. According to table 2, the remaining of , so is shorter than that of 2J 7J the work station 21W S 7J . is first processed by the free Acknowledgement Figure 6 Gantt results of Re-entrant Hybrid Flowshop This work was supported by Liaoning Provincial Science Foundation(No. 201602608), Project of Liaoning Province Education Department（No. LJZ2017015）, Shenyang Municipal Science and Technology Project(No. Z18-5-015) and Project of Sichuan Province Education Department (No. 17ZB0823). Reference 5 Conclusion the complete Based on multi-sticking crape masking in spraying workshop of automobile manufacture enterprises, this paper investigates the RHFS searching problems. In view of reentrant process system in practical society, this paper also raises RHFS dispatch searching problem with reentrant system and describes this problem. Mathematics models come into being with the target of minimizing time. This paper highlights the LDWPA algorithm. Improving the scouting behaviors of WPA with the Levy flight and evolution power of algorithm with dynamic regenerating population are used to improve WPA algorithm and resolve this problem. Any scales of simulation results are analyzed by experiments and compared with other algorithms. The simulation results show the proposed LDWPA algorithm can acquire near-optimal solutions in reasonable time. Our future work will focus on decreasing algorithm operation time and increasing control requirements for the quality of big scale problems. Kumar, P, R . (1993) 'Re-entrant Lines', Queuing Systems, Vol.13,No,1-3,pp.87-110. Wang, C.,Li, J.(2010) 'Approximate Analysis of Reentrant Lines With Bernoulli Reliability Model', IEEE Transactions on Automation Science and Engineering, Vol.7,No.3,pp.708-715. Han, Z., Zhu, Y., Ma, X., Chen, Z. (2016) 'Multiple rules with game theoretic analysis for flexible flow shop scheduling problem with component altering times', International Journal of Modeling Identification and Control, Vol.26, No.1,pp.1-18. Han, Z., Ma, X., Yao, L., Shi, H. (2012) 'Cost Optimization Problem of Hybrid Flow-Shop Based on PSO Algorithm', Advanced Materials Research, Vol. 532, pp. 1616-1620. Han, Z., Wang, S., Dong, X., Ma, X. (2017) 'Improved NSGA-II algorithm for multi-objective scheduling in hybrid flow shop', 2017 The 9th problem International Conference on Modeling, Identification and Control (ICMIC), Kunming, China, pp.740-745. Sun, Y., Lin, S., Li, T., Ma, X. (2017) 'Improved Imperialist Competitive Algorithm for Flexible Flow Shop Scheduling', 2017 The 9th International Conference on Modeling, Identification and Control (ICMIC),Kunming, China, pp. 169-174. Han, Z., Sun, Y., Ma, X., Lv, Z. (2018) 'Hybrid Flow Shop Scheduling with Finite Buffer', International Journal of Simulation and Process Modelling, Vol.13,No.2, pp.156-166. Ying, K., Lin, S., Wan, S. (2015) 'Bi-objective reentrant iterated Pareto of hybrid flowshop scheduling: an greedy Production Research, Vol.54,No.12,pp.1-18. International algorithm', Journal Hekmatfar. M., Ghomi, S.M.T.F., Karimi, B. (2011) 'Two stage re-entrant hybrid flow shop with setup times and the criterion of minimizing makespan', Applied Soft Computing Research, Vol.11, No.8, pp.4530-4539. Liu, X., Lin, J., Deng, K. 'Scheduling Optimization in Re-entrant Lines Based on a GA and PSO Hybrid Algorithm', Journal of Tongji University (natural science), Vol.39,No.5,pp.726-730. (2011) Fan, J., Yan, H., Zhou, K. 'Optimizing Aeroengine Assembly Shop Schedule Based on Genetic Algorithm', Computer Technology and Development, Vol.22,No.9,pp.205-209. (2012) Lin, G., Liiu, J., Mao, N. (2016) 'Dynamic batch re-entrant mould heat-treatment Integrated Manufacturing scheduling flow-shop', Computer System, Vol.22,No.4, pp.1046-1058. for Sangsawang, C., Sethnanan, K., Fujimoto, T., Gen, M. (2015) 'Met heuristics optimization approaches for two-stage reentrant flexible flow shop with blocking constraint', Expert Systems with Applications, Vol.42,No.5,pp.1046-1058. Collart. C., Verschueren. K. (2014) 'Re-entrant flow shop scheduling problem with time windows using hybrid genetic algorithm based on auto-tuning strategy', International Journal of Production Research, Vol.52, No.9,pp.2612-2629. Lin, D., Lee, C.K.M.., Wu, Z. (2012) 'Integrating analytical hierarchy process to genetic algorithm for problem', re-entrant International Journal of Production Research, Vol.50,No.7,pp.1813-1824. scheduling Yang, C., Tu, X., Chen, J. (2007) 'Algorithm of marriage in honey bees optimization based on the wolf pack search', Computing, International Conference on(IPC), Jeju Island, Korea. Intelligent Pervasive Zhou, Q., and Zhou, Y. (2013) 'Wolf colony search algorithm based on leader strategy', Application No.9, of Research pp.2629-262932. Computers, Vol.30, shop flow Harada, Y., Fujimoto, K., Fukuhara, M., Yoshida M. 'Minimum Hamming Distance Search (2017) Associative Memory Using Neuron CMOS Inverters', Japan, Electronics & Communications Vol.100,No.1, pp.36-42. in Atallah, M.J. and Duket, T.W. (2011) 'Pattern matching in thresholds', Information Processing Letters, Vol.111,No.14 pp.674-677. the Hamming distance with Haider, B., Suresh, D. (2015) 'A Hamming distance based binary particle swarm optimization (HDBPSO) algorithm for high dimensional feature selection, classification and validation', Pattern Recognition Letters, Jharkhand, India. Palyulin, VV., Chechkin, AV., Metzler, R. (2014) 'Levy flights do not always optimize random blind search for sparse targets', Proceedings of the National Academy of Science of the United States of America,Vol.111,No.8,pp.2931-2936. Ehsan, V., Elham, V. (2013)'A cuckoo search algorithm by Levy flights for solving reliability redundancy allocation problems' Engineering Optimization, Vol.45,No.11,pp.1273-1286. Ibrahim, A. (2016) 'Cost optimization of reinforced concrete cantilever retaining walls under seismic loading using a biogeography-based optimization flights' algorithm with Engineering Optimization, Vol.49,No.3,pp.381-400. Levy Carlier, J., Néron, E (2000) 'An exact method for solving RAIRO-Operations multi-processor Research, Vol.34, No.1, pp.1-25. flows-shop' Ribas, I., Listen, R., Framinan, J.M. (2010) 'Review and classification of hybrid flow shop scheduling problems from a production sys-tem and a solutions procedure perspective' Computer & Operations Research, Vol .37, No.8, pp.1439-1454.
1004.4450	1	1004	2010-04-26T09:37:46	Improving Supply Chain Coordination by Linking Dynamic Procurement Decision to Multi-Agent System	[ "cs.MA" ]	The Internet has changed the way business is conducted in many ways. For example, in the field of procurement, the possibility to directly interact with a trading partner has given rise to new mechanisms in the supply chain management. One such interactive dynamic procurement, which lets both buyer and seller software agents bid by potential buyer agents instead of static procurement by vendors. Dynamic procurement decision could provide the buying and selling channel to buyer, to avoid occurring condition that seller could not deliver on the contract promise. Using NYOP(Name Your Own Price) to be the core of dynamic procurement negotiation algorithm sets up multi-agent dynamic supply chain system, to present the DSINs(Dynamic Supply Chain Information Networks) by JADE, and to present the dynamic supply chain logistic simulation by eM-Plant. Finally, evaluating supply chain performance with supply chain performance metrics (such as bullwhip, fill rate), to be the reference of enterprise making deciding in the future.	cs.MA	cs	JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Improving Supply Chain Coordination by Linking Dynamic Procurement Decision to Multi-Agent System Yee M ing Chen 42 Abstract—The Internet has changed the way business is conducted in many ways. For example, in the field of procurement, the possibility to directly interact with a trading partner has given rise to new mechanisms in the supply chain management. One such interactive dynamic procurement, which lets both buyer and seller software agents bid by potential buyer agents instead of static procurement by vendors. Dy- namic procurement decision could provide the buying and selling channel to buyer, to avoid occurring condition that seller could not deliver on the contract promise. Using NYOP(Name Your Own Price) to be the core of dynamic procurement negotiation algorithm sets up multi-agent dynamic supply chain system, to present the DSINs(Dynamic Supply Chain Information Networks) by JADE, and to present the dynamic supply chain logistic simulation by eM-Plant. Finally, evaluating supply chain performance with supply chain performance metrics (such as bullwhip, fill rate), to be the reference of enterprise making deciding in the future. Index Terms—Multi-Agent System、Bullwhip effect、 Fill rate 、NYOP(Name Your Own Price) ——————————  —————————— 1 INTRODUCTION N owadays agent technology is utilized in many in- dustrial applications such as production planning, collaborative engineering, and more recently supply chain management(SCM). One of the most difficult but critical issues in SCM is to improve the efficiency of supply chains from the perspective of the whole supply chain, not individual companies. More specifically, the formation, optimisation and minimizing the bullwhip effect(BWE) (the magnification of demand fluctuations as orders move up the supply chain.) [3] are considered crit- ical issues for efficient supply chain management [14]. Current approaches for reducing the BWE can be distin- guished in (1) information sharing (2) collaborative plan- ning. Although it is well-known that sharing information among supply chain members can lead to improved effi- ciency, information sharing is not always possible, often because of the limitations in information systems[1][12]. Collaborating production and orders of supply chain members by a single company or a decision-making unit is in many cases infeasible, because supply chain mem- bers are usually independent companies. Considering the limitations of information sharing and collaborative plan- ning, this paper proposes an agent-based Dynamic Supply Chain Information Networks (DSINs) in which each participant decides about procurement and produc- tion based on local information only. The interaction with other participants adopt a special form of Name Your Own Price(NYOP). The NYOP channel, exemplified by Priceline, is a popular online alternative to other, more traditional channels, through which service providers ————————————————  Yee Ming Chen is with the Department of Industrial Engineering and Management, Yuan Ze University, Taoyuan, Taiwan. such as airlines, hotels, and car rental companies offer their products to customers. This paper is motivated by the importance of developing an understanding of the market implications of this pricing and distribution mod- el. Therefore,NYOP plays a dominating and in many cas- es an exclusive role for coordinating supply and demand. The objective of our work is to design a dynamic pro- curement mechanism in the multi-agent system. Each SC participant is represented by an autonomous software agents that negotiation for coordinating supply and de- mand to reduce the BWE. The remainder of our paper is structured as follows: In section 2, we review existing work. Then we describe a dynamic procurement scheme that integrates concepts of NYOP. In section 4, we eva- luate our proposal by conducting a simulation study us- ing a multi-agent system(Dynamic Supply Chain Infor- mation Networks; DSINs). Section 5 discusses the pros and cons of DSINs. Finally, section 6 draws conclusions, and points out avenues of future research. 2. RELATED WORK The related work to ours can be grouped into three major areas: using multiagent technology and information shar- ing in supply chains, countering the BWE, and adopting dynamic procurement for coordination in supply chains. Each area is briefly discussed below. (1) Multi-agent technology and sharing information Coordination plays a pivotal role in successful design and implementation of supply chains, especially for those that are formed by independent and autonomous compa- nies[2]. In SCM, multiagent technology is at the brink of being integrated into real-world applications. An example JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 43 panies to interact with customers in ways that may have been prohibitively costly using traditional chan- nels[10][11]. In particular, the Internet has facilitated the emergence of name-your-own-price (NYOP) auctions. In such auctions, as popularized by Priceline, buyers bid for a product or service[8]. If a bid exceeds the seller＇s con- cealed threshold price, the buyer receives the product at the price of her bid. This selling mechanism has been used predominantly in the SC. Several normative models have been proposed regarding the optimal bid sequence [9][13]. In this section, we describe procurement decisions in SC with local information. This research employs a stylized model to identify and understand key tradeoffs driving the decision by a service provider in the SC to employ an NYOP channel, assuming that such a channel is available. 3. DYNAMIC PROCUREMENT SCHEME In this section, we describe dynamic procurement scheme with NYOP in SC. The adaptation of NYOP for dynamic procurement calls for specifications of the three steps : Step 1: Buyer agent’s bid: For calculating the bid of the buyer agent we refer to the demand function. A linear demand curve, thus an increasing price P causes a de- creasing demand Q Q d a bP   (1) where a>0 and b>0, the demand function describes the bidding behavior of the buyer agent. Step 2: Seller agent’s minimum price: For calculating the minimum price of the supplier we refer to the supply function. The linear form is (2) cQs  dP where c>0 and d>0 . Step 3: Matching: By comparing bid price and minimum price for the requested quantity, the price been decided. Step 1 and 2 require that equilibrium price P* and equili- brium quantity Q* as well as estimates for the price elas- ticity of demand Ed and supply Es are available. Then we can determine the parameters a and b of the demand function as well as c and d of the supply function. Figure 1 shows the relationship between demand and supply as described. is MASCOT, which is a reconfigurable, multilevel, agent- based architecture for planning and scheduling, aimed at improving SC agility [13]. Chen and Wei [5] analyzed the effects of negotiation-based information sharing in a dis- tributed make-to-order manufacturing supply chain in a multi-period,multi-product types environment, which is modeled as a multi-agent system. Verdicchio and Colom- betti [6] also stress that information sharing as a critical factor for successful business process management. Supply chain decisions are improved with access to glob- al information. However, supply chain partners are fre- quently hesitant to provide full access to all the informa- tion within an enterprise [10]. A mechanism to make de- cisions based on global information without complete access to that information is required for improved supply chain decision making. Cigoloni and Marco [2] carried out a study of the Collaborative Planning Fore- casting and Replenishment (CPFR) process for trading partners (belonging to the same supply chain) who are willing to collaborate in exchanging sales and order fore- casts. The two major out come of the literature survey is that information sharing is most important requirement of efficient supply chain and multi agent modelling is most suitable for designing of supply chains. (2) Bullwhip effect The term Bullwhip Effect was coined by Procter & Gam- ble management who noticed an amplification of infor- mation distortion as order information travelled up the supply chain [3]. The Bullwhip Effect (or Whiplash Effect) is an observed phenomenon in forecast-driven distribu- tion channels. To address the Bullwhip Effect, many tech- niques are employed to manage various supply chain processes, such as order information sharing, demand forecasting, inventory management, and shipment sche- duling[ 3]. Factors contributing to the Bullwhip Effect: forecast errors, overreaction to backlogs, lead time (of information – orders and of material) variability, no communication and no coordination up and down the supply chain, delay times for information and material flow, batch ordering (larger orders result in more va- riance), rationing and shortage gaming, price fluctuations, product promotions, free return policies, inflated orders. Bullwhip Effect is also attributed to the separate owner- ship of different stages of the supply chain. Each stage in such a structured supply chain tries to amplify the profit of the respective stages, thereby decreasing the overall profitability of the supply chain [13 ]. (3) Dynamic procurement Dynamic procurement in a capacitated supply chain faces uncertain demand. Burnetas and Gilbert [1] study reser- vation or forward purchase decisions by providing the buyer a second means of capacity procurement after the demand uncertainty is revealed [7]. Fredergruen and Heching [6] study a model similar to ours and show that dynamic pricing/procurement redistributes the power between the supply chain partners and allows the buyer to impact the wholesale prices. The Internet allows com- JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 44 Figure 1 Demand and supply function 4, FORMATION OF BUYER-SUPPLIER AGENTSHIP WITH DYNAMIC PROCUREMENT SCHEME Fig 1 shows an example supply chain of a Taiwam LCD panel assembly company which is used throughout this paper. LCD modules are widely used for display devices such as mobile phones, PDAs , and notebook PCs. In the supply chain, for a company like Cell Plant demand is distorted along the multiple paths from the markets to the company due to bullwhip effects and accumulation of erroneous estimations. Figure 2 An example supply chain model of a LCD panel assembly manufacturing company DSINs(Dynamic Supply Chain Information Networks) takes a simple but practical approach to address this problem. In an ADINS, agents autonomously form an information network by only local collaboration and in- formation sharing. Fig 3 shows the architecture of a pro- posed DSINs. To increase the practicality of DSINs, all the interactions among participating agents in a supply chain are implemented with Java Agent Development Frame- work (JADE) which based on the specifications from FIPA ACL (Agent Communication Language) format [4]. Figure 3 The architecture of DSINs Using the network, they perform order and production planning in a synchronized way without any central con- trolling entities. For this, we assume that agents are able to observe market demands directly rather than relying on the possibly distorted demand figures that are re- ceived by their companies. By doing this, DSINs is able to reduce bullwhip effects and improve service quality such as fill rates. In this multi-agent-based system, each SC participant is represented by an autonomous software agent. The cen- tral component of the JADE agent platform is the agent management system, which keeps supervisory control over access and use of the agent platform. Agents coordi- nate their ordering behavior by exchanging messages using the JADE message transport system. DSINs agents are generated using the JADE agent class and by imple- menting the appropriate agent control behavior on top of it. DSINs agents can send orders to other agents and re- ceive goods. In this context, we distinguish two types of scenarios: Agents with/without NYOP scheme determine their orders according to the dynamic procurement nego- tiation algorithm (Figure 4.), and present the dynamic supply chain logistic simulation by eM-Plant. JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 45 5. EXPERIMENTS SETTING AND RESULTS We defined a four-tier SC consisting of four agents (see also figure 2). Only the last agent ( k = 4 ) knows the final retailer demand D . We set the following parameters: (1) Final retailer demand D : normally distributed with expected value  =100 and standard deviation =10 . (2) The price elasticity of demand sE =1.56. (3) T between 5 and 15 (time periods of the moving av- erage forecast). dE =-0.75 and supply (4) k between 1 and 4 (same lead time for all tiers) Table 1 BWE for k=4 (Retailer) (SD: standard deviation , T：number of periods) Figure 4. Dynamic procurement negotiation algorithm An agent interacts with other agents by placing orders for goods. An example of this interaction is shown in Figure 5: In each period T, an FSA agent ( forecasting & structur- ing agent) observes the current inventory level and sends an order (calculated by algorithm ) to PSA (production & scheduling agent). After the order is placed, PSA agent observes and fills its demand for this period. Since our objective is to quantify the BWE, we must determine the variance of placed orders, relative to the variance of De- mand. Figure 5 presents message exchanges captured in the experiment with the help of a JADE provided sniffer agent. Table 2 BWE for k=3 (Wholeseller) Table 3 BWE for k=2 (Manufactuer) Figure 5 Screen captures showing our NYOP scheme in transaction. JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Table 4 BWE for k=1 (Supplier) Table 6 Fill Rate for k=3 (Wholeseller) 46 In the first set of experiments, we determined the BWE un- der variation of T. From Table 1 to Table 4 present the data for the comparison of the conventional procurement decision and the NYOP scheme yields a reduction in the order BWE of 8.03 % to 26.86 %. In general, the BWE decreases with an increasing T (see also figure 6). Table 7 Fill Rate for k=2 (Manufactuer) Table 8 Fill Rate for k=1 (Supplier) Figure 6 BWE and variation T In the second set of experiments, we determined the Fill rate (FR) under variation of T. Where is the standardized loss function. k=1,2,3,4 (3) zG )( u QVar ( ) k k 1 )( zGu FR k 1  L k In Table 5 ~ 8, the fill rate improve from 1.70 % to 3.74 % with NYOP scheme against the conventional procurement decision in four tier supply partners. The FR increases with an increasing T (figure 7). . Table 5 Fill Rate for k=4 (Retailer) (SD: standard deviation , T：number of periods) Figure 7 Fill rate and variation T JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 47 6. CONCLUSIONS AND FUTURE WORK We aimed at reducing the bullwhip effect and improve fill rate in multi-tier SCs by means of NYOP scheme. In par- ticular, the multiagent based behaviour model for each agent was developed to clarify the states, transitions, and commu- nication requirements of agents and also to facilitate the de- rivation of concrete procedures for agent behaviours that can be used for actual development of agent systems. In order to show the practical feasibility of the approach, the conversa- tions among agents were also modelled with FIPA’s stan- dard interaction protocols and messages and a prototype DSINs system was constructed using a FIPA-compliant agent platform JADE. A major limitation is that we assumed linear supply and demand functions. This assumption is, however, coherent with basic concepts of market economics, we believe that this paper has contributed to proving the ever-growing po- tential of agent technology for practical supply chain man- agement where analytic or optimization results from multi- tier supply chains cannot be easily applied or global infor- mation sharing or central coordination is impossible. Further rsearch issues include the full implementation of DSINs, relaxing the assumptions on supply chains, and analysing the bidding strategies for the iterative relaxation Contract Net in the supply chain management. REFERENCES [1] A. Burnetas and S. Gilbert, “ Future Capacity Procurements Under Unknown Demand and Increasing Costs”. Management Science, Vol 47, pp979–992, 2001. [2] M.,R.Cigoloni and D. De Marco,” Improving supply-chain collaboration by linking intelligent agents to CPFR”, Interna- tional Journal of Production Research, Vol. 43, No. 20, pp 4191–4218,2005. [3] F. Chen, Z. Drezner, J. K. Ryan, and D. Simchi-Levi, “Quan- tifying the Bullwhip Effect in a Simple Supply Chain: The Impact of Forecasting Lead Times, and Information”, Man- agement Science, Vol 46, No 3, pp 436 -443 ,2000. [4] FIPA web site – http://www.fipa.org/. [5] Y.M. Chen and C.W. Wei, ”Multi-Agent-Oriented Approach to Supply Chain Planning and Scheduling in Make-to-Order Manufacturing, International Journal of Electronic Business, Vol. 5, pp. 427-454, 2007. [6] A. Fredergruen and A. Heching, “Combined Pricing and Iven- tory Control under Uncertainty”, INFORMS,Operations Re- search, Vol. 47, pp. 454-475, 1999. [7] J. B. Kim, ”Multi-Component Contingent Auction (MCCA): A Procurement Mechanism for Dynamic Formation of Supply Networks”, Proceedings of International Conference on e- Commerce, Pittsburgh, USA, ,2003. [8] P. K. Kanna, and P. K. Kopalle, “Dynamic pricing on the In- ternet: Importance and implications for consumer behavior”, J. Electronic Commerce , Vol 5, No 3, pp 63– 83, 2001. [9] Y. Mujaj, J. Leukel, and S. Kirn, S. “A Reverse Pricing Model for Multi-Tier Supply Chains,” Proceedings of the IEEE Joint Conference on E-Commerce Technology (CEC’07) and Enterprise Computing, E-Commerce and E-Services (EEE’07), IEEE Computer Society, 2007. [10] Y. H. Park and E. Bradlow, “ An integrated model for bidding behavior in Internet auctions: Whether, who, when, and how much”. J. Marketing Res. Vol 42, No 4, pp 470–482, 2005. [11] C. Terwiesch, S. Savin, and I.H. Hann. “ Online haggling at a name-your-own-price retailer: Theory and application. Man- agement Science, Vol 51, No 3, pp 339–351, 2005. [12] Z. Yu, H. Yan and T. C. E. Cheng, “ Benefits of information sharing with supply chain partnerships”, Industrial Manage- ment & Data Systems, Vol 101, No 3, pp 114-119, 2001. [13] J. M. Swaminathan, “ Modeling Supply Chain Dynamics: A Multiagent Approach, Decision Sciences ,Vol 29 No 3, pp 607-632, 1998. [14] R. Zimmermann, S. Winkler, and F. Bodend , “ Agent-based Supply Chain Event Management – Concept and Assessment, Proceedings of the 39th Hawaii International Conference on System Sciences, pp 1-10, 2006. Yee Ming Chen is a professor in the Department of Industrial Engi- neering and Management at Yuan Ze University, where he carries out basic and applied research in agent-based computing. His cur- rent research interests include soft computing, supply chain man- agement, and pattern recognition.
1808.07975	1	1808	2018-08-24T00:53:43	A Communication Protocol for Man-Machine Networks	[ "cs.MA" ]	One of the most challenging coordination problems in artificial intelligence is to achieve successful collaboration across large-scale heterogeneous systems that include Robots, Agents, and People (RAP). In the best case, these RAP systems are potentially capable of leveraging the strengths of the individual entities to achieve complex distributed tasks. However, without intelligent communication protocols, man-machine partnerships are likely to fail as the humans become overloaded with irrelevant information. This paper introduces a communication protocol for man machine systems and demonstrates that its message routing performance approaches the central optimized solution in a simulated smart environment scenario.	cs.MA	cs	A Communication Protocol for Man-Machine Networks Neda Hajiakhoond Bidoki Department of Computer Science University of Central Florida Orlando, FL USA [email protected] Gita Sukthankar Department of Computer Science University of Central Florida Orlando, FL USA [email protected] Abstract -- One of the most challenging coordination problems in artificial intelligence is to achieve successful collaboration across large-scale heterogeneous systems that include Robots, Agents, and People (RAP). In the best case, these RAP systems are potentially capable of leveraging the strengths of the individual entities to achieve complex distributed tasks. However, without intelligent communication protocols, man-machine partnerships are likely to fail as the humans become overloaded with irrelevant information. This paper introduces a communication protocol for man machine systems and demonstrates that its message routing performance approaches the central optimized solution in a simulated smart environment scenario. I. INTRODUCTION The potential of man-machine teams has tantalized researchers for over a decade. Scerri et al. wrote a seminal paper introducing the acronym, RAP, to describe systems consisting of Robots Agents and People [1]. RAP systems leverage the strengths of the heterogeneous components, drawing from the common sense knowledge of the human, the robots' ability to perform repetitive physical tasks, and the ability of software agents to solve specialized artificial intelligence problems. They augment large-scale participatory sensor networks composed of humans carrying mobile devices with additional autonomous robot and software agents. Scerri et al. envisioned an architecture in which software agent proxies running on mobile devices could be used to coordinate the man-machine teams. in urban areas. There have been RAP systems are valuable for a variety of problems, including command and control, sensor networks, urban rescue, and personal assistance. In the future, RAP systems are likely to become an integral part of smart cities, serving the function of proactively helping humans limited demonstrations of HRI (human-robot interaction systems) as museum tour guides [2] and as building receptionists [3]. To extend these systems to include multiple robots and humans requires solving coordinated task allocation and scheduling which are NP-hard problems [4]. Glas et al. [5] introduced a general framework for networked robots that supports different social robot services including the observation of human behavior using environmental sensor networks, structured knowledge sharing, centralized resource and service allocation, global path planning for coordination between robots, and support for selected recognition and decision tasks by a human operator. In this paper, we propose new communication protocols to support this type of man machine system that includes networked robots cooperating with humans. We demonstrate that our new communication protocols are valuable for reducing communication costs in a simulated Netlogo scenario inspired by the Glas et al. shopping assistance system. II. PROBLEM STATEMENT the team In a man-machine team, humans, robots, and agents must cooperate to achieve the joint goal. In our smart environment shopping assistance problem, customers are aided by a combination of fellow shoppers, mobile robots, and software agents who help locate a desired set of items. We define the problem as consisting of an environment defined by a map, a set of robots, 𝑅, a set of human-service assistants, 𝐻, a set of customers 𝐶, and a set of additional system constraints. The set of robots on the team is defined as 𝑅 ∶= {𝑟1, 𝑟2, … , 𝑟𝑁} where 𝑁 is the number of robots on the team. The set of human is defined as 𝐻 ∶= service assistants on {ℎ1, ℎ2, … , ℎ𝑀} where 𝑀 is the number of human service assistants on the team. To make it simpler, we consider one customer or request sender as c. Our communication protocol must connect the customer with the best set of RAP assistants such that both human and non- human agents assist to accomplish his requirements while minimizing cost. Finding a balance between reward and recruitment effort remains a challenge. We assume that humans' willingness to collaborate changes over time, and that the customer uses a monetary offer or tips to motivate other humans to provide assistance. According to the incentive theory of motivation, if people receive a positive profit from performing a task, there is a higher chance that they will successfully complete it. An analysis of workplace incentive programs suggests that correctly employed incentives are able to enhance participants' performance [6], [7] [8]. Additionally, prompt awards enhance participants' motivation even more; an instant award, combined with repetitive actions, can create new behavioral habits. the customer III. METHEDOLOGY Our proposed protocol (History-based Financial Incentive) leverages the history of incentive acceptance to determine the best message routing. We compare our protocol against a centralized optimization algorithm to calculate the best possible agent allocation as well as Directed Diffusion protocol which is designed for robustness, scaling and energy efficiency in wireless sensor networks . Centralized optimization algorithms are undesirable for RAP systems since they rely on centralized information as well as a single computational node. These characteristics reduce their ability to deal with large scale problems and datasets, due to the high computational complexity. Moreover, it is inefficient to collect and store data in a centralized manner especially when the communication is multicast. In such scenarios, collecting all the necessary information through a central node is both time-consuming and incurs a high communication cost due to the large amount of packet exchange. Having a single point of failure also jeopardizes the inherently resilient nature of RAP systems. As an alternative to the centralized optimization algorithm, we implemented two protocols: The first is a Directed Diffusion algorithm and the second is our proposed history-based algorithm that tracks successful assistants.  Directed Diffusion protocol The customer requests as interests for named resources. Agents satisfying the interest can be found by flooding the message. Confirmation is exchanged by intermediate agents and resources are shipped when confirmed.  History-based Financial Incentive protocol (HFI) In our proposed protocol, the history of previously successful assists is recorded. Customers can make requests to the set of agents who are stored in his records. We believe that this protocol can reduce communication cost in many applications, especially when customers make repeated requests for similar types of assistance. We have implemented these two algorithms along with the incentive strategy used for motivating human agents. A. Human Behavior Modeling To account for differences between the agents and humans, we created a separate human behavior model. Skill is often a major determinant of human success, yet it can be sensitive to situational factors. In human-robot interaction tasks requiring physical control, human skills have been found to change over both the short and long term. Over the long term body movements slowed down and/or became less accurate; simple control skills may exhibit different kinematics and dynamics and are affected by microgravity [9]. Also, time-of-day affects the individual's performance; for instance, circadian rhythms such as morningness or eveningness can impact productivity. Creating a physically realistic human behavior model is complex and beyond the scope of our work. Instead in our scenario, we assume that skill is a negligible factor but that the human's current ability to perform the task can be modeled by a normal distribution, centered on the human's preferred time of day; this preferred time is when they are most available to render assistance. Peak time differs for each individual in our simulated scenario. Although robot performance can also fluctuate over time due to causes such as mechanical malfunctions and improper maintenance, we do not expect these situations to occur frequently in the short term. Thus, we assume that the non-human agents use a greedy task acceptance model; whenever a robot is not busy with other tasks, it always renders assistance. B. Mathematical Model We present a mathematical model for the problem that is used to calculate the optimal allocation that serves as our comparison benchmark. 𝐻, 𝑅 and 𝐶 represent the set of human agents, robot agents and customers (and their locations) respectively. ℎ𝑖, 𝑟𝑖 refers to 𝑖𝑡ℎ human and 𝑖𝑡ℎ robot agents. For simplicity, we consider one customer represented as 𝑐. Variables and parameters are as follows: Variables       𝑙ℎ𝑖 : Binary variable assuming the value 1 if human agent 𝑖 has been selected to assist the customer; 0 otherwise. 𝑜ℎ𝑖 : Binary variable assuming the value 1 if human agent 𝑖 has been selected to assist the customer; 0 otherwise. 𝑙𝑟𝑖 : Binary variable assuming the value 1 if robot agent 𝑖 has been selected to assist the customer; 0 otherwise. 𝑐ℎ𝑖 : Number of human agents have been requested by current customer. 𝑐𝑟𝑖 : Number of robot agents have been requested by current customer. 𝑐𝑡: Current customer monetary offer. Parameters  𝑁ℎ and 𝐻ℎ: are the size of human, robot, agents sets respectively. We assume each robot agent, human agent as well as customer is equipped with an IoT device with an omni- directional halfduplex antenna [10], [7], [11]. C. Integer Linear Programming Formulation The objective function minimizes the communication cost of agents and customers through the process of customer agent resource assembling. For the robot costs we consider the shortest path the robot can take to reach the customer. For human agents we also include the reward costs required to motivate response. 𝐸ℎ𝑖 and 𝐸𝑟𝑖 represent the total cost for human agent 𝑖 and robot agent 𝑖 necessary to reach current customer and assist with his demand. Therefore the objective function can be written as follows:  minimize 𝐸 = ∑ ℎ=1 𝑙ℎ𝑖𝐸ℎ𝑖 + ∑ 𝑟=1 𝑙𝑟𝑖𝐸𝑟𝑖   subject to the following constraints:   ∑ ℎ=1 𝑙ℎ𝑖 = 𝑐𝑖 and ∑ 𝑟=1 = 𝑐𝑟𝑖 𝑙𝑟𝑖 𝑙ℎ𝑖 ∗ 𝑜ℎ𝑖 <= 𝑐𝑡   The first constraint ensures that the total number of human and robot agents that are selected is equal to the number of human and robot agents that the customer has requested. The second constraint guarantees that the minimum offer value of the selected human is less than what the customer has offered, thus ensuring that the human agent is motivated to assist the customer. D. Implementation in Netlogo which We implemented our shopping assistance scenario and routing protocols is a multi-agent programmable modeling environment [12]-[13]. Fig. 1 shows the simulator interface, and Fig. 2 shows a flow chart of its operation. All variables are set when the simulation is initialized. Then a random customer creates a message with the required resource and reward info. This message is either broadcast to all surrounding agents (the flooding protocol) or unicast to the agents who have assisted the customer with previous The DD protocol does not perform well in comparison to the HFI algorithm as it constantly broadcasts to the surrounding agents. Obstacles such as walls do not block signal reception of agents but do block movement. Therefore, an agent may receive a request quickly while needing take a long path in order to reach the customer, due to the existence of solid obstacles. A matched paired t-test on the mean values of the same scenarios under different algorithms yields no significant difference between our proposed FIH protocol and the optimal ILP solution. Comparing DD protocol communication cost with HIF communication cost yielded t equal to -7.937337, indicating that the result is significant at 𝑝 ≤ 0.01. demands. After a response, the message will be updated with the remaining required resources. If any required resources exist, the message is broadcast. This process will continue until no resources are needed or all the agents have been contacted. In the latter case, the customer can increase the reward to attract more help, and the process will repeat. Algorithm 1 History-based routing algorithm Result: Message routing Contact agents in history Update required resources If more resources are needed then While more resource needed or all agents have not received message do to other nodes Upon receiving a message from an agent forward it if an agent exhibits interest then Update the needed resources Forward message else Forward message end end end IV. RESULTS The results for the centralized optimization model were obtained by solving the ILP model using AIMMS run on a Windows-based 64-bit core-i7 computer with 24GB of RAM. In all the scenarios we considered, LP model executions were fast, never lasting more than few seconds. We implemented the heuristics in a home-grown software framework written in NetLogo. Their executions were similar, lasting only a few minutes according to the number of agents in the simulated environment. After initializing the map, we determine the number of agents (including the number of human agents, robot agents and customers) as well as our budget and the maximum number of human and robot agents that a customer can request as inputs. Our simulation then randomly places humans, robots, and customers on the map of a building, assuming constant sensor radio coverage. For each scenario created two communication protocols (DD and HFI). The map and model info were loaded directly into our AIMMS program in order to execute the optimization procedure. In this way we are able to calculate the results of all three models on the same scenario. Initialization parameters (including number of agents) were varied; for each set of parameters, 20 different scenarios were generated. The average cost of all 20 different scenarios is used as the communication cost. Fig. 3 shows the results; as expected the History-based Financial Incentive algorithm (marked as HFI) had a better performance, closely matching the optimum selection of agents. This occurs due to several facts. First agents who previously participated on a team are likely to be around, having recently finished their previous task. the NetLogo simulation, we run the in Fig. 2. Simulation logic flow chart Fig. 3. Communication cost for each routing protocol Fig. 1. Netlogo simulation of shopping assistance scenario V. CONCLUSION for man-machine This paper introduces a history-based financial incentive communication algorithm systems. NetLogo was used to simulate a shopping assistance scenario in which a smart store environment summons help for the shopper in the form of robots and other humans to help locate items. Although non-human agents have no reason not to respond if available, humans are likely to be performing other shopping tasks and need to be incentivized to render assistance. In our simulation, they are modeled as having time availability preferences and as being less willing to respond to lower incentives outside their peak availability period. We demonstrate that the agent allocation solution reached our proposed algorithm results in an insignificant cost increase over a centralized solution calculated with an ILP solver. Lower communication costs are particularly important in man-machine systems to avoid bombarding the human with unwanted messages. In future work, we plan to implement our communication protocol in ROS (the Robot Operating System) for use in coordinating quadcopters with humans for autonomous photography tasks. REFERENCES [1] P. Scerri, D. Pynadath, L. Johnson, P. Rosenbloom, M. Si, N. Schurr, and M. Tambe, "A prototype infrastructure for distributed robot-agent- person teams," in Proceedings of the Second International Joint Conference on Autonomous Agents and Multiagent Systems, 2003, pp. 433 -- 440. [2] S. Thrun, M. Bennewitz, W. Burgard, A. B. Cremers, F. Dellaert, D. Fox, D. Hahnel, C. Rosenberg, N. Roy, J. Schulte, and D. Schulz, "Minerva: a second-generation museum in Proceedings of the IEEE International Conference on Robotics and Automation, vol. 3, 1999, pp. 1999 -- 2005. tour-guide robot," [3] S. Sabanovic, M. P. Michalowski, and R. Simmons, "Robots in the wild: observing human-robot social interaction outside the lab," in IEEE International Workshop on Advanced Motion Control, 2006, pp. 596 -- 601. [4] M. Koes, I. Nourbakhsh, and K. Sycara, "Heterogeneous multirobot coordination with spatial and temporal constraints," in AAAI, vol. 5, 2005, pp. 1292 -- 1297. [5] D. F. Glas, S. Satake, F. Ferreri, T. Kanda, H. Ishiguro, and N. Hagita, "The network robot system: enabling social human-robot interaction in public spaces," Journal of Human-Robot Interaction, vol. 1, no. 2, pp. 5 -- 32, 2012. [6] "Increase customer effective motiation strategies,"https://rewardstream.com/blog/increase-customer referrals- usingmotivation/, accessed: 2018-02-19. referral using [7] A. Mayle, N. H. Bidoki, S. S. Bacanli, L. Boloni, and D. Turgut, "Investigating the value of privacy within the internet of things," in Proceedings of IEEE GLOBECOM'17, December 2017. [8] M. Baghbahari, A. Behal, "Real-time policy generation and its application to robot grasping," FCRAR 2018, May 2018. [9] F. Steinberg, M. Kalicinski, M. Dalecki, and O. Bock, "Human performance in a realistic instrument-control task during short-term microgravity," PloS One, vol. 10, no. 6, p. e0128992, 2015. [10] N. H. Bidoki, M. B. Baghdadabad, G. R. Sukthankar, and D. Turgut, "Joint value of information and energy aware sleep scheduling in wireless sensor networks: A linear programming approach," in IEEE ICC'18, May 2018. [11] N. H. Bidoki and M. D. T. Fooladi, "Linear programming-based model for joint routing and sleep scheduling in data-centric wireless sensor networks," in Information and Knowledge Technology (IKT), 2014 6th Conference on. IEEE, 2014, pp. 73 -- 78. [12] U. Wilensky, "Netlogo," Center for Connected Learning and Computerbased Modeling, Northwestern University, Evanston, IL, Tech. Available: http://ccl.northwestern.edu/netlogo/. Rep.,1999. [Online]. [13] S. J. Russell and P. N. A. I. A. Modern, "Approach," Prentice Hall Pearson Education Inc., Upper Saddle River, New Jersey, vol. 7458, pp. 116 -- 119, 2003.
1311.6233	1	1311	2013-11-25T08:42:59	Agent Based Negotiation using Cloud - an Approach in E-Commerce	[ "cs.MA" ]	Cloud computing allows subscription based access to computing. It also allows storage services over Internet. Automated Negotiation is becoming an emerging, and important area in the field of Multi Agent Systems in ECommerce. Multi Agent based negotiation system is necessary to increase the efficiency of E-negotiation process. Cloud computing provides security and privacy to the user data and low maintenance costs. We propose a Negotiation system using cloud. In this system, all product information and multiple agent details are stored on cloud. Both parties select their agents through cloud for negotiation. Agent acts as a negotiator. Agents have users details and their requirements for a particular product. Using users requirement, agents negotiate on some issues such as price, volume, duration, quality and so on. After completing negotiation process, agents give feedback to the user about whether negotiation is successful or not. This negotiation system is dynamic in nature and increases the agents with the increase in participating user.	cs.MA	cs	Agent Based Negotiation using Cloud - an Approach in E-Commerce Amruta More1, Sheetal Vij1, Debajyoti Mukhopadhyay2, 1 Department of Computer Engineering , 2Department of Information Technology, Maharashtra Institute of Technology, Pune- 411038, India, {moreamruta930, sheetal.sh, debajyoti.mukhopadhyay} @gmail.com Abstract. 'Cloud computing' allows subscription based access to computing. It also allows storage services over Internet. Automated Negotiation is becoming an emerging, and important area in the field of Multi-Agent Systems in E- Commerce. Multi-Agent based negotiation system is necessary to increase the efficiency of E-negotiation process. Cloud computing provides security and privacy to the user data and low maintenance costs. We propose a Negotiation system using cloud. In this system, all product information and multiple agent details are stored on cloud. Both parties select their agents through cloud for negotiation. Agent acts as a negotiator. Agents have user’s details and their requirements for a particular product. Using user’s requirement, agents negotiate on some issues such as price, volume, duration, quality and so on. After completing negotiation process, agents give feedback to the user about whether negotiation is successful or not. This negotiation system is dynamic in nature and increases the agents with the increase in participating user. Keywords: Cloud computing, negotiation, multi-agent, E-Commerce 1 Introduction In business negotiation two or more parties come together to find mutually agreeable contractual decision. For negotiation process both parties must show their interest, thus, negotiation can be complicated and lengthy process. When all parties have to come to final decision, negotiation will be stopped. In negotiation, each individual aim to achieve the best possible outcome for their organization. Negotiator is an individual representing an organization which listens to all the parties’ decision carefully and takes his own decision which gives profit to his organization. In negotiation process organization profit depends on organization's negotiator so that negotiator needs to understand situation and all other organization's negotiator. Negotiator must know how to negotiate well to successfully close deals, avoid conflicts, and establish better relations among the other organization's negotiators making the organization a better place to work. For successful negotiation individuals or negotiator must learn to compromise and stop f inding faults in each other. Today cloud computing is widely used and is becoming a popular technology. Cloud is a remote server, where the user can store their data and access the data remotely whenever it's required. Cloud computing provides security and privacy to the user data. User has no burden in maintaining huge amount of data stored on cloud. Cloud computing is sometimes referred to as "on-demand resources" and is usually based on pay-per-use basis. If business owner requires more space as compared to his previous space on cloud, owner can easily request for additional data storage on the cloud. Also owner can easily request for additional bandwidth, processing speed, and additional licenses. Using cloud computing, user can access information from any device like desktop, minicomputer, mobile etc. anywhere and anytime. Amazon, Microsoft, Openstack, Google all these are the cloud providers. Google Apps is one example of cloud. In Agent based negotiation system, we propose a system for negotiation between a provider and a consumer using cloud. In this system, all product information and multiple agent details are stored on the cloud. Both provider and consumer will select their agents through cloud for negotiation. An agent acts as a negotiator. Agent has user’s details and their requirements for a particular product. Using user’s requirement, agents negotiate on certain features or issues. 2 Features of Cloud for E-Negotiation 1. Rapid Scalability: Cloud computing has the ability of scaling the resources both ways for the consumers and as per the need. Cloud is infinite and one can buy the computing power as per the need. Negotiation system is dynamic, so that if more data storage is required, it can be easily made available by cloud. 2. Security: It is the core feature of cloud computing. Security is much stricter in cloud computing. Data is shared within a server hence, the provider must ensure that each account is secured, and only authorized users in one account can access it. Any product or negotiation process information is stored in a secure manner. Only authorized agents have access to the product information and negotiation process. 3. No Need of Maintenance: User can store all types of data on the cloud, and do not have to worry about maintenance of data. The product data can be stored on cloud, so that organizations do not require any server and maintenance of that server. Simultaneously maintenance cost is also reduced. 4. No Need of Backup: Business owner do not need to worry about the backup responsibilities, as the supplier has already taken effort to put up a great system for backup. Disk failure, server crash or system failure won’t create much problem as the supplier can easily restore the latest backup from the cloud. 5. Device and Location Independent: User can access cloud from any device like mobile, mini-computer, and desktop etc. anywhere and anytime. 6. Transparent Software Updates: Softwares which are necessary with any e- commerce system, are updated, transparently and require minimum download time. 3 Technical Literature Survey Li Pan [1] introduced a framework for automated service negotiation in cloud computing environments. In this framework, software agents negotiate with each other on behalf of service consumer and provider. This system also used a bilateral multi-step monotonic concession negotiation protocol for service negotiation in cloud computing environments. Service provider and consumer agents interact with each other due to the negotiation process and they make decisions according to the negotiation protocol. Miguel A. Lopez-Carmona, Ivan Marsa-Maestre and Mark Klein [2] says that, consensus policy based mediation framework is used to perform multi -agent negotiation. This paper also proposed a mediation mechanism which is used to perform the exploration of negotiation space in the multiparty negotiation setting. The performance of mediator mechanism is under guidance of aggregation of agent performance and on the set of alternatives the mediator proposes in each n egotiation round. Mikoto Okumura, Katsuhide Fujita proposed [3], a collaborative park-design support system which is an example of collective collaboration support systems based on multi-agent systems. In this system, agents collect user information, many alternatives and reach optimal decision using automated negotiation protocol. Especially, in this paper, the attribute space and utility space of user in real world is decided. At the end of the system user gives feedback. According to the user’s feedback, if most of the users agree on some alternative, then this alternative is final or optimal. Amir Vahid Dastjerdi and Rajkumar Buyya [4] described SLA negotiation challenges in a cloud computing environment. This system also proposed time dependent negotiation which solves negotiation challenges. To increase the dependability of negotiation process, this system has included reliability assessment. Cloud providers can accommodate more requests and thus increase their profit by discriminating regarding the pattern of concession Ivan Marsa-Maestre, Miguel A. Lopez-Carmona and Mark Klein [5] presented a framework for characterization and generation of negotiation process. Considering both the structural properties of agent utility functions, and the complexity due to relationships between utility functions of the different agents, a set of metrics to measure high-level scenario parameters is provided. Then a framework is presented to generate scenarios in a parametric and reproducible way. The basis of generator is the aggregation of hyper volumes which is used to generate utility functions. Generator is also based on the use of shared hyper volumes and nonlinear regression which is used to generate negotiation scenarios. Bo An, Victor Lesser, David Irwin, Michael Zink [6] designed a system for dynamic resource allocation problem and implements a negotiation system. In negotiation model, multiple sellers and buyers are allowed to negotiate with each other concurrently. At the same time an agent is allowed to de-commit from an agreement at the cost of paying a penalty. This system also presents negotiation strategies for both seller and buyer. Moustapha Tahir Ateib [7] has presented a fuzzy logic based negotiation modeling that can be used to overcome the complexity of automation negotiation processes. This system uses fuzzy logic to deal with ambiguity and uncertainties. Yan Kong, Minjie Zhang [8] proposed a negotiation-based method which is used for task allocation under time constraints in an open, dynamic grid environment. In this environment, both consumers and provider agents can enter into or , exit the environment freely at any time. There is no central controller so that agents are negotiating with each other for task allocation based only on local views. Hsin Rau, Chao-Wen Chen, and Wei-Jung Shiang [9] developed a negotiation model which is used for a supply chain with one supplier and one buyer. This model is useful to achieve coordination under incomplete information environment. To find an optimal solution, an objective programming approach is applied. Liu Xiaowen, Yu Jin [10] introduced automated negotiation model for tourism industry. To improve the negotiation efficiency and success rate, this system proposed RBR and CBR. The model employs CBR method to support an automated negotiation by past successful negotiation cases used for those negotiation partners that have no contract rule existing in each other. Mukhopadhyay et. al. recently proposed related solutions in negotiation over the Internet for efficient E-Commerce and negotiation prediction. [11] [12] 4 Proposed System 4.1 Scope of the System Due to cloud computing, negotiation system will be secured, user can access it any time on any device like desktop, mobile etc. and organization maintenance cost is also reduced. Further we can use features like rule based reasoning and case based reasoning [10]. These two features improve the efficiency and success rate of the negotiation process. 4.2 Purpose The objective of this system is to reduce maintenance cost of organizing data and provide security for data and negotiation process. So, it makes automated negotiation faster, flexible, secure and reliable. 4.3 Problem Definition In order to make some agreeable decision, two or more parties come together during the negotiation process. And there are organizations to maintain data of negotiation process and product data. But this maintenance is a very tedious job. In order to overcome this problem, all organizations’ product data is stored on cloud. Hence, security and maintenance cost of organizations’ data is reduced. 5 Proposed System Architecture In agent based negotiation system, agents are the negotiator. The negotiation process is done by agents through cloud. Cloud is used to store all product details and agents information. Further we go on case based reasoning and rule based reasoning, we will add two more databases in this system. Fig. 1. System Architecture of Agent Based System If sometimes buyer has time for doing negotiation process but at the same time seller is busy in his/her work. In this situation buyer has to wait until seller is free. For this reason we can use agent based system. In this system, firstly, agents are register ed on the cloud. All information related to agents (such as agent name, experience etc.) is stored on cloud database. Cloud database also has product details (for example product of company, price, features of product etc).Seller and Buyer of product selects agent using cloud database for negotiation . Agents have all requirement and details of seller and buyer respectively. Using these requirements agents negotiate. After completing negotiation, respective agents will give feedback to the seller and buyer through cloud. For this system, we can use Amazon Simple Storage Service (Amazon S3). Amazon S3 is used as huge storage area for the internet. It is designed to make easy development of web-scale computing. Amazon S3 offers a simple web services interface which can be used to store, access and retrieve any amount of data at any time. On Amazon S3, user can write, read, and delete objects containing from 1 byte to 5 terabytes of data each. User can store unlimited number of objects. 5.1 System Components For this system, we can use three modules. Using these components, system becomes easy to use and works efficiently. 1. Store data on cloud and Agent registration: For storing data, we can use Amazon S3 service. It makes our system fast, secure and highly reliable. In this component, agent detail and product information is stored on cloud in proper format. Only authorized agent can access product details. Fig. 2. Flowchart of the System 2. Negotiation process: For negotiation process, seller and buyer select their agents respectively. After that they can give their requirement to agent in encrypted format that is to generate the hash code of that requirement and encrypt that hash code using agent's public key. Agent's public key is known to sellers and buyers. Buyer or Seller Requirements = E {H (m), Apk}. (1) Where, for generating hash function MD5 algorithm is used. Apk is agent's public key. We can use encryption for security purpose, in this process, we can use digital signature concept same as seller. Buyer can do same process for generating hash code of his requirement. After getting requirement, agent decrypts the hash code using his private key. And calculate s the hash code for checking whether this message comes from appropriate seller or buyer and whether it is modified or not. Agent Receive Requirements = D {H (m), Apri}. (2) Where, Apri is agent's private key. When an agent confirms that this message comes from appropriate seller or buyer and message is not modified, then negotiation process will be start. For negotiation process, we can use The Bilateral Negotiation Model presented by Moustapha Tahir Ateib [7]. Let x (x €{x1, x2,…,xm})represents the buyer agent and y(y€{y 1,y2,…,yn}) be the supplier agent. And let then i (i€{i1,i2,…,in}) be the issues under negotiation, such as price, volume, duration, quality and so on. Each agent assigns to each issue i, weight Wi denoting the relative importance of that x represents the importance of issue i to agent x, issue to the agent. Hence, Wi therefore the overall utility function of an offer O is: . (3) Where U(O) is the overall utility for the offer O (=[O_1,...,Om] T) and ui(xi) is the individual utility function for issue i for ui € [0,1]$ and the preference degree of an agent to an issue i is denoted as Wi € [0,9]. Each agent also specifies a minimum acceptable utility level to determine if an offer is acceptable. Hence, for benefit-oriented, the utility function U i(xi) is computed as follows : For cost oriented however, the utility function can be written as: (4) (5) 3. Feedback: When negotiation process is finished, agent gives feedback to his appropriate seller or buyer about negotiation whether it is successful or not. Then the actual E-commerce part would start which is not part of our current research. 6 Conclusion Cloud computing provides various features such as security, scalability, reliability and low maintenance which are beneficial to negotiation process in E-Commerce. In this paper, we propose an agent based negotiation system, in which the agent uses cloud for storage of data and product information . Agents negotiate on some issues using the product information and seller’s or buyer’s requirement. After completing negotiation process, agents give feedback to user about whether the negotiation is successful or not. This negotiation system is dynam ic, if number of users is increased, then number of agents also increases automatically. Our future work is to make a faster, secure and flexible E-negotiation agent using rule based reasoning and case based reasoning [10]. References 1. Li Pan. Towards A Framework For Automated Service Negotiation In Cloud Computing. 61284-204-2/11 IEEE (2011) 2. Miguel A. Lopez-Carmona, Ivan Marsa-Maestre and Mark Klein. Consensus Policy Based Multi-Agent Negotiation. ANAC (2011) 3. Mikoto Okumura, Katsuhide Fujita. Implementation of Collective Collaboration Support System based on Automated Multi-Agent Negotiation. ANAC (2011) 4. Amir Vahid Dastjerdi and Rajkumar Buyya. An Autonomous Reliability -aware Negotiation Strategy for Cloud Computing Environments. 978-0-7695-4691-9/12 IEEE (2012) 5. Ivan Marsa-Maestre, Miguel A. Lopez-Carmona and Mark Klein. A Scenario Generation Framework for Consistent Comparison of Negotiation Approaches. ANAC (2011) 6. Bo An, Victor Lesser, David Irwin, Michael Zink. Automated Negotiation with Decommitment for Dynamic Resource Allocation in Cloud Computing. Proc. of 9th Int. Conf. on Autonomous Agents and Multiagent Systems AAMAS( 2010) 7. Moustapha Tahir Ateib. Agent Based Negotiation In E-commerce.978-1-4244-6716-7/101 IEEE (2010) 8. Yan Kong, Minjie Zhang, Xudong Luo, Dayong Ye. A Negotiation Method for Task Allocation with Time Constraints in Open Grid Environments. ANAC (2013) 9. Hsin Rau, Chao-Wen Chen, and Wei-Jung Shiang. Development of an Agent-based Negotiation Model for Buyer-supplier Relationship with Multiple Deliveries. 978-1-4244- 3492-3/09 IEEE(2009) 10. Liu Xiaowen, Yu Jin. Hybrid Approach Using RBR and CBR to Design an Automated Negotiation Model for Tourism Companies. 978-0-7695-4853-1/12 IEEE (2012) 11. Saurabh Deochake, Shashank Kanth, Subhadip Chakraborty, Suresh Sarode, Vidyasagar Potdar Debajyoti Mukhopadhyay. HENRI: High Efficiency Negotiation -based Robust Interface for Multi-party Multi-issue Negotiation over the Internet; ACM Digital Library, USA; pp.647-652; ISBN 978-1-4503-1185-4. (2012) 12. Debajyoti Mukhopadhyay, Sheetal Vij, and Suyog Tasare. NAAS: Negotiation Automation Architecture with Buyer’s Behavior Pattern Prediction Component; Springer-Verlag, Germany; pp.425-434; ISSN 1867-5662, ISBN 978-3-642-31513-8. (2012) 13. Xianrong Zheng, Patrick Martin, Wendy Powley, Kathryn Brohman. Applying Bargaining Game Theory to Web Services Negotiation. IEEE International Conference on Services Computing (2010)
1711.03966	1	1711	2017-11-10T11:44:23	Multi-agent based IoT smart waste monitoring and collection architecture	[ "cs.MA" ]	Solid waste management is one of the existing challenges in urban areas and it is becoming a critical issue due to rapid increase in population. Appropriate solid waste management systems are important for improving the environment and the well being of residents. In this paper, an Internet of Things (IoT) architecture for real time waste monitoring and collection has been proposed; able to improve and optimize solid waste collection in a city. Netlogo Multiagent platform has been used to simulate real time monitoring and smart decisions on waste management. Waste filling level in bins and truck collection process are abstracted to a multiagent model and citizen are involved by paying the price for waste collection services. Furthermore, waste level data are updated and recorded continuously and are provided to decision algorithms to determine the vehicle optimal route for waste collection to the distributed bins in the city. Several simulation cases executed and results validated. The presented solution gives substantial benefits to all waste stakeholders by enabling the waste collection process to be more efficient.	cs.MA	cs	International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 MULTI-AGENT BASED IOT SMART WASTE MONITORING AND COLLECTION ARCHITECTURE Eunice David Likotiko, Devotha Nyambo, Joseph Mwangoka Information and Communication Science and Engineering, The Nelson Mandela African Institution of Science and Technology (NM-AIST), Tanzania. ABSTRACT Solid waste management is one of the existing challenges in urban areas and it is becoming a critical issue due to rapid increase in population. Appropriate solid waste management systems are important for improving the environment and the well-being of residents. In this paper, an Internet of Things(IoT) architecture for real time waste monitoring and collection has been proposed; able to improve and optimize solid waste collection in a city. Netlogo Multi-agent platform has been used to simulate real time monitoring and smart decisions on waste management. Waste filling level in bins and truck collection process are abstracted to a multi-agent model and citizen are involved by paying the price for waste collection services. Furthermore, waste level data are updated and recorded continuously and are provided to decision algorithms to determine the vehicle optimal route for waste collection to the distributed bins in the city. Several simulation cases executed and results validated. The presented solution gives substantial benefits to all waste stakeholders by enabling the waste collection process to be more efficient KEYWORDS Solid Waste Management, Multiagent Simulation, IoT, WSN, Real Time Monitoring, Waste Collection 1. INTRODUCTION The amount of waste generated in urban areas is proportional to the rapid growth of cities. The increase of human activities and consumption patterns generate various types of waste that must be appropriately managed to ensure sustainable development and a decent standard of living for all urban residents. Managing Urban solid waste is complex especially to most of the resourced constrained regions. Conventional approach to Municipal Solid Waste Management (MSWM) suffers from lack of sustainable collection throughput [1], lack of data on collection time and location, lack of efficient monitoring systems of waste bin status, lack of efficient systems for waste collection and transportation, and delays in collection. Creativity of having reliable approach on solid waste management is needed, starting from the existing strengths of the city and build up on them with good involvement of all stakeholders in designing their own local models [2]. The use of technology can be adopted to improve the quality of waste data collection, the service availability and reliability in a city thus, building smart cities' world in the aspects of waste management. DOI : 10.5121/ijcseit.2017.7501 1 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 According to [3]"A Smart City is a city well performing in a forward-looking way in the following fundamental components (i.e., Smart Economy, Smart Mobility, Smart Environment, Smart People, Smart Living, and Smart Governance), built on the 'smart' combination of endowments and activities of self-decisive, independent and aware citizens". Therefore, smart city process for waste collection is a fundamental point in achieving green city environment and well-being of the citizen and its quality should be considered seriously. smart home applications are of trending. In this scenario, the application of Information and Communications Technologies (ICT) solutions through Internet of Things(IoT) is of interest.[3]and[4]defined IoT as: The Internet of Things allows people and things to be connected anytime, anywhere (AAA), with anything and anyone, ideally using any path/network and any service. The IoT concept enables easy access and interaction with a wide variety of devices such as, home appliances, surveillance cameras, monitoring sensors, actuators, displays, vehicles, and so on[5]. The idea of internet of things (IoT) was developed in parallel to Wireless sensor networks (WSNs). WSNs as an important aspect in IoT, changes and advances environmental monitoring scenarios and increasingly being adopted in different application[6]for instance, waste collection and monitoring activities. The real timewaste monitoring is now possible from the IoT enabled bins which can sense the status of waste level in a bin and update the status to the central server through the connections. In addition, citizens are now given a room to fully participate in the process of waste management, as a way to build a smart city in the aspect of smart waste management. Multi-agent based models presents tools and methods to visualize and abstract a given system. Multi-agent system (MAS) "is a set of software agents that interact to solve problems that are beyond the individual capacities or knowledge of each individual agent"[7].Many researches have been done using the agents in providing simulated solutions for different smart city scenarios, for instance intelligent waste collection[8], and generic model for smart city[9]. In this paper, a Multiagent based simulation model is used to simulate the functioning of the IoT and smart city in a waste management scenario. The aim of this work is to abstract real time solid waste monitoring and collections using multi- agent based model, to design architectural models for real time solid waste's bin monitoring and collection based on wireless sensor network (WSN) technology. WSN can be useful in smart waste management, to overcome the challenges on MSWM, providing optimal path for waste collection resulting into reduced operational costs. Furthermore, this paper highlights the importance of information flow between the authorities and waste collection points with a target of increasing collection throughput, optimize routes and continuously maintaining statistical records of the amount of waste collected, which are necessary for the sustainability of the operations. The remaining part of this paper has been divided into six sections: Related work, Methodology used with the study requirement analysis and Designs, Simulation set up, Results Discussion, Conclusion Future work and Acknowledgment. 2. RELATED WORK Several efforts have been invested in tackling the problem of Municipal Solid Waste Management (MSWM). Most of the initiatives use wireless communication technologies as one 2 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 of the solution in reducing amount of time and cost for waste collection and transportation and making smart waste management. A real time intelligent bin status monitoring and rule based decision algorithms[1] is among the invested efforts toward smart solid waste management. The monitoring application is based on decision algorithms for sensing solid waste data via a wireless sensor network. The system is built on a three-level architecture; smart bin, gateway and control station. The elementary concept is that, smart bins collect their status when any changes occur and transmit the status data to a server via an intermediate coordinator. A set of applications in server presents the updated bin status on real time. The strength of the system is the design and development of an automated bin status monitoring system that exercises a set of novel rules based decision algorithms. However, the system encounters technical challenges technical such as long range communication technologies on gateways, erroneous output of data from sensors readings, and lack of GPS for location detection, citizen involvement in the system for better interactions which is proposed by this paper. Automation and integrated system by[10]performed 55 tests run on the developed prototype. The authors provide the unique effort on waste monitoring by incorporating accelerometer, magnetic proximity, ultrasonic, weight sensing and long lived rechargeable battery respectively. The integrated sensing system is designed using rule-based decision procedure to offer a proficient and automatic bin status monitoring system. However, the designed system does not support the use of the obtained waste data for route optimization and how the citizen as the major waste stakeholders will play their role in making sure the smart waste management is achieved. Similarly,[11]used an open-source project Smart-M3 Platform which has smart-space for virtualization of the real environment used in simulating the real time waste collection. All components of the system; Proximity and Weight sensors, Raspberry PI, Xbee module were implemented using Python language and the ontology was developed using Protege. Authors argued that several simulation cases were run and a map of full and semi full bin was developed and the vehicle was sent to them. In additional a user android application integrated with the Google Map API was still to be a part of the implementation. Besides of the simulated tests the proposed approach didn't consider the development of the central system for users and their bins registration and authenticity also the graphical user interface for central system. Moreover, the system lacks the route decision making algorithm of vehicle for waste collection In [12]used RFID, cell load sensors technologies and Personal digital assistant (PDA) in the achievement of smart waste management. Authors detailed that RFID tag are embedded on the bin, the PDA (smart phone) consists of a reader which converts the radio waves reflected from the waste bin into digital information (i.e., bin ID) that is then recorded in a PDA, the bin is located on the waste collecting vehicle. The PDA-based RFID when robotic/lifting arms in the waste collector loaded onto the vehicle (truck), then the weighting system (e.g., SSC) measures the weight of each bin. The data (bin ID) is then used to calculate actual waste disposal charges for each individual household and sends it to the PDA for temporary storage after emptying each bin. In additional the security of components and customers information were considered, conversely the designed system lack prior planning waste collection scheduling route since bin is only detected when the vehicle is few meters away. On the other hand,[13]developed a wireless sensor network with three tier architecture the system was tested into three different bin samples, several sensors were involved temperature, humidity 3 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 and weight sensors, ZigBee, author contented that data from the sensor nodes are transmitted to the servers through GSM/GPRS. The control station with the database server continuously receive analyze the data and update the specific bin information, the developed web based programs run in the server to facilitate the management of data on the database and bin status monitoring purpose, also the user can monitor the bin status using a web browser. Minimum energy consumption less operation cost by avoiding GPRS in every bin were also thought out, but ignoring the part for citizen participation. Moreover, Optimal routing for efficient municipal solid waste transportation involving Geographical information system "ArcGIS" and Dijkstra algorithm for solving shortest path problem was done by [14].Authors were able to reduce the waste collection travelled distance by 9.93% of the 13 identified and selected wards,despite of the optimal route achievement also the time and cost was reduced. However, the authors didn't expound on how all type of waste stakeholders will be involved and the real time routing basis but authorities could use the solution to reduce cost and improve management. In addition,[8]proposed a model for multi agent simulation and GIS to abstract intelligent decision support for waste management. GIS maps representing domains are abstracted to 2D lattices in order to couple with the multi agent system. Waste bins are represented as points, and actual distances are converted to appropriate units, agent follows a route on the 2D lattice, collects waste bins up to the specified time and capacity limit. Yet the simulated design didn't not focus on wireless sensor network for real time waste status detection before the collection status, moreover author didn't reflect how the household found on the map are involved with the waste management. In general, this paper suggests the improvement for solid waste management by allowing citizens as major stakeholder of waste management to be involved during the operation. Real time monitoring and continuous reporting of waste status for wellbeing of the environment together with the prior planning for waste collection scheduling. Table 1: Summary of related work with their advantages and limitations s/n Related Citizen Real Time Scheduling Literature Involvement Monitoring and 1. 2. 3. 4. 5. 6. 7. [1] [10] [11] [12] [15] [8] [14] NOT NOT NOT YES NOT NOT NOT YES YES YES NOT YES NOT NOT optimization YES NOT NOT NOT YES NOT YES 4 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 3. METHODOLOGY 3.1 Requirement Analysis This work used the following software to model and simulate the real time waste collection, monitoring and generation of optimal paths for waste collections vehicles: Netlogo is an open source software, and a programmable modeling environment for simulating natural and social phenomena [16]. Netlogo 5.3.1 Multiagent Platform has been used to model and simulate the real time waste collection and generation of the optimal paths for collection. While UML diagrams were designed by using Microsoft Visio 2016 see decision logic Fig.4 and Fig.5, the general proposed architecture Fig.1 and central system architecture Fig. 2. 3.2 Functional Requirements The Functional requirements of the simulated multiagent based model are highlighted as follows; 1. Waste level in a bin shall be generated per simulation tick in minutes 2. Individual bin waste levels shall continuously be updated and recorded 3. Citizen shall pay the cost per waste unit to be collected 4. The truck shall follow the generated optimal path for waste collection 5. Bin and truck shall have the maximum capacity for carrying waste load while keeping the nonfunctional performance of the simulated model to have appropriate response time, and effective throughput during the waste collection. 3.3 System Design The design of the model for smart waste monitoring system has been split up into two sections which are conceptual designs and decision logic design. The sections of the designs are described in detail as follows. 3.3.1. Conceptual Design i) General architecture: The conceptual design gives an overview of the proposed general architectural implementation and the overview of the proposed central system architecture. As shown in Fig. 1, the general architecture reveals three actors; the citizen, system administrator and truck driver. The citizen buys smart bin from the company and registers it into the central system. The bin is located at the citizen intended place (home, church, mosque, hospital, school etc.). The citizen can access information and pay for waste collection through the web. The cover of the bin will be embedded with the integrated circuit board for continuous and real time waste monitoring purposes. The Arduino Wi-Fi Shield with GSM/GPRS connection is used as the gateway for transmitting waste status data into the central system database. The administrator manages citizens registration and payment information; also browses and retrieves the processed information for waste collection and transportation then assigns the optimal route to the trucks drivers. 5 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 Figure 1. General architectural implementation ii) Central system architecture: Fig.2 shows the central system, which is composed of three tier architectures: upper, middle and lower tiers. The upper tier comprises of a central database allied with the optimization model; the middle tier contains the gateway; and the lower tier consists of the sensor nodes. The central system receives updates, and stores the waste status via the gateways from different citizen's location after establishing connection with the server. Data are analyzed and optimal path for waste collection is established via the linked optimization model. The user graphical interface is provided with role based restrictions for both administrator and citizen from the central system. Figure 2.Smart bin and Central systems architecture. 6 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 3.3.2. Decision Logic The activity diagrams Fig.3 and Fig.4 describes the building blocks and logical decision involved from bin status recording to waste bin emptying. The diagrams provide an impression on how the multi-agent model will achieve real time monitoring and collection of solid waste. Activity diagrams consist of activities, states and transitions between activities and states [17]. These are essentially flowcharts that can be used to model dynamic aspect of a system. Simulation decision logic flow charts describe the working principles of the proposed architectures in Fig.1 and Fig.2 for IoT and WSN smart waste management embracing the principles of Multi-agent based models. Figure 3.Bin monitoring decision logic Figure 4. waste collection decision logic 7 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 4. SIMULATION SETUP The Multiagent model for real time waste monitoring and collection define three main agents(turtles breed) bin, truck and citizen with their attributes and procedures, the patches form the world of a city where, smart bins were randomly placed,smart bins data extracted into .cvs format from the Netlogo system. Agent Bin The Agent bin has capacity parameter with the maximum level of 25 waste units. In this simulation 25 bins are distributed in a city following the decisions rule defined in Fig.3. Initially, all bins are green and the level of the [bin_load = 0]. The levels are increased per 1 tick when simulation time starts [bin_load = bin_load + 1], when the [bin_load>=10] the color of the agent bin changes to yellow as the first alert, again when the level [bin_load>=25] it is assumed to be the maximum capacity of the bin no more level to be added so the agent [stop] color changes to red, the levels of the bin load are reported continuously and can be exported for analysis. The simulation model's monitor provides total number of distributed bins and when they are full the number of [full_Bins uncollected] is reported, the [full_Binsuncollected] shoulddecrease each time a truck is assigned into that path for collection, else the delay in collation is observed Fig.10. Agent Truck The agent truck follows the optimal path for waste collection to the full bins ([bin_load>=25] and color set red]). The truck is simulated with the maximum capacity of [truck_load>=100] waste units, the unloaded bins by a truck returns to green and level becomes zeroas shown in Fig.4. When [truck_load>=100] is emptied into a dump and the truck load return to zero. To reduce the delay in collections number of trucks can be added. Cost per trip can be calculated by identifying the distance between points of collections together with the defined fixed cost for the operations i.e. fuel, wages and the load carried by the truck. Although for the simulation purposes this cost is constant (1). Agent Citizen A citizen is another agent defined in the model in which, movement of the citizen portrays their activities which produce waste in a city. The citizen is involved in the simulated system by paying the price per unit waste they produce and collected from their own installed bins. Thus, total revenue for waste collection operation is given by: (cid:1)(cid:2) = (cid:4)(cid:2) ∗ (cid:6) (i) Where R is the Revenue, Ni is the weight of the full bin collected, and p is the price per unit weight of waste in units of currency (UC). 5. WASTE COLLECTION ROUTE OPTIMIZATION In order to reduce the operational cost for waste collections, route optimization is also considered in this work. The route optimization is achieved by using Dijkstra's algorithm[18]of the Netlogo Thus, for the set of vertices Q [B, C, D, E……T] inFig.5 the algorithm identifies the vertices (place or nodes) to be visited. The edges link the start and end of vertices and the shortest path can be drawn. 8 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 Figure 5. Set of vertices and edges linked by Dijkstra's algorithm. 6. RESULTS AND DISCUSSION The implementation of Multi-agent model for real time waste monitoring and optimal route for waste collection using the Netlogo platform was verified by running several simulation cases at different simulations time T in minutes. Initially the model presents 25 bins located randomly in a city, with the initial state of [bin_load=0] and color = green. The monitors provide the number of located bins and show no number of full_bins. Parameters of bins, trucks and cost per waste unit to be paid by the citizen are also shown in Fig.6. Simulation time T1 = 0 minutes Figure 6. Initial setup for real time waste monitoring multiagent model. 9 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 The Go button run the simulation and bin levels increased per 1 tick, the bins with [bin_load>=10] set color to yellow meaning an alert that the bin is nearly to be full. The full bins are red in color [bin_load>=25]. At simulation time T2 = 53 minutes, 8 out of 25 distributed bins were full, bin number 1, 5, 6, 10,18,19,20 and 24 see Fig.7 which, also reports the continuous recording of different waste level foreach number of bin per tick, the optimal path for waste collection by a truck was obtained Fig. 8to the bins which were full. Fig.8 also shows the change in color for the different waste level inside a bin, and truck optimal path for waste collections. From this simulation case, total revenue of 100,000UCs expected to be obtained by trucks for the collection operations to all 8 full bins if 500UCsbe an assumed amount for cost per unit waste, paid by the citizen for waste collection services: RT=∑ (cid:4)(cid:9) ∗ (cid:6) (cid:10) (cid:2) (ii) Where RT is the total revenue, Niis the weight of the ith bin, p is the price per unit weight of waste. Figure 7. Bins with different waste levels at time T = 53 minutes. At simulation time T = 93 minutes the continuous recording of different waste level for each number ofbin per tick is shown in Fig. 9. Number of uncollected waste increased to 12, new 8 full bins and other 4 which had delay in collection since T =53 minutes. The identification of vertices in the grid position for 4 bins which shown the delay in collection was done Fig.10. The obtained new optimal path included the four vertices which were uncollected from the first simulation time, vertices set for full bins with collection delay was given as; [(D [9,9], L[(-4.5, 0)], M[(0, 0)],O[(9,0)] 10 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 Figure 8. Color change for different bin's waste level and Truck optimal path for waste collection to 8 full bins Figure 9. Bins with different waste levels at time T = 93 minutes 11 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 Figure 10. Identified vertices which had delay in collection and truck optimal path 7. CONCLUSION AND FUTURE WORK In this paper, we have presented an overview of the proposed architectures and simulation of our on-going work towards smart solid waste management using multiagent based modeling approach. Real time and continuous reporting of waste level status from the bins ensures the clean and green environment and wellbeing of the citizen in a city. The use of WSN Technology assures efficiency of the MSW operations, despite of having technical challenges such as internet connectivity, security of components, power supply and error readings from the sensors, are to be overcome before the actual implementation. Further work would consider full development of the WSN for real time waste monitoring and collection. Real Geographical Information System (GIS) map of the case study area will be used to identify nodes for the located smart bins. The distance from one node to another will be used to calculate cost per trip by a truck. The profit or loss of the operation will be achieved by the deduction from the collected revenues. In additional optimal path for waste collection will be achieved using the developed mathematical model for route optimization which is appropriate for supporting decision making on municipal solid waste management. ACKNOWLEDGMENT The author would like to thank Ms. DevothaNyambo and Dr. Joseph Mwangoka of Department of Information Communication Science and Engineering at The Nelson Mandela African Institution of Science and Technology for the great supervision and technical guidance, and the German Academic Exchange Service (DAAD) for study fund support provided. 12 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 REFERENCES [1] M. A. Al Mamun, M. A. Hannan, A. Hussain, and H. Basri, "Theoretical model and implementation of a real time intelligent bin status monitoring system using rule based decision algorithms," Expert Syst. Appl., vol. 48, pp. 76–88, 2016. [2] D. C. Wilson, L. Rodic, A. Scheinberg, C. A. Velis, and G. Alabaster, "Comparative analysis of solid waste management in 20 cities," Waste Manag. Res., vol. 30, no. 3, pp. 237–254, 2012. [3] S. Balandin, S. Andreev, and Y. Koucheryavy, Internet of Things, Smart Spaces, and Next Generation Networks and Systems, vol. 9247, no. June. Cham: Springer International Publishing, 2015. [4] S. Kumar, "Ubiquitous Smart Home System Using Android Application," Int. J. Comput. Networks Commun., vol. 6, no. 1, pp. 33–43, 2014. [5] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, "Internet of things for smart cities," IEEE Internet Things J., vol. 1, no. 1, pp. 22–32, 2014. [6] A. Mesmoudi, M. Feham, and N. Labraoui, "Wireless Sensor Networks Localization Algorithms: A Comprehensive Survey," Int. J. Comput. Networks Commun., vol. 5, no. 6, pp. 45–64, 2013. [7] K. Potiron, A. El Fallah Seghrouchni, and P. Taillibert, From Fault Classification to Fault Tolerance for Multi-Agent Systems. London: Springer London, 2013. [8] N. V. Karadimas, G. Rigopoulos, and N. Bardis, "Coupling multiagent simulation and GIS - An application in waste management," in WSEAS Transactions on Systems, 2006, vol. 5, no. 10, pp. 2367–2371. [9] M. Longo, M. Roscia, and G. C. Lazaroiu, "Innovating Multi-agent Systems Applied to Smart City," Res. J. Appl. Sci. Eng. Technol., vol. 7, no. 20, pp. 4296–4302, 2014. [10] M. A. Al Mamun, M. A. Hannan, A. Hussain, and H. Basri, "Integrated sensing systems and algorithms for solid waste bin state management automation," IEEE Sens. J., vol. 15, no. 1, pp. 561– 567, 2015. [11] V. Catania and D. Ventura, "An Approch for Monitoring and Smart Planning of Urban Solid Waste Management Using Smart-M3 Platform," in Proceedings of 15th Conference of Open Innovations Association FRUCT, 2014, pp. 24–31. [12] B. Chowdhury and M. U. Chowdhury, "RFID-based real-time smart waste management system," in 2007 Australasian Telecommunication Networks and Applications Conference, 2007, pp. 175–180. [13] M. A. Al Mamun, M. A. Hannan, A. Hussain, and H. Basri, "Wireless Sensor Network Prototype for Solid Waste Bin Monitoring with Energy Efficient Sensing Algorithm," in 2013 IEEE 16th International Conference on Computational Science and Engineering, 2013, pp. 382–387. [14] V. Sanjeevi and P. Shahabudeen, "Optimal routing for efficient municipal solid waste transportation by using ArcGIS application in Chennai, India.," Waste Manag. Res., vol. 34, no. 1, pp. 11–21, 2016. [15] M. A. Al Mamun, M. A. Hannan, A. Hussain, and H. Basri, "Wireless sensor network prototype for solid waste bin monitoring with energy efficient sensing algorithm," Proc. - 16th IEEE Int. Conf. Comput. Sci. Eng. CSE 2013, pp. 382–387, 2013. 13 International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.7, No.5, October 2017 [16] F. Tancini et al., "1,1-Dicyano-4-[4-(diethylamino)phenyl]buta-1,3-dienes: Structure-Property Relationships," European J. Org. Chem., vol. 2012, no. 14, pp. 2756–2765, May 2012. is currently a master's student [17] H. Gomaa, "Activity Diagrams," Softw. Model. Des., pp. 89–92, 2011. [18] E. Millinocket and M. Dodge, "13 13.1 Shortest Paths," Algorithms, pp. 1–11, 2000. AUTHORS Eunice D. Likotiko in Information and Communication Science and Engineering at The Nelson Mandela African Institution of Science and Technology (NM-AIST)specializing in Information Technology and System Development, also employed as Tutorial assistant at Ardhi University in the Department of Computer Systems and Mathematics. She received BSc. (Hons) degree in Information Systems Management at Ardhi University in November 2012. Her interest includes, Modelling and Simulations, Internet of Things, smart and embeded systems development. Devotha Nyambo is a PhD candidate at the Nelson Mandela African Institution of Science and Technology Information and Communication Science and Engineering. She is a prestigious fellow of the African Women in Agricultural Research and Development (AWARD) through which she has been trained in scientific skills, leadership and mentorship. Devotha's research focus is on livestock informatics. Currently, she is using Multi-Agents Research and Simulation to understand evolvement of smallholder dairy farmers. She has developed strong capabilities in paperless data collection (tools development, database designs, data management, and field work), Data analysis (descriptive statistics and modeling) and e-documents design and development for business collaboration Dr. Joseph W. Mwangoka, is Lecturer at School of Computational and Communication Science and Engineering, at The Nelson Mandela African Institution of Science and Technology, Arusha, Tanzania. E- mail: [email protected] (NM-AIST), specializing in 14
1706.08046	1	1706	2017-06-25T07:35:00	An Algorithm for Supervised Driving of Cooperative Semi-Autonomous Vehicles (Extended)	[ "cs.MA", "cs.RO" ]	Before reaching full autonomy, vehicles will gradually be equipped with more and more advanced driver assistance systems (ADAS), effectively rendering them semi-autonomous. However, current ADAS technologies seem unable to handle complex traffic situations, notably when dealing with vehicles arriving from the sides, either at intersections or when merging on highways. The high rate of accidents in these settings prove that they constitute difficult driving situations. Moreover, intersections and merging lanes are often the source of important traffic congestion and, sometimes, deadlocks. In this article, we propose a cooperative framework to safely coordinate semi-autonomous vehicles in such settings, removing the risk of collision or deadlocks while remaining compatible with human driving. More specifically, we present a supervised coordination scheme that overrides control inputs from human drivers when they would result in an unsafe or blocked situation. To avoid unnecessary intervention and remain compatible with human driving, overriding only occurs when collisions or deadlocks are imminent. In this case, safe overriding controls are chosen while ensuring they deviate minimally from those originally requested by the drivers. Simulation results based on a realistic physics simulator show that our approach is scalable to real-world scenarios, and computations can be performed in real-time on a standard computer for up to a dozen simultaneous vehicles.	cs.MA	cs	An Algorithm for Supervised Driving of Cooperative Semi-Autonomous Vehicles (Extended) Florent Altch´e, Xiangjun Qian, and Arnaud de La Fortelle, 1 7 1 0 2 n u J 5 2 ] A M . s c [ 1 v 6 4 0 8 0 . 6 0 7 1 : v i X r a Abstract-Before reaching full autonomy, vehicles will gradu- ally be equipped with more and more advanced driver assistance systems (ADAS), effectively rendering them semi-autonomous. However, current ADAS technologies seem unable to handle complex traffic situations, notably when dealing with vehicles arriving from the sides, either at intersections or when merging on highways. The high rate of accidents in these settings prove that they constitute difficult driving situations. Moreover, intersections and merging lanes are often the source of important traffic congestion and, sometimes, deadlocks. In this article, we propose a cooperative framework to safely coordinate semi- autonomous vehicles in such settings, removing the risk of collision or deadlocks while remaining compatible with human driving. More specifically, we present a supervised coordination scheme that overrides control inputs from human drivers when they would result in an unsafe or blocked situation. To avoid unnecessary intervention and remain compatible with human driving, overriding only occurs when collisions or deadlocks are imminent. In this case, safe overriding controls are chosen while ensuring they deviate minimally from those originally requested by the drivers. Simulation results based on a realistic physics simulator show that our approach is scalable to real-world scenarios, and computations can be performed in real-time on a standard computer for up to a dozen simultaneous vehicles. Index Terms-Semi-autonomous driving, safety, supervisor, supervised driving. I. INTRODUCTION A DVANCED driver assistance systems (ADAS) are be- coming increasingly complex as they spread across the automotive market. Although adaptive cruise control (ACC) [1] and automated emergency braking (AEB) [2] are the best-known examples of such systems, applications of ADAS have been broadened and now include pedestrian [3], traffic light [4] or obstacle detection [5] as well as lane keeping assistance [6]. The development of this new equipment allows drivers to delegate part of the driving task to their vehicles. As these systems keep getting more efficient and able to handle more complex situations, vehicles will gradually progress towards semi-autonomous driving, where drivers remain in charge of their own safety, while their errors can be seamlessly corrected to prevent potential accidents. One of the challenges of semi-autonomous driving lies in efficiently handling vehicles on conflicting paths, for instance at an intersection or a highway entry lane. Traffic rules such as priority to the right can help determine whether to pass before F. Altch´e, X. Qian and A. de La Fortelle are with MINES ParisTech, PSL Research University, Centre for robotics, 60 Bd St Michel 75006 Paris, France (e-mail: [email protected].) F. Altch´e is also with ´Ecole des Ponts ParisTech, Cit´e Descartes, 6-8 Av Blaise Pascal, 77455 Champs-sur-Marne, France. or after another vehicle; however, many situations require driving experience to be handled efficiently. Learning-based approaches may eventually prove able to transfer driving ex- perience to a computer, but such knowledge is very hard to im- plement in a safety system. In this article, we consider another possible solution, consisting in using vehicle-to-vehicle or vehicle-to-infrastructure communication for cooperative semi- autonomous driving. In this setting, vehicles negotiate with one another, or receive instructions from a centralized computer, allowing them to drive safely and efficiently. In this article, we consider a method to ensure the safety of multiple semi-autonomous vehicles on conflicting paths, for instance crossing an intersection or entering a highway, while remaining compatible with the presence of human drivers. To this end, and inspired by earlier work in [7], [8], we propose a so-called Supervisor which monitors control inputs from each vehicle's driver, and is able to override these controls when they would result in an unsafe situation. More specifically, the role of the supervisor is twofold: first, knowing the current states of the vehicles, the supervisor should determine if the controls requested by the drivers would lead the vehicles into unsafe inevitable collision states [9]. In this case, the second task of the supervisor is to compute safe controls – maintaining the vehicles in safe states – which are as close as possible to those actually requested by the drivers. We say that such a control is minimally deviating. This paper provides two main contributions: from a prac- tical standpoint, we design and implement a mathematical framework allowing to simultaneously perform the safety verification of target control inputs, and the computation of minimally deviating safe controls if target inputs are unsafe. From a theoretical standpoint, we formally prove that verifying safety over a finite time horizon is enough to ensure infinite horizon safety, and we provide a sufficient condition on the verification horizon for this property to hold. Unlike previous work focusing on specific situations such as intersections [7], [8], our framework can be applied to a wide variety of driving scenarios including intersections, merging lanes and roundabouts. The rest of the article is structured as follows: in Section II, we provide a review of the related literature. In Section III, we present our modeling of semi-autonomous vehicles and introduce the Supervision problem of verifying the safety of drivers control inputs and finding a minimally deviating safe control if necessary. In Section IV, we present an infinite horizon formulation based on constraints programming to solve this problem. In Section V, we derive a finite horizon to the infinite formulation which we prove is equivalent horizon one. In Section VI, we use computer simulations to showcase the performance of the proposed supervisor in various driving situations. In Section VII, we present possible methods for real-world implementations of our approach. Finally, Section VIII concludes the study. II. RELATED WORK In the last decade, a lot of research has been focused on coordinating fully autonomous vehicles in challenging settings such as crossroads, roundabouts or merging lanes, with the ambition of improving both safety and traffic efficiency. Nau- mann et al. [10], followed by Dresner and Stone [11] have seemingly pioneered the work of adapting traffic intersections management methods to fully autonomous vehicles, designing so-called autonomous intersection management algorithms. They propose that each approaching autonomous vehicle re- serves a time interval to cross the intersection; collisions are prevented by ensuring that conflicting vehicles are assigned non-overlapping crossing times. Subsequent studies on this particular problem have led to other approaches. In [12], vehicles choose their control inputs based on navigation functions which include a collision avoidance term, allowing vehicles to react to maneuvers from other traffic participants. In [13], collision avoidance is ensured by assigning relative crossing orders to incoming vehicles; each vehicle then uses model predictive control to plan collision-free trajectories respecting these priorities. Other authors have considered different driving situations for autonomous vehicles, such as cooperative merging on a highway [14]–[16], or entering a roundabout [17]. By contrast, relatively little work has considered semi- autonomous driving assistance, possibly because the presence of human drivers brings a lot of additional complexity. The goal of a semi-autonomous driving assistant is to help the driver avoid collisions, either by notifying of a potential danger [18] or by taking over vehicle control in dangerous situations [19]–[21]. To be accepted by human drivers, such systems should be as unobtrusive as possible, and in particular should only intervene when necessary. Most of the currently existing literature on semi-autonomous driving mostly focuses on highway driving [19]–[21], which presents relatively low difficulty as vehicles trajectories remain mostly parallel. The aim of this article is to bring semi-autonomy one step fur- ther, to allow cooperative driving between semi-autonomous vehicles in more complex conflict situations. Some of these more complex problems have already been studied in the literature. In [22], the authors consider semi- autonomous driving at an intersection and propose that human drivers let an automated system control their vehicle while crossing said intersection. However, this scheme is rather intrusive as drivers completely relinquish control for a time, and handing back controls to a potentially distracted driver poses problems by itself. Colombo et al. [7], [8] introduced the idea of a supervisory instance (called supervisor) tasked with preventing the system of vehicles from entering undesirable states by overriding the controls of one or several vehicles. In this more human-friendly approach, overriding only occurs 2 when necessary, i.e. if an absence of intervention would result in a crash. The question of determining whether overriding is needed or not, called verification problem, is NP-hard [23]; under several simplifying assumptions, it is shown in [7] to be equivalent to a scheduling problem. In this reformulation, vehicles are each assigned a time slot during which they are allowed inside the intersection, and assigned slots are mutually disjoint. If, due to vehicle dynamics, no feasible schedule exists, the initial state is deemed unsafe. This allows the authors to design a so-called least restrictive supervisor, which verifies the safety of the desired inputs and overrides them if necessary. However, the proposed supervisor is only suitable to simple intersection geometries with a single conflict point. Moreover, no additional property is required from the safe controls used for overriding, which can widely deviate from the desired ones. Several variations have been proposed based on the equiva- lence demonstrated in [7]. Reference [24] designs a supervisor which is robust to bounded uncertainties by adding safety mar- gins. Reference [25] leverages job-shop scheduling to develop a supervisor that considers several possible conflict points inside the intersection; however, vehicle dynamics are only modeled as first-order integrators, which is not realistic in a real-world setting. Campos et al. [8] proposed a Pareto-optimal supervisor leading to a minimally deviating formulation by recursively finding the most constrained vehicle, reserving its optimal crossing time, and scheduling the crossing of the remaining vehicles using the previous schedule as constraints. This method allows to minimize the deviation between the overridden and desired controls, but may be computationally intensive. Indeed, one of the major difficulties of performing optimization in this context lies in the necessity to consider all the possible orderings of the vehicles. This problem is highly combinatorial; it has been shown that there exists up to 2n(n−1)/2 orderings for n vehicles [26]. Moreover, it is generally ignored by most authors studying motion planning problems, who either use simple heuristics such as first-come, first-served [11], [27] or rely on exhaustive search [8], [28]. A possible method to handle the combinatorial explosion is to use pruning techniques such as branch-and- bound, which avoid exploring branches of the decision tree that would provably yield suboptimal results. These methods are commonly used in mixed-integer linear (see, e.g., [29], [30] for applications to motion planning) or quadratic programming (see, e.g., [31]) problems, which combine continuous and discrete optimization. More general nonlinear methods have also been used in motion planning [32], [33], although their high computational difficulty generally requires linearization for effective resolution, as illustrated in [34]. To the best of the authors' knowledge, branch-and-bound methods have never been applied to semi-autonomous driving. This article significantly differs from references [7], [8], [25]. Instead of using a scheduling approach, we formulate the supervision problem as a Mixed Integer Quadratic Program- ming (MIQP) problem, which can handle various geometries with multiple collision points such as multi-lane intersections, merging lanes or roundabouts. Our formulation only requires to consider a small, finite planning horizon, while previous approaches [7], [8], [25] needed to schedule the crossing of all the considered vehicles. Furthermore, the MIQP formula- tion is highly flexible, allowing to take into account various constraints (e.g., maximal turning speed) and different cost functions. Finally, the resolution of MIQP can leverage highly- optimized solvers [35], allowing real-time implementations even for a relatively large number of vehicles. This article expands the results presented in the conference paper [36]; among the significant improvements made in this extension, we now give a more comprehensive model of our vision of semi-autonomous vehicles and adjust the modeling of the problem to handle bounded control errors. We provide a detailed discussion on how complex road geometries with multiply-intersecting paths can be handled, leading to a very versatile framework. Finally, we extend the theoretical results to continuous arrivals of vehicles, and provide possible ways for actual implementation as a roadside unit. III. SUPERVISION PROBLEM We consider the problem of safely coordinating multiple semi-autonomous vehicles on the road, in order to prevent collisions and deadlock situations where no vehicle is able to move forward. Since vehicles are human-driven, a form of outside supervision is necessary to prevent undesirable situations. This section presents our formulation of a so-called Supervision problem generalizing the work of Colombo et al. [7]; solving this problem yields a provably safe control, as close as possible to the original intentions of the drivers. A. Modeling 1) Supervision area: We consider an isolated portion of a road infrastructure used by semi-autonomous vehicles, where some form of coordination is required to ensure vehicles safety. For instance, this could be a classical road intersection, a roundabout or an entry or weaving lane on a highway. We call this bounded portion of infrastructure the supervision area and we assume that vehicles can travel safely outside of the collision area using only their ACC capacities. In a real-world setting, different critical portions of infrastructure which are far enough apart can be considered individually, but need to be treated jointly if traffic from one can influence another. Figure 1 shows examples of roads configurations and the corresponding possible choice for a supervision area. In this article, we present an embodiment of a Supervisor working over a spatially static supervision area over time, that can be thought of as a dedicated computer on the infrastructure or in the cloud. Vehicles are assumed to establish a connection to the supervisor when they enter the supervision area (using, for instance, V2I communication), and maintain it until they exit this region. We denote by Nt the set of vehicles currently inside the supervision area at a time t. semi- autonomous vehicles equipped with advanced driver assistance systems, many of which are already commercially available, and Vehicle to Infrastructure (V2I) communication capacities. In particular, vehicles are assumed to have advanced cruise control, automated braking and lane keeping assistance 2) Semi-autonomous vehicles: We consider 3 (a) Crossroads (b) Roundabout (c) Highway merging Fig. 1. Examples of considered road configurations, and corresponding supervision areas (interior of the dotted rectangles). systems such that accelerating, braking and steering can be actuated by an on-board computer. Moreover, we suppose that vehicles have access to reliable cartographic data and are capable of precisely measuring their current position, orientation and velocity with reference to a unique global frame, for instance using GNSS and inertial navigation. Since the vehicles are not assumed to have advanced environment-sensing capacities, for instance based on LIDAR data, they are not able to handle all situations and still require a human driver to safely navigate, for instance in the case of on-road obstacles or loss of GNSS signal. Moreover, lateral collisions or deadlock situations can happen due to human error, justifying the need for supervision. 3) Parametrization: In the remainder of this article, we only consider the two-dimensional kinematics and dynamics of the vehicles. We denote by Ei a bounding polygon for the shape of vehicle i ∈ Nt, and by ci the center of Ei. We assume that the geometry and lane markings of the roads inside the supervision area define a finite number of reference paths across this region, as exemplified in fig. 2. Due to the presence of a lane keeping assistance system, we assume that every vehicle is able to follow one of these reference paths with a small bounded lateral error. Noting γi the reference path of a vehicle i, we assume that the distance of ci from γi is bounded from above by ξi ≥ 0. Moreover, we assume that γi is at least C2-continuous, and that ξi is small enough to ensure, for all x ∈ R2, d(x, γi) ≤ ξi ⇒ ∃! y ∈ γi, x − y = d(x, γi). (1) This condition allows to use the curvilinear position of the point of γi closest to ci to uniquely encode the longitudinal position of vehicle i along γi. We denote by si this curvilinear position, with the convention that si = 0 when the front bumper of i first enters the supervision area and increases when i goes forward; we let sout be the longitudinal position at which the rear bumper of i fully exits the supervision area. i 4) Vehicle dynamics: In this article, we mostly focus on the longitudinal dynamics of the vehicles, and we let xi = (si, vi)T be the state of vehicle i, where si and vi are re- spectively its longitudinal position and longitudinal speed. We assume that vehicles follow second-order integrator dynamics with a bounded longitudinal error, and that the control input ui corresponds to the longitudinal acceleration as: 4 xi = Axi + Bui, 0 0 ) and B = ( 0 (2) where A = ( 0 1 1 ). Since we mostly consider situations with conflicting vehicles, we assume that human drivers maintain a relatively low speed (compared to the cur- vature of their path), which allows neglecting lateral dynamics and slip [37]. i∈Nt performance. At a given time t, we let Ut = (cid:81) To account for speed limitations on the vehicles, each vehicle i is supposed to have a bounded non-negative velocity, so that vi ∈ [0, vi] (with vi > 0) at all times. Moreover, we assume that the acceleration ui of each vehicle is bounded as ui ∈ [ui, ui], with ui < 0 < ui. These bounds can differ between vehicles, thus allowing heterogeneous vehicle [ui, ui] be the set of admissible accelerations for the vehicles of Nt. We denote bt boldface x = (xi)i∈Nt and u = (ui)i∈Nt the state and control for the system of vehicles. In what follows, we let vmax > 0 be a global upper bound for vi, ua > 0 a lower bound for ui and ub < 0 an upper bound for ui such that for all t ≥ tκ and all i ∈ Nt, vi ≤ vmax and ui ≤ ub < 0 < ua ≤ ui. Therefore, all vehicles are capable of braking with ub and accelerating with ua; finally, we let umax be a global upper bound for ui. 5) Collision regions: Finally, we assume that the angle between the orientation of vehicle i and the tangent to γi at its point closest to ci is also bounded. With these hypotheses, for any pair of vehicles (i, j), we can compute the bounded set Cij of curvilinear positions (si, sj) ⊂ [0, sout ] × [0, sout ] for which a collision could happen between i and j. Note that these sets are "inflated" to take into account the bounded control errors. We call Cij the collision region between i and j; fig. 2 shows examples of paths and corresponding computed collision regions for different driving situations. Note that collision regions can be empty or have one or multiple connected components. If Cij (cid:54)= ∅, we say that vehicles i and j are conflicting; when Cij has multiple connected components, we denote by Cp ij its p-th component, using the convention ij = Cp Cp ji. 6) No-stop regions: To prevent creating deadlock situa- tions, vehicles are not allowed to stop when doing so would (cid:1) block traffic in other directions. To this extent, we define a no-stop region (see fig. 3) Di for each vehicle i ∈ Nt as the ij for all t(cid:48) ≥ t, j ∈ Nt(cid:48) and all p such that (0, 0) /∈ Cp ij; in this formula, Πsi is the projection operator on the first coordinate. The no-stop region corresponds to the part of the supervision area where a vehicle may have to yield to another; if Cp ij contains (0, 0), then either i or j enters the supervision area behind the other, in which case the relative ordering of the vehicles is given and the Cp Note that, although this definition theoretically requires knowledge of all future vehicles, Di can be computed off-line as a finite intersection of intervals provided that there only exists a finite number of possible paths inside the supervision area. In what follows, we let vmin > 0 be a minimum allowed smallest interval Di = [s⊥i , s⊥i ] containing all min(cid:0)ΠsiCp ij does not count in Di. j i (a) Simple orthogonal intersection situation with example vehicle shapes (b) Roundabout situation with multiple connected components in Cij (c) Highway merging situation Fig. 2. Examples of paths (left) and corresponding collision regions (right) for vehicles with the polygonal shape shown in fig. 2a. Fig. 3. Illustration of the no-stop region Di and acceleration region Ai inside the supervision area (dotted rectangle). i i speed for any vehicle inside its no-stop region, and we assume that vmin ≤ vi for all vehicles. For a no-stop region Di, we define the corresponding accel- , s⊥i ] such that, if vehicle i is stopped eration region Ai = [sacc , it can reach a speed vmin before reaching s⊥i . More at sacc for all specifically, we require that 0 ≤ sacc i. Inside the acceleration region, vehicles are only allowed to accelerate; this condition prevents vehicles from stopping right before the entrance of the no-stop region, leaving them unable to proceed forward due to the minimum speed requirement. Figure 3 illustrates an example of the no-stop regions and the corresponding acceleration regions. i ≤ s⊥i − vmin 7) Time discretization: Drivers continuously change the control input of their vehicle; however, due to computational 2ua 2 γiγjCijsisjγiγjC1ijC2ijsisjγiγjCijsisjij1j2jn...Cij1Cij2CijnDiAiand communication constraints, it is impractical to handle functions of a continuous variable. In the remainder of this article, we choose a constant time step duration τ > 0, and we assume that all vehicles use piecewise-constant controls with step τ, typically 0.5 s. To simplify the formulation, we further assume that vehicles update their control simultaneously at times tκ = κτ for κ ∈ N, and we denote by Uτ (tκ) the set of piecewise-constant admissible controls for the vehicles of Ntκ. By definition, for all t ≥ tκ and all u ∈ Uτ (tκ), u(t) ∈ Utκ. B. Problem statement Before presenting the so-called supervision problem, we first define the safety criterion for the vehicles inside the supervision area at a given time. Definition 1 (Safe state). We say that the supervision area is in a safe state xκ at time tκ if there exists an admissible piecewise-constant control u ∈ Uτ (tκ) defined over [tκ, +∞[ such that, under this control and starting from xκ, for all t ≥ tκ and all i, j ∈ Ntκ, (si(t), sj(t)) /∈ Cij. Such a control is said to be a safe control. τ τ With this definition, the supervision area is in a safe state when all the vehicles inside this area can apply a dynamically admissible, infinite horizon control without a risk of collision. This safety condition corresponds to a contraposition of the notion of "inevitable collision state" proposed by Fraichard et (tκ) the set of al. [9]. In what follows, we denote by U saf e safe and dynamically admissible piecewise-constant controls for the vehicles in Ntκ; by definition, a control u ∈ U saf e (tκ) is a piecewise-constant function from [tκ, +∞[ to Utκ. We now define the safety condition for vehicles entering the supervision area. Definition 2 (Safe entry). Consider a safe state xκ at time tκ and let t1 > tκ be the first time at which a new vehicle enters the supervision area. We say that the vehicles of Nt1 \ Ntκ safely enter the supervision area with a margin τ if t1 ≥ tκ+τ, or if any safe control u ∈ U saf e • keeps the system of the vehicles of Nt1 safe at time tκ +τ • remains safe over [tκ, +∞[ for the vehicles of Ntκ, regardless of the control applied by the vehicles of Nt1 \ Ntκ over [t1, t1 + τ ]. (tκ): and τ This definition ensures that a safe control computed for the vehicles of Ntκ remains safe after new vehicles enter, i.e. the entry of new vehicles does not invalidate previously safe controls. Moreover, we assume that we can safely exclude vehicles departing the supervision area from the safety ver- ification problem, i.e. that drivers are able to safely follow the previously departed vehicles without supervision. We will show in Section IV-C that these hypotheses allow discrete-time supervision with continuous vehicle arrival. In the remainder of this article, we consider a centralized supervisor working in discrete time steps of duration τ, and we assume that new vehicles always enter safely with a margin τ. At the beginning of each time step κ, the supervisor receives an information about the desired longitudinal control of each 5 τ vehicle for the next time step, denoted by uκ i,des. The collection of these desired controls for the vehicles of Ntκ defines a des defined over [tκ, tκ + τ [. constant desired system control uκ This control may, or may not, lead the system of vehicles into an unsafe state. The supervisor is tasked with preventing the system from entering an unsafe state, by overriding the desired control if necessary. To remain compatible with human drivers, it is desirable that the supervisor has several properties, namely being least restrictive and minimally deviating. Letting τ,κ (tκ) be the restriction of the functions of U saf e (tκ) to U saf e [tκ, tκ +τ [, we define the least restrictive supervision problem: Definition 3 (Least restrictive supervision). Consider a safe state xκ at time tκ = κτ, a desired system control uκ des and assume that all new vehicles enter the supervision area safely with a margin τ. The least restrictive supervision problem (SP ) is that of finding a control uκ τ,κ (tκ) such that uκ τ,κ (tκ). des if uκ this definition corresponds to that of [7] in our generalized setting. Such a supervisor is least restrictive because overriding only occurs if the initially requested control would lead the vehicles in an unsafe state. However, it is also desirable that the control used for overriding is chosen close to the drivers' desired control. Extending the work in [8], we define the minimally deviating supervision problem as follows: Definition 4 (Minimally deviating supervision). Consider a safe state xκ at time tκ = κτ, a desired system control uκ des and assume that all new vehicles enter the supervision area safely with a margin τ. The minimally deviating supervision problem (SP ∗) is that of finding a constant control u∗κ such that: saf e ∈ U saf e des ∈ U saf e saf e = uκ Note that saf e u∗κ saf e = arg min u ∈ U saf e τ,κ (tκ)uκ − uκ des (3) where · is a norm defined over Utκ. Note that, from this definition, any solution to SP ∗ is a solution to SP . This concept of minimally deviating supervision follows a different fail-safety paradigm that could be found in, e.g., rail transportation where all trains in an area should perform an emergency braking when an incident occurs. The reasoning behind definition 4 is that, to improve efficiency without sacrificing safety, intervention is only performed on vehicles which are actually at risk, and does not necessarily result in a full stop. However, at individual vehicle level, the safe overriding control u∗κ saf e may differ greatly from the driver's input, e.g. braking instead of accelerating. IV. INFINITE HORIZON FORMULATION OF THE SUPERVISION PROBLEM In this section, we present an extension of the work in [36] allowing to reformulate the generalized minimally deviating supervision problem using mixed-integer quadratic program- ming (MIQP) in Section IV-A. As the supervisor works in discrete time steps of duration τ, we consider the beginning of a step κ, corresponding to a time tκ = κτ and formulate an infinite-horizon MIQP problem. Assuming the initial state 6 is safe, we will show in Section IV-B that this formulation can be used to find a minimally deviating safe control for the vehicles in Ntκ. We will show in Section IV-C that, if the vehicles of Ntκ follow the corresponding control, our formulation expressed at tκ+1 = (κ + 1)τ remains feasible for the vehicles of Ntκ+1, provided that all new vehicles enter safely with a margin τ. These properties ensure that our infinite horizon MIQP formulation can be solved in a receding horizon fashion, to ensure safety for all future vehicles. A. Model variables and constraints Fig. 4. Minimum bounding hexagon for the collision region presented in fig. 2a. sk+1 i − sk vk+1 i − vk 1 i = 2 i = uk i τ In what follows, we present the variables and constraints used in our model. Unless specified otherwise, these con- straints are enforced at all time steps k ≥ κ, and for all vehicles of Ntκ. 1) Vehicle dynamics: When they evolve inside the super- vision area, vehicles use a piecewise-constant control, which is updated every τ seconds. For a vehicle i ∈ Ntκ at step k, i ∈ [ui, ui], we introduce the variables sk respectively denoting its curvilinear position and longitudinal speed at tk, and longitudinal acceleration over [tk, tk +τ [. The following constraints enforce vehicle dynamics: i ∈ [0, vi] and uk i , vk (cid:0)vk (cid:1) τ i + vk+1 i (4) (5) 2) Logical constraints: In [38], we showed that it is pos- sible to enforce logical constraints on continuous and integer variables with linear inequalities using a "big-M" formulation. More specifically, if b is a binary variable and x a continuous or integer variable bounded so that x < M, then the logical constraint: (b = 0 ⇒ x ≤ a) is equivalent to the linear inequality constraint x ≤ a + bM. This method can be used to define indicator binary variables for a given semi- infinite interval: for a continuous variable x and a constant a ∈ R, we denote by b = χ[a,+∞[(x) the constraints (b = 0 ⇒ x ≤ a) ∧ (b = 1 ⇒ x ≥ a), where ∧ denotes the binary conjunction; we use ¬ to denote the binary negation. 3) Collision avoidance: As presented in Section III-A3, the collision region between two vehicles i, j ∈ Ntκ, Cij, can be computed off-line. As it was already presented in [38], it is possible to compute a minimal bounding convex polygon for each connected component Cp ij of Cij. A good compromise be- tween accuracy and complexity is to use a bounding hexagon with edges either parallel to the si = 0, si = sj or sj = 0 lines; such a polygon is uniquely defined by six parameters, as shown in fig. 4. To ensure that vehicles do not enter any of the collision regions, we introduce a set of binary variables to encode the discrete decisions arising from the choice of an ordering of vehicles, as presented in [38]. For all conflicting vehicles i, j ∈ Ntκ, we let πp ij = 1 if vehicle i passes the p-th collision region before j, and πp ij = 0 otherwise; moreover, we introduce the binary indicator variables for all k ≥ κ: εij,p(k) = χ[sij,p,+∞[(sk i ), ε⊥ij,p(k) = χ[s⊥ ij,p,+∞[(sk i ). We enforce the collision avoidance constraints for all con- flicting vehicles i, j ∈ Ntκ and k ≥ κ as: (cid:0)πp ij ∧ ¬ εij,p(k)(cid:1) (cid:0)πp ij ∧ εij,p(k) ∧ ¬ ε⊥ij,p(k)(cid:1) ij ∧ εij,p(k) ∧ ¬ ε⊥ij,p(k)(cid:1) (cid:0)πp j ≤ s⊥ji,p i ≥ sk+1 ⇒ sk+1 ⇒ sk+1 ⇒ (cid:0)vk+1 τ 2 sk+1 i ≥ sk+1 j +dij,p + j − vk+1 i j + dij,p (8) (9) (10) (cid:1) where dij,p = sij,p − s⊥ji,p. Constraint (8) corresponds to "crossing situations", where a vehicle has to wait for another to pass; constraints (9) and (10) correspond to "following situa- tions", where a vehicle needs to maintain a certain longitudinal distance from another. Note that constraints (8) to (10) use the values of the indicator variables at step k to force the positions of the vehicles at step k + 1 in order to avoid a "corner cutting" phenomenon; the additional constraint (10) prevents collisions between two time steps. These constraints are very slightly stronger than that of collision avoidance, i.e. for all t ≥ tκ, (si(t), sj(t)) /∈ Cij. Consequently, the results in the rest of this article are to be understood replacing the exact collision avoidance constraints in definition 1 by conditions (8)-(10). Finally, to ensure the consistency of the formulation, we add the mutual exclusion constraint for all conflicting vehicles i, j ∈ Ntκ: πp ij + πp ji = 1. (11) 4) Deadlock avoidance: As described in Section III-A6, we require all vehicles to maintain a minimum speed inside their no-stop region Di = [s⊥i , s⊥i ]. This requirement is enforced by defining additional binary variables, for all i ∈ Ntκ and all k ≥ κ: i ζ acc (k) =χ[sacc i ζ in i (k) =χ[s⊥ ζ out (k) =χ[s⊥ i ηi(k) =χ[vmin−uaτ,+∞[(vk i ) ,+∞[(sk i ) i ,+∞[(sk i ) i ,+∞[(sk i ) (12) (13) (14) (15) and using the constraints: (cid:0)ζ acc i (k) ∧ ¬ ηi(k)(cid:1) (k)(cid:1) (cid:0)ζ in i (k) ∧ ¬ ζ in (16) (17) As long as the acceleration regions Ai are large enough, constraint (16) prevents vehicles from remaining blocked due i ≥ vk i ≥ vmin. ⇒ vk+1 ⇒ vk i (k) ∧ ¬ ζ out i + uaτ, i (6) (7) s⊥ijsijs⊥ijs⊥jisjis⊥jiCijto the minimum speed requirement (17). We will show in the next section that these conditions effectively prevent deadlocks for all future times. 5) Initial conditions: The supervision problem is used in a receding horizon fashion, and we consider that the state of each vehicle of Ntκ at time tκ is known before solving the problem. Therefore, we use the following initial condition for all i ∈ Ntκ: (sκ i , vκ i ) = (si(tκ), vi(tκ)) (18) B. Objective function Any piecewise-constant control verifying constraints (4) to (18) for all k ≥ κ is dynamically admissible and prevents collisions for all future times, and is therefore in U saf e τ,κ (tκ). To remain compatible with human driving, we now formulate an objective function allowing to find a least restrictive and minimally deviating control given a desired control uκ des = i )i∈Ntκ be a set of (uκ strictly positive weights, X be the tuple of all the problem variables, and we define: i,des)i∈Ntκ . In what follows, we let (wκ J κ(X) = wκ i i − uκ i,des . (19) (cid:0)uκ (cid:88) i∈Ntκ (cid:1)2 Noting πuκ the projection operator such that πuκ(X) = uκ, we deduce the following theorem: Theorem 1. The solution of the optimization problem: u∗ = πuκ arg min J κ(X) subj. to ∀k ≥ κ, (4) − (18) X (IH-SP) is a solution to the minimally deviating supervision problem SP ∗ at time tκ, for the norm associated with J κ ◦ πuκ. Note that the weighting terms wκ i allow distinguishing between different types of agents, for instance to prioritize emergency services or high-occupancy vehicles. More com- plex cost functions can also be used, for instance to penalize a forced acceleration more than a forced braking. C. Receding horizon properties We now assume that there exists a solution to IH-SP at time tκ, that the vehicles of Ntκ follow this solution control over [tκ, tκ + τ ], and that the vehicles of Ntκ+1 enter safely with a margin τ. From Definitions 1 and 2, we have the following theorem: Theorem 2 (Recursive feasibility). Let τ > 0, κ ≥ 0, tκ = κτ and tκ+1 = tκ + τ. Assume that: • there exists a solution to IH-SP at time tκ for the vehicles • the vehicles of Ntκ follow this solution control over • the vehicles of Ntκ+1 \Ntκ enter safely with a margin τ. Then there exists a solution to IH-SP at time tκ+1 for the vehicles of Ntκ+1. Proof. From definitions 1 and 2, and using theorem 1, we know that the first two hypotheses guarantee that the vehicles of Ntκ, [tκ, tκ+1], 7 in Ntκ are in a safe state at time tκ+1. Moreover, the third hypothesis ensures that the vehicles in Ntκ+1 also are in a safe state at tκ+1 regardless of the control applied by the vehicles of Ntκ+1 \ Ntκ up to time tκ+1. By definition 1, there exists a feasible solution to IH-SP thus proving the theorem. We now state that the IH-SP formulation effectively prevents the apparition of deadlocksThe proof of this theorem can be found in Appendix A-A. Theorem 3 (Deadlock avoidance). Let κ ≥ 0 and assume that, for all κ ≤ k ≤ κ0, the conditions of theorem 2 remain satisfied at time tk. There exists a feasible solution of IH-SP at time tκ0 in which all the vehicles in Ntκ0 exit the supervision area in finite time. Note that theorem 3 only ensures that, at all times, there exists a solution where all the vehicles inside the supervision at this particular time eventually exit. However, there is no guarantee that such a solution will actually be selected, for instance if one driver wishes to stop although there is no other vehicle. There is also no fairness guarantee, i.e. it is possible that one vehicle is forced to remain stopped for an arbitrarily long time, for instance if there is a very heavy traffic coming from another direction. Future developments will focus devising more complex objectives function to take traffic efficiency and fairness into account. D. Multiple paths choices The above formulation assumes that the path of each vehicle is known in advance. However, this may not be realistic in the context of semi-autonomous cars where drivers can decide to change paths, for instance to avoid an obstacle on the road or use another itinerary. Using additional variables to indicate the path to which a vehicle is assigned, our formulation can be extended to handle multiple possible paths for each vehicle. Due to length limitations, this extension will be detailed in future work. V. FINITE HORIZON FORMULATION In Section IV-C, we presented an infinite horizon formu- lation to solve the minimally deviating supervision problem. However, due to the infinite number of variables, this formu- lation is not suitable for practical resolution. In this section, we derive an equivalent finite horizon formulation that can be implemented and solved using standard numerical techniques. In what follows, we let K ≥ 1 and we denote by FH-SPK the restriction of IH-SP at time tκ to the variables at steps k with κ ≤ k ≤ κ + K, and we only consider the constraints (4) to (18) up to step κ + K. The objective function is unchanged. A solution to FH-SPK at time tκ allows to compute a control preventing collisions up to time tκ + Kτ; however, due to the dynamics of the vehicles, the state reached at tκ + Kτ may not be safe. Since FH-SPK only has a subset of the constraints of IH-SP, we can formulate the following proposition: Proposition 1. Let K ≥ 1 and let X be a solution of IH-SP at step κ. The restriction of X to the first K + 1 time steps is a feasible solution to FH-SPK. Using the global bounds ua, ub, umax and vmax defined in section III-A4, we will now prove a reciprocal implication to proposition 1: if K is chosen large enough, any solution of FH-SPK can be used to construct a solution of IH-SP. As presented in [36], the key idea of the proof lies in the choice of a planning horizon long enough to allow any vehicle to fully stop. The structure of the demonstration is as follows: lemma 1 gives a lower bound on the time horizon to allow a single isolated vehicle to stop using discrete dynamics, although with a potential risk of rear-end collisions from following vehicles. In proposition 2, we give a slightly higher bound on the time horizon ensuring that all vehicles in a line can all safely stop without rear-end collisions. Finally, in proposition 3 we give a bound on K ensuring the recursive feasibility of FH-SPK; this allows formulating theorem 4, stating the equivalence of FH-SPK and IH-SP. In this section, we only present sketches of proofs for each result; detailed demonstrations can be found in Appendix A-B. Lemma 1. At a time tκ, consider a horizon T = Kτ with + τ. Let i ∈ Ntκ be a vehicle for which there exists T ≥ vmax ub a piecewise-constant control (uk i )κ≤k<κ+K such that, for all κ ≤ k < κ+K, uk i ∈ [ui, ui], corresponding to a dynamically feasible trajectory si(t) over [tκ, tκ + T + τ ]. i )κ≤k≤κ+K such that for all There exists a discrete control (uk i , and for which the κ ≤ k ≤ κ + K, uk corresponding dynamically feasible trajectory t (cid:55)→ xi(t) = (si(t), vi(t)) verifies si(tκ + T + τ ) ≤ si(t1 + T ) and vi = 0 over [tκ + T, tκ + T + τ ]. Sketch of proof. vmax is an upper bound on the required time ub for any vehicle to stop by applying a control ub, which by definition is dynamically feasible. The additional τ accounts for the fact that we require uκ i ∈ [ui, ui] and uκ i at the first time step. i = uκ i = uκ 1 + ub (cid:109)(cid:17) (cid:108) umax In the following proposition and noting (cid:100)·(cid:101) the ceiling function, we prove a bound ensuring that a line of vehicles can safely stop before the leader reaches its final computed position at the end of the time horizon, without risk of rear- end collisions: Proposition 2. At a time tκ, suppose that p vehicles of Ntκ (cid:16) (denoted by 1, . . . , p from rear to front) are following one another. Consider a horizon Tstop = Kstopτ ≥ vmax + (p − ub τ + τ, and assume that every vehicle i ∈ 1) i )κ≤k<κ+K such that, {1, . . . , p} has a safe discrete control (uk for all κ ≤ k < κ + K, uk i ∈ [ui, ui]. We let t (cid:55)→ xi(t) be the trajectory over [tκ, tκ + T ] for vehicle i under control (uk i ). For all i ∈ {1, . . . , p}, there exists a safe discrete control i )κ≤k≤κ+K such that for all κ ≤ k ≤ κ + K, uk i ∈ [ui, ui], (uk i and for which the corresponding dynamically feasible uκ i = uκ and safe trajectory t (cid:55)→ xi(t) = (si(t), vi(t)) verifies si(tκ + T + τ ) ≤ si(t1 + T ) and vi = 0 over [tκ + T, tκ + T + τ ]. Sketch of proof. The worst case that needs to be taken into account corresponds to a situation where the initial states of the vehicles require each of them to accelerate in order to avoid a rear-end collision from the vehicle behind. This rather extreme situation happens when a vehicle goes faster than the 8 one it is following, and the two are too close to allow a safe deceleration. In this case, the rearmost vehicle can always brake with the control from lemma 1, until it decelerates below the speed of the vehicle in front of it. The second rearmost vehicle can then decelerate, then the third and up to τ arises from the front-most vehicle. The term the piecewise-constant control hypothesis, and vanishes as τ goes to 0. Note that the condition Kstopτ ≥ vmax + τ ub also provides the same guarantees; depending on the value of p, this second bound might be more efficient. (cid:108) umax + vmax ua (cid:109)(cid:17) (cid:16) ub 1 + Remark 1. The bound from proposition 2 depends on the number of vehicles in a line, and can become quite high when p is large. It can be proven that the condition Kstopτ ≥ vmax + ub + τ also provides the same guarantees; depending on vmax ua the value of p, this second bound might be more efficient. i ua We can now prove the recursive feasibility of FH-SPK for vmin + d (cid:0)s⊥i − sacc (cid:1) and we let Tstop be the a large enough K, as follows: Proposition 3. Consider a time tκ, and assume that at most p vehicles are following one another at all times t ≥ tκ. We set d = maxt≥tκ,i∈Nt stopping horizon from proposition 2 for p vehicles; moreover, we define Trec = Krecτ ≥ Tstop + vmin + τ. We assume that all vehicles of Nt for all t > tκ enter safely with a margin τ. Problem FH-SPKrec is recursively feasible under the hy- potheses of theorem 2, i.e. if there exists a solution to FH- SPKrec at time tκ for the vehicles of Ntκ, there exists a solution at tκ + τ for the vehicles of Ntκ+τ . Sketch of proof. The idea between the choice of Trec is to ensure that each vehicle can either stop safely before entering its acceleration region (without generating rear-end collisions), or has already planned to exit its no-stop region safely. Moreover, the safe entering hypothesis ensures that the entry of new vehicles does not invalidate previously safe solutions, which can therefore be extended. We obtain the equivalence between IH-SP and FH-SPK: Theorem 4. Problems IH-SP and FH-SPK with Kτ ≥ Trec are equivalent, i.e. an optimal solution to one is also an optimal solution to the other. Proof. Proposition 1 ensures that any optimal solution to IH- SP is a feasible solution of FH-SPK. Proposition 3 shows that a solution to FH-SPK (with Kτ ≥ Trec) can be recursively extended to a solution of IH-SP; therefore, the optimal solution of FH-SPK is feasible for IH-SP. Using these two results, we deduce the stated theorem. An important corollary of theorems 1, 3 and 4 is that the control obtained by solving FH-SPK with K large enough is also a solution to the minimally deviating supervision problem, and ensures deadlock avoidance as well. Contrary to IH-SP, FH-SPK is relatively easy to solve with dedicated mixed- integer quadratic programming solvers, as will be demon- strated in the following section. VI. SIMULATION RESULTS A. Simulation environment 9 The presented Supervisor framework has been validated using extensive computer simulations on various test scenar- ios. In the absence of standardized test situations and since no open-sourced implementation of comparable methods [7], [8] is available, this section does not aim at a quantitative comparison with existing algorithms. Since our Supervisor is by design guaranteed to output an optimal1 safe control, the major evaluation criterion is rather its ability to handle a wider variety of traffic scenarios than existing techniques, which is demonstrated in the rest of this section. Due to implementation reasons, the resolution of the su- pervision problem is performed off-line and simulations are run in two successive phases. In the first phase, we define the geometry of the roads inside the supervision area and the corresponding possible paths, and compute the collision and acceleration regions information for each pair of paths. Since these sets only depend on the geometry of vehicles and paths, the corresponding parameters are computed off-line. In the second phase, we run the simulation by coupling the high-fidelity vehicle physics simulator PreScan [39] with an external Python implementation of our supervisor. The actual resolution uses the commercial MIQP solver GUROBI [35]; the Python program runs a coarse simulation over a set time horizon with a fixed time step duration. Vehicles are generated using random Poisson arrivals, with a predefined arrival rate for each possible path, while respecting the safe entering con- dition; the initial velocity of each generated vehicle is chosen randomly according to a truncated Gaussian distribution. At each time step, the finite horizon supervision problem FH-SP is solved for the vehicles inside the supervision area, and yields the best safe control for the set of vehicles. The state of these vehicles at the next time step is then computed according to equations (4) and (5). In parallel, we use PreScan to validate the consistency of this output: from the safe controls computed in the Python supervisor and knowing the reference paths of the vehicles, we compute a target state comprising a desired position, heading and longitudinal velocity for each vehicle. This target state is fed into a low-level controller which outputs a steering and an acceleration or braking control. The vehicle model used in the validation phase takes into account engine response as well as chassis and suspensions dynamics, but does not model road- tire friction. PreScan's collision detection and visualization capacities are then used to validate the absence of collision or dangerous situations. Note that vehicles controllers are designed to ensure a bounded positioning error for any vehicle, relative to their prescribed path and velocity profile. This error is taken into account in the computation of the collision regions, so that the system is robust to control imperfections. B. Test scenarios In the rest of this section, we consider three test scenarios – chosen to represent a wide variety of driving situations – (a) Longitudinal positions; the shaded area is the collision region. (b) Longitudinal velocities Fig. 5. Vehicles positions and velocities in the merging scenario; solid lines correspond to vehicles on the entry lane, dashed lines to vehicles starting on the highway. The thick colored portions show overriding intervals. consisting of merging on a highway, crossing an intersection or driving inside a roundabout. To showcase the performance of our framework in avoiding accidents and deadlocks, we assume that drivers are "oblivious" and focused on tracking a desired speed, regardless of the presence of other vehicles. A video of the presented simulations is available online2. 1) Highway merging: We first consider a very simple highway merging scenario, where an entry lane merges into a single-lane road; the possible paths for the vehicles are the same as in fig. 2c. The collision region between a vehicle i in the entry lane and a vehicle j on the highway have a single connected component given as s⊥ij = s⊥ji = 89 m and sij = sji = 94 m, taking control errors into account. To illustrate the action of the supervisor, we consider a set of six vehicles, three of which are on the highway and three on the entry lane. All vehicles are assumed to have "oblivious" drivers maintaining a constant speed, thus resulting in potential collisions. This admittedly unrealistic behavior has been chosen to generate a higher probability of collisions in absence of supervision. Figure 5 shows the longitudinal trajectories of the supervised vehicles; colored (thick) portions of the lines represent intervals of time during which overriding occurs. The area in gray corresponds to the collision region between entering vehicles and vehicles on the highway; thanks to the action of the supervisor, all collisions are successfully avoided. 2) Intersection crossing: The second scenario is the cross- ing of a + shaped intersection by a total of eight vehicles, with two vehicles per branch. In each branch, the front vehicle goes straight, and the rear vehicle turns left; moreover, all vehicles in front start at the same distance from the center of the intersection, and the same is true for the vehicles in the 1Among the set of piecewise-constant controls with a given time step duration and in the sense of Definition 4 2Available at https://youtu.be/JJZKfHMUeCI 024681012050100024681012050100Time(s)Position(m)0246810120102002468101201020Time(s)Speed(m/s)(a) Longitudinal positions; the shaded area is the collision region. (b) Longitudinal velocities Fig. 6. Vehicles positions and velocities in the intersection crossing scenario; solid lines correspond to vehicles on the entry lane, dashed lines to vehicles starting on the highway. The thick colored portions show overriding intervals. 10 vehicles, a different class of solution is chosen. A video of a longer, one hour simulation is also available online4. 4) Computation time: Due to the relatively short time hori- zon needed to ascertain infinite horizon safety, computation time remains reasonable despite the NP-hardness of the MIQP formulation. Figure 8 shows the evolution of the computation time in the intersection crossing and roundabout scenarios; the limited available space in the merging scenario does not allow enough vehicles for a similar diagram. These measurements have been obtained on a computer equipped with an Intel Core i7-6700K CPU clocked at 4 GHz with 16 GB of RAM, using the GUROBI solver in version 7.0. It can be seen that computation time remains below the duration of a time step in 90% of cases for up to approximately ten simultaneous vehicles, thus allowing real-time computation at 2 Hz. Note that the MIQP problem only loosely depends on the paths geometry, but rather on the average number of conflicts per vehicle which is higher in the case of roundabout driving, thus explaining the longer times reported in fig. 8b. Moreover, the implemented algorithm has been devised for readability over efficiency, and can be optimized by removing redundant variables to further reduce computation time. In practice, this refresh rate means that vehicles could apply a new acceleration every 0.5 s, which is faster than the typical reaction time of one second for a human driver, and should therefore be barely perceived. Note that for practical implementation purposes, the input of the supervisor should be predicted states at the end of the computation period instead of current states; since the acceleration of each vehicle is assumed to be known to the supervisor, these predictions can be easily performed by forward integration. Fig. 7. Illustration of three possible classes of trajectories found by the solver, depending on the initial states of the vehicles. Trajectories 1 and 2 correspond to one vehicle passing the two collision points before the other. Trajectory 3 corresponds to the case where the vehicle on the inner lane enters after the other, and overtakes it inside the roundabout. rear. This scenario illustrates the symmetry-breaking capacities of our framework, which handles this perfectly symmetrical scenario well, as shown in fig. 6. The area in gray corresponds to the collision region between vehicles on different branches. A video of a longer, one hour simulation is available also online3. 3) Roundabout driving: Finally, the third scenario consists of vehicles driving inside a two-lanes roundabout. The par- ticularity of this situation is that collision regions can have multiple connected components, for instance for the paths shown in fig. 2b. Since our formulation explicitly distinguishes each of these connected components, the supervisor is able to choose an ordering for each point of conflict, as illustrated in fig. 7: depending on the initial states and control targets of the VII. DISCUSSION ON IMPLEMENTATION In the previous sections, we presented an optimization-based algorithm for the supervision of semi-autonomous vehicles; we now briefly discuss obstacles and possible solutions for actual implementation. First and foremost, not all vehicles will be equipped with the required communication capacities at the same time; therefore, the ability to deal with unequipped vehicles and other traffic participants is key to envision actual applications. Second, this work assumes perfect communica- tion and control, and in general ignores uncertainties arising from real-world constraints. A. Dealing with unequipped vehicles As with all innovations, the penetration rate of our system would gradually increase overtime, but remain below 100 % for years, yet the formulation proposed in Section IV re- quires all vehicles to be equipped with supervision capacities. Although a detailed study on the integration of unequipped vehicles in our framework is out of the scope of this paper, we present a possible technique to handle these vehicles provided that they can avoid longitudinal collisions with the leading vehicle, and have a bounded reaction time. First, note that it is always possible to consider unequipped vehicles conservatively as proposed in [40]: at a given step k, 3https://youtu.be/cl32nbceZvw 4https://youtu.be/pLoG32wFnkE 0246810121402040608002468101214020406080Time(s)Position(m)02468101214051015200246810121405101520Time(s)Speed(m/s)4060801001202040608010040608010012020406080100Positionvehicle1(m)Positionvehicle2(m)12311 into account by adjusting the lower bound on the longitudinal acceleration of vehicle ie. Note that this approach still guarantees that no collision can happen between an unequipped and an equipped vehicle; moreover, as the penetration rate of equipped vehicles in- creases, additional rules may be enforced to reduce the number of occurrences in which conflicting unequipped vehicles are simultaneously allowed in the conflict region, thus increasing safety even for the unequipped vehicles. Future work will study the impact of penetration rate on safety and efficiency for both equipped and unequipped vehicles. B. Practical implementation (a) Intersection scenario We propose a centralized implementation, where a roadside computer (supervisor) with communication capacities is added to the infrastructure, and is tasked with repeatedly solving FH-SPK. Note that resolution could also be performed using cloud computing, possibly providing much faster computations without necessitating fully dedicated hardware. The supervisor is also assumed to be equipped with a set of sensors (e.g., cameras), so that the arrival of new vehicles in the supervision area can be monitored (in order to account for unequipped vehicles and other traffic participants). Equipped vehicles are supposed to regularly communicate their current state, includ- ing position, velocity and driver's control input, and receive instructions (the safe acceleration sequence (uk i ) solution of FH-SPK) from the roadside supervisor. The vehicle's on-board computer then uses these instructions to override the driver's control inputs when needed. We argue that the main sources of uncertainty, i.e. communication, sensing and control errors, can be taken into account by using safety margins when computing collision regions. Communications are assumed to have similar performance to current 802.11p specifications; we use the figures provided in [43], [44] as reference, with latency below 20 ms, and packet loss probability of less than 30 % under 300 m. To ac- count for network congestion, we use more conservative values than those reported experimentally in [44]. Moreover, using the additional roadside sensors, we estimate that uncertainty in each vehicle's localization could be reduced to below 1 m longitudinally. First, the 20 ms latency corresponds to less than 1 m at highway speed. Second, since they do not require exchanging a lot of data, such messages can be sent much more frequently than the refresh rate of the supervisor. Considering messages can be sent at 20 Hz, the probability of a message not being received in 0.25 s is roughly 0.2 %, and 6 × 10−6 after 0.5 s. Since they receive a whole sequence of safe accelerations, individual vehicles can keep executing this sequence until a new one is successfully received. A worst-case scenario would be having one vehicle using acceleration ua (maximum acceleration) where it should have used ub (maximum brak- ing): after a duration t, the corresponding positioning error is 2 t2(ua +ub), which is roughly 30 cm after 0.25 s and 1.3 m 1 after 0.5 s for typical values of ua and ub of 5 m s−2. More robust contingency protocols could likely be developed, and will be the subject of future work, but these values can be used as safety margins without compromising performance. (b) Roundabout scenario Fig. 8. Distribution of computation times depending on the number of vehicles, for τ = 0.5 s. Shaded areas represent the [0, 90%] percentiles. we compute the minimum and maximum curvilinear position that can be reached at time tk by the unequipped vehicle iu, denoted by sk iu,max respectively. Using the same notations as in Section IV, we then define: iu,min and sk εiuj,p(k) = χ[siuj,p,+∞[(sk ε⊥iuj,p(k) = χ[s⊥ iuj,p,+∞[(sk iu,max), iu,min). (20) (21) Therefore, the unequipped vehicle is considered as occupying the conflict region at step k when there exists a control (maximum acceleration) for which it could be inside this region at step k. Similarly, the vehicle is only considered as liberating the conflict region when, even by applying a maximum braking, it would exit it. The collision avoidance constraints (9) and (10) are also modified to use sk+1 iu,min and vk+1 iu,min, where vk+1 iu,min is the minimum speed reachable by iu at k + 1. Other traffic participants such as cyclists (and, to a lesser extent, pedestrians) could also be taken into account in this fashion. Recently proposed "non-conservatively defensive strategies" [41] could also be applied. A limitation of this simple approach is that it can lead equipped vehicles to often yield right-of-way to unequipped vehicles, which may be problematic and can slow the accep- tance of the system. A possible method (introduced in [42]) to reduce this problem while improving the global level of safety is to use the existing equipped vehicles to force the unequipped ones to stop when required. Suppose that an unequipped vehicle (denoted by iu) follows an equipped one (ie), both crossing the path of another equipped vehicle je. By setting πieje = 0 (thus requiring je to pass before ie), we effectively force the unequipped vehicle iu to also pass after je; the reaction time of the unequipped vehicle can be taken 510152000.511.52NumberofvehiclesComputationtime(s)Min.Medianp90[0,90%]2468101214161800.511.52NumberofvehiclesComputationtime(s)Min.Medianp90[0,90%]Similarly, positioning and control uncertainty can be ac- counted for as margins in the collision regions, provided they can be bounded. In this work, we assume that vehicle self- positioning can be improved using the roadside sensors from the supervisor (which can be precisely calibrated), which could provide relatively tight bounds on error. VIII. CONCLUSION In this article, we designed a framework allowing safe semi- autonomous driving of multiple cooperative vehicles in various traffic situations. We first introduce a set of linear constraints ensuring infinite horizon safety for a group of human-driven vehicles, traveling inside predefined corridors with the help of existing lane-keeping technologies. Based on this set of constraints, a discrete-time Supervisor is allowed to override the drivers' longitudinal control inputs if they would lead the vehicles into an inevitable collision state. In this case, the control used for overriding is chosen as close as possible to the one originally requested by the drivers. These two properties ensure that intervention only occurs when strictly necessary to maintain safety, thus facilitating the acceptation of the system by human drivers. Theoretical considerations prove this supervisor guarantees both safety and deadlock avoidance, and can be applied with- out distinction to multiple situations such as traffic intersec- tion, highway entry lanes or roundabouts. Using the realistic vehicle physics simulator PreScan, we demonstrated that our algorithm can handle complex situations over an arbitrary duration, with continuous arrivals of vehicles. Moreover, the proposed formulation can be solved in real-time on a standard desktop computer for up to ten vehicles, which makes it suitable for practical applications. Additionally, this work opens up many perspectives for future research. First and foremost, the current framework does not deal with non-equipped vehicles or other traffic participants such as cyclists or pedestrians, nor does it take sensor and communication uncertainties into account. Before considering an actual implementation, the system should be more robust to these various sources of noise. Moreover, our formulation has been designed in a mostly centralized fashion; various approaches need to be explored to design a more realistic decentralized system, that could be implemented in actual cars. APPENDIX A DEMONSTRATIONS A. Proofs for Section IV Before proving theorem 3, we introduce the following lemma stemming from graph theory: Lemma 2. Let G = (V, E) a directed graph with vertices set V and edges set E. All cycles in G can be removed by reversing a set of edges, each of them contributing to at least one cycle. Proof. The proof is based on the existence of minimum feedback arc sets [45], i.e. a minimum set Ef eedback ⊂ E such that G(cid:48) = (V, E \ Ef eedback) is acyclic. By minimality 12 of Ef eedback, any e ∈ Ef eedback belongs to at least one cycle of G. Moreover, it can be seen that reversing the edges of Ef eedback also leads to an acyclic graph, thus proving the lemma. Proof of Theorem 3. Note that the only constraints requiring a vehicle to stop are (8) to (10), forcing a vehicle j to wait for a vehicle i with πp ij = 1. From the hypotheses and theorem 2, there exists a solution X to IH-SP at time tκ0. We define a directed priority graph GX = (V, E) with V = Ntκ0 and where an edge i → j belongs to E if there exists p such that πp ij = 1. Using this representation, a cycle in GX corresponds to a chain of q conflicting vehicles i1, i2, . . . , iq, iq+1 = i1 for which there exists a connected component Cpn inin+1 such that πpn inin+1 = 1 for all n = 1 . . . q. If GX is acyclic, it defines a (partial) topological order, and it is always possible to admit the vehicles one by one in that order. Therefore, there exists a feasible solution where all the vehicles of Ntκ0 exit the supervision area in finite time. We now assume that there exists at least one cycle in GX. If all the vehicles involved in the cycle can exit in finite time, the result of the theorem is proven. Otherwise, we note Ndead ⊂ Ntκ0 a set of vehicles corresponding to a cycle in GX: all of these vehicles are stopped at infinity, and are prevented to move further by a constraint of form (8), for a certain j ∈ Ndead. Moreover, the no-stop condition (17) ensures that, for all i ∈ Ndead and all k ≥ κ, sk From lemma 2, we know that it is possible to change the values of the variables πp ij for i, j ∈ Ndead to render GX acyclic. Using the fact that sk for all of these vehicles, we know that modifying these priorities does not violate constraints (8) to (10). Therefore, we can build a solution X(cid:48) for which the corresponding priority graph is acyclic, which proves the theorem. i ≤ s⊥i . i ≤ s⊥i is trivial B. Proofs for Section V Proof of Lemma 1. The proof if we consider continuous-time dynamics, as a single vehicle can always apply a control lower or equal to ub starting from time tk + τ, which ensures it is stopped for t ≥ tk + τ + vmax . A slight additional complexity happens at the final braking time step when considering piecewise-constant controls, applying ub for a duration τ might result in a negative velocity, which is not allowed in our framework. We now proceed to the formal proof, as below. ub Let (uk let us define a control (wk for κ < k < κ + K, and wκ+K (uk i ) be the control corresponding to trajectory si, and i , wk i = min(ub, uk i ) = ub. We construct i = i is the speed of vehicle i and, for k ≥ κ + 1, uk , where vk i ) as: wκ i = uκ i i = uk i τ ≥ 0 (cid:40) i + wk if vk otherwise i ) iteratively as uκ wk i − vk As vκ+1 = vκ+1 i τ i at time tk under control (uk i ). i i ≤ vmax ≤ (K − 1)τub from the hypothesis, there exists a minimal value of k0 ≥ κ such that vk0 i ≤ ubτ; moreover, the condition on K ensures that vκ+1 i − (K − 2)ub∆t ≤ ub∆t, and so k0 ≤ (κ + 1) + (K − 2) = κ + K − 1. 13 0 i i i i i j ≤ uk i ≤ uk We now let i ≥ 2 and assume that every vehicle j ∈ j ). We note {1, . . . , i− 1} follows its corresponding control (uk j the position and speed of vehicle j at step k under j and vk sk this control. Since uk j for these vehicles, we deduce from the monotony of the system that the original control solution for vehicle i (uk i )κ≤k<κ+K prevents rear-end collisions if vehicle i − 1 applies (uk i−1). Therefore, any dynamically feasible extension of (uk i ) is safe over [κτ, (κ + K + 1)τ [. As a result, the set U saf e i , [κ, κ + K]) of all admissible (uκ controls (uk i and i )κ≤k≤κ+K for vehicle i such that uκ i for κ ≤ k < κ + K is not empty. We note i ≤ uk uk i ) a minimum element of this set (and so uk i ); we (uk will prove that vκ+Ki If for all k ≥ κ, vk i−1, we conclude that vehicle i stops before vehicle i−1 which proves the result from the induction hypothesis. Otherwise, we let ki 0 ≥ κ be the minimum k such i−1, and we know that vki 0−1 that vk i−1 + umaxτ. i ≥ vk 0, we know from the monotony of the system For all k ≥ ki that the control min(uk i , ub) prevents rear-end collisions; i−1, uk i ≤ vk−1 we deduce that, for k ≥ ki i−1 + umaxτ. Therefore, 0, vk (cid:109) vκ+Ki−1+1 i−1 + umaxτ and we deduce from the induc- tion hypothesis that vκ+Ki−1+1 ≤ umaxτ + ubτ. Therefore, we obtain the recursion relation Ki = Ki−1 + 1 + which yields the announced result. (cid:108) umax i ≤ vki ≤ vKi−1 ≤ ubτ. i ≤ uk i ≤ vk i = uκ ub (cid:109) + vmax ua (cid:108) umax i , we deduce that sκ+K ub i ≤ uk which proves the proposition. Finally, we conclude that vehicle i can fully stop (without rear-end collisions) at step κ + Ki + 1; therefore the set of p vehicles can safely stop before the beginning of step κ + K if K ≥ Kp = vmax + 1. Since the recursion ubτ + (p − 1) ensures that for all i and k, uk ≤ sκ+K i Note that the time needed for vehicles to match speeds can also be bounded by vmax regardless of the value of p. ua Therefore, all vehicles can also fully stop within the time horizon if T = Kτ ≥ vmax + 2τ. Depending on ub the value of p, this bound may be better than the previously demonstrated one. Proof of Proposition 3. We consider a time tκ = κτ, and we let Trec = Krecτ ≥ Tstop + vmin + τ. Consider a solution X of FH-SPKrec for the vehicles of Ntκ, defined for steps κ ≤ k ≤ κ + K. We will first show that this solution can be extended to a solution of FH-SPK+1 for the vehicles in Ntκ. Note that the only constraints which can be unfeasible are the safety constraints (8)-(10) and the minimum velocity constraints (16) and (17). Consider a vehicle i ∈ Ntκ: using the control corresponding to this solution, two cases can arise: , in which case proposition 2 ensures that i and all the vehicles behind it can fully stop before , and can remain stopped up to step Krec+1. reaching sacc Since we also require that sacc if j follows i, this ensures that keeping i and its followers stopped satisfies all the above constraints; • si(Tstop) ≤ sacc j ≥ sacc + d vmin ua i i i • otherwise, si(Trec) ≥ s⊥i , in which case the crossing and minimum velocity constraints (8), (16) and (17) are satisfied for all conflicting vehicle j up to step Krec. The requirement Trec ≥ Tstop and proposition 2 ensure that the safe following constraints (9) and (10) involving Fig. 9. spond to integer multiples of the time step. Illustration of the "overshoot" phenomenon; the vertical grid corre- i = 0. Since uk From the definition of (uk k ≤ κ + K, vk we also know that sk0 the minimal admissible control starting from xk0 i sk0+1 i i ), we know that for all k0 + 1 ≤ i for κ ≤ k ≤ k0 − 1, i ≤ vk0 is ; therefore, which proves the above lemma. . Finally, uk0 i i ≤ uk and vk0 i ≤ sk0 = sκ+K i i i ≤ sk0+1 i Proof of Proposition 2. We first consider the continuous-time case to give an intuition of the proof. We build upon the fact that the rearmost vehicle in a line can always brake with acceleration ub until it fully stops. However, some the initial conditions may require vehicles in front to accelerate in order to avoid collisions, for instance if the rearmost vehicle is too fast. However, even in this case, we know that the second rearmost vehicle can brake with ub as soon as it has matched the speed of the rearmost vehicle, and by induction this is true for all the vehicles in the line. In the continuous-time case, all of these vehicles can therefore stop in a time bounded by vmax . ub When considering piecewise-continuous controls, an addi- tional complexity arises from the fact that vehicles may match speed between two time steps, resulting in an "overshoot" in velocity. We use the hypotheses on the acceleration to bound this overshoot, as illustrated in fig. 9: we consider a fast vehicle (noted 1, blue curve) following a slower vehicle (noted 2, red curve). To avoid collisions, vehicle 2 is required to accelerate; vehicles match speed at time t2→1; however, due to the time discretization, the overshoot phenomenon can occur. Using the bounds on the acceleration, noting k2→1 the time step immediately following t2→1, we know that 1 + τ (umax + ub). Moreover, after step k, vehicle 2 ≤ vk vk 2 can brake with a control ub up to time t1, corresponding to the first integer time step k1 when vk1 1 ≤ τub. Since both vehicles 1 and 2 have the same acceleration of [t2→1, t1], we know that vk1 2 ≤ τub + τ (ub + umax). Therefore, noting 2 ≤ τub. The k2 = k1 + 1 + same reasoning can then be repeated for the vehicles preceding vehicle 2. We will now formalize this recursion, as below. , we know that vk2 (cid:108) umax (cid:109) ub We will prove by induction that, for i ∈ {1, . . . , p}, there exists a dynamically feasible control (uk i and (cid:16) i for k ≥ κ such that the corresponding vehicle speed i ≤ uk uk i ) verifies vκ+Ki (vk + (i − τ. First, for the rearmost vehicle i = 1, the ≤ ubτ with Kiτ = (cid:108) umax (cid:108) vmax 1) proof of lemma 1 provides the result with uk i ) with uκ (cid:109)(cid:17) i = uκ (cid:109) ub ub 1 + i . i = uk i t1≤umaxτ≤ubτt2→1≤τ(umax+ub)t1timespeedi d vehicle i remain satisfiable for the vehicles of Ntκ up to step Krec + 1. Indeed, if s⊥i ≥ si(Tstop) > sacc , condition (16) ensures that vehicle i accelerates at least with acceleration ua until reaching speed vmin, which takes at most a time vmin . The vehicle is ua then required to maintain speed vmin until reaching s⊥i , which . Therefore, vehicle i necessarily takes at most a time reaches s⊥i by time Trec; the additional τ accounts for vehicles reaching doing so between two time steps. vmin The above considerations ensure that the solution of FH- SPK at time tκ can be prolongated to a solution of FH-SPK+1 for the vehicles of Ntκ. Finally, noting that the safe entry hypothesis ensures that this solution remains safe even when taking the vehicles of Ntκ+τ \ Ntκ into consideration. By definition, there also exists a safe control (and therefore a solution to FH-SPK) for these vehicles at time tκ + τ. As a result, there exists a solution to FH-SPK at time tκ + τ for the vehicles of Ntκ+τ which proves the stated result. REFERENCES [1] A. Vahidi and A. Eskandarian, "Research advances in intelligent col- lision avoidance and adaptive cruise control," IEEE Transactions on Intelligent Transportation Systems, vol. 4, no. 3, pp. 143–153, sep 2003. [2] E. Coelingh, A. Eidehall, and M. Bengtsson, "Collision Warning with Full Auto Brake and Pedestrian Detection - a practical example of Automatic Emergency Braking," in 13th International IEEE Conference on Intelligent Transportation Systems. IEEE, sep 2010, pp. 155–160. [3] D. Geronimo, A. M. Lopez, A. D. Sappa, and T. Graf, "Survey of Pedestrian Detection for Advanced Driver Assistance Systems," IEEE Transactions on Pattern Analysis and Machine Intelligence, no. 7, pp. 1239–1258, jul. [4] M. Diaz-Cabrera, P. Cerri, and P. Medici, "Robust real-time traffic light detection and distance estimation using a single camera," Expert Systems with Applications, vol. 42, no. 8, pp. 3911–3923, may 2015. [5] D. Kim, J. Choi, H. Yoo, U. Yang, and K. Sohn, "Rear obstacle detection system with fisheye stereo camera using HCT," Expert Systems with Applications, vol. 42, no. 17-18, pp. 6295–6305, oct 2015. [6] Z. Kim, "Robust Lane Detection and Tracking in Challenging Scenar- ios," IEEE Transactions on Intelligent Transportation Systems, vol. 9, no. 1, pp. 16–26, mar 2008. [7] A. Colombo and D. Del Vecchio, "Least Restrictive Supervisors for Intersection Collision Avoidance: A Scheduling Approach," IEEE Trans- actions on Automatic Control, vol. 60, no. 6, pp. 1515–1527, jun 2014. [8] G. R. Campos, F. Della Rossa, and A. Colombo, "Optimal and least restrictive supervisory control: Safety verification methods for human- driven vehicles at traffic intersections," in 2015 54th IEEE Conference on Decision and Control. IEEE, dec 2015, pp. 1707–1712. [9] T. Fraichard and H. Asama, "Inevitable Collision States. A Step Towards Safer Robots?" Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems, vol. 1, no. October, pp. 388–393, 2003. [10] R. Naumann, R. Rasche, and J. Tacken, "Managing autonomous vehi- cles at intersections," IEEE Intelligent Systems and their Applications, vol. 13, no. 3, pp. 82–86, 1998. [11] K. Dresner and P. Stone, "Multiagent traffic management: a reservation- based intersection control mechanism," in Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems, 2004. IEEE Computer Society, July 2004, pp. 530–537. [12] L. Makarem and D. Gillet, "Fluent coordination of autonomous vehi- cles at intersections," in Conference Proceedings - IEEE International Conference on Systems, Man and Cybernetics, vol. 18, no. 1. IEEE, oct 2012, pp. 2557–2562. [13] X. Qian, J. Gregoire, A. de La Fortelle, and F. Moutarde, "Decentralized model predictive control for smooth coordination of automated vehicles at intersection," in 2015 European Control Conference (ECC). IEEE, jul 2015, pp. 3452–3458. [14] K. Mu, F. Hui, and X. Zhao, "Merging Driver Assistance Decision Sys- tem Using Occupancy Grid-Based Traffic Situation Representation," in 2015 IEEE 18th International Conference on Intelligent Transportation Systems. IEEE, sep 2015, pp. 231–237. 14 [15] I. A. Ntousakis, I. K. Nikolos, and M. Papageorgiou, "Optimal vehicle trajectory planning in the context of cooperative merging on highways," Transportation Research Part C: Emerging Technologies, vol. 71, pp. 464–488, oct 2016. [16] A. Mosebach, S. Rochner, and J. Lunze, "Merging control of cooperative vehicles," IFAC-PapersOnLine, vol. 49, no. 11, pp. 168–174, 2016. [17] V. Desaraju, H. Ro, M. Yang, E. Tay, S. Roth, and D. Del Vecchio, "Partial order techniques for vehicle collision avoidance: Application to an autonomous roundabout test-bed," in 2009 IEEE International Conference on Robotics and Automation. IEEE, may 2009, pp. 82–87. [18] R. Vasudevan, V. Shia, Yiqi Gao, R. Cervera-Navarro, R. Bajcsy, and F. Borrelli, "Safe semi-autonomous control with enhanced driver modeling," in 2012 American Control Conference. IEEE, jun 2012, pp. 2896–2903. [19] A. Gray, Y. Gao, J. K. Hedrick, and F. Borrelli, "Robust Predictive Control for semi-autonomous vehicles with an uncertain driver model," in 2013 IEEE Intelligent Vehicles Symposium, no. 1239323. IEEE, jun 2013, pp. 208–213. [20] A. Gray, Y. Gao, T. Lin, J. K. Hedrick, and F. Borrelli, "Stochastic predictive control for semi-autonomous vehicles with an uncertain driver model," in 16th International IEEE Conference on Intelligent Transportation Systems. IEEE, oct 2013, pp. 2329–2334. [21] C. Liu, A. Gray, C. Lee, J. K. Hedrick, and J. Pan, "Nonlinear stochastic predictive control with unscented transformation for semi-autonomous vehicles," in 2014 American Control Conference. IEEE, jun 2014, pp. 5574–5579. [22] T.-C. Au, S. Zhang, and P. Stone, "Autonomous Intersection Manage- ment for Semi-Autonomous Vehicles," Handbook of Transportation, pp. 88–104, 2015. [23] S. A. Reveliotis and E. Roszkowska, "On the Complexity of Maximally Permissive Deadlock Avoidance in Multi-Vehicle Traffic Systems," IEEE Transactions on Automatic Control, vol. 55, no. 7, pp. 1646–1651, jul 2010. [24] L. Bruni, A. Colombo, and D. Del Vecchio, "Robust multi-agent collision avoidance through scheduling," in 52nd IEEE Conference on Decision and Control. IEEE, dec 2013, pp. 3944–3950. [25] H. Ahn and D. Del Vecchio, "Semi-autonomous intersection collision avoidance through job-shop scheduling," in Proceedings of the 19th International Conference on Hybrid Systems: Computation and Control. ACM, 2016, pp. 185–194. [26] J. Gregoire, "Priority-based coordination of mobile robots," Ph.D. [Online]. Available: http: dissertation, MINES ParisTech, 2014. //arxiv.org/abs/1410.0879 [27] K. D. Kim and P. R. Kumar, "An MPC-based approach to provable system-wide safety and liveness of autonomous ground traffic," IEEE Transactions on Automatic Control, vol. 59, no. 12, pp. 3341–3356, dec 2015. [28] T. Sim´eon, S. Leroy, and J. P. Laumond, "Path coordination for multiple mobile robots: A resolution-complete algorithm," IEEE Transactions on Robotics and Automation, vol. 18, no. 1, pp. 42–49, 2002. [29] J. Peng, "Coordinating Multiple Robots with Kinodynamic Constraints Along Specified Paths," The International Journal of Robotics Research, vol. 24, no. 4, pp. 295–310, apr 2005. [30] E. R. Muller, R. C. Carlson, and W. K. Junior, "Intersection control for automated vehicles with MILP," IFAC-PapersOnLine, vol. 49, no. 3, pp. 37–42, 2016. [31] J. Park, S. Karumanchi, and K. Iagnemma, "Homotopy-Based Divide- and-Conquer Strategy for Optimal Trajectory Planning via Mixed- Integer Programming," IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1101–1115, 2015. [32] F. Borrelli, D. Subramanian, A. Raghunathan, and L. Biegler, "MILP and NLP Techniques for centralized trajectory planning of multiple unmanned air vehicles," in 2006 American Control Conference. IEEE, 2006, pp. 5763–5768. [33] S. Cafieri and N. Durand, "Aircraft deconfliction with speed regulation: New models from mixed-integer optimization," Journal of Global Op- timization, vol. 58, no. 4, pp. 613–629, 2014. [34] N. Murgovski, G. R. de Campos, and J. Sjoberg, "Convex modeling of conflict resolution at traffic intersections," in 2015 54th IEEE Conference on Decision and Control (CDC). IEEE, dec 2015, pp. 4708–4713. [35] Gurobi Optimization, Inc., "Gurobi optimizer reference manual," 2015. [Online]. Available: http://www.gurobi.com [36] F. Altch´e, X. Qian, and A. De La Fortelle, "Least Restrictive and Minimally Deviating Supervisor for Safe Semi-Autonomous Driving at an Intersection: An MIQP Approach," in 2016 IEEE 19th International Conference on Intelligent Transportation Systems, Nov. 2016. [37] P. Polack, F. Altch´e, B. d'Andr´ea-Novel, and A. de La Fortelle, "The Kinematic Bicycle Model : a Consistent Model for Planning Feasible Trajectories for Autonomous Vehicles ?" 2017 IEEE Intelligent Vehicles Symposium (IV), no. IV, 2017. [38] F. Altch´e, X. Qian, and A. De La Fortelle, "Time-optimal Coordination of Mobile Robots along Specified Paths," in 2016 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems, Oct. 2016. [39] TASS International. [Online]. Available: http://www.tassinternational. com/prescan [40] G. Schildbach, M. Soppert, and F. Borrelli, "A collision avoidance system at intersections using Robust Model Predictive Control," in 2016 IEEE Intelligent Vehicles Symposium (IV). IEEE, jun 2016, pp. 233– 238. [41] W. Zhan, C. Liu, C.-y. Chan, and M. Tomizuka, "A non-conservatively defensive strategy for urban autonomous driving," in 2016 IEEE 19th International Conference on Intelligent Transportation Systems (ITSC). IEEE, nov, pp. 459–464. [42] X. Qian, J. Gregoire, F. Moutarde, and A. De La Fortelle, "Priority-based coordination of autonomous and legacy vehicles at intersection," in 17th International IEEE Conference on Intelligent Transportation Systems (ITSC). IEEE, pp. 1166–1171. [43] D. Jiang, V. Taliwal, A. Meier, W. Holfelder, and R. Herrtwich, "Design of 5.9 ghz dsrc-based vehicular safety communication," IEEE Wireless Communications, no. 5, pp. 36–43, oct. [44] S. Demmel, A. Lambert, D. Gruyer, A. Rakotonirainy, and E. Monacelli, "Empirical IEEE 802.11p performance evaluation on test tracks," in 2012 IEEE Intelligent Vehicles Symposium. IEEE, jun, pp. 837–842. [45] D. Younger, "Minimum feedback arc sets for a directed graph," IEEE Transactions on Circuit Theory, vol. 10, no. 2, pp. 238–245, Jun 1963. 15 Arnaud de La Fortelle (M10) received the M.S. degree in engineering from ´Ecole Polytechnique and ´Ecole des Ponts et Chausses, Paris, France, in 1994 and 1997 respectively and the Ph.D. degree from ´Ecole des Ponts et Chausses in 2000, with a spe- cialization in applied mathematics. He is professor at MINES ParisTech, Paris, France. From 2003 to 2005, he investigated communica- tions for cooperative systems and the architecture required in distributed systems at INRIA, participat- ing in the CyberCars project. Since 2006, he has been the director of the Joint Research Unit LaRA (La Route Automatis´ee) of INRIA and MINES ParisTech and, since 2008, has also served as the Director of the Center of Robotics in MINES ParisTech. His research inter- ests include cooperative systems (communication, data distribution, control, and mathematical certification) and their applications (Autonomous vehicles, collective taxis...). He coordinates the international research chair Drive for All (with partners UC Berkeley, Shanghai JiaoTong University and EPFL). Dr. de La Fortelle has been elected to the Board of Governors of IEEE Intelligent Transportation System Society in 2009 and is member of the Board of the French Automotive Engineers Society. He is President of the French ANR evaluation committee for sustainable transport and mobility since 2015 and served as expert in European H2020 program. Florent Altch´e received the M.S. degree in engi- neering from French ´Ecole Polytechnique in 2014 and a Specialized Master degree of public policies from ´Ecole des Ponts ParisTech in 2015. He is currently working towards a Ph.D. degree at the Centre for Robotics at Mines ParisTech. His research interests include coordination and cooperation of autonomous vehicles or robots, as well as planning in an uncertain environment. Xiangjun Qian received the B.S. degree in com- puter science from Shanghai Jiao Tong University, Shanghai, China in 2010, the Dip-Ing degree from MINES ParisTech, Paris, France, in 2012, and re- ceived his Ph.D. degree in robotics and automation at MINES ParisTech in 2016. His main research interest lies in the control and coordination of au- tonomous vehicles. He is also interested in the application of machine-learning techniques on the analysis of large-scale transportation networks.
1901.08492	1	1901	2019-01-24T16:44:16	Feudal Multi-Agent Hierarchies for Cooperative Reinforcement Learning	[ "cs.MA", "cs.AI", "cs.LG" ]	We investigate how reinforcement learning agents can learn to cooperate. Drawing inspiration from human societies, in which successful coordination of many individuals is often facilitated by hierarchical organisation, we introduce Feudal Multi-agent Hierarchies (FMH). In this framework, a 'manager' agent, which is tasked with maximising the environmentally-determined reward function, learns to communicate subgoals to multiple, simultaneously-operating, 'worker' agents. Workers, which are rewarded for achieving managerial subgoals, take concurrent actions in the world. We outline the structure of FMH and demonstrate its potential for decentralised learning and control. We find that, given an adequate set of subgoals from which to choose, FMH performs, and particularly scales, substantially better than cooperative approaches that use a shared reward function.	cs.MA	cs	Feudal Multi-Agent Hierarchies for Cooperative Reinforcement Learning Sanjeevan Ahilan 1 Peter Dayan 2 9 1 0 2 n a J 4 2 ] A M . s c [ 1 v 2 9 4 8 0 . 1 0 9 1 : v i X r a Abstract We investigate how reinforcement learning agents can learn to cooperate. Drawing inspiration from human societies, in which successful coordination of many individuals is often facilitated by hierar- chical organisation, we introduce Feudal Multi- agent Hierarchies (FMH). In this framework, a 'manager' agent, which is tasked with maximis- ing the environmentally-determined reward func- tion, learns to communicate subgoals to multiple, simultaneously-operating, 'worker' agents. Work- ers, which are rewarded for achieving managerial subgoals, take concurrent actions in the world. We outline the structure of FMH and demonstrate its potential for decentralised learning and control. We find that, given an adequate set of subgoals from which to choose, FMH performs, and partic- ularly scales, substantially better than cooperative approaches that use a shared reward function. reinforcement 1. Introduction In cooperative multi-agent learning, simultaneously-acting agents must learn to work together to achieve a shared set of goals. A straightforward approach is for each agent to optimise a global objective using single-agent reinforcement learning (RL) methods such as Q-Learning (Watkins & Dayan, 1992) or policy gradients (Williams, 1992; Sutton et al., 2000). Unfortunately this suffers various problems in general. First, from the perspective of any one agent, the environ- ment is non-stationary. This is because as other agents learn, their policies change, creating a partially unobservable influ- ence on the effect of the first agent's actions. This issue has recently been addressed by a variety of deep RL methods, for instance with centralised training of ultimately decen- 1Gatsby Computational Neuroscience Unit, University College London, London, United Kingdom 2Max Planck Institute for Bio- logical Cybernetics, 72076 Tubingen, Germany. Correspondence to: Sanjeevan Ahilan <[email protected]>. tralised policies (Lowe et al., 2017; Foerster et al., 2018). However, this requires that the centralised critic or critics have access to the actions and observations of all agents during training, which may not always be possible. A second problem is coordination; agents must learn how to choose actions coherently, even in environments in which many optimal equilibria exist (Matignon et al., 2012). Whilst particularly challenging when agents are completely independent, such problems can be made more feasible if a form of communication is allowed (Vlassis, 2007). Nev- ertheless, it is difficult for agents to learn how to communi- cate relevant information effectively to solve coordination problems, with most approaches relying on further helpful properties such as a differentiable communication channel (Sukhbaatar et al., 2016; Foerster et al., 2016) and/or a model of the world's dynamics (Mordatch & Abbeel, 2018). Third, multi-agent methods scale poorly -- the effective state space grows exponentially with the number of agents. Learn- ing a centralised value function therefore suffers the curse of dimensionality, whilst the alternative of decentralised learn- ing often appears inadequate for addressing non-stationarity. Optimising a global objective also becomes challenging at scale, as it becomes difficult to assign credit to each agent (Wolpert & Tumer, 2002; Chang et al., 2004). A clue for meeting this myriad of challenges may lie in the way in which human and animal societies are hierarchically structured. In particular, even in broadly cooperative groups, it is frequently the case that different individuals agree to be assigned different objectives which they work towards for the benefit of the collective. For example, members of a company typically have different roles and responsibili- ties. They will likely report to managers who define their objectives, and they may in turn be able to set objectives to more junior members. At the highest level, the CEO is responsible for the company's overall performance. Inspired by this idea, we propose an approach for the multi- agent domain, which organises many, simultaneously acting agents into a managerial hierarchy. Whilst typically all agents in a cooperative task seek to optimise a shared re- ward, in Feudal Multi-agent Hierarchies (FMH) we instead only expose the highest-level manager to this 'task' reward. Feudal Multi-Agent Hierarchies The manager must therefore learn to solve the principal- agent problem (Jensen & Meckling, 1976) of communicat- ing subgoals, which define a reward function, to the worker agents under its control. Workers learn to satisfy these sub- goals by taking actions in the world and/or by setting their own subgoals for workers immediately below them in the hierarchy. FMH allows for a diversity of rewards. This can provide individual agents with a rich learning signal, but necessarily implies that interactions between agents will not in general be fully cooperative. However, the intent of our design is to achieve collective behaviours which are apparently cooperative, from the perspective of an outside observer viewing performance on the task objective. Our idea is a development of a single-agent method for hier- archical RL known as feudal RL (Dayan & Hinton, 1993; Vezhnevets et al., 2017), which involves a 'manager' agent defining subgoals for a 'worker' agent in order to achieve temporal abstraction. Feudal RL allows for the division of tasks into a series of subtasks, but has largely been investi- gated with only one worker per manager acting at any one time. By introducing a feudal hierarchy with multiple agents acting simultaneously in FMH, we not only divide tasks over time but also across worker agents. Furthermore, we em- brace the full multi-agent setting, in which observations are not in general shared across agents. We outline a method to implement FMH for concurrently acting, communicating agents trained in a decentralised fash- ion. Our approach pre-specifies appropriate positions in the hierarchy as well as a mapping from the manager's choice of communication to the workers' reward functions. We show how to facilitate learning of our deep RL agents within FMH through suitable pretraining and repeated communication to encourage temporally extended goal-setting. We conduct a range of experiments that highlight the advan- tages of FMH. In particular, we show its ability to address non-stationarity during training, as managerial reward ren- ders the behaviour of workers more predictable. We also demonstrate FMH's ability to scale to many agents and many possible goals, as well as to enable effective coordina- tion amongst workers. It performs substantially better than both decentralised and centralised approaches. 2. Background 2.1. Markov Decision Processes Single-agent RL can be formalised in terms of Markov Decision Processes, which consist of sets of states S and actions A, a reward function r : S×A → R and a transition function T : S × A → ∆(S), where ∆(S) denotes the set of discrete probability distributions over S. Agents act future reward, defined as Eπ[(cid:80)∞ according to a stochastic policy π : S × A → [0, 1]. One popular objective is to maximise the discounted expected t=0 γtr(st, at)] where the expectation is over the sequence of states and actions which result from policy π, starting from an initial state distribution ρ0 : S → [0, 1]. Here γ ∈ [0, 1) is a discount factor and t is the time step. This objective can be equivalently expressed as Es∼ρπ,a∼π[r(s, a)], where ρπ is the discounted state distribution induced by policy π starting from ρ0. 2.2. Deterministic Policy Gradient Algorithms Deterministic policy gradients (DPG) is a frequently used single-agent algorithm for continuous control using model- free RL (Silver et al., 2014). It uses deterministic policies µθ : S → A, whose parameters θ are adjusted in an off- policy fashion using an exploratory behavioural policy to perform stochastic gradient ascent on an objective J(θ) = Es∼ρµ,a∼µθ [r(s, a)]. We can write the gradient of J(θ) as: ∇θJ(θ) = Es∼ρµ[∇θµθ(s)∇aQµ(s, a)a=µθ(s)]. (1) Deep Deterministic Policy Gradients (DDPG) (Lillicrap et al., 2015) is a variant with policy µ and critic Qµ being represented via deep neural networks. Like DQN (Mnih et al., 2015), DDPG stores experienced transitions in a re- play buffer and replays them during training. 2.3. Markov Games A partially observable Markov game (POMG) (Littman, 1994; Hu et al., 1998) for N agents is defined by a set of states S, and sets of actions A1, ...,AN and observations O1, ...,ON for each agent. In general, the stochastic policy of agent i may depend on the set of action-observation histories Hi ≡ (Oi × Ai)∗ such that πi : Hi × Ai → [0, 1]. In this work we restrict ourselves to history-independent stochastic policies πi : Oi × Ai → [0, 1]. The next state is generated according to the state transition function T : S × A1 × ... × An → ∆(S). Each agent i obtains deterministic rewards defined as ri : S ×A1× ...×An → R and receives a private observation oi : S → Oi. There is an initial state distribution ρ0 : S → [0, 1] and each agent aims to maximise its own discounted sum of future rewards. 2.4. Centralised and Decentralised Training In multi-agent RL, agents can be trained in a centralised or decentralised fashion. In decentralised training, agents have access only to local information: their own action and observation histories during training (Tan, 1993). Agents are often trained using single-agent methods for RL, such as Q-Learning or DDPG. In centralised training, the action and observation histories Feudal Multi-Agent Hierarchies of all agents are used, effectively reducing the multi-agent problem to a single-agent problem. Although it may appear restrictive for agents to require access to this full informa- tion, this approach has generated recent interest due to the potential for centralised training of ultimately decentralised policies (Lowe et al., 2017; Foerster et al., 2018). For exam- ple, in simulation one can train a Q-function, known as a critic, which exploits the action and observation histories of all agents to aid the training of local policies for each one. These policies can then be deployed in multi-agent systems in the real world, where centralisation may be infeasible. 2.5. Multi-Agent Deep Determinstic Policy Gradients Multi-agent deep deterministic policy gradients (MADDPG) (Lowe et al., 2017) is an algorithm for centralised training and decentralised control of multi-agent systems. It uses deterministic polices, as in DDPG, which condition only on each agent's local observations and actions. MADDPG handles the non-stationarity associated with the simultane- ous adaptation of all the agents by introducing a separate centralised critic for each agent. Lowe et al., trained history- independent feedforward networks on a range of mixed cooperative-competitive environments, significantly improv- ing upon agents trained independently with DDPG. 3. Methods We propose FMH, a framework for multi-agent RL which addresses the three major issues outlined in the introduction: non-stationarity, scalability and coordination. In order to increase the generality of our approach, we do so without requiring a number of potentially helpful features: access to the observations of all agents (as in centralised training), a differentiable model of the world's dynamics or a differen- tiable communication channel between agents. 3.1. Hierarchies We begin by defining the type of hierarchical structure we create for the agents. In its most straightforward setting, this hierarchy corresponds to a rooted directed tree, with the highest level manager as the root node and each worker reporting to only a single manager. Our experiments use this structure, although, for simplicity, using just two-level hierarchies. However, we also note the possibility, in more exotic circumstances, of considering more general directed acyclic graphs in which a single worker reports to multiple managers, and allowing for more than one manager at the highest level of the hierarchy. Acyclicity prevents poten- tially disastrous feedback cycles for the reward. Appropriate assignment of agents to positions in a hierarchy should depend on the varied observation and action capa- bilities of the agents. Agents exposed to key information associated with global task performance are likely to be suited to the role of manager, whereas agents with more narrow information, but which can act to achieve subgoals are more naturally suited to being workers. As we show in our experiments, identifying these differences can often make assigning agents straightforward. Here, we specify positions directly, to focus on demonstrating the resulting effectiveness of the hierarchy. However, extensions in which the hierarchy is itself learned would be interesting. 3.2. Goal-Setting In FMH, managers set goals for workers by defining their rewards. To arrange this, managers communicate a special kind of message to workers that defines the workers' reward functions according to a specified mapping. For example, if there are three objects in a room, we might allow the manager to select from three different messages, each defin- ing a worker reward function proportional to the negative distance from a corresponding object. Our experiments ex- plore variants of this scheme, imposing the constraint that the structure of the reward function remain fixed whilst the target is allowed to change. In this context and task, mes- sages therefore correspond to 'goals' requiring approach to different objects. The manager must learn to communicate these goals judiciously in order to solve complex tasks. In turn, the workers must learn how to act in the light of the re- sulting managerial reward, in addition to immediate rewards (or costs) they might also experience. Figure 1. An example of a worker-computed Feudal Multiagent Hierarchy with one manager agent and two worker agents. Worker rewards are goal-dependent and computed locally, the manager's reward is provided by the environment. Since a communicated goal is simply another type of action for the manager, our approach is consistent with the formal- ism of a (partially-observable) Markov game, for which the ManagerWorker 1Worker 2EnvironmentFeudal Multi-Agent Hierarchies reward for agent i is defined as ri : S ×A1 × ...×An → R. However, the worker's reward need not be treated as part of the environment. For instance, each worker agent i could compute its own reward locally ri : Oi×Ai → R, where Oi includes the observed goal from the manager. We illustrate this 'worker-computed' interpretation in Fig. 1. 3.3. Pretraining and Temporally Extended Subgoals We next consider the issue of non-stationarity. This fre- quently arises in multi-agent RL because the policies of other agents may change in unpredictable ways as they learn. By contrast, in FMH we allow manager agents to determine the reward functions of workers, compelling the workers towards more predictable behaviour from the perspective of the manager. However, the same issue applies at the starting point of learning for workers: they will not yet have learned how to satisfy the goals. We would therefore expect a manager to underestimate the value of the subgoals it se- lects early in training, potentially leading it sub-optimally to discard subgoals which are harder for the worker to learn. Thus, it would be beneficial for worker agents already to have learned to fulfill managerial subgoals. We address this issue practically in two steps. First, we introduce a bottom-up 'pretraining' procedure, in which we initially train the workers before training the manager. Although the manager is not trained during this period, it still acts in the multi-agent environment and sets subgoals for the worker agents. As subgoals are initially of (approximately) equal value, the manager will explore them uniformly. If the set of possible subgoals is sufficiently compact, this will enable workers to gain experience of each potential subgoal. This pretraining does not completely solve the non- stationarity problem. This is because the untrained manager will, with high probability, vacillate between subgoals, pre- venting the workers under its command from having any reasonable hope of extracting reward. We therefore want managers not only to try out a variety of subgoals but also to commit to them long enough that they have any hope of being at least partially achieved. Thus, the second compo- nent of the solution is a communication-repeat heuristic for the manager such that goal-setting is temporally extended. We demonstrate its effectiveness in our experiments. 3.4. Coordination Multi-agent problems may have many equilibria, and good ones can only be achieved through effective coordination of agent behaviour. In FMH, the responsibility of coordination is incumbent upon the manager, which exerts control over its workers through its provision of reward. We show in the simplified setting of a matrix game, how a manager may coordinate the actions of its workers (see Suppl. Mat. C). 3.5. FMH-DDPG FMH provides a framework for rewarding agents in multi- agent domains, and can work with many different RL algo- rithms. In our experiments, we chose to apply FMH in com- bination with the single-agent algorithm DDPG, trained in a fully decentralised fashion. As we do not experiment with combining FMH with any other algorithm, we frequently refer to FMH-DDPG simply as FMH. 3.6. Parameter Sharing We apply our method to scenarios in which a large number of agents have identical properties. For convenience, when training using a decentralised algorithm (FMH-DDPG or vanilla DDPG) we implement parameter sharing among identical agents, in order to train them efficiently. We only add experience from a single agent (among those sharing parameters) into the shared replay buffer, and carry out a single set of updates. We find this gives very similar results to training without parameter sharing, in which each agent is updated using its own experiences stored in its own replay buffer (see Suppl. Mat. A.2). 4. Experiments and Results We used the multi-agent particle environment1 as a frame- work for conducting experiments to test the potential of our method to address the issues of non-stationarity, scala- bility and coordination, comparing against MADDPG and DDPG. We will release code for both the model and the environments after the reviewing process ends. The RL agents have both an actor and a critic, each cor- responding to a feedforward neural network. We give a detailed summary of all hyperparameters used in Suppl. Mat. A.1. 4.1. Cooperative Communication We first experiment with an environment implemented in Lowe et al. (2017) called 'Cooperative Communication' (Figure 2a), in which there are two cooperative agents, one called the 'speaker' and the other called the 'listener', placed in an environment with many landmarks of differing colours. On each episode, the listener must navigate to a randomly selected landmark; and in the original problem, both lis- tener and speaker obtain reward proportional to the negative distance2 of the listener from the correct target. However, whilst the listener observes its relative position from each of the differently coloured landmarks, it does not know to which landmark it must navigate. Instead, the colour of the 1https://github.com/openai/multiagent-particle-envs 2The original implementation used the negative square distance; we found this slightly worse for all algorithms. Feudal Multi-Agent Hierarchies Figure 2. Cooperative Communication with 1 listener and 12 landmarks. (a) The speaker (grey circle) sees the colour of the listener (green circle) , which indicates the target landmark (green square). It communicates a message to the listener at every time step. Here there are 12 landmarks and the agent trained using FMH has successfully navigated to the correct landmark. (b) FMH substantially outperforms MADDPG and DDPG. The dotted green line indicates the end of pretraining for FMH. (c) FMH worker reward and the probability of the manager correctly assigning the correct target to the worker. The manager learns to assign the target correctly with probability 1. target landmark can be seen by the speaker, which cannot observe the listener and is unable to move. The speaker can however communicate to the listener at every time step, and so successful performance on the task corresponds to the speaker enabling the listener to reach the correct target. We also note that, although reward is provided during the episode, it is used only for training agents and not directly observed, which means that agents cannot simply learn to follow the gradient of reward. Although this task seems simple, it is challenging for many RL methods. Lowe et al. (2017) trained, in a decentralised fashion, a variety of single-agent algorithms, including DDPG, DQN and trust-region policy optimisation (Schul- man et al., 2015) on a version of this problem with three landmarks and demonstrated that they all perform poorly on this task. Of these methods, DDPG reached the highest level of performance and so we use DDPG as the strongest commonly used baseline for the decentralised approach. We also compare our results to MADDPG, which combines DDPG with centralised training. MADDPG was found to perform strongly on Cooperative Communication with three landmarks, far exceeding the performance of DDPG. For our method, FMH, we also utilised DDPG, but reverted to the more generalizable decentralized training that was previously ineffective. Crucially, we assigned the speaker the role of manager and the listener the role of worker. The speaker can therefore communicate messages which corre- spond to the subgoals of the different coloured landmarks. The listener is not therefore rewarded for going to the correct target but is instead rewarded proportional to the negative distance from the speaker-assigned target. The speaker itself is rewarded according to the original problem definition, the negative distance of the listener from the true target. We investigated in detail a version of Cooperative Commu- nication with 12 possible landmarks (Figure 2a). FMH per- formed significantly better than both MADDPG and DDPG (Figure 2b) over a training period of 100 epochs (each epoch corresponds to 1000 episodes). For FMH, we pretrained the worker for 10 epochs and enforced extended communication over 8 time steps (each episode is 25 time steps). In Figure 2c, the left axis shows the reward received by the FMH worker (listener) over training. This increased during pretraining and again immediately after pretraining concludes. Managerial learning after pretraining resulted in decreased entropy of communication over the duration of an episode (see Suppl. Mat. A.3), allowing the worker to optimise the managerial objective more effectively. This in turn enabled the manager to assign goals correctly, with the rise in the probability of correct assignment occurring shortly thereafter (right axis), then reaching perfection. Our results show how FMH resolves the issue of non- stationarity. Initially, workers are able to achieve reward by learning to optimise managerial objectives, even whilst the manager itself is not competent. This learning elicits robust behaviour from the worker, conditioned on manage- rial communication, which makes workers more predictable from the persepective of the manager. This then enables the manager to achieve competency -- learning to assign goals so as to solve the overall task. Our implementation of FMH used both pretraining and ex- tended goal-setting with a communication repeat heuristic. In Suppl. Mat. A.4, we show that pretraining the worker improved performance on this task, although even with- out pretraining FMH still performed better than MADDPG and DDPG. The introduction of extended communication is however more critical. In Figure 3 we show the performance of FMH with goal-setting over various number of time steps (and fixed pretraining period of 10 epochs). When there (a)(b)(c)NUMBER OF LISTENERS LANDMARKS 1 1 1 3 5 10 3 6 12 6 6 6 Feudal Multi-Agent Hierarchies FINAL REWARD MADDPG DDPG FMH −6.63 ± 0.05 −6.58 ± 0.03 −14.26 ± 0.07 −17.28 −6.91 ± 0.07 −6.69 ± 0.06 −18.10 ± 0.07 −18.95 −7.79 ± 0.06 −15.96 ± 0.09 −19.32 ± 0.11 −19.56 −7.10 ± 0.04 −11.13 ± 0.03 −18.90 ± 0.05 −18.95 −7.17 ± 0.03 −18.47 ± 0.04 −19.73 ± 0.06 −18.95 −8.96 ± 0.03 −19.80 ± 0.06 −21.19 ± 0.03 −18.95 COM FMH MADDPG 24 66 − − 75 59 EPOCHS UNTIL CONVERGENCE DDPG 55 42 36 50 53 32 56 57 − 77 79 − Table 1. Performance of FMH, MADDPG and DDPG for versions of Cooperative Communication with different numbers of listeners and landmarks. Final reward is determined by training for 100 epochs and evaluating the mean reward per episode in the final epoch. We indicate no convergence with a − symbol. For further details see Sup. Mat. A.6. were no communication repeats (Comm 1), performance was similar to MADDPG, but introducing even a single repeat greatly improved performance. By contrast, introduc- ing communication repeats to MADDPG did not improve performance (see Suppl. Mat. A.5). Figure 3. Communication sent by the manager is re- peated for extended goal-setting over var- ious numbers of time steps. 4.1.1. SCALING TO MANY AGENTS We next scaled Cooperative Communication to include many listener agents. At the beginning of an episode each listener is randomly assigned a target out of all the possible landmarks and the colour of each listener's target is ob- served by the speaker. The speaker communicates a single message to each listener at every time step. To allow for easy comparison with versions of Cooperative Communi- cation with only one listener, we normalised the reward by the number of listeners. As discussed in the methods, we shared parameters across listeners for FMH and DDPG. In Table 1 we show the performance of FMH, MADDPG and DDPG for variants of Cooperative Communication with different numbers of listener agents and landmarks. Con- sider the version with 6 landmarks, which we found that MADDPG could solve with a single listener within 100 epochs (unlike for 12 landmarks). On increasing the num- ber of listeners up to a maximum of 10, we found that FMH scales much better than MADDPG; FMH was able to learn an effective policy with 5 listener agents whereas MAD- DPG could not. Further, FMH even scaled to 10 listeners, although it did not converge over the 100 epochs. To aid interpretation of the reward values in Table 1 we also compare performance to a policy which simply moves to the centroid of the landmarks. This 'Centre of Mass' (CoM) agent was trained using MADDPG until convergence on a synthetic task in which reaching the centroid is maximally rewarded, and then evaluated on the true version of the task. We find that both MADDPG and DDPG do not perform better than the CoM agent when there are 10 listeners and 6 landmarks. In Figure 4 we show the final state achieved on an example episode of this version of the task, for agents trained using MADDPG and FMH. After training for 100 epochs, MAD- DPG listeners do not find the correct targets by the end of the episode whereas FMH listeners manage to do so. Figure 4. Scaling Cooperative Communication to 10 listeners with 6 landmarks - final time step on example episode. 4.2. Cooperative Coordination We then designed a task to test coordination called 'Coop- erative Coordination' (Figure 5a). In this task, there are 6 landmarks. At the beginning of each episode 3 landmarks are randomly selected to be green targets and 3 blue decoys. A team of 3 agents must navigate to cover the green targets whilst ignoring the blue decoys, but they are unable to see the colours of the landmarks. A fourth agent, the speaker, can see the colours of the landmarks and can send messages to the other agents (but cannot move). All agents can see MADDPGFMHFeudal Multi-Agent Hierarchies Figure 5. Cooperative Coordination (a) Three listeners (light grey agents) must move to cover the green landmarks whilst ignoring the blue landmarks. However, only the speaker (dark grey agent) can see the landmarks' colours; it communicates with the listeners at every time step. In this example, FMH agents have successfully coordinated to cover the three correct targets. (b) FMH performs significantly better than MADDPG and DDPG. The dotted green line indicates the end of pretraining for FMH. (c) Agents trained using FMH cover on average more targets, by the end of the episode, than MADDPG and DDPG (d) Agents trained using FMH avoid collisions more effectively than MADDPG and DDPG over the duration of an episode. each others' positions and velocities, are large in size and face penalties if they collide with each other. The task shares aspects with 'Cooperative Navigation' from (Lowe et al., 2017) but is considerably more complex due to the greater number of potential targets and the hidden information. We apply FMH to this problem, assigning the speaker agent the role of manager. One consideration is whether the man- ager, the worker, or both should receive the negative penal- ties due to worker collisions. Here we focus on the case in which the manager only concerns itself with the 'task' re- ward function. Penalties associated with collisions are there- fore experienced only by the workers themselves, which seek to avoid these whilst still optimising the managerial objective. In Figure 5b we compare the performance of FMH to MAD- DPG and DDPG. As with Cooperative Communication, FMH does considerably better than both after training for 150 epochs. This is further demonstrated when we evaluate the trained policies over a period of 10 epochs: Figure 5c shows the mean number of targets covered by the final time step of the episode, for which FMH more than doubles MADDPG. Figure 5d shows the mean number of collisions (which are negatively rewarded) during an episode. FMH collides very rarely whereas MADDPG and DDPG collide over 4 times more frequently. We also implement a version of Cooperative Coordination in which the manager is responsible not only for coordinating its workers but must also navigate to targets itself. We find that it can learn to do both roles effectively (see Suppl. Mat. B.2.1). 4.2.1. EXPLOITING DIVERSITY One role of a manager is to use the diversity it has available in its workers to solve tasks more effectively. We tested this in a version of Cooperative Coordination in which one of the listener agents was lighter than the other two and so could reach farther targets more quickly. We trained FMH (without parameter sharing due to the diversity) on this task and then evaluated the trained policies on a 'Two-Near, One-Far' (TNOF) version of the task in which one target is far away and the remaining two are close (see Suppl. Mat. B.2.2). We did this to investigate whether FMH, which was trained on the general problem, had learned the specific strategy of assigning the farthest target to the fastest agent. We found this this to be true 100 percent of the time (evaluating over 10 epochs); we illustrate this behaviour in Figure 6. Figure 6. FMH solves the TNOF task (example episode). Left: Agents are initialised at the bottom of the environment, two targets are close by, and one far away. Right: By the end of the episode, the faster (red) agent covers the farther target on the right, despite starting on the left. (a)(b)(c)(d) Feudal Multi-Agent Hierarchies 5. Discussion We have shown how cooperative multi-agent problems can be solved efficiently by defining a hierarchy of agents. Our hierarchy was reward-based, with manager agents setting goals for workers, and with the highest level manager opti- mising the overall task objective. Our method, called FMH, was trained in a decentralised fashion and showed consider- ably better scaling and coordination than both centralised and decentralised methods that used a shared reward func- tion. Our work was partly inspired by the feudal RL architecture (FRL) (Dayan & Hinton, 1993), a single-agent method for hierarchical RL (Barto & Mahadevan, 2003) which was also developed in the context of deep RL by Vezhnevets et al. (2017). In particular, FMH addresses the 'too many chiefs' inefficiency inherent to FRL, namely that each manager only has a single worker under its control. A much wider range of management possibilities and benefits arises when multiple agents operate at the same time to achieve one or more tasks (Busoniu et al., 2008). We focused on the cooperative setting (Panait & Luke, 2005); however, unlike the fully-cooperative setting, in which all agents optimise a shared reward function (Boutilier, 1996) our approach introduces a diversity of rewards, which can help with credit- assignment (Wolpert & Tumer, 2002; Chang et al., 2004) but also introduces elements of competition. This competition need not always be deleterious; for example, in some cases, an effective way of optimising the task objective might be to induce competition amongst workers, as in generative adversarial networks (Goodfellow et al., 2014). We followed FRL (though not all its successors; (Vezhn- evets et al., 2017) in isolating the workers from much of the true environmental reward, making them focus on their own narrower tasks. Such reward hiding was not complete -- we considered individualised or internalised costs from collisions that workers experienced directly, such that their reward was not purely managerially dependent. A more complete range of possibilities for creating and decompos- ing rewards between managers and workers when the objec- tives of the two are not perfectly aligned, could usefully be studied under the aegis of principal-agent problems (Jensen & Meckling, 1976; Laffont & Martimort, 2009). Goal-setting in FMH was achieved by specifying a relation- ship between the chosen managerial communication and the resulting reward function. The communication was also incorporated into the observational state of the worker; how- ever, the alternative possibility of a goal-embedding would be worth exploring (Vezhnevets et al., 2017). We also speci- fied goals directly in the observational space of the workers, whereas Vezhnevets et al. specified goals in a learned hidden representation. This would likely be of particular value for problems in which defining a set of subgoals is challenging, such as those which require learning directly from pixels. More generally, work on the way that agents can learn to construct and agree upon a language for goal-setting would be most important. To train our multi-agent systems we leveraged recent ad- vances in the field of deep RL; in particular the algorithm DDPG (Lillicrap et al., 2015), which can learn in a de- centralised fashion. Straightforward application of this al- gorithm has been shown to achieve some success in the multi-agent domain (Gupta et al., 2017) but also shown it to be insufficient in handling more complex multi-agent problems (Lowe et al., 2017). By introducing a managerial hierarchy, we showed that FMH-DDPG has the potential to greatly facilitate multi-agent learning whilst still retaining the advantage of decentralised training. Our proposed ap- proach could also be combined with centralised methods, and this would be worth further exploration. Other work in multi-agent RL has also benefitted from ideas from hierarchical RL. The MAXQ algorithm (Dietterich, 2000) has been used to train homogenous agents (Makar et al., 2001), allowing them to coordinate by communicat- ing subtasks rather than primitive actions, an idea recently re-explored in the context of deep RL (Tang et al., 2018). A meta-controller which structures communication between agent pairs in order to achieve coordination has also been proposed (Kumar et al., 2017) (and could naturally be hy- bridized with FMH), as well as the use of master-slave architecture which merges the actions of a centralised mas- ter agent with those of decentralised slave agents (Kong et al., 2017). Taken together, these methods represent inter- esting alternatives for invoking hierarchy which are unlike our primarily reward-based approach. There are a number of additional directions for future work. First, the hierarchies used were simple in structure and spec- ified in advance, based on our knowledge of the various information and action capabilities of the agents. It would be interesting to develop mechanisms for the formation of complex hierarchies. Second, in settings in which workers can acquire relevant task information, it would be worth- while investigating how a manager might incentivise them to provide this. Third, it would be interesting to consider how a manager might learn to allocate resources, such as money, computation or communication bandwidth to en- able efficient group behaviour. Finally, we did not explore how managers should train or supervise the workers beneath them, such as through reward shaping (Ng et al., 1999). Such an approach might benefit from recurrent networks, which could enable managers to use the history of worker performance to better guide their learning. Feudal Multi-Agent Hierarchies 6. Acknowledgements We would like to thank Heishiro Kanagawa, Jorge A. Menen- dez and Danijar Hafner for helpful comments on a draft ver- sion of the manuscript. Sanjeevan Ahilan received funding from the Gatsby Computational Neuroscience Unit and the Medical Research Council. Peter Dayan received funding from the Max Planck Society. References Barto, A. G. and Mahadevan, S. Recent advances in hier- archical reinforcement learning. Discrete event dynamic systems, 13(1-2):41 -- 77, 2003. Boutilier, C. Planning, learning and coordination in multia- gent decision processes. In Proceedings of the 6th confer- ence on Theoretical aspects of rationality and knowledge, pp. 195 -- 210. Morgan Kaufmann Publishers Inc., 1996. Hu, J., Wellman, M. P., et al. Multiagent reinforcement In learning: theoretical framework and an algorithm. ICML, volume 98, pp. 242 -- 250. Citeseer, 1998. Jensen, M. C. and Meckling, W. H. Theory of the firm: Man- agerial behavior, agency costs and ownership structure. Journal of financial economics, 3(4):305 -- 360, 1976. Kong, X., Xin, B., Liu, F., and Wang, Y. Revisiting the master-slave architecture in multi-agent deep reinforce- ment learning. arXiv preprint arXiv:1712.07305, 2017. Kumar, S., Shah, P., Hakkani-Tur, D., and Heck, L. Fed- erated control with hierarchical multi-agent deep rein- forcement learning. arXiv preprint arXiv:1712.08266, 2017. Laffont, J.-J. and Martimort, D. The theory of incentives: the principal-agent model. Princeton university press, 2009. Busoniu, L., Babuska, R., and De Schutter, B. A comprehen- sive survey of multiagent reinforcement learning. IEEE Transactions on Systems, Man, And Cybernetics-Part C: Applications and Reviews, 38 (2), 2008, 2008. Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. Chang, Y.-H., Ho, T., and Kaelbling, L. P. All learning is local: Multi-agent learning in global reward games. In Advances in neural information processing systems, pp. 807 -- 814, 2004. Dayan, P. and Hinton, G. E. Feudal reinforcement learning. In Advances in neural information processing systems, pp. 271 -- 278, 1993. Dietterich, T. G. Hierarchical reinforcement learning with the maxq value function decomposition. Journal of Arti- ficial Intelligence Research, 13:227 -- 303, 2000. Foerster, J., Assael, I. A., de Freitas, N., and Whiteson, S. Learning to communicate with deep multi-agent rein- forcement learning. In Advances in Neural Information Processing Systems, pp. 2137 -- 2145, 2016. Foerster, J. N., Farquhar, G., Afouras, T., Nardelli, N., and Whiteson, S. Counterfactual multi-agent policy gradi- In Thirty-Second AAAI Conference on Artificial ents. Intelligence, 2018. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672 -- 2680, 2014. Gupta, J. K., Egorov, M., and Kochenderfer, M. Cooperative multi-agent control using deep reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems, pp. 66 -- 83. Springer, 2017. Littman, M. L. Markov games as a framework for multi- agent reinforcement learning. In Machine Learning Pro- ceedings 1994, pp. 157 -- 163. Elsevier, 1994. Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, O. P., and Mordatch, I. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pp. 6379 -- 6390, 2017. Makar, R., Mahadevan, S., and Ghavamzadeh, M. Hierarchi- cal multi-agent reinforcement learning. In Proceedings of the fifth international conference on Autonomous agents, pp. 246 -- 253. ACM, 2001. Matignon, L., Laurent, G. J., and Le Fort-Piat, N. Inde- pendent reinforcement learners in cooperative markov games: a survey regarding coordination problems. The Knowledge Engineering Review, 27(1):1 -- 31, 2012. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidje- land, A. K., Ostrovski, G., et al. Human-level control through deep reinforcement learning. Nature, 518(7540): 529, 2015. Mordatch, I. and Abbeel, P. Emergence of grounded com- positional language in multi-agent populations. In Thirty- Second AAAI Conference on Artificial Intelligence, 2018. Ng, A. Y., Harada, D., and Russell, S. Policy invariance under reward transformations: Theory and application to reward shaping. In ICML, volume 99, pp. 278 -- 287, 1999. Feudal Multi-Agent Hierarchies Panait, L. and Luke, S. Cooperative multi-agent learning: The state of the art. Autonomous agents and multi-agent systems, 11(3):387 -- 434, 2005. Schulman, J., Levine, S., Abbeel, P., Jordan, M., and Moritz, In International P. Trust region policy optimization. Conference on Machine Learning, pp. 1889 -- 1897, 2015. Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., and Riedmiller, M. Deterministic policy gradient algorithms. In ICML, 2014. Sukhbaatar, S., Fergus, R., et al. Learning multiagent com- munication with backpropagation. In Advances in Neural Information Processing Systems, pp. 2244 -- 2252, 2016. Sutton, R. S., McAllester, D. A., Singh, S. P., and Mansour, Y. Policy gradient methods for reinforcement learning with function approximation. In Advances in neural in- formation processing systems, pp. 1057 -- 1063, 2000. Tan, M. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth inter- national conference on machine learning, pp. 330 -- 337, 1993. Tang, H., Hao, J., Lv, T., Chen, Y., Zhang, Z., Jia, H., Ren, C., Zheng, Y., Fan, C., and Wang, L. Hierarchical deep multiagent reinforcement learning. arXiv preprint arXiv:1809.09332, 2018. Vezhnevets, A. S., Osindero, S., Schaul, T., Heess, N., Jader- berg, M., Silver, D., and Kavukcuoglu, K. Feudal net- works for hierarchical reinforcement learning. In Interna- tional Conference on Machine Learning, pp. 3540 -- 3549, 2017. Vlassis, N. A concise introduction to multiagent systems and distributed artificial intelligence. Synthesis Lectures on Artificial Intelligence and Machine Learning, 1(1): 1 -- 71, 2007. Watkins, C. J. and Dayan, P. Q-learning. Machine learning, 8(3-4):279 -- 292, 1992. Williams, R. J. Simple statistical gradient-following algo- rithms for connectionist reinforcement learning. Machine learning, 8(3-4):229 -- 256, 1992. Wolpert, D. H. and Tumer, K. Optimal payoff functions for members of collectives. In Modeling complexity in eco- nomic and social systems, pp. 355 -- 369. World Scientific, 2002. Feudal Multi-Agent Hierarchies A. Experimental Results A.1. Parameter settings In all of our experiments, we used the Adam optimizer with a learning rate of 0.001 and τ = 0.01 for updating the target networks. γ was 0.75. The size of the replay buffer was 107 and we updated the network parameters after every 100 samples added to the replay buffer. We used a batch size of 1024 episodes before making an update. For our feedforward networks we used two hidden layers with 256 neurons per layer. We trained with 10 random seeds (except otherwise stated). Hyperparameters were optimised using a line search centred on the experimental parameters used in (Lowe et al., 2017). Our optimised parameters were found to be identical except for a lower value of γ (0.75) and of the learning rate (0.001), and a larger replay buffer (107). We found these values gave the best performance for both MADDPG and FMH on a version of Cooperative Communication with 6 landmarks evaluated after 50 epochs (an epoch is defined to be 1000 episodes). S2). Agents are trained on Cooperative Communication with 12 landmarks and a single listener. As the target does not change during the middle of an episode, we expect the entropy to decrease as agents learn. For FMH, during pretraining, entropy is high as all goals are sampled approximately uniformly (but with enforced ex- tended communication over 8 time steps). However, shortly after pretraining ends the entropy of managerial communi- cation rapidly decreases. This is in contrast to MADDPG which decreases in entropy more steadily. Figure S2. Entropy of managerial com- munication over the duration of an episode at different stages in training A.2. Parameter Sharing A.4. Pretraining We implemented parameter sharing for the decentralised algorithms DDPG and FMH. To make training results ap- proximately similar to implementations which do not use parameter sharing we restrict updates to a single agent and add experience only from a single agent to the shared re- play buffer (amongst those sharing parameters). We find in practice that both approaches give very similar results for FMH, whereas parameter sharing slightly improves the performance of DDPG -- we show this for a version of Cooperative Communication with 3 listeners and 6 targets (Figure S1). In general sharing parameters reduces train- ing time considerably, particularly as the number of agents scales. does Figure S1. Parameter sharing not affect performance for FMH but slightly improves DDPG. Agents trained using FMH were pretrained for 10 epochs across all experiments. During pretraining all agents act in the multi-agent environment. Although all experiences are added appropriately into the replay buffers, the manager does not update its parameters during pretraining whereas the workers do. We show the benefits of pretraining on Cooperative Com- munication with 12 landmarks and a single listener agent in Figure S3. Figure S3. Pretraining improves FMH, al- though agents can still learn effectively without it. We use extended communi- cation (goal-setting) for 8 time steps. A.3. Entropy of Communication We show the change in entropy of communication over the duration of an episode for the various algorithms (Figure A.5. Extended Communication In the main text we showed that extended goal-setting in FMH lead to substantially improved performance. We also considered whether a similar approach would benefit meth- ods which do not treat communication as goals. We found that extended communication did not help MAD- DPG on the same task, with learning curves being in all Feudal Multi-Agent Hierarchies cases virtually identical (Figure S4). Figure S4. Extended commu- nication does not significantly improve the performance of MADDPG. A.6. Further details on Table 1 Values in the table were determined using 10 random seeds in all cases, except for the one exception of MADDPG with 10 listeners and 6 landmarks, which used 3 random seeds (training time is substantially longer as we do not share parameters). The CoM agent was trained on the synthetic task with different numbers of landmarks. Performance of the trained CoM policies was then evaluated over a period of 10 epochs on the corresponding true tasks. Convergence was determined by comparing the mean perfor- mance in the final 5 epochs with the mean performance of a sliding window 5 epochs in width (we also take the mean across random seeds). If the mean performance within the window was within 2 percent of the final performance, and remained so for all subsequent epochs, we defined this as convergence, unless the first time this happened was within the final 10 epochs. In such a case, we define the algorithm as not having converged. For assessing the exact time of convergence in the case of FMH we report values which include the 10 epochs of pretraining. A.7. Cooperative Communication with 3 landmarks For reference we show performance of the various algo- rithms on Cooperative Communication with 3 landmarks. Both MADDPG and FMH perform well on this task, al- though MADDPG reaches convergence more rapidly (Fig- ure S5). Figure S5. Cooperative Communication with 1 listener and 3 landmarks. B. Environments B.1. Cooperative Communication We provide further details on our version of Cooperative Communication (see main text for original description). In general, we keep environment details the same as Lowe et al., including the fact that the manager only has access to the target colour. However, we also scale up the number of coloured landmarks, which we do by taking the RGB values provided in the multi-agent particle environment, [0.65, 0.15, 0.15], [0.15, 0.65, 0.15], [0.15, 0.15, 0.65], and adding 9 more by placing them on the remaining vertices of the corresponding cube and at the centre-point of four of the faces (in RGB space). The colours for the 12 landmarks are therefore: [0.65, 0.15, 0.15], [0.15, 0.65, 0.15], [0.15, 0.15, 0.65], [0.65, 0.65, 0.15], [0.15, 0.65, 0.65], [0.65, 0.15, 0.65], [0.65, 0.65, 0.65], [0.15, 0.15, 0.15], [0.40, 0.40, 0.65], [0.40, 0.40, 0.15], [0.15, 0.40, 0.40], [0.65, 0.40, 0.40]. The particular colour values used for the landmarks influ- ences the performance of RL algorithms as landmarks which have similar colours are harder for the speaker to learn to distinguish. B.2. Cooperative Coordination We provide further details on our version of Cooperative Coordination (see main text for original description). The task provides a negative reward of -1 to each agent involved in a collisions. For DDPG and MADDPG this penalty is shared across agents, whereas in FMH only the agents involved in the collision experience this penalty. We also evaluated performance of trained policies in Figures 5c and 5d with slight modifications to the overall task. In the case of Figure 5c, to ensure that targets were never impossible to achieve by overlapping with the immobile manager, we moved the manager off-screen. For Figure 5d we ensured that agent positions were never initialised in a way such that they would automatically collide (such cases are rare). B.2.1. COORDINATION WITH A MOBILE MANAGER We implemented a version of Cooperative Coordination with 2 listeners and 1 speaker which together need to cover the 3 green targets. The manager must multi-task, directing workers to the correct targets whilst also covering these targets itself. We find that the manager learns to do this, outperforming both MADDPG and DDPG (Figure S6) Feudal Multi-Agent Hierarchies Figure S6. FMH performs well, even when the manager is required to move to cover targets whilst also setting goals for workers B.2.2. THE TWO-NEAR, ONE-FAR TASK This task is not used for training but to evaluate the per- formance of agents trained on a version of Cooperative Coordination in which one agent is twice as light as normal and the remaining two are twice as heavy. Evaluating the optimal assignments on this task can be dif- ficult, so we assess it in the more easily interpreted TNOF environment. In TNOF, the agents start at the bottom of the environment. Two green targets are located nearby (in bot- tom 40 percent of screen) whereas one target is far away (in top 30 percent of screen). The x-coordinates are randomly sampled at the beginning of each episode and blue decoys are also added (one nearby, two far). C. Solving a Coordination Game Some of the most popular forms of multiagent task are coordination games from microeconomics in which there are multiple good and bad Nash equilibria, and it is necessary to find the former. It is intuitively obvious that appointing one of the agents as a manager might resolve the symmetries inherent in cooperative coordination game in which agents need to take different actions to receive reward: Player Y A B Player X A (0, 0) B (1, 1) (1, 1) (0, 0) This game has two pure strategy Nash equilibria and one mixed strategy Nash equilibrium which is Pareto dominated by the pure strategies. The challenge of this game is for both agents to choose a single Pareto optimal Nash equilibrium, either (A, B) or (B, A). For a matrix game, we define the feudal approach as allowing Player X, the manager, to specify the reward player Y will receive for its actions. This is a simplification when compared to the more general setting of a Markov game in which the feudal manager can reward not only actions but also achievement of certain states.3. In order to specify the reward, we assume that Player X communicates a goal, either gA or gB, prior to both players taking their actions. If Player X sends gA it means that action A is now rewarded for Player Y and action B is not. Player X's rewards are unchanged, and so together this induces the following matrix game: Player Y A B Player X A (0, 1) B (1, 1) (1, 0) (0, 0) Action A for player Y is now strictly dominant and so a rational Player Y will always choose it. By iterated elimi- nation of strictly dominated strategies we therefore find the resulting matrix game: Player Y A (0, 1) (1, 1) Player X A B And so a rational Player X will always choose B, resulting in an overall strategy of (B, A) conditioned on an initial communication of gA. By symmetry, we can see that con- ditioned on gB, rational players X and Y will play (A, B). The feudal approach therefore allows the manager to flexibly coordinate the pair of agents to either Nash equilibrium. For games involving N players, coordination can be achieved by the manager sending out N-1 goals. 3Typically we prefer the manager to choose distal states as targets rather than actions as this requires the manager to micro- manage less and so supports temporal abstraction
1604.04180	2	1604	2017-05-26T09:23:13	Job Selection in a Network of Autonomous UAVs for Delivery of Goods	[ "cs.MA", "cs.RO" ]	This article analyzes two classes of job selection policies that control how a network of autonomous aerial vehicles delivers goods from depots to customers. Customer requests (jobs) occur according to a spatio-temporal stochastic process not known by the system. If job selection uses a policy in which the first job (FJ) is served first, the system may collapse to instability by removing just one vehicle. Policies that serve the nearest job (NJ) first show such threshold behavior only in some settings and can be implemented in a distributed manner. The timing of job selection has significant impact on delivery time and stability for NJ while it has no impact for FJ. Based on these findings we introduce a methodological approach for decision-making support to set up and operate such a system, taking into account the trade-off between monetary cost and service quality. In particular, we compute a lower bound for the infrastructure expenditure required to achieve a certain expected delivery time. The approach includes three time horizons: long-term decisions on the number of depots to deploy in the service area, mid-term decisions on the number of vehicles to use, and short-term decisions on the policy to operate the vehicles.	cs.MA	cs	Job Selection in a Network of Autonomous UAVs for Delivery of Goods Pasquale Grippa1, Doris A. Behrens2,3, Christian Bettstetter1,4, and Friederike Wall5 1Institute of Networked and Embedded Systems Alpen-Adria-Universitat Klagenfurt, Klagenfurt, Austria 2School of Mathematics, Cardiff University 23 Senghennydd Road, Cardiff CF24 4AG 3Aneurin Bevan University Health Board Lodge Road, Caerleon NP18 3XQ 4Lakeside Labs GmbH, Klagenfurt, Austria 5Institute for Business Management, Alpen-Adria-Universitat Klagenfurt, Klagenfurt, Austria Emails: [email protected], [email protected] Abstract 1 Introduction This article analyzes two classes of job selection poli- cies that control how a network of autonomous aerial vehicles delivers goods from depots to customers. Cus- tomer requests (jobs) occur according to a spatio-temporal stochastic process not known by the system. If job selec- tion uses a policy in which the first job (FJ) is served first, the system may collapse to instability by removing just one vehicle. Policies that serve the nearest job (NJ) first show such threshold behavior only in some settings and can be implemented in a distributed manner. The timing of job selection has significant impact on delivery time and stability for NJ while it has no impact for FJ. Based on these findings we introduce a methodological approach for decision-making support to set up and operate such a system, taking into account the trade-off between mone- tary cost and service quality. In particular, we compute a lower bound for the infrastructure expenditure required to achieve a certain expected delivery time. The approach includes three time horizons: long-term decisions on the number of depots to deploy in the service area, mid-term decisions on the number of vehicles to use, and short-term decisions on the policy to operate the vehicles. Manuscript submitted on 30 January 2017. Small unmanned aerial vehicles (UAVs) have successfully found their way to civil applications. A broad variety of UAV models has been commercialized in the past few years. They fly routes in an autonomous manner, carry cameras for aerial photography, and may transport goods from one place to another. The range of applications is broad, including aerial monitoring of plants and agricul- ture fields as well as support for first time responders in disasters [21, 35, 27, 1]. Delivering goods via a network of UAVs becomes an option if classical means of trans- portation - like trucks, trains, and planes - are inappro- priate. First, this comes about if roads, railway tracks, or landing facilities do not exist, if weather conditions make it impossible to use them, or if their use is too danger- ous or time-consuming. In this context, a compelling ser- vice would be the delivery of medicine, vaccinations, or laboratory samples for patients in remote areas and cri- sis regions. Second, such a service is also worthwhile in densely populated metropolitan areas, when congestion makes roads nearly impassable. The main objective of this paper is to provide theo- retical insight for the architectural setup and control of a UAV-based delivery system. We analyze policies that con- trol how an interconnected team of UAVs resolves service requests that are randomly distributed in space and time. 1 The entities of the system are goods, customers, vehicles, and depots. Customers request goods that are stored in depots and delivered by vehicles. Service requests, also denoted as jobs or customer demands, are not known in advance and arrive over time at certain locations accord- ing to a space-time stochastic process. We analyze both centralized and distributed policies. In the latter, the sys- tem's "intelligence" is literally embedded into each vehi- cle, i.e., each vehicle decides by its own which job to se- lect next, thus raising the autonomy of vehicles to a level that goes beyond autonomous flying. If the vehicles are capable of consecutively serving sev- eral customers before they return to a depot, e.g., if goods are lightweight and total distances are small, the problem of selecting customer requests to be served falls into the domain of dynamic vehicle routing with stochastic de- mands, dating back to [5, 6, 7]. In contrast to this, ve- hicles in our system serve no more than one customer per trip for capacity reasons, and we use the term job selec- tion to emphasize the difference to routing. By focusing on non-partitioning job selection for delivery, we comple- ment research that focused on partitioning routing policies for wide-area surveillance (see [18, 37, 16, 31, 10]). Our performance measure is delivery time, i.e., the time it takes for a customer from requesting to obtaining a good. The stability of such a service is linked to the queu- ing of jobs, i.e., customers may have to wait until other customers have received the goods they requested. The system becomes unstable if the average number of waiting customers persistently increases over time. We show fea- tures of distributed non-partitioning job selection policies that are found by simulation for M/G/K queues, with K > 1, and described in terms of delivery time, which is related to system stability [5, 6, 7]. Using simulations, we deduce some results of general validity on the system behavior. A key insight is that the timing of job selection matters for distributed non-partitioned policies but has no impact on the centralized ones. If, for example, a policy is applied in which the job that first comes in is first served, denoted as FJ-policy, the selection can occur as soon as possible (just after the service of the last cutomer) or as late as possible (just before loading a good). If, in con- trast, a policy is applied in which vehicles decide that the currently nearest job is first served, denoted as NJ-policy, the decision on job selection should be made as early as possible. Moreover, we find that a shift in the timing of job selection even alters the structure of the policy in case of NJ-policies (altering also its vulnerability with respect to changes in customer demand). These insights into system behavior are accompanied by findings about the relation between job selection poli- cies and system setup, the total volume of demand, and system performance. It is known, for example, that FJ- policies outperform NJ-policies for low system loads, while the opposite is true for higher loads, e.g., caused by higher customer arrival rates or a smaller number of vehicles [5]. So far, it is not emphasized in the literature that a system operating according to a FJ-policy will lit- erally collapse all of a sudden and tip into instability if the vehicle fleet is reduced by a single entity. We high- light this threshold effect, as it is essential for both system implementation and operation. Based on these findings, we derive decision making support for the investment in a delivery system and its op- eration. In particular, the infrastructure of depots is sub- ject of a long-term decision, the number of vehicles can be modified in mid-term, and the job selection policy is a short-term choice. The set-up of the system (number of depots and vehicles) shapes its monetary costs and, in conjunction with the job selection policy, the service qual- ity provided to the customers for which they are willing to pay for. Hence, for investing in such a delivery system an interesting question is which minimum expenditure is required to provide a certain service quality. For this, we compute a lower bound for the expenditure necessary to set up a stable system as a function of the targeted ser- vice quality in terms of average delivery time and exem- plarily illustrate the application of this service-possibility- frontier for parameters reported by the company Matter- net. 2 Related Work 2.1 Types of Vehicle Routing Problems (VRPs) The framework of the used model dates back to [13], who introduced their formulation of the vehicle routing prob- lem (VRP) as a generalization of the traveling salesman problem [17]. Ever since then, the operations research community has intensively studied how a central plan- ner determines optimal sets of routes for fleets of homo- geneous vehicles, supplying given sets of geographically dispersed customers with goods [19]. In the context of a "classical VRP", such an optimal set of routes accom- plishes that (i) all customers are supplied with the de- manded products, (ii) none of the vehicles exceeds its ca- pacity traveling along its route, (iii) no customer is visited more than once, (iv) all routes start and end at a central depot, and (v) the overall routing cost is minimized. In practical applications, VRPs have a broad diversity of ad- ditional requirements and operational constraints affect- ing the construction of the optimal set of routes. Among these are periodic VRPs [2], VRP with pickup and deliv- ery [14], VRP with split deliveries [15], and VRP with time windows within which customers have to be served [11], to mention but a few. For reviews of exact and ap- proximate methods of solving the classical VRP, we refer to [4, 12, 23, 24, 22, 40, 41], and, for an exhaustive bibli- ography on vehicle routing to [25]. VRPs are classified according to the nature of system input information. If all system input is known before the vehicles leave the depot(s) and does not change dur- ing mission execution, the problem of concern is like the one described in the paragraph above, and said to be both static and deterministic. For many real-world ap- plications, at least some input information, like customer arrivals, behaves according to a probability distribution rather than being known a priori. These VRPs are de- noted as stochastic. If some input information appears or changes during missing execution, which has to be imme- diately integrated into decision-making, the VRP is called dynamic ([33, 34], or, for a recent review, [32]). Then, designing sets of routes has to be replaced by designing routing policies, which describe the evolution of motion paths as a function of newly arriving input. 2.2 Stochastic and Dynamic Vehicle Rout- ing in Robotics and Aeronautics The papers [5, 6, 7] were the first to comprehensively an- alyze the stochastic and dynamic VRP. A generalization of the VRP is the pickup and delivery problem (PDP) [39, 42]. While the VRP was originally investigated for applicability in classical forms of transportation and logis- tics, there has been a shift towards applications in robotics and aeronautics in the last decade. Particular attention was devoted to the motion coordination of mobile robots, which includes, among others, a VR-based development of spatially-distributed (surveillance) policies for UAVs that are adaptive to network changes [18, 37, 16, 31, 10]. Strategies were developed that ensure that a certain frac- tion of stochastically generated service requests is served before the jobs expire [29], that account for service priori- ties [38] and translating demands [8], and that accomplish an effective system management without explicit commu- nication [3]. These approaches utilize partitioning poli- cies to find ways to operate a distributed system that are scalable to large-vehicle networks [31]. These partition- ing policies applied to the control of dynamic and stochas- tic VRPs methodologically differ from distributed parti- tioning algorithms [30] that analyze how to partition the service area prior to applying a control rule. We deviate from both of above approaches and seek to find policies that are adaptive in an alternative way. For our application, not partitioning the service area (since de- pots are fixed) allows for greater system adaptability if vehicles need to endogenously pool in "hot spot" regions of the service area, e.g., whenever disasters or diseases shift demands to certain regions, and if depots run out of goods. Especially in the latter case, non-partitioning job selection policies allow for regarding fixed depots as "cus- tomers" that have to be supplied with whatever it is that is needed. 3 System Model 3.1 Entities of a UAV Delivery System The system is composed of K ∈ N vehicles moving in a bounded and convex service area A ⊂ R2 of size A := A, where · is the Euclidean norm. A vehicle is denoted by vk with identifier k∈ {1, . . . , K}. The cur- rent position of vehicle vk at time t is vk(t)∈ A with t ≥ 0. All vehicles travel at the same constant velocity ν ∈ R+ and are equipped with a battery, whose level at time t is represented by bk(t)∈ [ 0, 1]. The fact that bat- teries have to be recharged or exchanged is quantified by the parameter α ∈ (0, 1], which is the air-time ratio, i.e., α = air time/(air time + charge time). The arrival of customer demands, i.e., delivery requests for goods within A, is generated by a Poisson process with finite intensity λ∈ R+, where λ is the customer ar- rival rate. The demands, also called jobs, are indexed by the job identifier n ∈ N, which indicates the order of re- quest arrivals. The corresponding customer is called cn; his or her position is denoted by cn ∈ A and assumed to be independently and uniformly distributed in A. Cus- tomer demands do not only differ with respect to timing and locations but also with respect to the goods that are requested to be delivered. Goods are, in general, different but have the same ex- piration date and are treated with identical priority. There are L∈ N depots in the system, which store the goods. The depots are interconnected and provide a sufficient number of all goods and service activities, like recharg- ing batteries. We assume that, for capacity reasons, a ve- hicle cannot serve more than one customer demand per trip (see [6, p. 71]). The depots are set up at locations d = [d1 d2 ... dL]∈ AL, where d is chosen such that the expected distances between a random point q∈ A (poten- tial demand) generated according to a uniform distribu- tion over A and the closest depot are minimal: HL(d, A) := 1 A ·A min l:l∈{1,...,L}dl−qdq. (1) This corresponds to the solution of the continuous mul- timedian problem known from geometric optimization [28, 44]): d∗ = arg min d∈AL HL(d, A), with H ∗ L(A) := HL(d∗, A). (2) (3) A depot is a storage but not a permanent home base for particular vehicles. Whenever a vehicle has delivered a good, it either approaches the nearest depot or another one that is more appropriate to handle the next customer demand. Appropriateness is determined by the job selec- tion rule that is implemented (see Section 5). This non- restricted movement of all K vehicles to all corners of an L-depot system is a major difference to related work in [5, 6, 7, 18, 37, 16, 31, 10]. 3.2 Service Operations and Delivery Time The delivery time for the customer cn (n∈ N) is denoted by the stochastic variable Tn = Wn + Rn + Sn, where c . ( d i) ( d e g r e t u r n i n c h g if n g a r g i n d i n a l o i) g eli n v tr a a n l o u ( c n − ) g a r g i n l o g if n d i n a g h g d i n r e t u r n i n c j ) e c . ( d j ) ( d eli n v tr a u n ) ( c g d i n a g n l o τ0 τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8 τ9 τ10 t Wn Rn Sn Tn Figure 1: Time intervals involved in a customer service: delivery time T , waiting time W , return time R, and ser- vice time S. Wn is the waiting time, Sn is the service time, and Rn is the return time. Fig. 1 illustrates all operations involved in the service of cn from arrival at t = τ0 to the service completion at t = τ10. The waiting time is Wn = τ5−τ0. The return time is Rn = τ7−τ5, and depends on the posi- tion of the customer cn− being served before cn and on the system's load. Rn is included in [0, R′ n]: It is null if the vehicle is ready at the depot, and maximum if the service of cn− is not completed at the arrival of cn. The service time is Sn = τ10−τ7. This definition differs Bert- simas and van Ryzin's on-site service time [5, 6, 7] which corresponds to the time a vehicle is "unloading" goods at a customer. It also differs from the notion of service time Bn in classical queuing theory: Sn = Bn − Rn. 3.3 Queuing Phenomena and Stability A job selection policy, called π = (π1, π2, ..., πK ), has to restrain the outstanding jobs [5, 6, 7, 18, 37, 16, 31, 10]. Policy π is said to be stabilizing if the expected number of pending jobs (customers waiting for service) stays con- fined over time, i.e., if there exists an arbitrary constant κ <∞ such that ¯Nπ := lim t→∞ E[N(t)π] ≤ κ, (4) where N(t) denotes the number of pending jobs at time t. We assume that N (0) = 0, i.e., no customer is waiting at t = 0. The return and service times are crucial for the stability of the system. A necessary condition for stability is [6, p. 63]: ¯D ν ≤ K λ (5) if the on-site service time is null. ¯D is the average Eu- clidean distance between two customers served in se- quence, λ is the arrival intensity, and ν is the vehicle speed. This condition applies to our problem with two changes: First, the distance becomes the distance function customer-depot-customer, which divided by the speed gives ¯R′ + ¯S (Fig. 1), where ¯R′ is the average return time in high load and steady state, and ¯S is the average service time in steady state. Second, the effective number of used vehicles is on average αK, where α is the airtime ratio. Therefore, our stability condition is ¯R′ + ¯S ≤ αK λ . (6) The problem analyzed in this paper can be modeled as an M/G/K queue with interdependent service times Bn. M indicates Poisson distributed customer arrivals, G indi- cates service times distributed according to a generic dis- tribution, and K is the number of servers. For M/G/K queues with independent service times [20], the load fac- tor is defined as ρ := . (7) λ ¯B αK The system is said to be in light load condition if ρ → 0 and in heavy load condition if ρ → 1. A necessary con- dition for the stability of the system is ρ < 1. In case of stability, ρ can be interpreted as the expected value of the fraction of busy servers. This definition does not apply to our case because ¯B depends on the system state [20], which includes the number of waiting customers. Nev- ertheless, the condition of stability is still valid: ¯B ap- proaches ¯R′ + ¯S for ρ → 1, leading to (6). In words, to stabilize the system, it is necessary that the average time between two successfully completed service requests is not larger than the average time between two customer arrivals multiplied by the average number of available ve- hicles. 4 Expenditure for Minimum Infras- tructure We start to derive the minimum expenditure for infras- tructure necessary to build a stable system as a func- tion of system performance by deriving a lower bound on average delivery time, ¯T . This time is bounded by the minimum service time, ¯T ≥ ¯Smin. The latter can be expressed in terms of the multi-median function (3), i.e., ¯Smin = H ∗ L(A) can be bounded by 2√A/3√πL [44], which implies that L(A)/ν. H ∗ ¯Tmin≥ L(A) H ∗ ν > 2 3ν A πL . (8) By rephrasing (8) we derive a condition for the minimum number of depots necessary to yield a certain average de- livery time (larger than minimum time), i.e., L ≥ 4A 9πν2 ¯T 2 min . (9) From the definition of the load factor (given by (7)) and in consideration of ¯S, ¯R ≥ ¯Tmin together with ρ < 1, we derive a condition for the minimum number of vehicles necessary to yield a certain average delivery time, i.e., K > 2λ α · ¯Tmin. (10) For Cv and Cd denoting the costs of a vehicle and a de- pot, respectively, total infrastructure expenditure is deter- mined by C(K, L) = Cd ·L + Cv ·K. (11) With the necessary conditions for a "stabilizing" infras- tructure required to yield a particular delivery time, given by (9) and (10), we use (11) to compute an auxiliary func- tion, i.e., g( ¯Tmin) = Cd · 4A 9πν2 ¯T 2 min + Cv · 2λ α · ¯Tmin, (12) and intend to develop a lower bound for total infrastruc- ture expenditure as a function of ¯T . In this context, it is important to make aware that the values of ¯Tmin in (12) are contained in a countable set: one value for every L, i.e., ¯Tmin = ¯Tmin(L). If we changed ¯Tmin into a continu- ous variable, ¯τ ∈ R, the auxiliary function g would have a minimum at ¯τ ∗ = 3 4αA·Cd 9πλν ·Cv . (13) Assume that there exists an L = L′ for which the corre- sponding minimum delivery time is larger than the one min > ¯T ′′ min := ¯Tmin(L′) > ¯τ ∗. Then, the defined by (13), i.e., ¯T ′ configuration L = L′, with L′ < L′′ but ¯T ′ min, may lead to a higher total expenditure for infrastructure than the configuration L = L′′. I.e., decreasing the number of depots increases minimum delivery time and may, there- fore, increase the number of vehicles necessary to stabi- lize the system. Depending on the ratio between Cd and Cv this may increase total infrastructure expenditure nec- essary to build a stable system, denoted by Imin. Yet, Imin has to be a non-increasing function of ¯T . By construction this is accomplished in the following way: min > ¯T ′′ For any two configurations of depots such that L′ < L′′, ¯T ′ min, and g(L′) > g(L′′), we are able to obtain delivery time ¯T ′ min and reduce total expenditure by employing L = L′′ instead of L = L′ depots along- side delaying any delivery by ¯T ′ min time units. This makes Imin a piecewise constant function. Additionally, our construction of Imin has to account for 2λ· ¯Tmin/α being an integer. Altogether, this yields min− ¯T ′′ Imin( ¯T ) = min l:l∈{L,...,∞} Cd · 4A 9π(H ∗ l (A))2 + Cv · 2λH ∗ l (A) αν for H ∗ H ∗ ν ν L(A) L(A) H ∗ L−1(A) ν ≤ ¯T < ≤ ¯T L = 2, 3, . . . L = 1 (14) these two policies, we seek to gain insight to the follow- ing question: Is it worth to delay the decision on job se- lection to obtain more information about new customer requests in order to reduce the delivery time? For both classes we evaluate two extreme cases: when selection is made as soon as possible (just after the previous service), and when selection is made as late as possible (just before loading the good). A first issue in this context is whether the vehicles con- sider the arrival order of customers in the waiting queue. A second issue is the timing of job selection. In particu- lar, it seems plausible to assume that decisions should be made "as late as possible" to utilize the most recent infor- mation about the system status, basically postponing the individual job selection until loading goods in the depot. Such delayed decisions could also have a negative effect. This relates to the fact that, for L ≥ 2, one of the most important decisions to be made is where a vehicle should return to after satisfying a customer demand. If a non- postponed decision about the next job includes a "clever" selection of the depot to travel to, while a postponed de- cision leads to being at a suboptimal place at the moment of job selection, the advantage gained by processing more recent information is counteracted by the disadvantage of the extra distance traveled to the customer to be served next. We will investigate this effect in Section 7. A third issue is the coordination of job selections to avoid that more than one vehicle selects the same job. Such a scenario where a particular customer, say cn, is selected by K ′ vehicles with 2≤ K ′≤ K, can easily hap- pen if the system is not fully utilized, i.e., for ρ→ 0. FJ- policies are centralized, i.e. there is an inherent coordina- tion mechanism among vehicles. We refer to this mech- anism as assortative: a central entity selects the next job and assigns it to the nearest vehicle. Hence, the selec- tion is based on time arrival and positions of waiting cus- tomers, positions of depots, and positions of all vehicles. NJ-policies are distributed. Every vehicle makes selec- tions based on its own position with respect to the po- sitions of customers and depots, and whenever a vehicle selects a job this is removed form the list of waiting jobs. Every vehicle only knows its own position, the positions of depots, and the positions of waiting customers. For these policies, the coordination mechanism is random- ized: Every time that more than one vehicle has the same time of job selection, a random priority order is assigned which constitutes the minimum infrastructure expenditure necessary for building a stable system that meets a tar- geted performance ¯T . Note that (14) does not constitute the amount of financial resources sufficient to guarantee system stability because the bound on delivery time is not tight for ρ → 1, and, in general, policies do not stabilize the system for all ρ < 1. 5 Job Selection Policies Job selection goes beyond merely picking the next de- mand. It has to specify all decisions needed to operate a delivery system, including which customer demand to serve first, which vehicle to let serve the next customer demand, at which depot to let vehicles load up goods, which paths to let vehicles follow, and where to let ve- hicles return to if no customers are waiting. We analyze two classes of job selection policies: nearest job first (NJ) and first job first (FJ). NJ-policies select jobs based on the location of the customer; FJ-policies select jobs based on the arrival time of the customer's job. By comparing to the vehicles involved. The randomize mechanism cor- responds to a real case where vehicles are not coordinated at all and the connection delay to the list of waiting cus- tomers is not predictable because it depends on the net- work conditions. Table 1 shows the job selection policies used in this ar- ticle and classifies them according to three features: the order of jobs to be done, the timing of job selection, and the coordination of jobs in case of ambiguous selection. 1) Policies in which vehicles select jobs according to their current distance from the customer can be named "near- est job first served" and are called NJ-policies. Policies in which vehicles select the smallest n with cn∈ N(t) (where t indicates the instance of job selection) are named "first job first served" and are called FJ-policies. 2) If ve- hicles make their individual job selection immediately af- ter completion of a service, the policy is marked with sub- script +. If vehicles make their job selection later in time, when they are ready to upload goods at a depot, the policy is marked with a subscript −. 3) The subscripts r and a correspond to random and assortative job coordination in case of an ambiguous job selection process, respectively. The following four policies are used as reasonable com- binations: πNJ −a. A feature included in all policies is the effect of a vehicle vk's battery level, bk(t), on whether or not a vehicle selects a job at time t at all. In particular, if at the instance of job selection the battery level is less than 30 % the vehicle approaches the nearest depot to recharge and does not select a job until its battery level has again reached 80 %. If there is no newly upcoming service request, vk goes on with charging until the battery is full. +a, and πFJ +r , πNJ −r, πFJ 5.1 NJ-Policies With Randomized Job Co- ordination Policy πNJ +r is called Do Nearest Job. After completion of a service, the vehicle vk selects the next customer such that the travel distance to this customer via a generic depot is minimized. In mathematical terms, we solve min l:l∈{1,...,L} n:cn ∈N(τ ) vk(τ ) − dl + dl − cn , (15) where τ is the time instant of completion of the previous job, which is here equal to the time instant of the new job selection. Such selection is made if the battery level at time τ is large enough, i.e., bk(τ )≥ 0.3; otherwise a ve- hicle will return to the nearest depot. Every vehicle has to make L · N(τ ) comparisons to conclude individual job selection. If a job minimizes the traveling distance for more than one vehicle, according to the random coordina- tion described above, only one vehicle randomly chosen will serve the job. Policy πNJ −r is called Rush to Depots. After completion of a service, the vehicle vk first travels to the closest de- pot. The vehicle then selects the nearest customer, i.e., it solves: min n:cn∈N (τ+R)vk(τ +R) − cn, (16) where τ+R is the time instant vk arrives at the depot. The vehicle remains in the depot if there is no waiting cus- tomer. If a job minimizes the traveling distance for more than one vehicle, the supplying vehicle is determined ran- domly. This policy requires L+N(τ + R) comparisons. 5.2 FJ-Policies With Assortative Job Coor- dination Following a first come first served (FCFS) policy, a cen- tral unit always selects the "oldest" demand request still unaddressed, i.e., the smallest n with cn∈ N(t), where t indicates the instance of job selection. Let this demand be called n(t). In particular, policy πFJ +a (also called FCFS by Nearest Vehicle; see Tab. 1) implies that whenever one or more than one vehicles complete a service, at time t = τ , a central controller selects the next customer in line, i.e., demand n(τ ), and chooses the vehicle vk∗ and the path through the depot, dl∗ , that minimizes the total distance between its current position vk(τ ), a depot's position dl, and the customer's position cn(τ ): l:l∈{1,...,L} min k:k∈K′ vk(τ ) − dl + dl − cn(τ ), (17) where K′ is the number of vehicles ready for the service. Obeying (17) a total of K ′ ·L computations is needed to come up with a vehicle's individual selection of whom to next serve. Obeying the rules of policy πFJ −a (or FCFS by First Ve- hicle at Depot; see Tab. 1), whenever one or more than one vehicle is available to a depot, the central unit selects Table 1: Features of Job Selection Policies Features Do Nearest NJ +r ) Job (π Rush to Depots (π NJ −r ) FCFS by Nearest Vehicle (π FJ +a) by First FCFS Vehicle at Depot (π FJ −a) Jobs are selected in first come first served (FCFS) order from a shared queue. There is no specific selection order of jobs. Decisions are made at a customer, just after completing a service. Decisions are made at a depot, just before loading goods. If more than one vehicle seeks to select a job, a random one does. If more than one vehicle seeks to select a job, the nearest one does. If there are no jobs a vehicle goes to the nearest depot. If a vehicle's battery level is less that 30% the vehicle approaches the nearest depot. If due to recharging a vehicle's battery reaches a level higher than 80% the vehicle is ready again for service. If there is no customer the vehicle fully recharges its battery. - Yes Yes - Yes - Yes Yes - Yes - Yes Yes - Yes Yes Yes - Yes - - Yes Yes Yes Yes - - Yes - Yes Yes Yes the demand n(τ + R) and performs K ′ computations to determine the vehicle nearest to that job. Subsequently L computations are performed to find the depot nearest to the selected customer where the vehicle has to return to after the delivery: k:k∈K′ vk(τ +R)−cn(τ +R)+ min min l:l∈{1,...,L}cn(τ +R)−dl. (18) 6 Simulation Setup −r, πFJ +a, πFJ +r, πNJ For π ∈ {πNJ −a}, we simulate the move- ment of K ∈ {1, . . . , 24} vehicles in a square area of A = 16 km2. The L∈ {1, 4, 9, 16} depots are located such that they minimize the average distance of any poten- tial demand in A from the nearest depot. Customer de- mands are assumed to randomly arrive over time at a rate λ = 0.65 requests/minute. We consider the non-variation of λ as feasible, since with K we vary λ's counterpart in affecting the load factor, defined by (7), so to derive suffi- cient information about the performance of the system for varying levels of resource utilization. All vehicles travel at a constant velocity of ν = 30 km/h, neglecting acceleration and deceleration phases and neglecting the extra time for starting and landing. Ve- hicles' air time are limited due to their finite battery ca- pacity, which is assumed to be full at t = 0. We choose the ratio between air time and charge time as being 1/3, i.e., α = 0.25. Moreover, we assume that at t = 0 every vk is placed at one of the depots. Finally, we neglect the loading time (goods). In the simulations, the system presented in Section 3 is observed every δ time units. Interarrival times are gen- erated according to a geometric distribution so that the number of customer arrivals per time unit follows a bino- mial distribution with parameters h = 1/δ and p = λδ. The Poisson distribution is a sufficient approximation of the binomial distribution if p ≤ 0.08 and h≥ 1500p [9]. Therefore, for all of our simulation runs, we choose δ such that δ≤ min{0.08/λ, 1/√1500λ}. The data collected from the simulations are used to compute the average delivery time. Specifically, we firstly estimate the length of the warm-up phase using Welch's method [43]. Secondly, we compute the average deliv- ery time with the replication/deletion method [26], which is equivalent to the independent replications method [43]. These methods are used for statistical analysis of data and enable us to estimate expected values and confidence in- tervals. For every parameter setup, we perform 10 simulations with a minimum of 4,000 customer requests for every simulation. If it is impossible to evaluate whether the sys- tem reaches the steady state within this period, we sim- ulate 10,000 customers requests, which is more than 10 days of continuous operation. 7 Simulation Results 7.1 Warm-Up Phase We discuss the computation of the warm-up phase for the policies Do Nearest Job and FCFS by Nearest Vehi- cle for K ∈ {11, 12, 14, 16} vehicles and L = 16 depots. ) w ( n ¯T , s e t u n m n i i e m i t y r e v i l e d e g a r e v A 6 5 4 3 2 1 ¯Tmin 0 0 Do Nearest Job (πNJ +r ) FCFS by Nearest Vehicle (πFJ +a) K = 11 K = 12 K = 14 K = 16 K = 11 K = 12 K ∈ {14, 16} 500 1,000 1,500 2,000 2,500 3,000 Demand, n Figure 2: Delivery time averaged over 10 simulation runs and 2 w + 1 = 1001 demands vs. demand index n for K vehicles. Fig. 2 reveals that the length of the transient phase de- creases with an increasing number of vehicles, i.e., with a decreasing load factor. We conclude that delivery sys- tems provided with fewer vehicles must be simulated for a longer period (higher number of demands n) to reach steady state conditions. Being conservative, we estimate that the length of the warm-up phase is nwu = 3000 for the case K = 11, while being nwu = 2000 for the case K = 12, and nwu = 500 for all other cases. Moreover, Fig. 2 introduces a result addressed when discussing Fig. 3d below: FCFS by Nearest Vehicle (πFJ +a) can yield a substantially better system performance than Do Nearest Job (πNJ +r). The only drawback observed is that, while πFJ +a gets very close to the minimum average delivery time ¯Tmin for K ∈ {12, 14, 16}, it cannot stabi- lize the system for K = 11. 7.2 Evaluating Job Selection Policies Fig. 3 shows the average delivery time ¯T for different job selection policies as a function of the number of vehi- cles K and the number of depots L. The shaded areas in- dicate "impossible regions", i.e., all ordered pairs (K, ¯T ) for which either ρ > 1 or ¯T < ¯Tmin or both. Depot num- bers are in the set {2n; n = 0, 1, 2, 4} so that the multi- median problem has traceable mathematical solution for a squared service area: The area is divided into L iden- tical squares, and the depots are located in the centroids of the squares. The following basic phenomena can be observed. Increasing the number of ve- Tipping Point Behavior hicles K has no effect on system performance once K exceeds a certain threshold K. This observation holds for all studied policies. It results from the fact that, for any K larger than K, there is a vehicle in the chosen depot ready to serve a new demand, such that the overall delivery time T just consists of the service time S. In such light load conditions, FJ-policies (πFJ −a) yield better system performance than NJ-policies (πNJ −r). Such policy differences in system performance are known from [5, 6], and they are amplified by the differences in job coordina- tion in case of ambiguity. +r and πNJ +a and πFJ Decreasing the number of vehicles K improves re- source utilization but in turn increases the load factor, which eventually implies system instability. The deliv- ery time of Do Nearest Job (πNJ +r) increases slowly with a decreasing K, i.e., the system shows a gradual transi- tion from the region of system underuse to the region of instability. In contrast to this, for all other policies, the system tips into instability all of a sudden. This tipping point behavior of πNJ −a can be explained by the fact that these policies resemble the behavior of the K stochastic queue median (K-SQM) policy introduced by [6, p. 65], while πNJ +r in heavy load behaves similar to the nearest neighbor (NN) policy presented by [5, p. 612]. +a, and πFJ −r, πFJ Timing Matters The timing of the job selection deci- sion has an impact on delivery time and system stability. Let us compare two extreme cases for both FJ and NJ- policies: the decision is made as soon as possible (just upon the service at the previous customer), or the deci- sion is made as late as possible (just before loading the good at the depot). Using FJ-policies, timing has no effect on delivery time and system stability (see Fig. 3). In contrast, using NJ- policies, an early decision is beneficial for stability, i.e., Do Nearest Job πNJ +r is more robust than Rush To Depots πNJ −r. The positive effect of more recent information out- weighs the negative effect of returning to a depot located sub-optimally (on average) for the next job. The gap be- tween the two NJ-policies widens with more depots, i.e., the effect of "being at a suboptimal place" is more pro- nounced in systems with many depots. This happens be- ¯T , s e t u n m n i i e m i t y r e v i l e d e g a r e v A ¯T , s e t u n m n i i e m i t y r e v i l e d e g a r e v A 20 15 10 Do Nearest Job (πNJ +r ) Rush to Depots (πNJ −r ) FCFS by Nearest V. (πFJ +a) FCFS by First V. at Depot (πFJ −a) 5 ¯Tmin 0 0 4 8 Number of vehicles, K 12 16 ¯T , s e t u n m n i i e m i t y r e v i l e d e g a r e v A 20 15 10 5 ¯Tmin 0 Rush to Depots (πNJ −r ) FCFS by First V. at Depot (πFJ −a) Do Nearest Job (πNJ +r ) FCFS by Nearest V. (πFJ +a) 20 24 0 4 8 Number of vehicles, K 12 16 (a) Number of depots L = 1 (b) Number of depots L = 4 20 15 10 5 ¯Tmin 0 Do Nearest Job (πNJ +r ) Rush to Depots (πNJ −r ) FCFS by Nearest V. (πFJ +a) FCFS by First V. at Depot (πFJ −a) ¯T , s e t u n m n i i e m i t y r e v i l e d e g a r e v A 20 15 10 5 ¯Tmin 0 Do Nearest Job (πNJ +r ) Rush to Depots (πNJ −r ) FCFS by Nearest V. (πFJ +a) FCFS by First V. at Depot (πFJ −a) 0 4 8 Number of vehicles, K 12 16 20 24 0 4 8 Number of vehicles, K 12 16 20 24 20 24 (c) Number of depots L = 9 (d) Number of depots L = 16 Figure 3: Average delivery time (with 90% confidence) vs. number of vehicles for L depots. cause, in heavy load conditions, randomized job coordina- tion is rarely necessary (if at all), and πNJ +r behaves similar to a NN policy (as mentioned above), which serves the nearest demand available after every service completion. In summary, the timing of job selection has fundamen- tal impact on the qualitative behavior of NJ-policies: it changes the vulnerability with respect to variation in cus- tomer demand. This phenomenon does not occur for our FJ-policies: both maintain structurally equivalent to K- SQM policies. 7.3 Decision Making Support We have seen that the number of depots must be chosen in relation to the size of the service area, and that this long- term choice on depot infrastructure must be coordinated with the mid-term choice on vehicles and the short-term choice on the job selection policy. For company-specific parameter values, a diagram like Fig. 4 translates the in- sights derived in the subsection above into the monetary domain. It relates a company's expenditure I for depots and vehicles to average delivery time ¯T . The purpose of such a plot is to provide decision making support for com- panies that set up an airborne delivery system equipped with small UAVs. To give an example, Fig. 4 is produced on the as- sumption that the cost of a UAV suitable to deliver two- kilogram packages is 1,000 US$ plus a maintenance cost of 100 US$ per annum, and the cost of a depot is 15,000 US$ plus a maintenance cost of 500 US$ per annum. Op- erating the system over ten years, the costs per vehicle and depots are Cv = 2,000 US$ and Cd = 20,000 US$, respec- tively. These parameter values are those assumed by the ¯T 10 , s e t u n m n i i e m i t y r e v i l e d e g a r e v a d e t e g r a T 9 8 7 6 5 4 3 2 1 0 Serve Closest Demand (πNJ +r ) FCFS by First Vehicle at Depot (πFJ −a) Lower bound on infrastructure expenditure (Imin) L = 1 L = 4 L = 9 L = 16 L = 1 2 3 4 5 · · · 9 · · · 16 · · · 0 50 100 150 200 250 300 350 400 Expenditure for infrastructure in 1,000 US$ Figure 4: Relation between infrastructure expen- diture and average delivery time; Cv = 2, 000 US$, Cd = 20, 000 US$. company Matternet [36]. A lower bound on the expenditure required to build a stable system, denoted by Imin, is derived in Sec. 4 and is given by (14). Fig. 4 plots this bound for the given pa- rameters, where all values below the bound are shown as a shaded area. The bound has a "staircase" shape with tread levels being the minimum average delivery time achiev- able with a particular number of depots. No operable sys- tem exists for parameters in this area, while every combi- nation of I and targeted ¯T located above fulfills ρ < 1 and ¯T > ¯Tmin. The bound corresponds to a service possibility frontier; it gives a necessary but not a sufficient condition for infrastructure expenditure. The actual performance is, policy depended, and more financial resources than Imin may be needed to operate the system in a stable manner and to meet the targeted performance. A company that wants to operate a delivery service and serve a customer within a certain average delivery time can employ such a diagram as follows: If the average de- livery time should be no more than τ , the company has to look for feasible combinations of infrastructure and sta- bilizing policies, i.e., squares and triangles, that are lo- cated as close as possible to the origin and below the ( ¯T = τ )-line. If the customers' willingness to pay for several levels of service quality is given, it is possible to quantify the company's marginal revenues of increasing performance. From Fig. 4 we know the marginal cost of decreasing delivery time. Then, we are able to determine the combination of infrastructure and job selection policy that maximizes the company's profit. For the parameters of Matternet we find that a system with L = 1 depot can serve a customer in about three minutes on average in an area of 16 km2, which comes at a cost of 60,000 US$. This time can be reduced to less than 1.5 minutes if the company spends more than 100,000 US$ for infrastruc- ture (associated with L = 4 depots). If the company's marginal revenue of reducing delivery time by 1.5 min- utes (50 %) is larger than approximately 40,000 US$, the four-depot configuration is better than the one-depot con- figuration. In other words, Fig. 4 informs about the finan- cial resources required for achieving a certain quality of service with a certain policy and about the volume of ad- ditional monetary resources obligatory to "buy" a shorter delivery time. 8 Conclusions We addressed the architectural setup and mechanisms of self-control in a network of UAVs for delivery of goods. We combine short-term decisions (e.g., job selection pol- icy), mid-term decisions (e.g., number of vehicles), and long-term decisions (e.g., number of depots) into an in- tegrated decision making model. Taking into account the short and mid-term horizons, the delivery time of the sys- tem experiences a threshold behavior: a stable system could lead to instability all of a sudden if one vehicle fails; adding vehicles has almost no positive influence on deliv- ery time once the number of vehicles exceeds a certain threshold. The timing of job selection has significant im- pact on delivery time and stability for NJ-policies while it has no relevance for FJ-policies. We also introduced a methodological approach for decision making support to setup a stable delivery system for efficiently resolving the tradeoff between expenditure and service quality. This approach reflects all three time horizons and includes an analytically derived service possibility frontier. Acknowledgement This work was funded in part by ERDF, BABEG, and KWF under grant number KWF-20214/23793/35529. It was carried out in the SOSIE project by Lakeside Labs GmbH. References [1] Torsten Andre, Karin A. Hummel, Angela P. Schoel- lig, Evsen Yanmaz, Mahdi Asadpour, Christian Bettstetter, Pasquale Grippa, Hermann Hellwagner, Stephan Sand, and Siwei Zhang. Application-driven design of aerial communication networks. IEEE Commun. Mag., 52(5):129–137, May 2014. [2] Enrico Angelelli and Maria G. Speranza. A periodic vehicle routing problem with intermediate facilities. European J. Oper. Res., 137:233–247, Mar. 2002. [3] Alessandro Arsie, Ketan Savla, and Emilio Fraz- zoli. Efficient routing algorithms for multiple ve- hicles with no explicit communication. IEEE Trans. Autom. Control, 54(10):2302–2317, Oct. 2009. [4] Roberto Baldacci, Paolo Toth, and Daniele Vigo. Recent advances in vehicle routing exact algorithms. 4OR, 5(4):269–298, Dec. 2007. [5] Dimitris J. Bertsimas and Garrett J. van Ryzin. A stochastic and dynamic vehicle routing problem in the Euclidean plane. Oper. Res., 39(4):601–615, July/Aug. 1991. [6] Dimitris J. Bertsimas and Garrett J. van Ryzin. Stochastic and dynamic vehicle routing in the Eu- clidean plane with multiple capacitated vehicles. Oper. Res., 41(1):60–76, Jan./Feb. 1993. [7] Dimitris J. Bertsimas and Garrett J. van Ryzin. Stochastic and dynamic vehicle routing with general demand and interarrival time distributions. Adv. in Appl. Probab., 25:947–978, Dec. 1993. [8] S. D. Bopardikar, Stephen L. Smith, Franceso Bullo, and J. P. Hespanha. Dynamic vehicle rout- ing for translating demands: Stability analysis and receding-horizon policies. IEEE Trans. Autom. Con- trol, 55(11):2554–2569, Nov. 2010. [9] Ilja N. Bronshtein, Konstantin A. Semendyayev, Gerhard Musiol, and Heiner Muhlig. Handbook of Mathematics. Springer, Berlin, DE, 5 edition, 2007. [10] Francesco Bullo, Emilio Frazzoli, Marco Pavone, Ketan Savla, and Stephen L. Smith. Dynamic ve- hicle routing for robotic systems. Proc. IEEE, 99 (9):1482–1504, Aug. 2011. [11] Jean-Franc¸ois Cordeau, Guy Desaulniers, Jacques Desrosiers, Marius M. Solomon, and Franc¸ois Soumis. The VRP with time windows. In The Vehicle Routing Problem, chapter 7, pages 157– 172. Society for Industrial and Applied Mathemat- ics, Philadelphia, PA, 2002. [12] Jean-Franc¸ois Cordeau, Gilbert Laporte, Mar- tin W.P. Savelsbergh, and Daniele Vigo. Vehicle routing. In Cynthia Barnhart and Gilbert Laporte, editors, Transportation, volume 14 of Handbooks Oper. Res. Management Sci., chapter 6, pages 367– 428. Elsevier B.V. (North Holland), Amsterdam, NL, 1 edition, 2007. [13] George B. Dantzig John H. Ramser. The truck dispatching problem. Manage. Sci., 6(1):81–90, Oct. 1959. and [14] Guy Desaulniers, Jacques Desrosiers, Andreas Erd- mann, Marius M. Solomon, and Franc¸ois Soumis. The VRP with pickup and delivery. In The Vehicle Routing Problem, chapter 9, pages 225–242. Society for Industrial and Applied Mathematics, Philadel- phia, PA, 2002. [15] Moshe Dror, Gilbert Laporte, and Pierre Trudeau. Vehicle routing with split deliveries. Discrete Appl. Math., 50(3):239–254, May 1994. [16] John J. Enright, Ketan Savla, Emilio Frazzoli, and Francesco Bullo. Stochastic and dynamic rout- ing problems for multiple uninhabitated aerial vehi- cles. J. of Guid. Control Dynam., 32(4):1152–1168, July/Aug. 2009. [17] Merrill M. Flood. The traveling-salesman problem. Oper. Res., 4(1):61–75, Feb. 1956. [18] Emilio Frazzoli and Francesco Bullo. Decentralized algorithms for vehicle routing in a stochastic time- varying environment. In Proc. IEEE Conf. Deci- sion and Control, pages 3357–3363, Paradise Island, BHA, Dec. 2004. [19] Bruce L. Golden and Arjang A. Assad, editors. Ve- hicle Routing: Methods and Studies. Elsevier B.V. (North Holland), Amsterdam, NL, 1988. [20] Leonard Kleinrock. Queueing Systems, Volume I: Theory. Wiley Interscience, New York, 1975. [21] Michael A. Kovacina, Daniel Palmer, Guang Yang, and Ravi Vaidyanathan. Multi-agent control algo- rithms for chemical cloud detection and mapping us- ing unmanned air vehicles. In Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pages 2782– 2788, Lausanne, CH, Oct. 2002. [22] Gilbert Laporte. The vehicle routing problem: An overview of exact and approximate algorithms. Eu- ropean J. Oper. Res., 52(3):345–358, June 1992. [23] Gilbert Laporte. What you should know about the vehicle routing problem. Nav. Res. Log., 54(8):811– 819, Dec. 2007. [24] Gilbert Laporte. Fifty years of vehicle routing. Transport. Sci., 43(4):408–416, Nov. 2009. [25] Gilbert Laporte and Ibrahim H. Osman. Routing problems: A bibliography. Ann. Oper. Res., 61(1): 227–262, 1995. [26] Averill M. Law. Simulation Modeling and Analysis. McGraw-Hill, Boston, MA, 4 edition, 2007. [27] Pedro Lima, Dario Floreano, Felix S. Schill, and Meysam Basiri. Audio-based localization system for swarms of micro air vehicles. In Proc. IEEE Int. Conf. Robotics and Automation, page forthcoming, Hong Kong, CHN, June 2014. [30] Marco Pavone, Alessandro Arsie, Emilio Fraz- zoli, and Francesco Bullo. Distributed algorithms for environment partitioning in mobile robotic net- works. IEEE Trans. Autom. Control, 56(8):1834– 1847, Aug. 2011. [31] Marco Pavone, Emilio Frazzoli, and Francesco Bullo. Adaptive and distributed algorithms for ve- hicle routing in a stochastic and dynamic environ- ment. IEEE Trans. Autom. Control, 56(6):1259– 1274, June 2011. [32] Victor Pillac, Michel Gendreau, Christelle Gu´eret, and Andr´es L. Medaglia. A review of dynamic ve- hicle routing problems. European J. Oper. Res., 225 (1):1–11, Feb. 2013. [33] Harilaos N. Psaraftis. Dynamic vehicle routing problems. In Vehicle Routing: Methods and Stud- ies, pages 223–248. Elsevier B.V. (North Holland), Amsterdam, NL, 1988. [34] Harilaos N. Psaraftis. Dynamic vehicle routing: Sta- tus and prospects. Ann. Oper. Res., 61(1):143–164, 1995. [35] Markus Quaritsch, Emil Stojanovski, Christian Bettstetter, Gerhard Friedrich, Herrmann Hellwag- ner, Bernhard Rinner, Michael Hofbaur, and Mubrak Shah. Collaborative microdrones: Applications and research challenges. In Proc. Int. Conf. Autonomic Compu. and Commun. Sys., pages 38:1–38:7, Turin, IT, Sept. 2008. [36] Andreas Raptopoulos. Physical (Solve for X talk), Feb. 2012. port. https://www.youtube.com/watch?v=7B3E4OOuGDk. trans- URL [28] Christos H. Papadimitriou. Worst-case and prob- abilistic analysis of a geometric location problem. SIAM J. Comput., 10(3):542–557, 1981. [29] Marco Pavone, Nabhendra Bisnik, Emilio Frazzoli, and Volker Isler. A stochastic and dynamic vehi- cle routing problem with time windows and cus- tomer impatience. Mobile Netw. Appl., 14(3):350– 364, June 2009. [37] Ketan Savla, Emilio Frazzoli, and Francesco Bullo. Traveling salesperson problems for the dubins vehi- cle. IEEE Trans. Autom. Control, 53(6):1378–1391, July 2008. [38] Stephen L. Smith, Marco Pavone, Francesco Bullo, and Emilio Frazzoli. Dynamic vehicle routing with priority classes of stochastic demands. SIAM J. Con- trol Optim., 48(5):3224–3245, Jan. 2010. [39] Michael R. Swihart and Jason D. Papastavrou. A stochastic and dynamic model for the single-vehicle pick-up and delivery problem. Eur. J. Oper. Res., 114(3):447–464, 1999. [40] Paolo Toth and Daniele Vigo. Models, relaxations and exact approaches for the capacitated vehicle routing problem. Discrete Appl. Math., 123(1-3): 487–512, Nov. 2002. [41] Paolo Toth and Daniele Vigo, editors. The Vehicle Routing Problem. Society for Industrial and Applied Mathematics, Philadelphia, PA, 2002. [42] Holly A. Waisanen, Devavrat Shah, and Munther A. Dahleh. A dynamic pickup and delivery problem in mobile networks under information constraints. Automatic Control, IEEE Transactions on, 53(6): 1419–1433, July 2008. [43] Peter D. Welch. The statistical analysis of simula- tion results. In Stephen S. Lavenberg, editor, Com- puter Performance Modeling Handbook, chapter 6. Academic Press, Inc., New York, NY, 1983. [44] Eitan Zemel. Probabilistic analysis of geometric lo- cation problems. Ann. Oper. Res., 1(3):215–238, 1984.
1802.07187	1	1802	2018-02-18T04:05:41	Leader-follower based Coalition Formation in Large-scale UAV Networks, A Quantum Evolutionary Approach	[ "cs.MA", "cs.GT" ]	The problem of decentralized multiple Point of Interests (PoIs) detection and associated task completion in an unknown environment with multiple resource-constrained and self-interested Unmanned Aerial Vehicles (UAVs) is studied. The UAVs form several coalitions to efficiently complete the compound tasks which are impossible to be performed individually. The objectives of such coalition formation are to firstly minimize resource consumption in completing the encountered tasks on time, secondly to enhance the reliability of the coalitions, and lastly in segregating the most trusted UAVs amid the self interested of them. As many previous publications have merely focused on minimizing costs, this study considers a multi-objective optimization coalition formation problem that considers the three aforementioned objectives. In doing so, a leader-follower- inspired coalition formation algorithm amalgamating the three objectives to address the problem of the computational complexity of coalition formation in large-scale UAV networks is proposed. This algorithm attempts to form the coalitions with minimally exceeding the required resources for the encountered tasks while maximizing the number of completed tasks. The proposed algorithm is based on Quantum Evolutionary Algorithms(QEA) which are a combination of quantum computing and evolutionary algorithms. Results from simulations show that the proposed algorithm significantly outperforms the existing coalition formation algorithms such as merge-and-split and a famous multi-objective genetic algorithm called NSGA-II.	cs.MA	cs	Leader-follower based Coalition Formation in Large-scale UAV Networks, A Quantum Evolutionary Approach 8 1 0 2 b e F 8 1 ] A M . s c [ 1 v 7 8 1 7 0 . 2 0 8 1 : v i X r a Sajad Mousavi∗, Fatemeh Afghah∗, Jonathan D. Ashdown† and Kurt Turck† ∗School of Informatics, Computing and Cyber Systems, Northern Arizona University, Flagstaff, AZ, United States †Air Force Research Laboratory, Rome, NY, United States Email: {sajadmousavi, fatemeh.afghah}@nau.edu Email: {jonathan.ashdown, kurt.turck}@us.af.mil Abstract-The problem of decentralized multiple Point of Interests (PoIs) detection and associated task completion in an unknown environment with multiple resource-constrained and self-interested Unmanned Aerial Vehicles (UAVs) is studied. The UAVs form several coalitions to efficiently complete the compound tasks which are impossible to be performed individually. The objectives of such coalition formation are to firstly minimize resource consumption in completing the encountered tasks on time, secondly to enhance the reliability of the coalitions, and lastly in segregating the most trusted UAVs amid the self interested of them. As many previous publications have merely focused upon minimizing costs, this study considers a multi- objective optimization coalition formation problem that considers the three aforementioned objectives. In doing so, a leader- follower-inspired coalition formation algorithm amalgamating the three objectives to address the problem of the computational complexity of coalition formation in large-scale UAV networks is proposed. This algorithm attempts to form the coalitions with minimally exceeding the required resources for the encountered tasks while maximizing the number of completed tasks. The proposed algorithm is based on Quantum Evolutionary Algo- rithms (QEA) which are a combination of quantum computing and evolutionary algorithms. Results from simulations show that the proposed algorithm significantly outperforms the existing coalition formation algorithms such as merge-and-split and a famous multi-objective genetic algorithm called NSGA-II 1. Keywords-Unmanned aerial vehicles, coalition formation, mis- sion completion, evolutionary algorithms. I. INTRODUCTION A single agent system is often unable to perform complex tasks considering the limited individual capabilities of such an agent. Therefore, cooperative multi-agent systems (MASs) can offer a practical solution to this problem by ensembling a complementary set of different capabilities/resources from several agents. However, one of the key challenges in such cooperative MASs is forming optimal sub-groups of agents (i.e., coalition formation) in order to efficiently perform the existing tasks, especially in a distributed case where no central controller is available. That being said, the coalition formation problem concerns how different coalitions can be formed considering the tasks' requirements and the agents' capabilities so much so that the collective goals of the tasks are reached in the most effective manner. 1DISTRIBUTION A. Approved for public release: distribution unlimited. Case Number: 88ABW-2018-0096 . Dated 10 Jan 2018. A considerable amount of research has been recently carried out in solving coalition formation problems. This has spawned several classic methods to form stable coalitions that follow common stability concepts based on Core, Shapley value, Bargaining Set, and Kernel [1], [2], [3]. However, achieving such stability concepts often mandates high compu- tational complexity. Many researchers have attempted to deal with the problem of coalition formation in multi-agent systems by applying various approaches including genetic algorithms [4], dynamic programing methods [5], graph theory [6], [7], iterative processes [8], cooperative Multi-Agent Reinforcement Learning (MARL) [9], and temporal-spatial abstraction MARL [10], [11], [12]. The majority of these reported studies have focused mainly in reducing costs and only a few of them have focused upon multiple objectives in solving the problem. In this paper, a solution for coalition formation problem in a heterogeneous resource-constraint UAV network in which multiple objectives are considered to form the optimal coali- tions is proposed. The objectives of the coalition formation method proposed here are : minimizing the cost associated with consumption of resources of the coalitions formed, maximizing the reliability of the formed coalitions, and lastly to select the most trustworthy of UAVs among the available self-interested UAVs in the network. Finding the solution of such a multi-objective coalition formation problem involves an NP-hard problem. Many ap- proaches such as mixed integer linear programming [13], [14] and dynamic network flow optimization [15] have been utilized to provide an exact solution to this problem. However, since these approaches seek such a solution, applying them to a large-scale problem is computationally taxing. More recently, metaheuristic algorithms such as Particle Swarm Optimization (PSO) [16], Ant Colony Optimization [17], Genetic Algo- rithms (GAs) [18], and Simulated Annealing (SA) [19] have offered reasonable solutions in efficient times for a variation of multi-objective optimization problems. Inspired by the success of evolutionary approaches, this paper presents a novel leader- follower-based coalition formation algorithm using a quantum evolutionary approach whilst considering the aforementioned objectives in addressing the problem. As an unknown, dynamic environment is assumed where no prior knowledge is available about targets or Point of Interests (PoIs) and UAVs need to gain the knowledge of environment dynamically [20], [21]. Some potential applications of the proposed method are search and rescue, humanitarian relief, and public safety operations in unknown remote environments or military operations in remote fields. In such environments, it can be safely assumed that ground station does not have prior information about the PoIs and their positions. To evaluate the performance of the proposed method in such an unknown environment, several scenarios with different numbers of tasks ranging from 4 to 24, and a heterogeneous network of UAVs consists of a different number of UAVs ranging from 8 to 124 were simulated. The rest of the paper is organized as follows: Section II presents the problem statement and the formulation of the multi-UAV coalition formation as a multi-objective optimiza- tion problem. Section III describes the proposed coalition formation algorithm. Section IV reports the simulation results followed by concluding remarks in section V. II. PROBLEM STATEMENT ui , r2 ui , . . . , rNr ui A heterogeneous In this section, the decentralized task allocation problem in a network of autonomous UAVs by forming optimum coalitions are formulated. The possibility of selfish behavior of these self-interested UAVs are accounted for and reputation guidelines in selecting the most reliable of UAVs to participate in the formed coalitions are also defined. It is vital to note that the proposed algorithm considers minimizing the cost of coalition formation in terms of overspending the resources on particular tasks as well as enhancing the reliability of the formed coalitions as paramount. network of N UAVs U each UAV, ui } denotes the set of = {u1, u2, . . . , uN}, where can carry a different set of resources compared to another is considered. Rui = {r1 resources available at UAV, ui where Nr is the number of possible resources in the network. It is also assumed that there exists n tasks in the environment T = {T1, T2, . . . , Tn}. Each task, Ti requires a certain amount available resource in each of Rui in order to be completed. The vector of required resources for task i is defined as follows: 2, τ i (1) where τ i j is the required amount of resource j for task i. It is assumed that each task is associated to one PoI (target) and the PoIs can be located in different positions with a diverse set of resource requirements. All UAVs are able to search the unknown environment for new PoIs. The tasks are carried out by the formed coalitions S = {S1, S2, . . . , Sm}, where each coalition Si is responsible for one task. A large search space where the PoIs are distributed far apart from each other are considered. It can therefore be assumed that the formed coalitions are sufficiently far from one another that each UAV can only be a member of a single coalition, i.e., Sk ∩ Sl = ∅,∀Sk, Sl ∈ S. This also means that the coalitions are non-overlapping. The capability of each coalition Si to complete its encountered task is defined as the value of coalition v(Si), (v(Si) ∈ R) as described in the next section. Moreover, cost(Si, Ti) is defined as the cost of coalition Si in performing task Ti. The cost function for coalition Si captures the cost imposed on all UAV members of this coalition cost(Si, uj)) in which their resources have been (i.e.,(cid:80) τ i = {τ i 3, . . . , τ i Nr 1, τ i }, uj∈Si shared to accomplish task Ti. Another key contribution of the proposed model is consid- ering the reliability factor in forming the optimal coalitions. While in the majority of existing techniques, it is assumed that the UAV members of formed coalitions are perfectly operational during the mission lifetime, this is obviously not a realistic assumption as the UAVs' operation can be interrupted for several reasons (e.g., exhaustion of battery or a particular resource). A practical scenario is considered where the UAVs are assumed to be either fully operational or one of their capabilities (e.g., resources) are bound to fail during the mission. Such failures in various capabilities are considered to be statistically independent. Furthermore, involving different types of UAVs in a coalition may result in different execution times of accomplishing the sub-tasks as the UAVs have a different set of capabilities. For instance, a given UAV may be able to fulfill its duty in a shorter time than another. The cost and reliability of a coalition indeed depend on these execution times where lower execution times are favored for incurring lower execution costs. In the next section, we define and formulate these factors (i.e., cost, reliability and reputation). A. Definition of Cost, Reliability, and Reputation A set of UAVs in the form of a coalition collaborate with one another to carry out the encountered task. Participation in such coalitions involves a cost of sharing and consuming the resources for the member UAVs. For a given task Tk }, where with the required resources τ k = {τ k i ≥ 0, the amount of resource i for task k is denoted by τ k the execution cost of consuming resource j of UAV i is ∀i, j and µj is a constant denoted by eij, where eij = µjrj ui coefficient in order to convert the amount of resource j to a time dependent value to have the same unit as the execution time. Thus, the cost of coalition Sk, C(Sk) can be calculated as follows: 3 , . . . , τ k Nr 1 , τ k 2 , τ k C(Sk) = eijkij + aij, (2) i=1 j=1 where kij is the execution time if UAV i carrying resource j is involved in task k, and aij is the travel time of UAV i to task j. Possibility of potential defects in the UAVs' resources that may result in performance failure of these UAVs during the mission is also accounted for. Considering so, the reliability of coalition Sk for a give task Tk, denoted by R(Sk) is defined as follows: e−(cid:80)Nr j=1 λij kij , R(Sk) = (3) i=1 is the failure rate of ln(R(Sk)) = −(cid:80)Nk where λij resource j of UAV i. For simplicity, loge transfer of R(Sk) function as follows, j=1 λijkij is used. The formulation of the reliability has been inspired by works reported in [22] and [23]. Interested readers are referred to these papers for more details on the probability that a system can accomplish a particular task without failure. Similar to other cognitive agents, the UAVs are expected to be self-interested in the sense that they prefer to save their limited resources, and act selfishly by not consuming enough resources during the mission. To monitor the cooperative behavior of these UAVs, an accumulative cooperative reputation related to the amount of resources that each UAV shares during the mission is defined. During each mission t (i.e., accomplishing an assigned task), the cooperative reputation of each UAV i, ρi is updated as follows: ρt i = i, i + ∆ρt ρt−1 i , ∃kui ∈ Sk otherwise (4) (cid:26) ρt−1 Nk(cid:88) Nr(cid:88) Nk(cid:89) (cid:80)Nr i=1 Table I: Notations aij eij kij λij ρi Nk Nr N Travel time of UAV i to task j. Execution cost of involving resource j of UAV i in a coalition. It is computed per unit time. Execution time of involving resource j of UAV i in a coalition. It depends on the task and capability of the UAV. Failure rate of involving the resource j of UAV i in a coalition. Credit of UAV i. Number of UAVs in coalition k. Number of network resources. Number of network UAVs. i is the amount of contribution of UAV i to coalition where ∆ρt Sk in terms of sharing resources to carry out the assigned task k and can be defined as follows: Υk(cid:80) m∈Sk fi, fm ∆ρt i = (5) denoting as Υk =(cid:80)Nr contributions of UAV i, defined as fi =(cid:80)Nr where Υk is the sum of the resource requirements of task Tk j and fi is the sum of the resource . As such, the coalition reputation of all involved UAVs in the coalition Sk is computed as follows: j=1 τ k rj ui τ k j j=1 Nk(cid:88) Nk(cid:88) P (Sk) = ρt i (6) i=1 B. Formulation of Multi-objective Optimization To consider all aforementioned optimization criteria including reducing the coalition cost, and increasing the reliability, and reputation of the formed coalitions, Multi- Objective Optimization Problem (MOOP) as a weighted-sum of three objectives is defined. The multi-objective optimization and its required constraints are defined as follows. min O(Sk) = C(Sk) − η1 ln R(Sk) − η2P (Sk) subject to eij ≥ τ k j ∀j = 1, 2, 3, . . . , Nr, i=1 where η1 and η2 are weighting parameters to assign the desired importance to each objective and scale them to be in comparative ranges. Constraint (8) refers to the requirement to secure enough resources in the formed coalition to complete the encountered task Tk. In table I, a summary of the notations used throughout this paper is presented. III. PROPOSED ALGORITHM The objective function described in section II-B is a NP- hard problem. Standard approaches such as dynamic program- ming and exact algorithms involve computational complexity of O(n2) that could be intractable in large-scale networks. Hence, evolutionary algorithms such as genetic algorithm can be considered as potential options to find feasible solutions of this problem. In this paper, we propose a coalition for- mation algorithm based on a version of genetic algorithm called Quantum-Inspired Genetic Algorithm (QIGA) to find the solution of the multi-objective problem in (7). A. Review of Quantum-Inspired Genetic Algorithm (QIGA) The idea behind QIG algorithms is to take advantage of both GA and quantum computing mechanisms [22]. In quantum computation, the data representation is based on qubit is considered as a superposition of two different states 0(cid:105) and 1(cid:105) that can be denoted as: information unit. A qubit is the smallest that ψ(cid:105) = α0(cid:105) + β1(cid:105), (9) where α and β are complex numbers such that α2 +β2 = 1. α2 and β2 are the probability of amplitudes where the qubit can be at states 0(cid:105) and 1(cid:105), respectively. With m the model can represent 2m independent states. qubits, However, when the value of the qubit is measured, it leads 0(cid:105) or 1(cid:105)). to a single state of the quantum state (i.e., Similar to a standard genetic algorithm, a chromosome's representation is defined as a string of m information units. Thus, a chromosome can be defined as a string of m qubits as: (cid:26)(cid:18)α0 (cid:19) (cid:18)α1 (cid:19) (cid:18)α2 (cid:19) (cid:18)αm−1 (cid:19)(cid:27) , (10) where each pair (αi, βi), i = 1, 2, 3, . . . , m− 1 denotes a gene of the chromosome. βm−1 , . . . , β0 β1 β2 , , To evaluate a quantum chromosome, a transfer function called measure is applied to convert the quantum states to a classical chromosome representation. For example, each pair (αi, βi) is converted to a value ci ∈ {0, 1} so that the m-qubit chromosome may result in a binary string {c1, c2, c3, . . . , cm}. More specifically, each pair becomes 0 or 1 using the corre- sponding qubit probabilities αi2 and βi2. To perform this conversion, we use the measure function defined as follow: function MEASURE(α) set r to a random number between 0 and 1; if r > α2 then else return 0; Each binary string is a possible solution which is evaluated via a problem dependent fitness function. In the following section, the proposed fitness function is described. B. Fitness function The fitness function or evaluation function determines how to fit a solution with respect to the constrained optimization problem. To build the fitness function, the negative sign of the objective function O(Sk), defined in (7), is used. In addition, as the solutions which meet all of the constraints get higher fitness value, the solutions which violate some of the constrains should achieve a lower objective value from the fitness function. To prevent potential violations, a penalty function technique that penalizes the solutions according to amount of constraints' mismatches is applied. Considering these facts, the fitness function is proposed as: F (Sk) = −(O(Sk) + g(Sk)), (11) where g(Sk) is the penalty function such that if there is no violation, its value will be zero, and positive otherwise. g(Sk) (7) (8) return 1; end if end function is defined as: g(Sk) = γ × Nr(cid:88) j − Ns(cid:88) max(0, τ k i=1 i=1 eij), (12) where γ is the penalty coefficient that controls the weight of the amount of constraints violated. Evolutionary strategies (e.g., crossover and mutation op- erations) are often considered in GA algorithms to improve their performance. However, as stated in [22], applying the crossover and mutation operators do not significantly improve the performance of the QIGA. Therefore, in the QIGAs, usually a qubit rotation gates strategy is used. Qubit (αi, βi) of the m-qubit chromosome is updated according to the rotation gates in order to get more or less probabilities to states 0(cid:105) and 1(cid:105). Therefore, at each time step t, the update of qubit is performed based on the following rotation gate matrix L(θi): (cid:18)cos(θi) −sin(θi) (cid:19) (cid:18)αt (cid:19) (cid:18)αt−1 (cid:19) ; cos(θi) sin(θi) L(θi) = , (13) where θi is the amount of angle rotation of qubit gate i. Algorithm 1 presents the pseudo-code for QIGA as described in [22]. = L(θi) βt−1 i βt i i i the relative significance of the travel time compared to the expected credit. Algorithm 2 shows the pseudo-code for the multi-UAV coalition formation. Algorithm 2 Multi-UAV coalition formation using the leader- follower method and QIGA 1: Search PoIs in the search space 2: Initialize each leader as a singleton coalition committed to 3: coalition members ← [ ] 4: while there exists an idle UAV around do 5: 6: coal mems ← execute MOQIG method regarding for all unsatisfied PoIs so far do a single PoI algorithm 1 coalition members.append(coal mems) end for Send bids to UAVs as potential followers Calculate the utility values of the followers and receive Update coalition members of each leader with respect 7: 8: 9: 10: 11: bid responses to the bid responses 12: end while Algorithm 1 Quantum-Inspired Genetic Algorithm 1: t ← 0 2: Initialize Q(t) as the population 3: Make P (t) by measuring Q(t) 4: Evaluate P (t) 5: Store the best solution b among P (t) 6: repeat 7: 8: 9: 10: 11: 12: until the termination-condition t ← t + 1 Make P (t) by measuring Q(t − 1) Evaluate P (t) Update Q(t) using quantum gates L(t) Store the best solution b among P (t) C. Multi-UAV Coalition Formation To establish a multi-UAV coalition, a leader-follower coali- tion formation method is followed. Initially, the UAVs are uniformly distributed in a search space to look for the PoIs. When a PoI is detected by a UAV, this UAV computes the resource requirements of the detected PoI and serves as a leader to form an optimal coalition. After calculating the required resources, the leader UAV calls other UAVs within a certain distance to join it in forming a coalition. Then, the UAVs with at least one of the required resources can respond to this call by reporting the amount of resources that they are able to contribute to help accomplish the task. It is also assumed that the UAVs are self-interested, meaning that if a UAV receives multiple requests, it will consider joining the coalition which offers the highest benefit. The UAV i, ui measures the value of each request based on travel time to reach the task and the expected cooperative reputation credit received. Thus, its utility value can be defined as: U (ui, Sk) = ρi − δaij, (14) where ρi and ai are the cooperation credit and travel time of ui when it attempts to join coalition k. δ is the weight indicating IV. EXPERIMENTAL RESULTS To evaluate the performance of the proposed QIGA-based coalition formation method, several scenarios with different number of UAVs and PoIs were simulated. It is assumed that the UAVs and PoIs are uniformly distributed across the region and the closest UAV to the PoI is considered to be the one that first detects the PoI and form a coalition (as a coalition leader). It is also assumed that each UAV has five different types of resources. The values for the UAVs' resources are generated with a random uniform distribution and the UAV's resource failure rates are produced randomly in the range (5× 10−5, 10−4). Furthermore, the execution times of the identified tasks are computed randomly in the range between 10 and 20, depending on the task and the capability of the UAV. The performance of our proposed algorithm is compared with three well-known algorithms including: i) the distance- based coalition formation method in which the coalitions are formed with the closest UAVs to the leader (i.e., the leader only considers the UAVs in a certain distance of the PoI to be in the coalition and do not evaluate them in terms of cooperative reputation or available resources), ii) the common merge- and-split coalition formation [24], and iii) a Non-Dominated Sorting Genetic Algorithm (NSGA-II) which is a fast, elitist and heuristic-based multi-objective algorithm. Table II shows the corresponding parameters for these algorithms. Table III represents some statistics regarding the completed tasks and resources violations for different algorithms. It demonstrates that the performance of the proposed coalition formation method is quite better than other algorithms ad- dressed here in terms of percentage of completed tasks and average of resource violations. We also compared the proposed method against the merge-and-split coalition formation algo- rithm. The percentage of completed tasks for the merge-and- split method in 30 missions for different numbers of UAVs and tasks were between 46% and 50%, while, as shown in table III, our method could achieve rate of 90% in task completion. Table II: Algorithms' parameters Parameter Population size Maximum number of iterations Number of objectives Mutation probability Crossover probability Distribution index for crossover Distribution index for mutation NA: Not Available Distance − Based N SGA − II M OQGA Method N A N A 1 N A N A N A N A 200 500 3 10% 90% 20 100 200 500 3 N A N A N A N A (a) 8-2 (b) 16-4 (c) 32-8 (d) 64-16 Figure 1: Comparison the qualities of solutions provided by the proposed method (MOQGA) and NSGA-II. Table III: Percentage of completed tasks and average of resource violations for different algorithms in 30 missions (with different numbers of UAVs and tasks). Distance − Based Method NSGA − II MOQGA No. of UAVs and tasks Completed tasks Resource violations Completed tasks Resource violations Completed tasks Resource violations 8-2 16-4 32-8 64-16 128-24 67% 43% 39% 45% 47% 1.6 3.83 6.93 11.97 20.33 80% 86% 91% 92% 90% 0.43 0.6 0.73 2.03 4.33 90% 90% 97% 95% 94% 0.3 0.43 0.20 1.23 2.47 100150200250300350400objective function 1 (Cost)-70-60-50-40-30-20-10objective function 2 (Reliability)NSGA-IIMOQGA150200250300350400450500550600650objective function 1 (Cost)-140-120-100-80-60-40-20objective function 2 (Reliability)NSGA-IIMOQGA100150200250300350400450objective function 1 (Cost)-150-100-500objective function 2 (Reliability)NSGA-IIMOQGA020040060080010001200objective function 1 (Cost)-400-350-300-250-200-150-100-500objective function 2 (Reliability)NSGA-IIMOQGATable IV: The selected UAVs by leaders in 10 missions, where it is assumed the failure rate of the UAVs u5 and u6 is 90%. Time slot 1 2 3 4 5 6 7 8 8 9 10 L: Leader, F: Follower Coalition 1 L{u6}; F{u3, u4, u8} L{u2}; F{u3, u8, u4} L{u2}; F{u3, u5, u8} L{u2}; F{u4, u7, u3} L{u5}; F{u1, u3, u4} L{u1}; F{u3, u4} L{u5}; F{u2, u3, u8} L{u5}; F{u2, u3, u6} L{u5}; F{u2, u6, u7, u8} L{u5}; F{u2, u7, u8} L{u5}; F{u1, u3, u4} Satisfied Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes Coalition 2 L{u1}; F{u7, u2} L{u1}; F{u3, u7} L{u6}; F{u4} L{u1}; F{u5, u6, u8} L{u2}; F{u7, u8} L{u2}; F{u5, u6, u8, u7} L{u1}; F{u4, u7} L{u1}; F{u4, u7} L{u1}; F{u3, u4} L{u1}; F{u3, u4} L{u2}; F{u7, u8} Satisfied Yes Yes Yes No Yes No Yes Yes Yes Yes No Unreliable UAVs {} {} {u5} {u5, u6} {} {u5, u6} {} {u6} {u6} {} {} to completing a higher number of tasks and minimally over- spending the resources. VI. ACKNOWLEDGMENT OF SUPPORT AND DISCLAIMER The authors acknowledge the U.S. Government's support in the publication of this paper. This material is based upon work funded by AFRL. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the US government or AFRL. REFERENCES [1] T. W. Sandholm, "Distributed rational decision making," Multiagent systems: a modern approach to distributed artificial intelligence, pp. 201–258, 1999. [3] [2] S. P. Ketchpel, "Coalition formation among autonomous agents," in European Workshop on Modelling Autonomous Agents in a Multi-Agent World. Springer, 1993, pp. 73–88. J. Contreras, M. Klusch, and J. Yen, "Multi-agent coalition formation in power transmission planning: A bilateral shapley value approach." in AIPS, 1998, pp. 19–26. J. Yang and Z. Luo, "Coalition formation mechanism in multi-agent systems based on genetic algorithms," Applied Soft Computing, vol. 7, no. 2, pp. 561–568, 2007. [4] [5] F. Cruz-Menc´ıa, J. Cerquides, A. Espinosa, J. C. Moure, and J. A. Rodriguez-Aguilar, "Optimizing performance for coalition structure generation problems' idp algorithm," in Proceedings of the International Conference on Parallel and Distributed Processing Techniques and Applications (PDPTA). The Steering Committee of The World Congress in Computer Science, Computer Engineering and Applied Computing (WorldComp), 2013, p. 706. [6] F. Bistaffa, A. Farinelli, J. Cerquides, J. Rodr´ıguez-Aguilar, and S. D. Ramchurn, "Anytime coalition structure generation on synergy graphs," in Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems. International Foundation for Au- tonomous Agents and Multiagent Systems, 2014, pp. 13–20. [7] L. Sless, N. Hazon, S. Kraus, and M. Wooldridge, "Forming coalitions and facilitating relationships for completing tasks in social networks," in Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems. International Foundation for Autonomous Agents and Multiagent Systems, 2014, pp. 261–268. [8] P. Janovsky and S. A. DeLoach, "Multi-agent simulation framework for large-scale coalition formation," in Web Intelligence (WI), 2016 IEEE/WIC/ACM International Conference on. IEEE, 2016, pp. 343– 350. [9] B. Ghazanfari and N. Mozayani, "Enhancing nash q-learning and team q-learning mechanisms by using bottlenecks," Journal of Intelligent & Fuzzy Systems, vol. 26, no. 6, pp. 2771–2783, 2014. [10] B. Ghazanfari and M. E. Taylor, "Autonomous extracting a hierarchical structure of tasks in reinforcement learning and multi-task reinforcement learning," arXiv preprint arXiv:1709.04579, 2017. Figure 2: Changes in UAVs' cooperative reputation over time, where UAVs u5 and u6 are assumed to be selfish Figure 1 compares the qualities of solutions of the MO- QGA to NSGA-II algorithm. The plots demonstrate that the performance of the proposed method is better than the NSGA- II algorithm and result in superior quality solutions in terms of lower cost and higher reliability. Figure 2 shows changes in UAV's cooperative reputation over the course of time. It is assumed that UAVs, u5 and u6 are not trustworthy. As seen in the figure, the reputations of these UAVs decrease at each time slot, therefore it is less likely that these UAVs are selected by the leaders over time. The impact of reliability in coalition formation is also studied, where a pre-defined failure rate of 90% is considered for resources of UAVs u5 and u6. Table IV represents the selected UAVs (e.g., followers) by leaders where there are two unreliable UAVs. As shown in the table, the proposed method tries to not select unreliable UAVs in most cases. The reason for the selection of unreliable UAVs in some cases is that the problem is a MOOP and the method has to consider other objectives (i.e., cost and reputation) to optimize as well. V. CONCLUSION A leader-follower UAV coalition formation method is proposed to provide a practical solution for distributed task allocation in an unknown environment. Three critical aspects of cost minimization, reliability maximization, and the po- tential selfish behavior of the UAVs were considered in this coalition formation, and a quantum-inspired genetic algorithm is proposed to find the optimal coalitions with a low level of computational complexity. The proposed approach led to promising results compared to existing solutions with respect 51015202530Time Slots00.10.20.30.40.50.60.70.80.91CreditC1C3C4C5C6C8[11] B. Ghazanfari and N. Mozayani, "Extracting bottlenecks for rein- forcement learning agent by holonic concept clustering and attentional functions," Expert Systems with Applications, vol. 54, pp. 61–77, 2016. [12] S. S. Mousavi, B. Ghazanfari, N. Mozayani, and M. R. Jahed-Motlagh, "Automatic abstraction controller in reinforcement learning agent via automata," Applied Soft Computing, vol. 25, pp. 118–128, 2014. [13] C. Schumacher, P. Chandler, M. Pachter, and L. Pachter, "Uav task assignment with timing constraints via mixed-integer linear program- ming," in Proc. AIAA 3rd Unmanned Unlimited Systems Conference, 2004. [14] M. A. Darrah, W. M. Niland, and B. M. Stolarik, "Multiple uav dynamic task allocation using mixed integer linear programming in a sead mission," Infotech@ Aerospace, pp. 26–29, 2005. [15] C. Schumacher, P. R. Chandler, and S. R. Rasmussen, "Task allocation for wide area search munitions," in American Control Conference, 2002. Proceedings of the 2002, vol. 3. J. Kennedy, "Particle swarm optimization," in Encyclopedia of machine learning. Springer, 2011, pp. 760–766. IEEE, 2002, pp. 1917–1922. [16] [17] M. Dorigo and L. M. Gambardella, "Ant colony system: a cooperative learning approach to the traveling salesman problem," IEEE Transac- tions on evolutionary computation, vol. 1, no. 1, pp. 53–66, 1997. [18] D. E. Goldberg, Genetic algorithms. Pearson Education India, 2006. [19] P. J. Van Laarhoven and E. H. Aarts, "Simulated annealing," in Springer, 1987, pp. Simulated annealing: Theory and applications. 7–15. [20] A. Rovira-Sugranes and A. Razi, "Predictive routing for dynamic uav networks," in 2017 IEEE International Conference on Wireless for Space and Extreme Environments (WiSEE), Oct 2017, pp. 43–47. [21] A. Razi, F. Afghah, and J. Chakareski, "Optimal measurement policy for predicting uav network topology," arXiv preprint arXiv:1710.11185, 2017. [22] K.-H. Han and J.-H. Kim, "Genetic quantum algorithm and its appli- cation to combinatorial optimization problem," in Evolutionary Com- putation, 2000. Proceedings of the 2000 Congress on, vol. 2. IEEE, 2000, pp. 1354–1360. [23] P.-Y. Yin, S.-S. Yu, P.-P. Wang, and Y.-T. Wang, "Multi-objective task allocation in distributed computing systems by hybrid particle swarm optimization," Applied Mathematics and Computation, vol. 184, no. 2, pp. 407–420, 2007. [24] F. Afghah, M. Zaeri-Amirani, A. Razi, J. Chakareski, and E. Bentley, "A coalition formation approach to coordinated task allocation in heterogeneous uav networks," arXiv preprint arXiv:1711.00214, 2017.
1902.09359	1	1902	2019-02-25T15:24:19	Anytime Heuristic for Weighted Matching Through Altruism-Inspired Behavior	[ "cs.MA", "cs.AI" ]	We present a novel anytime heuristic (ALMA), inspired by the human principle of altruism, for solving the assignment problem. ALMA is decentralized, completely uncoupled, and requires no communication between the participants. We prove an upper bound on the convergence speed that is polynomial in the desired number of resources and competing agents per resource; crucially, in the realistic case where the aforementioned quantities are bounded independently of the total number of agents/resources, the convergence time remains constant as the total problem size increases. We have evaluated ALMA under three test cases: (i) an anti-coordination scenario where agents with similar preferences compete over the same set of actions, (ii) a resource allocation scenario in an urban environment, under a constant-time constraint, and finally, (iii) an on-line matching scenario using real passenger-taxi data. In all of the cases, ALMA was able to reach high social welfare, while being orders of magnitude faster than the centralized, optimal algorithm. The latter allows our algorithm to scale to realistic scenarios with hundreds of thousands of agents, e.g., vehicle coordination in urban environments.	cs.MA	cs	Anytime Heuristic for Weighted Matching Through Altruism-Inspired Behavior Panayiotis Danassis, Aris Filos-Ratsikas, Boi Faltings Artificial Intelligence Laboratory (LIA), École Polytechnique Fédérale de Lausanne (EPFL) {panayiotis.danassis,aris.filosratsikas,boi.faltings}@epfl.ch 9 1 0 2 b e F 5 2 ] A M . s c [ 1 v 9 5 3 9 0 . 2 0 9 1 : v i X r a ABSTRACT We present a novel anytime heuristic (ALMA), inspired by the human principle of altruism, for solving the assignment problem. ALMA is decentralized, completely uncoupled, and requires no com- munication between the participants. We prove an upper bound on the convergence speed that is polynomial in the desired number of resources and competing agents per resource; crucially, in the realistic case where the aforementioned quantities are bounded independently of the total number of agents/resources, the conver- gence time remains constant as the total problem size increases. We have evaluated ALMA under three test cases: (i) an anti- coordination scenario where agents with similar preferences com- pete over the same set of actions, (ii) a resource allocation scenario in an urban environment, under a constant-time constraint, and finally, (iii) an on-line matching scenario using real passenger-taxi data. In all of the cases, ALMA was able to reach high social wel- fare, while being orders of magnitude faster than the centralized, optimal algorithm. The latter allows our algorithm to scale to real- istic scenarios with hundreds of thousands of agents, e.g., vehicle coordination in urban environments. KEYWORDS Coordination and Cooperation; Resource Allocation; Multi-agent Learning 1 INTRODUCTION One of the most relevant problems in multi-agent systems (MAS) is finding an optimal allocation between agents. This pertains to role allocation (e.g., team formation for autonomous robots [GA13]), task assignment (e.g., employees of a factory, taxi-passenger match- ing [VCGA12]), resource allocation (e.g., parking spaces and/or charging stations for autonomous vehicles [GC13]), etc. What fol- lows is applicable to any such scenario, but for concreteness we will refer to the allocation of a set of resources to a set of agents, a set- ting known as the assignment problem, one of the most fundamental combinatorial optimization problems [Mun57]. When designing algorithms for assignment problems, a signif- icant challenge emerges from the nature of real-world applica- tions, which is often distributed and information-restrictive. For the former part, a variety of decentralized algorithms have been developed [GLM10, IS17, ZSP08, BNBA12], all of which, though, require polynomial in the problem size number of messages. How- ever, inter-agent interactions often repeat no more than a few hun- dreds of times. Moreover, sharing plans and preferences creates high overhead, and there is often a lack of responsiveness and/or communication between the participants [SKKR10]. Achieving fast convergence and high efficiency in such information-restrictive settings is extremely challenging. Yet, humans are able to routinely and robustly coordinate in similar everyday scenarios. One driv- ing factor that facilitates human cooperation is the principle of altruism [CR02, NS05, Gin00]. Inspired by human behavior, the proposed heuristic is modeled on the principle of altruism. This results to fast convergence to highly efficient allocations, without any communication between the agents. A distinctive characteristic of ALMA is that agents make deci- sions locally, based on (i) the contest for resources that they are interested in, (ii) the agents that are interested in the same re- sources. If each agent is interested in only a subset of the total resources, ALMA converges in time polynomial in the maximum size of the subsets; not the total number of resources. In particular, if the size of each subset is a constant fraction of the total number of resources, then the convergence time is constant, in the sense that it does not grow with the problem size. The same is not true for other algorithms (e.g., the optimal centralized solution) which require time polynomial in the total number of agents/resources, even if the aforementioned condition holds. The condition holds by default in many real-world applications; agents have only lo- cal knowledge of the world, there is typically a cost associated with acquiring a resource, or agents are simply only interested in resources in their vicinity (e.g., urban environments). This is important, as the proposed approach avoids having to artificially split the problem in subproblems (e.g, by placing bounds or spatial constraints) and solve those separately, in order to make it tractable. Instead, ALMA utilizes a natural domain characteristic, instead of an artificial optimization technique (i.e., artificial bounds). Coupled to the convergence time, the decentralized nature of ALMA makes it applicable to large-scale, real-world applications (e.g., IoT devices, intelligent infrastructure, autonomous vehicles, etc.). 1.1 Our Results Our main contributions in this paper are: (1) We introduce a novel, anytime ALtruistic MAtching heuristic (ALMA) for solving the assignment problem. ALMA is decentral- ized, completely uncoupled (i.e., each agent is only aware of his own history of action/reward pairs [Tal13]), and requires no com- munication between the agents. (2) We prove that if we bound the maximum number of resources an agent is interested in, and the maximum number of agents com- peting for a resource, the expected number of steps for any agent to converge is independent of the total problem size. Thus, we do not require to artificially split the problem, or similar techniques, to render it manageable. (3) We provide a thorough empirical evaluation of ALMA on both synthetic and real data. In particular, we have evaluated ALMA un- der three test cases: (i) an anti-coordination scenario where agents with similar preferences compete over the same set of actions, (ii) a resource allocation scenario in an urban environment, under a constant-time constraint, and finally, (iii) an on-line matching scenario using real passenger-taxi data. In all of the cases, ALMA achieves high social welfare (total satisfaction of the agents) as com- pared to the optimal solution, as well as various other algorithms. 1.2 Related Work The assignment problem consists of finding a maximum weight matching in a weighted bipartite graph and it is one of the best- studied combinatorial optimization problems in the literature. The first polynomial time algorithm (with respect to the total number of nodes, and edges) was introduced by Jacobi in the 19th century [BJ65, Oll09], and was succeeded by many classical algorithms [Mun57, EK72, Ber79] with the Hungarian algorithm of [Kuh55] being the most prominent one (see [Su15] for an overview). The problem can also be solved via linear programming [Dan90], as its LP formulation relaxation admits integral optimal solutions [PS82]. In Section 3.3, we will apply ALMA on a non-bipartite setting, which corresponds to the more general maximum weight matching problem on general graphs. To compute the optimal in this case, we will use the blossom algorithm of [Edm65] (see [LP09]). In reality, a centralized coordinator is not always available, and if so, it has to know the utilities of all the participants, which is often not feasible. In the literature of the assignment problem, there also exist several decentralized algorithms (e.g., [GLM10, IS17, ZSP08, BNBA12] which are the decentralized versions of the aforemen- tioned well-known centralized algorithms). However, these algo- rithms require polynomial computational time and polynomial number of messages (such as cost matrices [IS17], pricing infor- mation [ZSP08], or a basis of the LP [BNBA12], etc.). Yet, agent interactions often repeat no more than a few hundreds of times. To the best of our knowledge, a decentralized algorithm that requires no message exchange (i.e., no communication network) between the participants, and achieves high efficiency, like ALMA does, has not appeared in the literature before. Let us stress the importance of such a heuristic: as autonomous agents proliferate, and their num- ber and diversity continue to rise, differences between the agents in terms of origin, communication protocols, or the existence of sub-optimal, legacy agents will bring forth the need to collaborate without any form of explicit communication [SKKR10]. Finally, inter-agent communication creates high overhead as well. ALMA is inspired by the decentralized allocation algorithm of [DF19]. Using such a simple learning rule which only requires en- vironmental feedback, allows our approach to scale to hundreds of thousands of agents. Moreover, it does not require global knowl- edge of utilities; only local knowledge of personal utilities (in fact, we require knowledge of pairwise differences which are far easier to estimate). 2 ALTRUISTIC MATCHING HEURISTIC In this section, we define ALMA and prove its convergence proper- ties. We begin with the definition of the assignment problem, and its interpretation in our setting. 2 2.1 The Assignment Problem The assignment problem consists of finding a maximum weight matching in a weighted bipartite graph, G = {N ∪ R, E}. In the studied scenario, N = {1, . . . , N} agents compete to acquire R = {1, . . . , R} resources. We assume that each agent n is interested in a subset of the total resources, i.e., Rn ⊂ R. The weight of an edge (n, r) ∈ E represents the utility (un(r)) agent n receives by acquiring resource r. Each agent can acquire at most one resource, and each resource can be assigned to at most one agent. The goal is to maximize the social welfare (sum of utilities), i.e., max x≥0 subject to (n,r)∈E (1) un,r xn,r xn,r = 1,∀n ∈ N xn,r = 1,∀r ∈ R r (n,r)∈E n(n,r)∈E 2.2 Learning Rule This section describes the proposed heuristic (ALMA: ALtruistic MAtching heuristic) for weighted matching. We make the follow- ing two assumptions: First, we assume (possibly noisy) knowledge of personal preferences by each agent. Second, we assume that agents can observe feedback from their environment. This is used to inform collisions and detect free resources. It could be achieved by the use of visual, auditory, olfactory sensors etc., or by any other means of feedback from the resource (e.g., by sending an occupancy message). Note here that these messages would be between the requesting agent and the resource, not between the participating agents themselves, and that it suffices to send only 1 bit of informa- tion (e.g., 0, 1 for occupied / free respectively). ALMA learns the right action through repeated trials as follows. Each agent sorts his available resources (possibly Rn ⊆ R) in decreasing order of utility (r1, r2, . . . , ri , ri +1, . . . , rRn ). The set of available actions is denoted as A = {Y , Ar1 , . . . , ArRn }, where Y refers to yielding, and Ar refers to accessing resource r. Each agent has a strategy (дn) that points to a resource and it is initialized to the most preferred one. As long as an agent has not acquired a resource yet, at every time-step, there are two possible scenarios. If дn = Ar (strategy points to resource r), then agent n attempts to acquire that resource. If there is a collision, the colliding parties back-off with some probability. Otherwise, if дn = Y, the agent choses a resource r for monitoring. If the resource is free, he sets дn ← Ar . Alg. 1 presents the pseudo-code of ALMA, which is followed by every agent individually. The back-off probability and the next resource to monitor are computed individually and locally based on the current resource and each agent's utilities, as will be explained in the following section. Finally, note that if the available resources change over time, the agents simply need to sort again the currently available ones. 2.3 Back-off Probability & Resource Selection Let R be totally ordered in decreasing utility under ≺n, ∀n ∈ N. If more than one agent compete for resource ri (step 4 of Alg. 1), each of them will back-off with probability that depends on their utility n = lossi k − i (2) where k ∈ {i + 1, . . . , Rn} denotes the number of remaining re- sources to be considered. For k = i + 1, the formula only takes into account the utility loss of switching to the immediate next best resource, while for k = Rn it takes into account the average utility loss of switching to all of the remaining resources. In the remainder n = un(ri)−un(ri +1). The of the paper we assume k = i +1, i.e., lossi actual back-off probability can be computed with any monotoni- cally decreasing function f on lossn, i.e., Pn(ri , ≺n) = fn(lossi n). In the evaluation section, we have used two such functions, a linear (Eq. 3), and the logistic function (Eq. 4). The parameter ϵ places a threshold on the minimum / maximum back-off probability for the linear curve, while γ determines the steepness of the logistic curve.  f (loss) = 1 − ϵ, ϵ, 1 − loss, if loss ≤ ϵ if 1 − loss ≤ ϵ otherwise f (loss) = 1 1 + e−γ(0.5−loss) (3) (4) loss of switching to their respective remaining resources. The loss is given by Eq. 2. k j=i +1 un(ri) − un(rj) Using the aforedescribed rule, agents that do not have good al- ternatives will be less likely to back-off and vice versa. The ones that do back-off select an alternative resource and examine its avail- ability. The resource selection is performed in sequential order, i.e., Sn(rprev, ≺n) = rprev+1, where rprev denotes the resource selected by that agent in the previous round. We also examined the possi- bility of using a weighted or uniformly at random selection, but achieved inferior results. 2.4 Altruism-Inspired Behavior ALMA is inspired by the human principle of altruism. We would expect an altruistic person to give up a resource either to some- one who values it more, if that resulted in an improvement of the well-being of society [CR02], or simply to be nice to others [Sim16]. Such behavior is especially common in situations where the backing-off subject has equally good alternatives. For example, in human pick-up teams, each player typically attempts to fill his most preferred position. If there is a collision, a colliding player might back-off because his teammate is more competent in that role, or because he has an equally good alternative, or simply to be polite; the player backs-off now and assumes that role at some future game. From an alternative viewpoint, following such an al- truistic convention leads to a faster convergence which outweighs the loss in utility. Such conventions allow humans to routinely and robustly coordinate in large scale and under dynamic and unpre- dictable demand. Behavioral conventions are a fundamental part of human societies [Lew08], yet they have not appeared meaningfully in empirical modeling of multi-agent systems. Inspired by human behavior, ALMA attempts to reproduce these simple rules in an artificial setting. 3 r1, r2, . . . , ri , ri +1, . . . , rRn . Algorithm 1 ALMA: Altruistic Matching Heuristic. Require: Sort resources (Rn ⊆ R) in decreasing order of utility Require: Initialize дn ← Ar1, and rprev ← r1. 1: procedure ALMA 2: 3: 4: 5: 6: 7: 8: Agent n attempts to acquire resource r. Set rprev ← r. if Collision(r) then back-off (set дn ← Y) with probability Pn(r , ≺n). Agent n monitors r ← Sn(rprev, ≺n). Set rprev ← r. if Free(r) then set дn ← Ar . if дn = Ar then else (дn = Y) 2.5 Convergence Agents who have not acquired a resource (дn = Y) will not claim an occupied one. Additionally, every time a collision happens, there is a positive probability that some agents will back-off. As a result, the system will converge. The following theorem proves that the expected convergence time is logarithmic in the number of agents N and quadratic in the number of resources R. Theorem 2.1. For N agents and R resources, the expected number of steps until the system of agents following Alg. 1 converges to a complete matching is bounded by (5), where p∗ = f (loss∗), and loss∗ is given by Eq. 6. (cid:18) O R (cid:18) 1 p∗ log N + R (cid:19)(cid:19) (cid:18) 2 − p∗ 2(1 − p∗) r ∈R,n∈N(lossr min n), 1 − max r ∈R,n∈N(lossr n) (cid:19) (5) (6) loss ∗ = arg min loss r n Proof. To improve readability, we will only provide a sketch of the proof. Please see the appendix for the complete version. The proof is based on [CF11, DF19]. We first assume that every agent, on every collision, backs-off with the same constant probability p. We start with the case of having N agents competing for 1 resource and model our system as a discrete time Markov chain. Intuitively, this Markov chain describes the number of individuals in a decreasing population, but with two caveats: the goal (absorbing state) is to reach a point where only one individual remains, and if we reach zero, we restart. We prove that the expected number of steps until we reach a state log N where either 1 or 0 agents compete for that resource is O(cid:16) 1 (cid:16) 2(1−p) Moreover, we prove that with high probability, Ω agent will remain (contrary to reaching 0 and restarting the process of claiming the resource), no matter the initial number of agents. Having proven that, we move to the general case of N agents competing for R resources. (cid:17). (cid:17), only 1 At any time, at most N agents can compete for each resource. We call this period a round. During a round, the number of agents competing for a specific resource monotonically decreases, since that resource is perceived as occupied by non-competing agents. Let the round end when either 1 or 0 agents compete for the resource. (cid:17) steps. If all agents backed-off, it will This will require O(cid:16) 1 2−p log N p p p 2−p 2(1−p) take on average R steps until at least one of them finds a free resource. We call this period a break. In the worst case, the system will oscillate between a round and a break. According to the above, R = 1, as mentioned in the previous paragraph, in expectation there will be 2−p 2(1−p) oscillations. For R > 1 the expected number of one oscillation requires in expectation O(cid:16) 1 oscillations is bounded by O(cid:16) (cid:16) 1 matching is O(cid:16) (cid:17) steps. If (cid:17). Thus, we conclude that if (cid:17)(cid:17). all the agents back-off with the same constant probability p, the expected number of steps until the system converges to a complete log N + R 2−p 2(1−p) log N + R R R p ) and outer ( 2−p Next, we drop the constant probability assumption. Intuitively, the worst case scenario corresponds to either all agents having a small back-off probability, thus they keep on competing for the same resource, or all of them having a high back-off probability, thus the process will keep on restarting. These two scenarios correspond to the inner ( 1 2(1−p)) probability terms of bound (5) p respectively. Let p∗ = f (loss∗) be the worst between the smallest or highest back-off probability any agent n ∈ N can exhibit, i.e., having loss∗ given by Eq. 6. Using p∗ instead of the constant p, we bound the expected convergence time according to bound (5). □ It is worth noting that the back-off probability p∗ in bound (5) does not significantly affect the convergence time. For example, using Eq. 3 with a quite small ϵ = 0.01, the resulting quantities 2. Most importantly, though, would be at most 100R log N , and 50R this is a rather loose bound (e.g., agents would rarely back-off with probabilities as extreme as p∗). Apart from the convergence of the whole system, we are inter- ested in the expected number of steps any individual agent would require in order to acquire a resource. In real-world scenarios, there is typically a cost associated with acquiring a resource. For example, a taxi driver would not be willing to drive to the other end of the city to pick up a low fare passenger. As a result, each agent is typically interested in a subset of the total resources, i.e., Rn ⊂ R, thus at each resource there is a bounded number of competing agents. Let Rn denote the maximum number of resources agent n is interested in, and N r denote the maximum number of agents competing for resource r. By bounding these two quantities (i.e., we consider Rn and N r to be constant functions of N , R), Corollary 2.2 proves that the expected number of steps any individual agent requires in order to claim a resource is independent of the total problem size (i.e., N , and R), or, in other words, that the convergence time is constant in these quantities. Corollary 2.2. Let Rn = Rn, such that ∀r ∈ Rn : un(r) > 0, and N r = Nr , such that ∀n ∈ Nr : un(r) > 0. The expected number of steps until an agent n ∈ N following Alg. 1 successfully acquires a resource is bounded by (7), where p∗ n = f (loss⋆) and loss⋆ (cid:18) (cid:19)(cid:19) is given by Eq. 8, independent of the total problem size N , R. (cid:19) N r) + max Rn′ n′∈∪r∈Rn Nr 2 − p∗ n 2(1 − p∗ n) max Rn′ n′∈∪r∈Rn Nr log( max r ∈Rn (cid:18) 1 p∗ n (cid:18) (7) O min r ∈Rn,n∈Nr (lossr n), 1 − max r ∈Rn,n∈Nr (lossr n) loss⋆ = arg min loss r n (8) 4 Proof. The expected number of steps until an agent n ∈ N successfully acquires a resource is upper bounded by the total con- vergence time of the sub-system he belongs to, i.e., the sub-system consisting of the sets of Rn resources and ∪r ∈RnNr agents. In such scenario, at most maxr ∈Rn N r agents can compete for any resource. Using Theorem 2.1 for maxr ∈Rn N r agents, maxn′∈∪r∈Rn Nr Rn′ re- sources, and worst-case loss⋆ given by any agent in ∪r ∈RnNr (i.e., Eq 8) results in the desired bound. Note that agents do not compete for already claimed resources (step 8 of Alg. 1), thus the conver- gence of an agent does not require the convergence of agents of overlapping sub-systems. □ i optimal)/ i achieved − in social welfare is ( 3 EVALUATION In this section we evaluate ALMA under various test cases. For the first two, we focus on convergence time and relative difference in social welfare (SW), i.e., (achieved − optimal)/optimal. In every reported metric, except for the social welfare, we report the average value out of 128 runs of the same problem instance. Error bars repre- sent one standard deviation (SD) of uncertainty. As a measure of so- cial welfare, we report the cumulative regret of the aforementioned 128 runs, i.e., for i ∈ [1, 128] runs, the reported relative difference i optimal. This was done to improve visualization of the results in smaller problem sizes, where really small differences result in high SD bars (e.g., if achieved = 1 × 10−5, and optimal = 2 × 10−5, the relative difference would be −50% for practically the same matching). The optimal matchings were computed using the Hungarian algorithm 1. The third test case is an on-line setting, thus we report the achieved SW (not the relative difference to the optimal), and the empirical competitive ratio (average out of 128 runs, as before). All the simu- lations were run on 2x Intel Xeon E5-2680 with 256 GB RAM. In Section 3.1 we use the logistic function (Eq. 4) with γ = 2, while in Sections 3.2 & 3.3 we use the linear function (Eq. 3) with ϵ = 0.1. It is important to stress that our goal is not to improve the convergence speed of a centralized, or decentralized algorithm. Rather, the computation time comparisons of Sections 3.1 & 3.2 are meant to ground the actual speed of ALMA, and argue in favor of its applicability on large-scale, real-world scenarios. Given the nature of the problem, we elected to use a specialized algorithm to compute the optimal solution, rather than a general LP-based technique (e.g., the Simplex method). Specifically, we opted to use the Hungarian algorithm which, first, has proven polynomial worse case bound, and second, as our simulations will demonstrate, can handle sufficiently large problems. 3.1 Test Case #1: Uniform, and Noisy Common Preferences 3.1.1 Setting. As a first evaluation test case, we cover the ex- treme scenarios. The first pertains to an anti-coordination scenario in which agents with similar preferences compete over the same set of actions [DF18]. For example, autonomous vehicles would prefer the least congested route, bidding agents participating in multiple auctions would prefer the ones with the smallest number of partici- pants, etc. We call this scenario 'noisy common preferences' and 1We used Kevin L. Stern's O(N 3) implementation: https://github.com/KevinStern/. (a) (c) (b) (d) Figure 1: From left to right, top to bottom: (1a) Total convergence time (#steps), (1b) average time (#steps) for an individual agent to successfully acquire a resource, (1c) Computation time (ns), and (1d) Relative difference in social welfare (%), for increasing number of resources, and N = R. Fig. 1a, 1c, and 1b are in double log. scale, while Fig. 1d is in single log. scale. model the utilities as follows: ∀n, n′ ∈ N, un(r) − un′(r) ≤ noise, where the noise is sampled from a zero-mean Gaussian distribu- tion, i.e., noise ∼ N(0, σ 2) 2. In the second scenario the utilities are initialized uniformly at random (U aR) for each agent and resource. 3.1.2 Convergence Time. Starting with Fig. 1a, we can see that the convergence time for the system of agents following Alg. 1 is linear to the number of resources R. From the perspective of a single agent, Fig. 1b shows that on average he will successfully acquire a resource significantly (> 2×) faster than the total convergence time. This suggest that there is a small number of agents which take longer in claiming a resource and which in turn delay the system's convergence. We will exploit this property in the next section to present the anytime property of ALMA. Fig. 1c shows that ALMA requires approximately 4 to 6 orders of magnitude less computa- tion time than the centralized Hungarian algorithm. Furthermore, ALMA seems to scale more gracefully, an important property for real world applications. Note also that in real-world applications we would have to take into account communication time, communica- tion reliability protocols, etc., which create additional overhead for the Hungarian or any other algorithm for the assignment problem. 3.1.3 Efficiency. The relative difference in social welfare (Fig. 1d) reaches asymptotically zero as R increases. For a small number of resources, ALMA achieves the worst social welfare, approximately 11% worse than the optimal. Intuitively this is because when we have a small number of choices, a single wrong matching can have a significant impact to the final social welfare, while as the number of resources grow, the impact of an erroneous matching is mitigated. For 16384 resources we lose less than 2.5% of the optimal. As a reference, Fig. 1d depicts the centralized greedy, and the random solutions as well. The greedy solution goes through the partici- pating agents randomly, and assigns them their most preferred unassigned resource. In this scenario, the random solution loses up to 50% of the optimal SW, while the greedy solution achieves similar results to ALMA, especially in high noise situations. This is to be expected, since first, all agents are interested in all the resources, and second, as the noise increases, the agents' preferences become more distinguishable, more diverse. ALMA is of a greedy nature as well, albeit it utilizes a more intelligent backing-off scheme. Con- trary to that, the greedy solution does not take into account the utilities between agents, thus there are scenarios where ALMA would significantly outperform the greedy (e.g., see Section 3.3). Finally, recall that ALMA operates in a significantly harder domain with no communication, limited feedback, and time constraints. In contrast, the greedy method requires either a central coordinator or message exchange (to communicate users' preferences and resolve collisions). Adopting a simple rule allows the applicability of ALMA to large scale multi-agent systems. In this section we will analyze such a scenario. Specifically we are interested in resource allocation in urban environments (e.g., parking spots / charging stations for 3.2 Test Case #2: Resource Allocation in a Cartesian Map with Manhattan Distances 2Similar results achieved using uniform noise, i.e., ∼ U(−ν, ν). 5 (a) (c) (e) (g) (b) (d) (f) (h) Figure 2: Left to right, top to bottom: (2a) Average time (#steps) for an agent to acquire a resource, (2b) Total convergence time (#steps), (2c) Computation time (ns), (2d) Relative difference in SW (%), (2e) Percentage of 'winners', (2f) Relative difference in SW (%) in interrupted execution, (2g) Percentage of 'winners' in interrupted execution. The aforementioned are for increasing number of resources, and N = R. Fig. (2a), (2b), and (2c) are in double log scale, while the rest are in single log scale. Fig. (2h) √4 × N . presents an example of the studied resource allocation scenario in an urban environment. We assume grid length of 6 autonomous vehicles, taxi - passenger matchings, etc.). The afore- mentioned problems become ever more relevant due to rapid ur- banization, and the natural lack of coordination in the usage of resources [Var16]. The latter result in the degradation of response (e.g., waiting time) and quality metrics in large cities [Var16]. 3.2.1 Setting. Let us consider a Cartesian map representing a city on which are randomly distributed vehicles and charging stations, as depicted in Fig. 2h. The utility received by a vehicle n for using a charging station r is proportional to the inverse of their distance, i.e., un(r) = 1/dn,r . Since we are in an urban environment, let dn,r denote the Manhattan distance. Typically, there is a cost each agent is willing to pay to drive to a resource, thus there is a cut-off distance, upon which the utility of acquiring the resource is zero (or possibly negative). This is a typical scenario encountered in resource allocation in urban environments, where there are spatial constraints and local interactions. The way such problems are typically tackled, is by dividing the map to sub-regions, and solving each individual sub-problem. For example, Singapore is divided into 83 zones based on postal codes [CN11], and taxi drivers' policies are optimized according to those [NKL17, VCGA12]. On the other hand, not placing bounds means that the current solutions will not scale. To the best of our knowl- edge, we are the first to propose an anytime heuristic for resource allocation in urban environments that can scale in constant time without the need to artificially split the problem. Instead, ALMA ex- ploits the two typical characteristics of an urban environment: the anonymity in interactions and homogeneity in supply and demand [Var16] (e.g., assigning any of two equidistant charging stations to a vehicle, would typically result to the same utility). This results in a simple learning rule which, as we will demonstrate in this section, can scale to hundreds of thousands of agents. 3.2.2 Convergence Time. To demonstrate the latter, we placed a bound on the maximum number of resources each agent is in- terested in, and on the maximum number of agents competing for a resource, specifically Rn = N r ∈ {8, 16, 32, 64, 128}. According to Corollary 2.2, bounding these two quantities should result in convergence in constant time, regardless of the total problem size (R, N). The latter is corroborated by Fig. 2a, which shows that the average number of time-steps until an agent successfully claims a resource remains constant as we increase the total problem size. Same is true for the system's convergence time (Fig. 2b), which caps as R increases. The small increase is due to outliers, as Fig. 2b reports the convergence time of the last agent. This results to ap- proximately 7 orders of magnitude less computation time than the centralized Hungarian algorithm (Fig. 2c), and this number would grow boundlessly as we increase the total problem size. Moreover, as mentioned, in an actual implementation, any algorithm for the assignment problem would face additional overhead due to com- munication time, reliability protocols, etc. 3.2.3 Efficiency. Along with the constant convergence time, ALMA is able to reach high quality matchings, achieving less than 7.5% worse social welfare (SW) than the optimal (Fig. 2d). The latter refers to the small bound of Rn = 8. As observed in Section 3.1, with a small number of choices, a single wrong matching can have a sig- nificant impact to the final social welfare. By increasing the bound 7 to a more realistic number (e.g., Rn = 32), we achieve less than 2.5% worse SW. In general, for R > 2 resources and different values of Rn, ALMA achieves between 1.9− 12% loss in SW, while the greedy approach achieves 8.0 − 24% and the random 7.3 − 43.4%. The be- havior of the graphs depicted in Fig. 2d for Rn ∈ {8, 16, 32, 64, 128} indicate that, as the problem size (R) increases, the social welfare reaches its lowest value at R = 2 × Rn. To investigate the latter, we have included a graph for increasing Rn (instead of constant to the problem size), specifically Rn = R/2. ALMA achieves a constant loss in social welfare (approximately 11%). The greedy approach achieves loss of 14%, while the random solution degrades towards 44% loss. Compared to Test Case #1, this is a significantly harder prob- lem for a decentralized algorithm with no communication and no global knowledge of the resources. The set of resources each agent is interested in is a proper subset of the set of the total resources, i.e., Rn ⊊ R (or could be R < N ). Furthermore, the lack of commu- nication between the participants, and the stochastic nature of the algorithm can lead to deadlocks, e.g., in Fig. 2h, if vehicle 2 acquires resource 1, then vehicle 1 does not have an available resource in range. Nonetheless, ALMA results in an almost complete matching. Fig. 2e, depicts the percentage of 'winners' (i.e., agents that have successfully claimed a resource r such that un(r) > 0). The afore- mentioned percentage refers to the total population (N ) and not the maximum possible matchings (potentially < N ). As depicted, the percentage of 'winners' is more than 90%, reaching up to 97.8% for Rn = 128. We also employed ALMA in larger simulations with up to 131072 agents, and equal resources. As seen in Fig. 2e, the percentage of winners remains stable at around 98%. Even though the size of the problem prohibited us from running the Hungar- ian algorithm (or an out-of-the-box LP solver) and validating the quality of the achieved matching, the fact that the percentage of winners remains the same suggests that the relative difference in SW will continue on the same trend as in Fig. 2d. Moreover, the average steps per agent to claim a resource remains, as proven, constant (Fig. 2a). The latter validate the applicability of ALMA in large scale applications with hundreds of thousands of agents. 3.2.4 Anytime Property. In the real world, agents are required to run in real time, which imposes time constraints. ALMA can be used as an anytime heuristic as well. To demonstrate the latter, we compare four configurations: the 'full' one, which is allowed to run until the systems converges, and three 'constant time' versions which are given a time budget of 32, 256, and 1024 time-steps. In this scenario, we do not impose a bound on Rn, N r , but we assume a cut-off distance, upon which the utility is zero. The cut-off distance was set to 0.25 of the maximum possible distance, i.e., as the prob- lem size grows, so do the Rn, N r . On par with Test Case #1, the full version converges in linear time. As depicted in Fig. 2f, the achieved SW is less than 9% worse than the optimal. The inferior results in terms of SW compared to Test Case #1 are because this is a signifi- cantly harder problem due to the aforementioned deadlocks. On the other hand though, ALMA benefits from the spatial constraints of the problem. The average number of time-steps an individual agent needs to successfully claim a resource is significantly smaller, which suggest that we can enforce computation time constraints. Restrict- ing to only 32, 256, and 1024 time-steps, results in 1.25%, 0.12%, and 0.03% worse SW than the unrestricted version, respectively. Even in larger simulations with up to 131072 agents, the percentage of winners (Fig. 2g) remains stable at 98.6%, which suggests that the relative difference in SW will continue on the same trend as in Fig. 2f (we do not suggest that this is the case in any domain. For example, in the noisy common preferences domain of Test Case #1, the quality of the achieved matching decreases boundlessly as we decrease the alloted time. Nevertheless, the aforedescribed domain is a realistic one, with a variety of real-world applications). Finally, the repeated nature of such problems suggests that even in the case of a deadlock, the agent which failed to win a resource, will do so in some subsequent round. 3.3 Test Case #3: On-line Taxi Request Match In this section we present a motivating test case involving ride- sharing, via on-line taxi request matching, using real data of taxi rides in New York City. Ride-sharing (or carpooling), offers great potential in congestion relief and environmental sustainability in urban environments. In the past few years, several commercially successful ride-sharing companies have surfaced (e.g., Uber, Lyft, etc.), giving rise to a new incarnation of ride-sharing: dynamic ride- sharing, where passengers are matched in real-time. Ride-sharing, though, results to some passenger disruption due to loss in flexi- bility, security concerns, etc. Compensation comes in the form of monetary incentives, as it allows passengers to share the travel ex- penses, and thus reduce the cost. Ride-sharing companies account for a plethora of factors, like current demand, prime time pricing, the cost of the route without any ride-sharing, the likelihood of a match, etc. Yet, a fundamental factor of the cost of any shared ride, no matter if it is a company or a locally-organized car sharing scheme, is the traveled distance. In this test case, we attempt to maximize the total distanced saved, by matching taxi requests of high overlap. Fig. 3 provides an illustrative example. There are two passengers (depicted as yel- low and red) with high overlap routes. Each can drive on their own to their respective destinations (dashed yellow and red line respectively), or share a ride (green line) and reduce travel costs. Dynamic ride-sharing is an inherently on-line setting, as a match- ing algorithm is unaware of the requests that will appear in the future and needs to make decisions for the requests before they 'ex- pire' (a similar setting was studied in [ABD+18]). ALMA is highly befitting for such a scenario, as it involves large-scale matchings under dynamic demand, it is highly decentralized, and partially observable. 3.3.1 Setting. We use a dataset 3 of all taxi requests (ρ) in New York City during one week ({01-01-16 0:00 - 01-07-16 23:59}, 34077 requests in total). The data include pickup and drop-off times, and geolocations. Requests appear (become open) at their respective pickup time, and wait kρ time-steps to find a match. Let a time-step be one minute. After kρ time-steps we call request ρ, critical. If a critical request is not matched, we assume they drive off to their destination in a single passenger ride. Let open, critical denote the sets of open, and critical requests respectively, and let current = open ∪ critical. To compute kρ we assume the following: There is Figure 3: Example of the studied taxi request matching sce- nario. a minimum minW , and a maximum maxW waiting time set by the ride-sharing company, i.e., minW ≤ kρ ≤ maxW ,∀ρ. Moreover, since each passenger specifies his destination in advance, we can compute the trip time (lρ). Assuming people are willing to wait for a time that is proportional to their trip time, let kρ = q × lρ, where q ∈ [0, 1]. The parameters minW , maxW , and q can be set by the ride-sharing company. We report results on different values for all of the above parameters. For each pair ρ1, ρ2 of requests, we compute the driving distance (dρ1, ρ2 = min of all possible combinations of driving between ρ1, ρ2's pickup and drop-off locations) that would be traveled if ρ1, ρ2 are matched, i.e., if they share the same taxi. Subsequently, the utility of matching ρ1 to ρ2 (distance saved) is uρ1(ρ2) = dρ1, ρ2 (km). Given the on-line nature of the setting, it might be beneficial to use the following non-myopic heuristic: avoid matching low utility pairs, as long as the requests are not critical, since more valuable pairs might be presented in the future. Thus, if uρ1(ρ2) < dmin, and ρ1, ρ2 (cid:60) critical, we do not match ρ1, ρ2. In what follows, we select for each algorithm and for each simulation the value dmin ∈ {0, 500, 1000, 1500, 2000, 2500} that results in the highest score. To compute the actual trip time, and driving distance, we have used the Open Source Routing Machine 4, which computes shortest paths in road networks. 3.3.2 Computation of the optimal matching. In this scenario, each request has a dual role, being both an agent and a resource, i.e., we have a non-bipartite graph. For computing the optimal (maxi- mum weight) matching for the employed on-line heuristics, we use the blossom algorithm of [Edm65], which computes a maximum weight matching in general graphs. This enables us to compute the best possible matching among current requests, i.e., requests that have not 'expired' at the time of the computation. In fact, the follow- ing observation allows us to compute the optimal off-line matching as well, i.e., the best possible matching over the whole time interval. Let aρ be the pick-up time of request ρ and let eρ = aρ + kρ be the time when it becomes critical. Then, we can redefine the utility of a matching as: (cid:40) uρ1(ρ2) = dρ1, ρ2 , −1, if ai ≤ ej and ei ≤ ej otherwise (9) i.e., the utility is the distance saved if both requests are simultane- ously active when matched and −1 otherwise. The latter effectively results in this pair never being matched in an optimal solution. We remark that this clairvoyant matching is not feasible in the on-line setting and serves as a benchmark against which we can compare 3https://www.kaggle.com/debanjanpaul/new-york-city-taxi-trip-distance-matrix/ 4http://project-osrm.org/ 8 the performance of on-line algorithms, as is common in the litera- ture of competitive analysis [BEY05]. The measure of efficiency, as compared to the off-line optimal, will be the empirical competitive ratio, i.e., the ratio of the social welfare of the on-line algorithm over the welfare of the optimal, as measured empirically for our dataset. 3.3.3 The Blossom Algorithm vs Linear Programming. Contrary to the other test cases, the fact that the graph is not bipartite has implications on how linear programming can be used to compute the optimal solution. In particular, in all of the presented test cases (#1, #2, and #3), one can formulate the problem of computing the optimal solution as an Integer Linear Program (ILP) and then solve it using some general solver like CPLEX 5. Yet, the computational complexity of solving the aforementioned ILP varies amongst the different test cases. Solving integer linear programs is generally quite computation- ally demanding. Thus, it is common to resort to solving the LP relaxation (where the integrality constraints have been 'relaxed' to fractional constraints), which can be computed in polynomial time. In the case of bipartite graphs (such as test cases #1 and #2), the standard LP relaxation admits integer solutions, i.e., solutions to the actual maximum weight bipartite matching problem (because the constraint matrix is totally unimodular). In the case of general (non-bipartite) graphs, this is no longer the case, thus we have to resort to solving the LP relaxation. It is known that the (fractional) optimal solution to the LP relaxation might be better than the (integral) optimal solution to the original ILP formulation (i.e., it has an integrality gap which is larger than 1, in fact 2). In other words, solving the LP relaxation will not provide solutions for the maximum weight matching problem but for an 'easier version' with fractional matchings. Thus, comparing against that solution could only give very pessimistic ratios, when the real empirical competitive ratios are much better. One can derive a different ILP formulation of the problem using more involved constraints (called 'blossom' constraints [Edm65]), whose relaxation admits integer solutions (i.e., the integrality gap is now 1). However, the latter would result in an exponential number of constraints, and one would need to employ the Ellipsoid method with an appropriately chosen separation oracle to solve it in poly- nomial time (see [FOW02, page 4] for more details]. Overall, the employment of the combinatorial algorithm of [Edm65] for find- ing the maximum weight matching is a cleaner and more efficient solution. 3.3.4 Benchmarks. Each request runs ALMA independently. ALMA waits until the request becomes critical, and then matches it by running Alg. 1, where N = critical, and R = current. In this non-bipartite scenario, if an agent is matched under his dual role as a resource, he is immediately removed. As we explained earlier, it is infeasible for an on-line algorithm to compute the optimal match- ing over the whole period of time. Instead, we consider just-in-time and in batches optimal solutions. Specifically, we compare to the following [AESW11, ABD+18]: 5https://www.ibm.com/analytics/cplex-optimizer 9 (a) (b) Figure 4: Total distance saved (km) for various values of minW , maxW , q. (4a, left) Pragmatic scenario, (4a, right) Re- quests become critical in just one time-step, (4b) Various levels of waiting time (q) assuming no bounds, i.e., minW = 0, maxW = ∞, (double log. scale). • Just-in-time Max Weight Matching (JiTMWM): Waits until a request becomes critical and then computes a maximum- weight matching of all the current requests, i.e N = R = current. • Batching Max Weight Matching (BMWM): Waits x time- steps and then computes a maximum-weight matching of all the current requests, i.e N = R = current. • Batching Greedy (BG): Waits x time-steps and then greed- ily matches current requests, i.e N = R = current (ties are broken randomly). Unmatched open requests are removed. For batch size x = 1 we get the simple greedy approach where requests are matched as soon as they appear. 3.3.5 Efficiency. Starting with the social welfare, Fig. 4 presents the total distance saved (km) for various values of minW , maxW , and q. ALMA loses 8.3% of SW in the pragmatic scenario of Fig. 4a (left), and 6.5% when the requests become critical in just one time-step (Fig. 4a (right)). If no bounds are placed on the minimum and maximum waiting time (i.e., minW = 0, maxW = ∞), ALMA exhibits loss of 8 − 11.5%, for various values of q (Fig. 4b). The above are compared to the best performing benchmark on each scenario (JiTMWM, or BMWM, x = 5). Moreover, it significantly outperforms every greedy approach. In the first scenario the BGs lose between 35.8 − 53.5%, in the second between 24 − 68.2%, and in the third between 31.5 − 69%. Once more, it is worth noting that ALMA requires just a broad- cast of a single bit to indicate the occupancy of a resource, while the Table 1: Empirical Competitive Ratio. (minW , maxW , q) ALMA JiTMWM BMWM, x = 1 BMWM, x = 2 BMWM, x = 5 BG, x = 1 BG, x = 2 BG, x = 5 (1, 3, 0.1) 0.4603 (1, 1,−) 0.3038 (0,∞, 0.1) 0.4840 (0,∞, 0.2) 0.5688 (0,∞, 0.5) 0.5133 (0,∞, 1.0) 0.4606 0.5883 0.3810 0.6221 0.7616 0.7668 0.7706 0.7890 0.8491 0.7835 0.7546 0.6939 0.6695 0.8621 0.9243 0.8528 0.8207 0.7439 0.7390 0.3991 0.5112 0.3900 0.3399 0.2731 0.2440 0.5422 0.6891 0.5299 0.4604 0.3714 0.3306 0.8568 0.9190 0.8486 0.8158 0.7368 0.7254 0.8283 0.8663 0.8211 0.8000 0.7448 0.7343 compared approaches require either message exchange for sharing the utility table, or the use of a centralized authority. For example, the greedy solution would require message exchange to commu- nicate users' preferences and resolve collisions in a decentralized setting, and every batching approach would require a common centralized synchronization clock. Table 1 presents the empirical competitive ratio for the first day of the week of the employed dataset. As we can see, even in the extreme, unlikely scenarios where we assume that people would be willing to wait for more than 10 or 20 minutes (which correspond to large values of q), ALMA achieves high relative efficiency, compared to the off-line (infeasible) benchmark. These scenarios are favorable to the off-line optimal, because requests stay longer in the system and therefore the algorithm takes more advantage of its foreseeing capabilities (hence, the drop in the competitive ratios is observed in all of the employed algorithms). In particular, ALMA achieves an empirical competitive ratio of 0.67 for q = 1 and even better ratios for more realistic scenarios (as large as 0.85). The just-in-time and batch versions of the maximum weight matching perform slightly better, but this is to be expected, as they compute the maximum weight matching on the graphs of the current requests. In spite of the unpredictability of the on-line setting, and the dynamic nature of the demand, ALMA is consistently able to exhibit high performance, in all of the employed scenarios. 4 CONCLUSION Algorithms for solving the assignment problem, whether central- ized or distributed, have runtime that increases with the total prob- lem size, even if agents are interested in a small number of re- sources. Thus, they can only handle problems of some bounded size. Moreover, they require a significant amount of inter-agent communication. Humans on the other hand are routinely called upon to coordinate in large scale in their everyday lives, and are able to fast and robustly match with resources under dynamic and unpredictable demand. One driving factor that facilitates human co- operation is the principle of altruism. Inspired by human behavior, we have introduced a novel anytime heuristic (ALMA) for weighted matching in both bipartite and non-bipartite graphs. ALMA is de- centralized and requires agents to only receive partial feedback of success or failure in acquiring a resource. Furthermore, the running time of the heuristic is constant in the total problem size, under reasonable assumptions on the preference domain of the agents. As autonomous agents proliferate (e.g., IoT devices, intelligent in- frastructure, autonomous vehicles, etc.), having robust algorithms 10 that can scale to hundreds of thousands of agents is of utmost importance. The presented results provide an empirical proof of the high quality of the achieved solution in a variety of scenarios, including both synthetic and real data, time constraints and on-line settings. Furthermore, both the proven theoretical bound (which guarantees constant convergence time), and the computation speed comparison (which grounds ALMA to a proven fast centralized algorithm), argue for its applicability to large scale, real world problems. As future work, it would be interesting to identify meaningful domains in which ALMA can provide provable worst-case performance guarantees, as well as to empirically evaluate its performance on other real datasets, corresponding to important real-world, large- scale problems. REFERENCES [ABD+18] Itai Ashlagi, Maximilien Burq, Chinmoy Dutta, Patrick Jaillet, Amin Saberi, and Chris Sholley. Maximum weight online matching with deadlines. arXiv preprint arXiv:1808.03526, 2018. [AESW11] Niels Agatz, Alan L Erera, Martin WP Savelsbergh, and Xing Wang. Dy- namic ride-sharing: A simulation study in metro atlanta. Procedia-Social and Behavioral Sciences, 17:532 -- 550, 2011. [Ber79] DP Bertsekas. A distributed algorithmfor the assignment problem. Labo- ratory for Information and Decision Systems Working Paper, Massachusetts Institute of Technology, Cambridge, MA, 1979. [BEY05] Allan Borodin and Ran El-Yaniv. Online computation and competitive analysis. cambridge university press, 2005. [BJ65] CW Borchardt and CGJ Jocobi. De investigando ordine systematis aequa- tionum differentialium vulgarium cujuscunque. Journal für die reine und angewandte Mathematik, 64:297 -- 320, 1865. [BNBA12] Mathias Bürger, Giuseppe Notarstefano, Francesco Bullo, and Frank All- göwer. A distributed simplex algorithm for degenerate linear programs and multi-agent assignments. Automatica, 48(9):2298 -- 2304, 2012. [CF11] Ludek Cigler and Boi Faltings. Reaching correlated equilibria through multi-agent learning. In The 10th International Conference on Autonomous Agents and Multiagent Systems-Volume 2, pages 509 -- 516. International Foundation for Autonomous Agents and Multiagent Systems, 2011. [CN11] S. F. Cheng and T. D. Nguyen. Taxisim: A multiagent simulation platform for evaluating taxi fleet operations. In 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology, volume 2, pages 14 -- 21, Aug 2011. [CR02] Gary Charness and Matthew Rabin. Understanding social preferences with simple tests*. The Quarterly Journal of Economics, 117(3):817 -- 869, 2002. [Dan90] George B Dantzig. Origins of the simplex method. ACM, 1990. [DF18] Panayiotis Danassis and Boi Faltings. Learning in ad-hoc anti-coordination scenarios. AAAI Spring Symposium Series, 2018. [DF19] Panayiotis Danassis and Boi Faltings. Courtesy as a means to coordi- nate. In Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS '19. International Foundation for Autonomous Agents and Multiagent Systems, 2019. [Edm65] Jack Edmonds. Maximum matching and a polyhedron with 0, 1-vertices. Journal of research of the National Bureau of Standards B, 69(125-130):55 -- 56, 1965. [EK72] Jack Edmonds and Richard M. Karp. Theoretical improvements in algo- rithmic efficiency for network flow problems. J. ACM, 19(2):248 -- 264, April 1972. [FOW02] Uriel Feige, Eran Ofek, and Udi Wieder. Approximating maximum edge coloring in multigraphs. In International Workshop on Approximation Algorithms for Combinatorial Optimization, pages 108 -- 121. Springer, 2002. [GA13] Tyler Gunn and John Anderson. Dynamic heterogeneous team formation for robotic urban search and rescue. Procedia Computer Science, 19:22 -- 31, 2013. The 4th International Conference on Ambient Systems, Net- works and Technologies (ANT 2013), the 3rd International Conference on Sustainable Energy Information Technology (SEIT-2013). [GC13] Yanfeng Geng and Christos G Cassandras. New âĂIJsmart parkingâĂİ system based on resource allocation and reservations. IEEE Transactions on Intelligent Transportation Systems, 14(3):1129 -- 1139, 2013. [Gin00] Herbert Gintis. Strong reciprocity and human sociality. Journal of theoret- ical biology, 206(2):169 -- 179, 2000. [GLM10] Stefano Giordani, Marin Lujak, and Francesco Martinelli. A distributed algorithm for the multi-robot task allocation problem. In International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems, pages 721 -- 730. Springer, 2010. [IS17] S. Ismail and L. Sun. Decentralized hungarian-based approach for fast and scalable task allocation. In 2017 International Conference on Unmanned Aircraft Systems (ICUAS), pages 23 -- 28, June 2017. [Kuh55] Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics (NRL), 2(1-2):83 -- 97, 1955. American Mathematical Soc., 2009. [Lew08] David Lewis. Convention: A philosophical study. John Wiley & Sons, 2008. [LP09] László Lovász and Michael D Plummer. Matching theory, volume 367. [Mun57] James Munkres. Algorithms for the assignment and transportation prob- lems. Journal of the society for industrial and applied mathematics, 5(1):32 -- 38, 1957. [NKL17] Duc Thien Nguyen, Akshat Kumar, and Hoong Chuin Lau. Collective multiagent sequential decision making under uncertainty. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. [Nor98] James R Norris. Markov chains. Number 2. Cambridge university press, 1998. [NS05] Martin A Nowak and Karl Sigmund. Evolution of indirect reciprocity. Nature, 437(7063):1291, 2005. [Oll09] François Ollivier. Looking for the order of a system of arbitrary ordinary differential equations. Applicable Algebra in Engineering, Communication and Computing, 20(1):7 -- 32, 2009. [PS82] Christos H. Papadimitriou and Kenneth Steiglitz. Combinatorial Optimiza- tion: Algorithms and Complexity. Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 1982. [Reg92] Vernon Rego. Naive asymptotics for hitting time bounds in markov chains. 2015. Acta Informatica, 29(6):579 -- 594, 1992. [Sim16] Jay Simon. On the existence of altruistic value and utility functions. Theory and Decision, 81(3):371 -- 391, 2016. [SKKR10] Peter Stone, Gal A. Kaminka, Sarit Kraus, and Jeffrey S. Rosenschein. Ad hoc autonomous agent teams: Collaboration without pre-coordination. In Proceedings of the Twenty-Fourth Conference on Artificial Intelligence, July 2010. [Su15] Hsin-Hao Su. Algorithms for fundamental problems in computer networks. [Tal13] Mohammad Sadegh Talebi. Uncoupled learning rules for seeking equilibria [Var16] Pradeep Varakantham. Sequential decision making for improving effi- in repeated plays: An overview. CoRR, abs/1310.5660, 2013. ciency in urban environments. 2016. [VCGA12] Pradeep Reddy Varakantham, Shih-Fen Cheng, Geoff Gordon, and Asrar Ahmed. Decision support for agent populations in uncertain and congested environments. 2012. [ZSP08] Michael M Zavlanos, Leonid Spesivtsev, and George J Pappas. A distributed auction algorithm for the assignment problem. In Decision and Control, 2008. CDC 2008. 47th IEEE Conference on, pages 1212 -- 1217. IEEE, 2008. 11 A APPENDIX A.1 Proof of Theorem 2.1 Theorem 2.1. For N agents and R resources, the expected number of steps until the system of agents following Alg. 1 converges to a complete matching is bounded by (10), where p∗ = f (loss∗) and loss∗ is given by Eq. 11. (cid:18) (cid:18) (cid:19)(cid:19) (cid:18) 1 p∗ log N + R n), 1 − max R 2 − p∗ 2(1 − p∗) r ∈R,n∈N(lossr min r ∈R,n∈N(lossr n) O loss ∗ = arg min loss r n (cid:19) (10) (11) In this section we provide a formal proof of Theorem 2.1. 6 To facilitate the proof, we will initially assume that every agent, on every collision, backs-off with the same constant probability, i.e.,: Pn(r , ≺n) = p > 0,∀n ∈ N,∀r ∈ R (12) A.1.1 Case #1: Multiple Agents, Single resource (R = 1). We will describe the execution of the proposed learning rule as a discrete time Markov chain (DTMC) 7. In every time-step, each agent per- forms a Bernoulli trial with probability of 'success' 1 − p (remain in the competition), and failure p (back-off). When N agents com- pete for a single resource, a state of the system is a vector {0, 1}N denoting the individual agents that still compete for that resource. But, since the back-off probability is the same for everyone (Eq. 12), we are only interested in how many agents are competing and not which ones. Thus, in the single resource case (R = 1), we can de- scribe the execution of the proposed algorithm using the following chain: Definition A.1. Let {Xt }t ≥0 be a DTMC on state space S = {0, 1, . . . , N} denoting the number of agents still competing for the resource. The transition probabilities are as follows: Pr(Xt +1 = N Xt = 0) = 1 Pr(Xt +1 = 1Xt = 1) = 1 Pr(Xt +1 = jXt = i) = (cid:18)i (cid:19) j restart absorbing i > 1, j ≤ i pi−j(1 − p)j (all the other transition probabilities are zero) Intuitively, this Markov chain describes the number of individ- uals in a decreasing population, but with two caveats: The goal (absorbing state) is to reach a point where only one individual remains, and if we reach zero, we restart. Before proceeding with Theorem 2.1's convergence proof, we will restate Mityushin's Theorem [Reg92] for hitting time bounds in Markov chains, define two auxiliary DTMCs, and prove two auxiliary lemmas. Theorem A.2. (Mityushin's Theorem [Reg92]) Let A = {0} be the absorbing state of a Markov chain {Xt }t ≥0. If E(Xt +1Xt = i) < i β , ∀i ≥ 1 and some β > 1, then: 6The proof is an adaptation of the convergence proof of [CF11] and [DF19]. 7For an introduction on Markov chains see [Nor98] 12 E(T A i ) < ⌈logβ i⌉ + β β − 1 (13) where T A i. i denotes the hitting time of a state in A, starting from state Definition A.3. Let {Yt }t ≥0 be a DTMC on state space S = {0, 1, . . . , N} with the following transition probabilities (two ab- sorbing states, 0 and 1): Pr(Yt +1 = 0Yt = 0) = 1 Pr(Yt +1 = 1Yt = 1) = 1 Pr(Yt +1 = jYt = i) = (cid:18)i j (cid:19) pi−j(1 − p)j absorbing absorbing i > 1, j ≤ i (all the other transition probabilities are zero) Definition A.4. Let {Zt }t ≥0 be a DTMC on state space S = {0, 1, . . . , N} with the following transition probabilities (state 0 the only absorbing state): Pr(Zt +1 = 0Zt = 0) = 1 Pr(Zt +1 = jZt = i) = (cid:18)i j (cid:19) pi−j(1 − p)j absorbing i ≥ 1, j ≤ i (all the other transition probabilities are zero) Lemma A.5. The expected hitting time of the set of absorbing states A = {0}, starting from state Z0 = N , of the DTMC {Zt } of Definition A.4 is bounded by O(cid:16) 1 p (cid:17) log N . Proof. If the DTMC {Zt } is in state Zt = i, the next state Zt +1 is drawn from a binomial distribution with parameters (i, 1 − p). Thus, the expected next state is E(Zt +1Zt = i) = i(1 − p). Using Theorem A.2 with β = 1 1−p E(T A N ) = O results in the required bound: (cid:18) 1 log N (14) (cid:19) p □ Corollary A.6. The expected hitting time of the set of absorbing states A = {0, 1}, starting from state Y0 = N , of the DTMC {Yt } of Definition A.3 is bounded by O(cid:16) 1 log N (cid:17) . p Proof. The expected hitting time of the absorbing state of {Zt } is an upper bound of the expected hitting time of {Yt }. This is because any path that leads into state 0 in {Zt } either does not go through state 1 (thus happens with the same probability as in {Yt }), or goes through state 1. But, state 1 in {Yt } is an absorbing state, hence in the latter case the expected hitting time for {Yt } would be one step shorter. □ Let hA i denote the hitting probability of a set of states A, starting from state i. We will prove the following lemma. if i ∈ A if i (cid:60) A pijhA j , (16) By replacing pij with the probabilities of Definition A.3, the system of equations 16 becomes: (cid:0)i j (cid:1)pi−j(1 − p)jhA j , if i ∈ A if i (cid:60) A (17) = = 1, j∈S i hA i hA i = 1, = j=0 hA i hA i Lemma A.7. The hitting probability of the absorbing state {1}, starting from any state i ≥ 1, of the DTMC {Yt } of Definition A.3 is given by Eq. 15. This is a tight lower bound. {1} i h = Ω ,∀i ≥ 1 (15) (cid:18) 2(1 − p) (cid:19) 2 − p {1} i ∆= h Proof. For simplicity we denote hi . We will show that 2−p ,∀i ≥ 1 using induction. First note 2(1−p) for p ∈ (0, 1), hi ≥ λ = that since state {0} is an absorbing state, h0 = 0, h1 = 1 ≥ λ and that λ ∈ (0, 1). for a set of states A is the minimal non-negative solution to the system of linear equations 16: The vector of hitting probabilities hA = (hA i : i ∈ S = {0, 1, . . . , N}) Base case: 2p(1 − p) 1 − (1 − p)2 2 = h0 = ≥ λ h2 + 2p(1 − p)h1 + p h2 = (1 − p)2 2(1 − p) 2 − p Inductive step: We assume that ∀j ≤ i − 1 ⇒ hj ≥ λ. We will prove that hi ≥ λ,∀i > 2. (cid:19) i (cid:18)i pi−j(1 − p)jhj hi = j j=0 i−1 i−1 j=2 (cid:19) (cid:19) (cid:18)i (cid:18)i j pi−j(1 − p)jhj = pih0 + ipi−1(1 − p)h1 + + (1 − p)ihi ≥ pih0 + ipi−1(1 − p)h1 + + (1 − p)ihi = ipi−1(1 − p) + [1 − pi − (1 − p)i − ipi−1(1 − p)]λ + (1 − p)ihi pi−j(1 − p)j λ ipi−1(1 − p) + [1 − pi − (1 − p)i − ipi−1(1 − p)]λ j=2 j ⇒ hi = ⇒ hi = λ − pi 1 − (1 − p)i λ + We want to prove that: 1 − (1 − p)i ipi−1(1 − p) 1 − (1 − p)i (1 − λ) hi ≥ λ ⇒ ipi−1(1 − p) 1 − (1 − p)i (1 − λ) ≥ pi 1 − (1 − p)i λ ⇒ ipi−1(1 − p) ≥ [pi + ipi−1(1 − p)]λ ⇒ pi ipi−1(1 − p) + pi − pi ≥ λ ⇒ pi + ipi−1(1 − p) pi + ipi−1(1 − p) ≥ λ ⇒ 1 − pi + ipi−1(1 − p) ≥ 2(1 − p) 2 − p pi + ipi−1(1 − p) ≥ 1 − p 2 − p pi + ipi−1(1 − p) ≤ p ⇒ 2 − p 1 − 1 − pi pi pi ⇒ ⇒ pi(2 − p) ≤ p[pi + ipi−1(1 − p)] ⇒ pi(2 − p) ≤ pi[p + i(1 − p)] ⇒ 2 − 2p − i + ip ≤ 0 ⇒ 2 − i − p(2 − i) ≤ 0 ⇒ (2 − i)(1 − p) ≤ 0 ⇒ 2 − i ≤ 0 which holds since i > 2. The above bound is also tight since ∃i ∈ S : hi = λ, specifically □ h2 = λ. Now we can prove the following theorem that bounds the con- vergence time of the DTCM of Definition A.1, which corresponds to the proposed learning rule for the case of a single available resource (R = 1) and constant back-off probability. O (cid:19) Theorem A.8. The expected hitting time of the set of absorbing states A = {1} of the DTMC {Xt } of Definition A.1, starting from any initial state X0 ∈ {0, 1, . . . , N}, is bounded by: 2 − p 2p(1 − p) log N (cid:18) (cid:17) steps (Corollary A.6). Thus, the expected Proof. Using Lemma A.7 we can derive that the DTMC {Xt } 2−p 2(1−p) passes until it hits state 1. Each needs in expectation 1 λ log N pass requires O(cid:16) 1 hitting time of state A = {1} is O(cid:16) (cid:17). 2−p 2p(1−p) log N (18) □ = p A.1.2 Case #2: Multiple Agents, Multiple resources (R > 1). Theorem A.9. For N agents and R resources, assuming a constant back-off probability for each agent, i.e., Pn(r , ≺n) = p > 0,∀n ∈ N,∀r ∈ R, the expected number of steps until the system of agents following of Alg. 1 converges to a complete matching is bounded by (19). (cid:18) 2 − p 2(1 − p) R (cid:18) 1 p (cid:19)(cid:19) log N + R (19) O 13 log N Proof. At most N agents can compete for each resource. We call this period a round. During a round, the number of agents competing for a specific resource monotonically decreases, since that resource is perceived as occupied by non-competing agents. Let the round end when either 1 or 0 agents compete for the re- source. Corollary A.6 states that in expectation this will require p If all agents backed-off, it will take on average R steps until at least one of them finds a free resource. We call this period a break. In the worst case, the system will oscillate between a round and a break. According to the above, one oscillation requires in (cid:17) steps. O(cid:16) 1 expectation O(cid:16) 1 expected number of oscillations is bounded by O(cid:16) (cid:17) steps. If R = 1, Lemma A.7 states that (cid:17). Thus, 2−p 2(1−p) oscillations. For R > 1 the in expectation there will be 1 λ log N + R = p we derive the required bound (19). □ A.1.3 Dynamic back-off Probability. So far we have assumed a constant back-off probability for each agent, i.e., Pn(r , ≺n) = p > 0,∀n ∈ N,∀r ∈ R. In this section we will drop this assumption. Let ψ = max(log N , R). Bound (19) becomes: 2−p 2(1−p) R 2 2 − p 2(1 − p) p ψ + 1 (20) Intuitively, the worst case scenario corresponds to either all agents having a small back-off probability, thus they keep on com- peting for the same resource, or all of them having a high back-off probability, thus the process will keep on restarting. These two scenarios correspond to the inner ( 1 2(1−p)) probability p terms of bound (20) respectively. We can rewrite the right part of bound (20) as: ) and outer ( 2−p 2 − p 2(1 − p) 1 p 1 1 − p 1 2 = τ + = + + 1 lim p→1−τ = lim (21) As seen by Eq. 21, τ assumes its maximum value on the two extremes, either with a high (p → 1−), or a low (p → 0+) back- p→0+τ = ∞. Let p∗ = f (loss∗) be off probability, i.e., the worst between the smallest or highest back-off probability any agent n ∈ N can exhibit, i.e., having loss∗ given by Eq. 23. Using p∗ instead of the constant p, we bound the expected convergence (cid:18) 1 time according to bound (22). 2 − p∗ p∗ log N + R 2(1 − p∗) n), 1 − max r ∈R,n∈N(lossr min r ∈R,n∈N(lossr n) ∗ = arg min loss r n (cid:18) (cid:18) (cid:19)(cid:19) (22) (23) (cid:19) loss O R This concludes the proof of Theorem 2.1. (cid:18) O (cid:18) 1 (cid:19)(cid:19) (cid:19) (cid:18) 1 p 14
1310.6704	1	1310	2013-10-24T18:49:31	A Hierarchical Dynamic Programming Algorithm for Optimal Coalition Structure Generation	[ "cs.MA" ]	We present a new Dynamic Programming (DP) formulation of the Coalition Structure Generation (CSG) problem based on imposing a hierarchical organizational structure over the agents. We show the efficiency of this formulation by deriving DyPE, a new optimal DP algorithm which significantly outperforms current DP approaches in speed and memory usage. In the classic case, in which all coalitions are feasible, DyPE has half the memory requirements of other DP approaches. On graph-restricted CSG, in which feasibility is restricted by a (synergy) graph, DyPE has either the same or lower computational complexity depending on the underlying graph structure of the problem. Our empirical evaluation shows that DyPE outperforms the state-of-the-art DP approaches by several orders of magnitude in a large range of graph structures (e.g. for certain scalefree graphs DyPE reduces the memory requirements by $10^6$ and solves problems that previously needed hours in minutes).	cs.MA	cs	A Hierarchical Dynamic Programming Algorithm for Optimal Coalition Structure Generation Meritxell Vinyals, Thomas Voice, Sarvapali Ramchurn, Nicholas R. Jennings School of Electronics and Computer Science University of Southampton, UK Abstract We present a new Dynamic Programming (DP) for- mulation of the Coalition Structure Generation (CSG) problem based on imposing a hierarchical organiza- tional structure over the agents. We show the efficiency of this formulation by deriving DyPE, a new optimal DP algorithm which significantly outperforms current DP approaches in speed and memory usage. In the classic case, in which all coalitions are feasible, DyPE has half the memory requirements of other DP approaches. On graph-restricted CSG, in which feasibility is restricted by a (synergy) graph, DyPE has either the same or lower computational complexity depending on the underly- ing graph structure of the problem. Our empirical eval- uation shows that DyPE outperforms the state-of-the- art DP approaches by several orders of magnitude in a large range of graph structures (e.g. for certain scale- free graphs DyPE reduces the memory requirements by 106 and solves problems that previously needed hours in minutes). Introduction (i.e. this the last A key part of any coalition formation process involves partitioning the set of agents into the most effective coalitions, the optimal coalition structure). How- ever, this Coalition Structure Generation problem (CSG) is akin to the set partitioning problem and hence NP- Hard (Sandholm et al. 1999). Over few years, several optimal CSG algorithms have been designed to combat complexity (Service and Adams 2011; Rahwan, Michalak, and Jennings 2012). cases, these algorithms were formulated for the classic CSG model in which all coalitions are feasible. In contrast, in this paper, we tackle the problem in which coalition membership is restricted by some kind of (synergy) graph. Such restrictions have been widely studied in the context of cooperative game theory (Greco et al. 2011; Demange 2004) since they naturally reflect many real-life settings, such as communication networks (Myerson 1977) and logistic networks (Johnson and Gilles 2000). In most In these restricted settings, Dynamic Programming (DP) approaches are attractive since they can solve the CSG problem by simply assigning an infinite negative Copyright c(cid:13) 2020, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. value to non-feasible coalitions. To date, all DP algo- rithms build on the same DP formulation of the CSG problem, due to (Rothkopf, Pekec, and Harstad 1995). As noted in (Rahwan and Jennings 2008), this DP formula- tion leads to a redundant search of the CSG space al- though some of this unnecessary calculations are avoided by IDP, the fastest DP algorithm for classic CSG. For (sparse) graph-restricted CSG, the fastest algorithm is DyCE (Voice, Ramchurn, and Jennings 2012), that outper- forms IDP by several orders of magnitude in sparse graphs by restricting the DP formulation to feasible coali- tions in the graph. Despite these advances, to date, the CSG problem can only be solved optimally for up to 32 agents, even when considering graph restrictions (Voice, Ramchurn, and Jennings 2012). Against this background, this paper presents DyPE, a new optimal DP algorithm which significantly outperforms cur- rent DP approaches and scales to larger problems. DyPE op- erates on a novel formulation of the CSG problem that im- poses a hierarchical structure over the set of agents. In more detail, this paper makes the following contributions: • We introduce a new DP formulation for the CSG prob- lem that builds its search on a pseudotree hierarchy of the agents' synergy graph. We further show how this formu- lation enables a new search of the CSG space in which the coalitions an agent can join are conditioned on the coali- tions formed by agents in earlier positions; • We propose and prove the correctness of our new algo- rithm, DyPE (Dynamic programming Pseudotree-based optimal coalition structure Evaluation) which is an effi- cient implementation of the hierarchical DP formulation; • We analyse the complexity of DyPE showing that it has either the same or lower computational complexity (de- pending on the structure of the synergy graph) than the current state-of-the-art DP algorithms; • We empirically show that DyPE solves the CSG problem faster than both IDP and DyCE in a range of graph struc- tures, including the classic case (e.g. with a tree-restricted problem with 40 agents it is 104 times faster and reduces by 107 times the memory requirements). Moreover, for particular graph classes, DyPE solves the CSG problem for hundreds of agents in minutes. This paper is organised as follows. We proceed with a back- ground section, followed by formulation, algorithm, com- P[C] P [1] P [2] P [3] P [1, 2] P [1, 3] P [2, 3] P [1, 2, 3] n o i t a r e m u n E SC v(1) v(2) v(3) v(1, 2) P [1] + P [2] v(1, 3) P [1] + P [3] v(2, 3) P [2] + P [3] v(1, 2, 3) P [1] + P [2, 3] P [1, 3] + P [2] P [1, 2] + P [3] # s13 s12 s11 s9 s10 s7 s8 s5 s6 s1 s2 s3 s4 {1}{2}{3} s6 s1 0 s 8 {1, 2}{3} {1}{2, 3} {1, 3}{2} s4 s 3 s2 {1, 2, 3} (a) DP computations. Figure 1: a) Sequences of computations performed by DP and b) CS graph given A = {1, 2, 3}. plexity and empirical sections, and then conclusions. (b) CS graph. Background Basic Definitions Let A = {1, . . . , A} be a set of agents. A subset C ⊆ A is termed a coalition. We denote the coalition composed of all agents in A as the grand coalition and the coalition com- posed of a single agent i as its singleton. A CSG problem is completely defined by its characteristic function v : 2A → ℜ (with v(∅) = 0), which assigns a real value representing util- ity to every coalition. The CSG problem involves finding the exhaustive disjoint partition of the set of agents into coali- tions (or, Coalition Structure (CS)) CS = {C1, . . . , Ck} so i=1 v(Ci), is maximised. that the total sum of values,Pk Now consider a CSG problem in which not all coalitions are assumed feasible. Rather their feasibility is restricted by a synergy graph G = (A, E) where: (i) each node of the graph represents an agent; and (ii) a coalition C is al- lowed to form iff every two agents in C are connected by some path in the subgraph induced by C.1 We denote the set of feasible coalitions in a G-restricted CSG problem as F (G). The characteristic function of a G-restricted CSG problem returns minus infinity for any non-feasible coali- tion ∀C 6∈ F (G) : v(C) = −∞. We will denote as K3 the complete graph among three agents and as L3 the graph G = ({1, 2, 3}, {{1, 2}, {2, 3}}) in which the three agents interact in a line. Then, in a L3-restricted CSG problem agent 1 can not form a coalition with agent 3 without agent 2 (e.g. C = {1, 3} 6∈ F (G)). In preparation for the further description of DP algo- rithms, we will denote, for any subset C ⊆ A, Φ(C) as the connected components of the induced subgraph of C on G and ΠC k as the set of all partitions of C into k parts. Existing DP Approaches for CSG DP solves optimization problems recursively: a problem is solved by independently solving a collection of subprob- lems. In CSG, a subproblem P [C] stores the value of the best CS that can be formed among agents in C ⊆ A. 1It is noteworthy that this representation subsumes classic CSG: any classic CSG problem can be modeled as a graph-restricted one by assuming a complete graph among agents. Current DP algorithms compute P [C] by splitting the CSG search space into n subspaces (SC): one containing coalition C and one for each partition of C into two sets Ck, Cl: P [C] = max(cid:18)v(C), max (Ck ,Cl)∈ΠC 2 P [Ck] + P [Cl](cid:19) (1) Figure 1a shows the trace of this DP formulation specify- ing the set of evaluated subproblems (P [C]), the subspaces evaluated for each subproblem (SC) and the number of sub- space (#) over the classic CSG among three agents. To make sure that a problem is computed before its sub- problems in Equation 1, DP algorithms iteratively com- pute subproblems by size: subproblems corresponding to all coalitions of size 1, to all coalitions of size 2, and so on un- til it reaches the grand coalition (note the direction of the enumeration arrow in Figure 1a). IDP algorithm by (Rahwan and Jennings 2008) is an improved implementation of Equation 1 that prunes the evaluations of some subspaces that are proven to be redundant during the search. DyCE algorithm by (Voice, Ramchurn, and Jennings 2012), specifically devised for graph-restricted problems, implements a variant of this DP formulation that restricts coalitions in Equation 1 to be feasible in the graph (Ck, Cj ∈ F (G)). Thus, in Figure 1a, if the CSG problem is restricted to the L3 graph, DyCE omits the evaluation of subproblem P [1, 3], as well as of subspace s3, since both involve {1, 3} 6∈ F (G) whereas IDP goes through all subspaces independently of the graph. The operation of these algorithms is typically visualized on the coalition structure (CS) graph. In this graph, nodes stand for coalition structures and, following Equation 1, an edge connects two coalition structures iff one of the coalition structures can be obtained from the other by splitting one coalition into two. Figure 1b depicts the CS graph among three agents with edges numbered with the number of the corresponding subspace that generated it. A Hierarchical Formulation for CSG This section presents a novel hierarchical DP formulation of the CSG problem based on a pseudotree of the agents' synergy graph. This pseudotree structure allows us to de- fine a more efficient search of the CSG space in which the coalitions an agent can join are explored conditioned on the coalitions formed by agents in earlier positions in the hier- archy. We further show how this search can be visualised in a particular graph of coalition structures that we refer to as a hierarchical coalition structure (HCS) graph. Synergy Graph Pseudotree A pseudotree (PT) is a directed tree structure commonly used in search and inference procedures (?). A pseudotree P T of synergy graph G is a rooted tree with agents A as nodes and the property that any two agents that share an edge in G are on the same branch in P T . Here we restrict our at- tention to edge-traversal pseudotrees, namely those whose edges correspond to edges in G. An edge-traversal pseu- dotree of a graph G can be computed by running a depth-first traversal search (DFS) algorithm (?). Specifically, Figure 2 2 1 3 2 3 1 (a) PT for K3. (b) PT for L3. Figure 2: Pseudotrees for synergy graphs K3 and L3. P[C] P [3] P [1] P [1, 3] P [1, 2, 3] n o i t a r e m u n E SC v(3) v(1) v(1, 3) v(1) + P [3] v(1, 2, 3) v(2) + P [1, 3] v(1, 2) + P [3] v(2, 3) + P [1] s# s8 s7 s5 s6 s1 s2 s3 s4 {1}{2, 3} {1}{2}{3} 6 s {1, 3}{2} s3 2 s s4 {1, 2, 3} {1, 2}{3} (a) Hierarchical DP. (b) HCS graph. Figure 3: a) Hierarchical DP and b) HCS graph for the clas- sic CSG with O = {2, 1, 3}. Frontier coalitions are under- lined. P[C] P [3] P [1] P [1, 2, 3] n o i t a r e m u n E SC v(3) v(1) v(1, 2, 3) v(2) + P [1] + P [3] v(1, 2) + P [3] v(2, 3) + P [1] s# s6 s5 s1 s2 s3 s4 (a) Hierarchical DP. {1}{2, 3} {1}{2}{3} s3 s4 2 s {1, 2}{3} {1, 2, 3} (b) HCS graph. Figure 4: a) Hierarchical DP and b) HCS graph for the 3L- restricted CSG with O = {2, 1, 3}. Frontier coalitions are underlined. depicts two P T s where boldfaced edges are those included in the P T , dashed edges are those in G that are not included in the P T , and the boldfaced node is the root agent. Figure 2a shows a PT for the synergy graph of a 3-agent classic CSG problem rooted at agent 2. Similarly, Figure 2b shows a pseudotree for the L3 graph rooted at agent 2. Unlike in the K3 graph, agent 1 and 3 here can be in different branches since they are not directly connected by an edge. We define the ancestors of an agent i in the P T as all the agents in the path between i and the root. Then, a pseu- dotree defines a partial hierarchical ordering among agents in which any agent should be placed before any of its an- cestors in the graph. We will denote OP T as one ordering for pseudotree P T . Notice that for the P T in Figure 2a, OP T = {2, 1, 3} is the unique ordering that satisfies the P T . In contrast, for the P T in Figure 2b, OP T = {2, 1, 3} and OP T = {2, 3, 1} are both valid orderings. Yet, OP T not only defines an ordering among agents but also on the set of feasible coalitions. Let i be the agent with lowest order included in a coalition C. Then, we define the order of C, O(C), as the position of i in O. Formally, O(C) = min{iO(i) ∈ C}. Thus, given OP T = {2, 1, 3} the order of {1, 2, 3} is 1 whereas of {1, 3} is 2. Hierarchical DP Formulation We next present the hierarchical DP formulation for classic CSG illustrating its operation with two simple examples. First, consider the classic CSG problem among three agents A = {1, 2, 3}. Solving this problem involves com- paring the value of five CSs (depicted as nodes in the CS graph in Figure 1b). Now, suppose that agents are organised as in the PT of Figure 2a with an ordering OP T = {2, 1, 3}. Given the lowest agent in the ordering, agent 2, the CSG space can be divided into four subspaces, one for each fea- sible coalition that contains this agent; namely: s1 contain- ing {{1, 2, 3}}; s2 containing CSs that include {2} and any CS among 1, 3; s3 containing CSs that include {1, 2} and any CS among 3; and s4 containing CSs that include {2, 3} and any CS among 1. Notice that finding the best CS in each subspace involves solving a CSG subproblem and that the agents in this subproblem depend on the particular coali- tion that agent 2 formed in this subspace. Thus, for example, the solution of s2 involves computing the best CS among agents not present in coalition {2}, i.e. computing subprob- lem P [1, 3]. Again, we can solve this problem by taking the agent with lowest order (agent 1), and repeating the above process. Figure 3a shows a complete trace of this exam- ple. It is noteworthy that we reduced the number of oper- ations with respect to the current DP operation: we solved 4 subproblems by evaluating 8 subspaces whereas the current DP operation solves 7 problems by evaluating 13 subspaces (compare Fig. 3a with Fig. 1a). This hierarchical DP search can be similarly applied to any graph-restricted problem. For example, we can follow the same approach in the CSG problem when restricted by the L3 graph (a complete trace is given in Figure 4a). In this case when computing subspace s2, P [1, 3] can be de- composed into two independent subproblems, namely P [1] and P [3], since agents 1, 3 do not interact (are disconnected) given agent 2 formed a coalition without including them. Again, note that this search is more efficient than the cur- rent DP operation: it solves 3 subproblems by evaluating 6 subspaces where the graph-restricted current DP operation solves 6 problems by evaluating 10 subspaces (compare Fig. 4a with Fig. 1a). In general, given an ordering O among agents, the so- lution of a G-restricted CSG problem can be computed as comparing (maximising) over n subspaces where the value of each subspace can be solved by evaluating a feasible coalition and a set of subproblems corresponding to the con- nected components2 of the rest of agents not present in this coalition. Formally, we can define our hierarchical DP for- mulation for the graph-restricted CSG problem as: P [C] = max Ck⊆C: O(Ck)=O(C), Ck∈F (G)  v(Ck) + XCl∈Φ(C\Ck)  P [Cl]  (2) Hierarchical Coalition Structure Graph To discuss how to construct a HCS graph we need first to define the notion of frontier coalitions. Let us define the frontier coalitions of a coalition structure CS as the set of coalitions that are not connected to any other coalitions of 2The connected components of a graph G are the set of the largest subgraphs of G that are connected. higher order in CS. That is, the frontier coalitions are the C ∈ CS such that if O(C′) > O(C) for C′ ∈ CS it implies C ∪ C′ is disconnected. In the HCS graph, nodes stand for feasible coalition structures and, following Equa- tion 2, an edge connects two feasible coalition structures if and only if one of the coalition structures can be obtained from the other by the evaluation of some subspace of one of its frontier coalitions. That is CS is linked to CS′ if and only if there exists C ∈ CS, C′ ∈ CS′ with C′ ⊆ C, CS′ = (CS \ {C}) ∪ {C′, Φ(C \ C′)} and C′ is a frontier coalition of CS. Figure 3b depicts the HCS graph of the 3-agent classic CSG with O = {2, 1, 3} and Figure 4b the graph when the problem is restricted by the L3 graph. Frontier coalitions are underlined in the graph. Notice that frontier coalitions cor- respond to subproblems that would be evaluated during the operation of the hierarchical DP formulation. DyPE We can now describe the operation of DyPE. First, we lay its formal foundations, namely how its search is efficiently derived from the hierarchical DP formulation, then we move to its algorithmic details and, finally, prove its correctness. Formal Foundations DyPE implements the hierarchical recursive formulation in- troduced in the former section. Accordingly, DyPE: (i) enu- merates all subproblems in a bottom-up order (from sub- problems that appear last in the recursion to the grand coali- tion); and (ii) computes the value of each enumerated sub- problem. In the hierarchical recursion formulation a subproblem is computed using the results of a set of subproblems of higher order (a subspace of a subproblem C contains a coalition of the same order as C and a set of subproblems of higher order). Thus, to guarantee a valid exploration order (so no problem is evaluated before one of its subproblems) DyPE evaluates subproblems corresponding to feasible coalitions by its order, from highest to lowest. However, not all fea- sible coalitions are required during the recursion. For ex- ample, the DP execution in Figure 3a requires enumerating subproblems {3}, {1}, {1, 3}, and {1, 2, 3} but not {2, 3}, although it corresponds to a feasible coalition. Detecting which of the subproblems will actually be needed3 is crucial for the performance of the DP implemen- tation. DyPE exploits the fact that the ordering is based on a synergy pseudotree to detect a necessary condition that any feasible coalition needs to satisfy in order to be evaluated as a subproblem during the recursion. In particular, let C be a feasible coalition. If C contains the root of P T , DyPE will only evaluate C if it is the grand coalition (C = A). Other- wise, if C does not include the root, DyPE will only evaluate C if the set of remaining agents, A \ C, is connected. The correctness of these claims is formally proved in next sec- tions, by proving the correctness of DyPE. 3Although a simple solution for only solving subproblems that are actually needed is memoization (?), the exponential overhead that incurs recursion in this case makes it not applicable. P [C] ← −∞; while C ′ ← nextConnectedSet(C ′, i, C) do Algorithm 1 DyPE( v(·), A, G, O) 1: C ← ∅; S ← ∅; /Current subproblem, current subspace/ 2: i ← A; /Start exploring the last agent in the ordering O/ 3: while (C, i) ←nextSubproblem(C, i) do 4: 5: 6: 7: 8: 9: 10: 11: 12: end while 13: return bestCS(A); V ← v(C ′) + PC′′∈Φ(C\C′) P [C ′′]; if P [C] < V /Compare the value of subspaces/ then P [C] ← V ; /Update subproblem value/ B[C] ← C ′; /Update the best subspace/ end if end while return (A, i); /Only the grand coalition / end if return ∅; /All agents explored, return empty set./ while C ← nextConnectedSet(C, i, {i, . . . , A}) do if C = ∅ then Algorithm 2 nextSubproblem(C, i) 1: if i = 1 /For the root agent / then 2: 3: 4: 5: 6: else 7: 8: 9: 10: 11: 12: if A \ C is connected then return (C, i); 13: end if end if end while return nextSubproblem(∅,i − 1); /Recursively call with the previous agent in the ordering/ The Algorithm For notational convenience, we use nextConnectedSet(·, ·, ·) as an iterator function of a connected subgraph enumeration (CSE) algorithm (Voice, Ramchurn, and Jennings 2012; ?). That is, for any feasible coalitions C′ ⊆ C with i ∈ C′ nextConnectedSet(C′, {i}, C), returns the subset of C that would follow C′ during the process of the chosen CSE al- gorithm as it iterates through all feasible subcoalitions of C that contain i. If C′ is the last subset to be enumerated by the CSE, the function returns the empty set. The pseudocode of DyPE is provided in Algorithm 1. As can be seen, DyPE takes as an input a G-restricted CSG problem and an ordering O that satisfies a synergy pseu- dotree of G. Let us assume, without loss of generality, that agents are numbered according to an ordering that satis- fies P T , so the root is 1. After initialisation, DyPE pro- ceeds to enumerate subproblems, using the iterator func- tion nextSubproblem(·, ·). For each agent i in the ordering O, DyPE goes through all subproblems that need to be eval- uated where i is the agent with lowest order. For each subproblem C, Algorithm 1 computes the value of the CSG problem over agents in C (lines 4-11). To do so, it goes over all subspaces of C that need to be evaluated (SC) by iteratively calling function nextConnectedSet(·, ·, ·) (line 5). In this way, DyPE evaluates one subspace for each feasible subcoalition C′ of C that contains agent i. The value of the subspace is computed as the value of coalition C′, v(C′), plus the value of each subproblem corresponding to each connected component in C \C′, Φ(C \C′) (line 6). The value of P [C] is computed in Algorithm 1 as the maximum between the values of the evaluated subspaces (lines 7-10). At the end of this process, the solution of the subprob- lem corresponding to the grand coalition (P [A]), contains the value of the best CS explored during the execution of the algorithm. To recover the best CS, Algorithm 1 also stores the subspace that maximizes each subproblem C in B[C] (line 12) and at the end of its execution calls a recur- sive procedure bestCS(·) over the grand coalition, where bestCS(C) returns {C} if C \ B[C] = ∅; bestCS(B[C]) ∪ SC ′∈Φ(C\B[C]) bestCS(C′) otherwise. The definition of nextSubproblem(·, ·) is given in Algo- rithm 2. For agents i = A . . . 2 (i.e. excluding the root) DyPE, uses nextConnectedSet(·, ·, ·) to enumerate as sub- problems every feasible coalition C consisting of agent i and any subset of agents placed after i in the ordering, {i, . . . , A} (line 7). DyPE evaluates subproblem C only if the rest of the graph, A \ C, remains feasible (lines 8-10). Lastly, for the root agent (i = 1) , DyPE evaluates a single subproblem corresponding to the grand coalition (lines 1-5). Correctness of DyPE The next theorem proves the correctness of DyPE. Theorem 1 (Correctness) For any given graph-restricted CSG, DyPE calculates an optimal coalition structure. Proof. Since the value of the best coalition structure CS∗ returned by DyPE is equal to P [A], it is sufficient to show that, for every feasible coalition structure CS, on completion of the algorithm, P [A] ≥ v(CS). Given such CS, let L be the order of the coalition with highest order in CS. Let C≤l be the set of coalitions in CS with order equal or lower than l (C≤l = {C ∈ CSO(C) ≤ l}). We prove the result by showing that for all l = 1 . . . L P [A] ≥ Vl where Vl contains the accumulated value (until step l) of an exploration "path", Vl = XC∈C≤l v(C) + XC ′∈Φ(A\C≤l) P [C′], and so VL = v(CS). We prove this by induction on l. In the base case, l = 1 and C≤1 is composed of a sin- gle coalition, C1, the coalition that contains the root agent in CS. Notice that all the subspaces of the grand coali- tion A corresponding to feasible coalitions that contain the root are evaluated. Thus, the subspace C1 of A (v(C1) + PC∈Φ(A\C1) P [C]) is evaluated and so P [A] ≥ V1. In the inductive step, consider the coalition in CS whose level is l + 1, Cl+1, (if there is a coalition with order l + 1, this coalition is unique since it is the one that contains the agent with order l + 1) and that the induction hypothesis holds for all coalitions in CS whose level is less or equal than l. Thus, CS≤l = {C≤l, Φ(A \ C≤l)} is connected through a path to the grand coalition A. Then, there must be one CC ∈ Φ(A \ C≤l) such that Cl+1 ⊆ CC. Since the ordering follows a pseudotree, the union of coalitions in C≤l forms a connected subgraph, A \ CC must be feasible and thus, CC is evaluated as a subproblem. As all agents in CC have higher order than l, O(Cl+1) = O(CC) and v(Cl+1) +PC ′∈Φ(CC\Cl+1) P [C′] must be evaluated in the computation of P [CC]. Thus, Vl+1 ≤ Vl, and so, by the inductive hypothesis, Vl+1 ≤ P [A]. Complexity Analysis 2 is 2A−1 − 1. In tree-restricted CSG, ΠA Next, we determine the complexity of DyPE and compare it to those of DyCE and IDP. Notice that each of these algo- rithms, for each evaluated subproblem: stores its value (and possibly its best space) and evaluates a number of subspaces (with a linear number of operations per subspace). Accord- ingly, we assess their complexity based on the number of subproblems (memory requirements) and the number of sub- spaces evaluated (computational requirements). Memory. DyPE evaluates, in addition to the grand coalition, one subproblem for C ∈ F (G) such that A \ C is feasible and C does not contain the root. This number is equal to the number of feasible coalition structures composed of one (the 2 ). Thus, the memory grand coalition) or two coalitions (ΠA requirements of DyPE are within O(ΠA 2 ). In classic CSG, 2 is equal to ΠA the number of edges in the tree (removing exactly one edge is the only way to disconnect the tree into two connected subsets), so DyPE has memory requirements within O(A). Table 1 shows how the memory requirements of DyPE compares to those of IDP and DyCE on graph-restricted CSG. Table 1 also highlights the particular cases of classic and tree-restricted CSG. Observe that independently from the graph the complexity of IDP is exponential in the num- ber of agents, whereas of DyCE is linear in the number of feasible coalitions. Since the number of feasible coalitions will always be greater than the number of coalition struc- tures composed of two coalitions, the memory requirements of DyPE are bounded above by those of DyCE. In classic CSG, although of the same order, the memory requirements of DyPE are one half those of IDP or DyCE as among sub- problems that contain the root DyPE only stores the grand coalition. Computation. DyPE evaluate subspaces which are sub- sets of subproblems. Thus, the computational complexity is bounded by a constant times the number of pairs of subsets C′, C with C′ ⊆ C, which is O(3A). This is the same or- der of complexity as IDP (and of DyCE in the classic CSG). In classic CSG, DyPE omits the evaluation of all subsets of subproblems that include the root node with exception of those of the grand coalition. Thus, we cannot hope to do better than O(3A). In tree-restricted CSG, for each agent i, DyPE evaluates exactly one subproblem with i as its lowest order agent. Thus, each feasible coalition can only generate a subspace in one subproblem, namely the one that has the same order. Conversely, the subproblem with i as its lowest order agent contains agent i and all agents reachable from i with higher level, and so all feasible coalitions generate ex- actly one subspace. Thus, the computational complexity of DyPE in this case is within O(F (G)), and thus expected to be much lower than those of DyCE since the latter evaluates a potentially large set of subspaces for each feasible coali- tion. Indeed, in the next section we show empirically that this holds true for a wide range of graph structures. Graph (G) O(2A) Classic Tree (T) O(2A) O(2A) IDP DyCE O(F (G)) O(2A) O(F (T )) DyPE O(ΠA O(2A) O(A) 2 ) Table 1: Memory requirements for the graph-restricted, clas- sic, and tree-restricted CSG problems. Experimental Evaluation We evaluate DyPE and compare its performance against IDP and DyCE on a variety of different synergy graph topolo- gies. We then go on to examine the issue of scalability. Benchmarking DyPE In our comparison, we take a similar approach to (Voice, Ramchurn, and Jennings 2012), and investigate per- formance over the following graph classes: random trees (RT), scalefree graphs (SF) (using the standard Barabasi- Albert preferential attachment generation model, with pa- rameters k = 1, 2, 3) and complete graphs (CG). Due to long runtimes we extrapolated the results as follows: from 23 agents onwards for IDP, from 27 onwards for DyCE on RT and SF k = 1, and on 24 onwards for DyCE on k = 2. For each configuration, we run 50 instances recording the number of evaluated subproblems and the running time of each algorithm. We now present the results of this compari- son. Figures 5 (a)-(b) show the results of our performance evaluation over random trees. The memory requirements for DyPE are up to 7 orders of magnitude lower than for DyCE and up to 11 orders of magnitude lower than for IDP (for 40 agents). This is because as the number of agents in- creases, memory requirements grow exponentially for IDP and DyCE, and only linearly for DyPE. In terms of run- time, DyPE can solve problems of 40 agents in about 20 minutes compared against 20 days for DyCE, and years for IDP. These results are in line with the intuition given by our complexity analysis section. The results of our performance evaluation over scalefree graphs are depicted in Figures 5 (c)-(d). For k = 1 the mem- ory requirements of DyPE for 30 agents are up to 6 orders of magnitude lower than for DyCE and up to 9 orders lower than for IDP. For k = 2, these savings are reduced, but still significant; 2 orders of magnitude better with respect to DyCE and 3 orders of magnitude better with respect to IDP in graphs with 30 agents. These results follow the intuition given in our complexity analysis that the computational sav- ings provided by DyPE more significant on sparse graphs. Turning to execution time, DyPE can solve problems with 30 agents in minutes (or hours when k = 2) instead of hours (or days for k = 2) for DyCE. Finally, our results over complete graphs are in line with the complexity analysis, which predicted a similar performance for all algorithms, excepting that: (i) DyCE takes more time than IDP and DyPE due to its less effective pruning; and (ii) DyPE uses half of the memory of the other approaches. Scalability of DyPE We have seen that DyPE performs well on sparse graphs (e.g., trees or scale free with k = 1). However, as argued Figure 6: Runtimes for DyPE on random trees (d=2,3,4). in (Voice, Ramchurn, and Jennings 2012), if the degree of agents is not bounded, even trees can lead to an exponential number of coalitions (e.g., a star has 2A−1 − 1). Based on this, we evaluated DyPE on random trees with bounded de- gree. In particular, Figure 6 shows the execution time where the degree of agents is bounded by d, for d = 2, 3, 4. Ob- serve that for d = 3 and d = 4, DyPE is able to run to completion for problems with 50 agents within 15 minutes and 3 hours respectively. For the particular case of d = 2, DyPE solves problems with 1000 agents within minutes. Conclusions We presented DyPE, a DP algorithm that implements a novel hierarchical DP formulation for CSG using a hierarchy based on pseudotrees. We proved that DyPE is optimal and that it improves upon current DP approaches with savings that go from linear to exponential, depending on the struc- ture of the underlying synergy graph. Our empirical results showed, that DyPE greatly improves on the state-of-the-art, in some cases by several orders of magnitude. Concretely, for random trees with bounded degree, DyPE managed to quickly find the optimal coalition structure for 1000 agents, when even 50 would be intractable for other DP approaches. in (Rahwan, Michalak, and Jennings 2012; the inter- Service and Adams 2011), we ested in enhancing the presented hierarchical DP approach with anytime properties. future work, field particularly following current trends As are (a) Memory requirements RT. (b) Execution Time RT. (c) Memory requirements SF. (d) Execution Time SF. Figure 5: Results for random tree (RT) graphs (a) (b) and scale free (SF) graphs (c)(d). manageable combinatorial auctions. Management Science 44(8):1131 -- 1147. [Sandholm et al. 1999] Sandholm, T.; Larson, K.; Anders- son, M.; Shehory, O.; and Tohm´e, F. 1999. Coalition struc- ture generation with worst case guarantees. Artif. Intell. 111(1-2):209 -- 238. [Service and Adams 2011] Service, T. C., and Adams, J. A. 2011. Constant factor approximation algorithms for coali- tion structure generation. Autonomous Agents and Multi- Agent Systems 23(1):1 -- 17. [Voice, Ramchurn, and Jennings 2012] Voice, T.; Ramchurn, S. D.; and Jennings, N. R. 2012. On coalition formation with sparse synergies. In AAMAS, 223 -- 230. References [Demange 2004] Demange, G. 2004. On Group Stability in Hierarchies and Networks. Journal of Political Economy 112(4):754 -- 778. [Greco et al. 2011] Greco, G.; Malizia, E.; Palopoli, L.; and Scarcello, F. 2011. On the complexity of compact coalitional games. In IJCAI, 147 -- 152. [Johnson and Gilles 2000] Johnson, C., and Gilles, R. 2000. Spatial social networks. Review of Economic Design (5):273300. 1977. Graphs and co- [Myerson 1977] Myerson, R. B. operation in games. Mathematics of Operations Research 2(3):225 -- 229. [Rahwan and Jennings 2008] Rahwan, T., and Jennings, N. R. 2008. An improved dynamic programming algorithm for coalition structure generation. In AAMAS, 1417 -- 1420. [Rahwan, Michalak, and Jennings 2012] Rahwan, Michalak, T. P.; and Jennings, N. R. algorithm for coalition structure generation. In AAAI. [Rothkopf, Pekec, and Harstad 1995] Rothkopf, M. H.; Pekec, A.; and Harstad, R. M. 1995. Computationally T.; 2012. A hybrid This figure "partialResults_ScaleFree_median.jpg" is available in "jpg"(cid:10) format from: http://arxiv.org/ps/1310.6704v1 This figure "partialResults_Trees_median.jpg" is available in "jpg"(cid:10) format from: http://arxiv.org/ps/1310.6704v1 This figure "solvingTimes_ScaleFree_largescalemedian.jpg" is available in "jpg"(cid:10) format from: http://arxiv.org/ps/1310.6704v1 This figure "solvingTimes_ScaleFree_median.jpg" is available in "jpg"(cid:10) format from: http://arxiv.org/ps/1310.6704v1 This figure "solvingTimes_Trees_median.jpg" is available in "jpg"(cid:10) format from: http://arxiv.org/ps/1310.6704v1
1508.04531	1	1508	2015-08-19T05:58:36	Modeling emergence of norms in multi-agent systems by applying tipping points ideas	[ "cs.MA" ]	Norms are known to be a major factor determining humans behavior. It's also shown that norms can be quite effective tool for building agent-based societies. Various normative architectures have been proposed for designing normative multi-agent systems (NorMAS). Due to human nature of the concept norms, many of these architectures are built based on theories in social sciences. Tipping point theory, as is briefly discussed in this paper, seems to have a great potential to be used for designing normative architectures. This theory deals with the factors that affect social epidemics that arise in human societies. In this paper, we try to apply the main concepts of this theory to agent-based normative architectures. We show several ways to implement these concepts, and study their effects in an agent-based normative scenario.	cs.MA	cs	Modeling emergence of norms in multi-agent systems by applying tipping points ideas Francisco Lopez School of Computer Science University of Manchester Manchester, United Kingdom [email protected] 5 1 0 2 g u A 9 1 ] A M . s c [ 1 v 1 3 5 4 0 . 8 0 5 1 : v i X r a ABSTRACT Norms are known to be a major factor determining humans behavior. It's also shown that norms can be quite effective tool for building agent-based societies. Various normative architectures have been proposed for designing normative multi-agent systems (NorMAS). Due to human nature of the concept norms, many of these architectures are built based on theories in social sciences. Tipping point theory, as is briefly discussed in this paper, seems to have a great poten- tial to be used for designing normative architectures. This theory deals with the factors that affect social epidemics that arise in human societies. In this paper, we try to apply the main concepts of this theory to agent-based normative architectures. We show several ways to implement these con- cepts, and study their effects in an agent-based normative scenario. Keywords norms, multi-agent systems, tipping points INTRODUCTION 1. Human societies are simultaneously frustratingly unchang- ing and yet susceptible to "epidemics" that sweep across the social fabric causing people to adopt previously rare prac- tices. Tipping point theories attempt to explain the subtle triggers behind these social processes. In 2000, Malcolm Gladwell [25] produced a popular science book summariz- ing three key factors which trigger tipping points: 1) scale- free networks (the Law of the Few); 2) effective messaging (the Stickiness Factor) and 3) environmental influences (the Power of Context). This paper relates tipping point theory to the process of norm emergence in multi-agent systems; we propose that normative agent architectures can serve an ex- cellent computational model for expressing many contagious social phenomena, including tipping points and information cascades. Social norms are known to be a major factor governing hu- mans' behavior; unbeknownst to us, many of our everyday behaviors are influenced by these implicit standards [10]. Various normative architectures have been proposed for de- signing normative multi-agent systems (NorMAS) capable of reasoning about norm adoption. Some of these systems have been grounded in social science theory, but the aim of many architectures is simply to effectively address standard multi-agent system challenges, including agreement forma- tion, coordination and conflict resolution [9]. Despite recent research progress in the area, the complete life-cycle of norms is far from fully understood. The complex nature of human decision-making makes comprehending the rationale behind social interactions difficult, since people are notoriously bad at self-reporting their motivations [15]. The field of agent-based modeling aims to create agents in the image of humans. These agents typically have cognitively- inspired decision-making components, and are situated in life-like scenarios. In both standard multi-agent systems and cognitively-inspired models, existing social theories have been employed toward the construction of normative models [5]. Various stages of the norm life-cycle including recogni- tion, adoption, compliance and emergence are often modeled on similar concepts in social sciences. This paper proposes a unified model of how norm emergence in networked agent societies can be used to predict the ef- fects of common tipping point triggers. Previous work on norm emergence in networks has investigated the effects of social network topology in static [34, 31] and also dynamic networks [29]. Yu et al. [39] presented an evaluation of dif- ferent learning methods on norm emergence in networked systems. In our work, we simply employ network structures as a medium to apply ideas from tipping point theory re- lating to the Law of the Few. Therefore, the structure of agents' network is not of interest by itself, other than mak- ing it congruent with human social networks. Our main contribution is showing the role and significance of tipping point principles in normative agent systems, and evaluating the potential impact of this model on NorMAS design. The next section provides an overview of related work in this area. 2. RELATED WORK Hollander and Wu [24] refer to three categories of normative studies in the social sciences: 1) social function of norms [15], 2) impact of social norms [7], 3) mechanisms leading to the emergence and creation of norms [14]. In the con- text of social function, norms are often concerned with the oughtness and expectations of agent behavior; where ought- ness refers to the condition where an agent should or should not perform an action, and expectation refers to the behav- ior that other agents expect to observe from that agent [13, 2]. An example of work belonging to this category is Boella and Torre's architecture containing separate subsystems for counts-as conditionals, conditional obligations, and condi- tional permissions [16]. Within the second category, social impact, norms are con- sidered in terms of cost provided to or imposed on the par- ties involved in a social interaction [4]. For instance, pun- ishment and sanctions are introduced as two enforcement mechanisms used to achieve the necessary social control re- quired to impose social norms [33]. Here they demonstrate a normative agent that can punish and sanction defectors and also dynamically choose the right amount of punishment and sanction to impose [12]. As noted in [24], the third category is concerned with the how of norms more than the why. Relevant work in Nor- MAS domain that falls into this category is divided into the following groups [11]. The first group is composed of re- search in which norms are practically hard-wired into the system or dictated directly to the agents. In the second set, norms emerge from social interactions among agents. Sen and Airiau's work [32] in which agent interactions are modeled using payoff matrices, focuses on norm emergence through social learning in agent societies. Conte et al.'s EMIL was designed as an architecture for modeling norm emergence [17]. The EMIL architecture includes a dynamic cognitive model of norm emergence and innovation [6]. Much of the existing work in normative multi-agent sys- tems explicitly or implicitly relies on social science theo- ries [8]. In a recent work, some of the well-known theories of philosopher David Hume were evaluated using an agent- based model called HUME2.0 [17]. This work demonstrates how social justice concepts can even emerge from heteroge- neous agents that are not endowed with norm representa- tions. Self-determination theory is also referenced by some of the normative works [3]. Here the focus is on the agents' mo- tivation and the extent to which the motivation is intrinsic or extrinsic. Neumann studied existing normative architec- tures to see how much they comply with self-determination theory [27]. Practice theory is an example drawn from anthropology; this theory describes how changes in the society are based on the interactions between the human agents and social structure. For instance, an agent-based model for energy demand and supply social practices is presented in [1], which shows how energy consumption norms form and evolve in urban soci- eties. The next section provides background on tipping point theory, primarily from a sociology perspective. 3. TIPPING POINT THEORY The term, "tipping point" was initially coined in physics to describe the situation in which the state of an object rapidly changes from one stable equilibrium to another different equilibrium. Morton Grodzins was the first to use term in social sciences to describe an interesting phenomenon he ob- served in the US cities, known as white flight [23]. His obser- vation was that in some metropolitan areas the percentage of African-American people would increase up to a certain point. After that point, those with white ethnicity immi- grated from those cities in large numbers. Thomas Schelling presented the general theory of tipping, which describes how individuals' micromotives and microbehavior can aggregate in the big picture [30]. The model of collective behavior introduced by Mark Granovetter [22] uses thresholds to de- termine the path of social events. This model was initially used to describe how fads are created. In normative studies, tipping points are usually denoted as the point of maximum return at which time the behavior has the highest level of acceptability from the population. For instance, in a certain group of friends, the number of times they shower in a week may vary, but a specific value has the highest acceptability by group members as the conventional pattern of behavior. In this paper, we study the impact of Gladwell's three factors on norm emergence in agent-based normative systems and demonstrate practical ways to apply this versatile theory. 4. EXPERIMENTAL SETUP For our experiments, we employ the classic scenario, rules of the road, that is frequently used to study normative be- havior in multi-agent systems. In this scenario, there exists a population of agents that do not have any preference to- ward driving on the left or right side of a two-way road. No rules or higher enforcement exist to determine the preferred side. This scenario represents a two-action stage game that models the situation where agents need to agree on one of several equally desirable alternatives. The societal norms that we would like to evolve are either driving on the left or driving on the right [32]. In this scenario agents receive a fixed value reward and pun- ishment based on the following payoff matrix shown in Ta- ble 1. Table 1: Payoff matrix for rules of the road scenario As Yu et al. [39] note, although this payoff matrix appears simple, the coordination game poses a very challenging puz- zle for human beings to solve efficiently. The game has two pure Nash-equilibria: both agents drive left or both agents drive right. Classical game theory, however, does not give a coherent account of how people would play a game like this. The conundrum is that there is nothing in the struc- ture of the game itself that allows the players (even purely rational players) to infer what they ought to do. In reality, people can play such games because they can rely on some contextual cues to agree on a particular equilibrium [38]. In similar studies on normative systems, usually the cumula- tive payoff (reward) of the whole population of agents is used as a measure of comparing various methods (see [32] and [39] left right left 1,1 -1,-1 right -1,-1 1,1 This equation shows the weight of the link connecting neigh- bor j to node i. C refers to the corresponding centrality value (degree, betweenness and closeness). Also, Degi de- notes the number of neighbors for node i. The top 10 percent of the population of agents with the greatest centrality val- ues are assumed to be the key elements of a society. At the beginning of our experiments, all of the agents follow a sin- gle norm; in other words, all of them have learned (through social learning [32]) to always drive on one side of the road. In our implementation, each agent has a utility value defined for each of four possible cases: Up-Left, Up-Right, Down- Left and Down-Right, where Up and Down determine the section of road, and Left and Right determine the direction an agent drives. These values are updated while receiving payoffs based on the matrix shown in Table 1. In our experiments, we compare the penetration of norm changing behaviors that emanate from key members of a society vs. other cases. We compare emanation from the top to emanation from the middle and bottom 10 percent of the population. At the beginning of the simulation, the agents (nodes) are ranked based on their centrality value to determine the top, middle and bottom agents. The utility value of these agents is kept fixed. Neighbors of these agents continue updating their behavior until a new norm emerges in the system. Figure 2, Figure 3, and Figure 4 show the number of iterations required for each case to converge. The population of agents contained 100 agents, and the reported results show the average values over 20 runs. Figure 2: Average number of iterations until the emergence of one norm, when using degree centrality to determine key agents. for examples). Instead, we opt to use the norm emergence time for each method as an evaluation metric. This is func- tionally equivalent since the payoff received by all agents post norm emergence is the same, hence a method which leads to faster norm emergence will also yield the higher cumulative payoff. 5. KEY FEW MEMBERS In this section, we study the effects of key members of an agent society on the rate of norm emergence. These key members are selected using standard heuristics for measur- ing influence within a network; we evaluate the performance of three centrality measures: degree, closeness, and between- ness. Degree centrality measures the number of edges con- nected to a node. Closeness is calculated based on the total distance to all other nodes. Nodes with a high betweenness centrality fall on a large proportion of the shortest paths (geodesics) in the graph. To model the characteristics of a real social network, we use an algorithm introduced in [35] to create a synthetic network which follows power law degree distribution and exhibits ho- mophily, a greater number of link connections between sim- ilar nodes.1 The network generator uses link density (ld ) and homophily (dh) to govern network formation. A simpli- fied version of the pseudo-code for this method is shown in Figure 1. For our model, we assumed predefined values for ld and dh. The nodes of the graph represent the individu- als (agents) in the simulation, who can be considered as car drivers. G = Null repeat sample r from uniform distribution U (0, 1) if r ≤ ld then randomChooseSource(G) determineCandidateSink(dh,G) pickSink() connect(source,sink) (cid:46) based on power-law distribution else add a new node to G end if until desired number of nodes added to the network Figure 1: Synthetic friendship network generator We use a weighted voting approach (also known as a struc- ture based method) to determine an agent's decision with regard to its neighbors. The weight for each of an agent's neighbors is computed using a normalized value of that neigh- bor's centrality value as shown in Equation 1. weighti,j = Cj ΣDegi k=1 Ck (1) Figure 3: Average number of iterations until the emergence of one norm, when using betweenness centrality to determine key agents. 1Commonly described as "birds of a feather flock to- gether" [26] The pattern observed in all of three cases was very similar. 40608010005001000150020002500% Norm BehavingTime (Ticks)Top 10%Middle 10%Bottom 10%4060801000500100015002000% Norm BehavingTime (Ticks)Top 10%Middle 10%Bottom 10%Figure 4: Average number of iterations until the emergence of one norm, when using closeness centrality to determine key agents. When the norm propagation starts from the top 10% of the population, the norm emerges much faster compared to the other cases. Moreover, there is a fairly sizable difference among top, middle and bottom agents. The magnitude of difference between the top and middle 10% is more than the difference between the middle and bottom. These results are consistent with the role of connectors in tipping point theory. 6. STICKINESS FACTOR According to the tipping point theory, the extent and rate of emerging social norms in a society is not only related to the members of the society, but also related to the content of the message. An effective message needs to be interesting or "sticky" enough to remain in agents' minds. This fac- tor is almost completely independent of the society and its structure, and is a property of the idea. As Gladwell [21] points out, it is potentially very compli- cated to determine if a certain message has the necessary stickiness or not, but one characteristic that is usually com- mon to sticky ideas is that it frequently returns to a person's mind. This could be in the form of a desire to sit and watch a popular TV show every night, or in a more extreme case, a clinical addiction to smoking or gambling. Conventional marketing and advertising domains refer to this phenomenon as rule of 27. According to this rule, a message (advertise- ment) should be seen at least 27 times, if the message is going to stick [28]. Figure 5: Average number of iterations until the emergence of one norm, when 2 out of 4 agents with fixed utility values play twice with each agent that they encounter. Figure 6: Average number of iterations until the emergence of one norm, when 2 out of 4 agents with fixed utility values go (drive) faster. In order to model this property, we assume that the stick- iness is represented by the number of games that an agent plays with another agent. Therefore a higher number of games will result in the same effect as a stickier belief. In our experiments, this idea is evaluated in two different ways. The first way is to increase the number of games that a cer- tain set of agents play. The second way is to have a certain number of agents driving faster than other agents to be ex- posed to more cars. 4060801000500100015002000250030003500% Norm BehavingTime (Ticks)Top 10%Middle 10%Bottom 10%406080100050010001500% Norm BehavingTime (Ticks)Normal caseSome playing more406080100060012001800% Norm BehavingTime (Ticks)Normal caseSome driving fasterFigures 5 and 6 show results related to these two cases. In both cases, we have our original 100 agents plus a group of 2 agents which have a fixed preference to drive on either the left or right. In the first scenario, one group of agents plays two games each time it encounters another agent. In the sec- ond scenario, one group of agents moves faster. Both of these scenarios lead to the same effect: increasing the number of times that an agent is exposed to an idea. This simulates the property of frequently returning to a person's mind. In both cases, when the stickiness factor is implemented, the entire system converges to a single norm faster. 7. POWER OF CONTEXT The third element of the tipping point theory refers to the power of context. As Gladwell points out: it is possible to be a better person on a clean street or in a clean subway, than in one littered with trash and graffiti [21]. The idea is mostly based on what known in criminology as the theory of broken windows [37]. According to this theory, slight changes in the environment could result in tipping effects over the whole society. In order to apply this part of the tipping point theory, we use ideas from the methods for studying fads and cascading effects in networks [36]. First, we build a network using the same approach described in Section 5. Then, we assign a threshold value for each agent. Similar to the probabilis- tic information cascade models, if the cumulative value of the perceived cascade is less than the threshold, nothing will change. If it's higher, the agent will change its cur- rent behavior, which in our scenario would result in driving on the other side of the road. Figure 7 shows the percent- age of times that a norm emerged in the system for a set of threshold values. The columns show the average results over 20 runs. Agents were selected randomly as a source of a small initial shock in the network, which results in negat- ing the current payoff values for driving on each side of the road. The frequency of shocks is determined randomly. The system runs until it reaches some fixed iteration number (50,000), unless a different norm is observed. This experi- ment illustrates how minor shocks can shape a population fad, resulting in a population-level behavior change. The shocks (pulses) in this model can be viewed as any of the small changes that tipping point theory predicts can result in large changes in the whole society. According to the re- sults presented in Figure 7, thresholds as small as 0.02 can lead to the emergence of norms in the system in almost 5 percent of the experiments. The computed values for each agent are compared to its tipping point value (normalized between 0 and 1). There is a second aspect to the power of context, which refers to the number of people in groups. The Rule of 150 says that the size of groups is a subtle contextual factor that makes a big difference. This number is referred as Dunbar's number [18], after the anthropologist who originally proposed the idea. In groups with fewer than 150 members, people will cooperate relatively easily and rapidly become infected with the community ethos. Once that threshold is crossed people begin to behave very differently. 150 is our social chan- nel capacity as determined on the basis of personal loyalties and 1-on-1 contacts. Beyond the tipping point of 150 the group dynamics simply become too complex. For the aver- Figure 7: Percentage of times that a norm emerges in the population, when agents have different threshold values for activating. Figure 8: Average number of iterations until the emergence of one norm, when the network structure of agents follows a power-law distribution and when the network is a complete clique. age person there are just too many relationships to manage. The group then becomes divided and alienated, and usually splits into two. Smaller groups have been shown to be more effective at tasks than larger groups. This may be due to biological limitations of humans which make it very difficult for them to handle a larger community. With the growth of virtual social media sites and the spread of online groups, there has been renewed interest in eval- uating the importance of this limit on Facebook [19], Mys- pace [20] and within massively multiplayer online role-playing games (MMORPGs). The pivotal issue here is that a person cannot maintain a close relationship with all of the mem- bers of a larger group which ultimately sabotages its suc- cess. Having a direct connection with each member of the group is a necessary component to having a positive social relationship. We propose using a clique structure to illustrate this idea. In a clique each node has a direct edge to all of other nodes. There are n∗ (n− 1) edges in the resulting graph. We opt to use a directed graph, as that seems to be the general assump- tion for friendship networks. We compare the emergence of driving norms in a network generated using the method de- scribed in Section 5. It should be noted that having more edges does not result in faster convergence. More connec- tions makes the diffusion of ideas easier, while it makes it 0204060801000.020.040.060.080.10.120.140.160.180.20.220.240.26% of times norm emergedThreshold value406080100050010001500% Norm BehavingTime (Ticks)Power lawCliqueharder for the agents to find an idea that all agents like. In a clique structure the major voting approach and the weighted voting approach (using the number of edges) are the effectively same, so neither of them elicits earlier norm emergence. Figure 8 shows the number of iterations that were required on average for the two cases to reach norm emergence. The driving norm emerged faster in case of the clique structure than in the power-law degree distribution network. This shows the potential benefit of such a struc- ture in constructing agent systems, at least for ideal cases. 8. CONCLUSION AND FUTURE WORK Norms are complex social behaviors that have been exten- sively studied in sociology, psychology, and other related fields. Most normative architectures draw upon theories from the social sciences. The theory of tipping points has in- spired much research in different disciplines. For this paper, we model some of the well-known elements of this theory, as applied to networked agent populations. We illustrate how three of principle ideas including key few members, sticki- ness factor, and the role of environment can affect the pro- cess of norm emergence. This paper is meant to be an initial step for convincing the NorMAS community of the impor- tance of tipping point theory concepts. For future work, we are interested in mapping the performance of our normative model to a real-dataset. 9. REFERENCES [1] T. Balke, T. Roberts, M. Xenitidou, and N. Gilbert. Modelling energy-consuming social practices as agents. In Social Simulation Conference, Barcelona, 2014. [2] R. Beheshti, A. M. Ali, and G. Sukthankar. Cognitive social learners: an architecture for modeling normative behavior. 2015. [3] R. Beheshti, M. Analui, and B. Minaei-Bidgoli. A pairwise classification method with support vector machines, based on game theory and alpha-beta algorithm. In The 2nd International Conference of Iranian Operation Research Society. Civilica, 2009. [4] R. Beheshti, R. Barmaki, and N. Mozayani. Negotiations in holonic multi-agent systems. In The Seventh International Workshop on Agent-based Complex Automated Negotiations (ACAN2014), 2014. [5] R. Beheshti and N. Mozayani. Predicting opponents offers in multi-agent negotiations using artmap neural network. In Future Information Technology and Management Engineering, 2009. FITME'09. Second International Conference on, pages 600 -- 603. IEEE, 2009. [6] R. Beheshti and N. Mozayani. A new mechanism for negotiations in multi-agent systems based on artmap artificial neural network. In Agent and Multi-Agent Systems: Technologies and Applications, pages 311 -- 320. Springer, 2011. [7] R. Beheshti and N. Mozayani. Homan, a learning based negotiation method for holonic multi-agent systems. Journal of Intelligent and Fuzzy Systems, 26(2):655 -- 666, 2014. [8] R. Beheshti, N. Mozayani, and A. T. Rahmani. A new hybrid evolutionary method with ant colony and pso algorithms based on fuzzy decision making. In The 2nd International Conference of Iranian Operation Research Society. Civilica, 2009. [9] R. Beheshti and A. Rahmani. A multi-objective genetic algorithm method to support multi-agent negotiations. In Future Information Technology and Management Engineering, 2009. FITME'09. Second International Conference on, pages 596 -- 599. IEEE, 2009. [10] R. Beheshti and G. Sukthankar. Extracting agent-based models of human transportation patterns. In Social Informatics (SocialInformatics), 2012 International Conference on, pages 157 -- 164. IEEE, 2012. [11] R. Beheshti and G. Sukthankar. An agent-based transportation simulation of the ucf campus. In SwarmFest 2013: 17th Annual Meeting on Agent-Based Modeling & Simulation, 2013. [12] R. Beheshti and G. Sukthankar. Analyzing agent-based models using category theory. In Web Intelligence (WI) and Intelligent Agent Technologies (IAT), 2013 IEEE/WIC/ACM International Joint Conferences on, volume 2, pages 280 -- 286. IEEE, 2013. [13] R. Beheshti and G. Sukthankar. Improving markov chain monte carlo estimation with agent-based models. In Social Computing, Behavioral-Cultural Modeling and Prediction, pages 495 -- 502. Springer, 2013. [14] R. Beheshti and G. Sukthankar. A hybrid modeling approach for parking and traffic prediction in urban and memory effect on convention emergence. In Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology-Volume 02, pages 233 -- 240. IEEE Computer Society, 2009. [35] X. Wang, M. Maghami, and G. Sukthankar. Leveraging network properties for trust evaluation in multi-agent systems. In Proceedings of the IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology-Volume 02, pages 288 -- 295. IEEE Computer Society, 2011. [36] D. J. Watts. A simple model of fads and cascading failures on sparse switching networks. In Economics with Heterogeneous Interacting Agents, pages 13 -- 25. Springer, 2001. [37] J. Q. Wilson and G. L. Kelling. Broken windows. Atlantic Monthly, 249(3):29 -- 38, 1982. [38] H. P. Young. The economics of convention. The Journal of Economic Perspectives, pages 105 -- 122, 1996. [39] C. Yu, M. Zhang, F. Ren, and X. Luo. Emergence of social norms through collective learning in networked agent societies. In Proceedings of the International Conference on Autonomous Agents and Multi-agent Systems, pages 475 -- 482. International Foundation for Autonomous Agents and Multiagent Systems, 2013. simulations. AI & SOCIETY, pages 1 -- 12, 2014. [15] R. Beheshti and G. Sukthankar. A normative agent-based model for predicting smoking cessation trends. In Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems, pages 557 -- 564. International Foundation for Autonomous Agents and Multiagent Systems, 2014. [16] G. Boella and L. van der Torre. An architecture of a normative system: counts-as conditionals, obligations and permissions. In Proceedings of the International Conference on Autonomous Agents and Multiagent Systems, pages 229 -- 231. ACM Press, 2006. [17] R. Conte, G. Andrighetto, and M. Campennl. Minding norms: Mechanisms and dynamics of social order in agent societies. Oxford University Press, 2013. [18] R. I. Dunbar. Neocortex size as a constraint on group size in primates. Journal of Human Evolution, 22(6):469 -- 493, 1992. [19] R. I. Dunbar. Primates on facebook. 2009. Retrieved from: http://econ.st/1qgCrDL. [20] J. W. Gibbons. Modeling Content Lifespan in Online Social Networks Using Data Mining. PhD thesis, University of Kansas, 2014. [21] M. Gladwell. The tipping point: How little things can make a big difference. Hachette Digital, Inc., 2006. [22] M. Granovetter. Threshold models of collective behavior. American Journal of Sociology, 83(6), 1978. [23] M. Grodzins. Metropolitan segregation. Scientific American, 1957. [24] C. D. Hollander and A. S. Wu. The current state of normative agent-based systems. Journal of Artificial Societies and Social Simulation, 14(2):6, 2011. [25] G. Malcolm. The tipping point. Little Brown, New York, 2000. [26] M. McPherson, L. Smith-Lovin, and J. M. Cook. Birds of a feather: Homophily in social networks. Annual Review of Sociology, 27(1):415 -- 444, 2001. [27] M. Neumann. Norm internalisation in human and artificial intelligence. Journal of Artificial Societies and Social Simulation, 13(1):12, 2010. [28] D. Pepper. How to Start and Grow Your Lawn Care Maintenance Business. Lulu Enterprises Incorporated, 2008. [29] B. T. R. Savarimuthu, S. Cranefield, M. K. Purvis, and M. A. Purvis. Norm emergence in agent societies formed by dynamically changing networks. Web Intelligence and Agent Systems, 7(3):223 -- 232, 2009. [30] T. C. Schelling. Micromotives and macrobehavior. WW Norton & Company, 2006. [31] O. Sen and S. Sen. Effects of social network topology and options on norm emergence. In Coordination, Organizations, Institutions and Norms in Agent Systems V, pages 211 -- 222. Springer, 2010. [32] S. Sen and S. Airiau. Emergence of norms through social learning. In IJCAI, volume 1507, page 1512, 2007. [33] D. Villatoro, G. Andrighetto, J. Sabater-Mir, and R. Conte. Dynamic sanctioning for robust and cost-efficient norm compliance. In IJCAI, volume 11, pages 414 -- 419, 2011. [34] D. Villatoro, S. Sen, and J. Sabater-Mir. Topology
1810.10591	1	1810	2018-10-24T19:52:01	Reasoning about Norms Revision	[ "cs.MA" ]	Norms with sanctions have been widely employed as a mechanism for controlling and coordinating the behavior of agents without limiting their autonomy. The norms enforced in a multi-agent system can be revised in order to increase the likelihood that desirable system properties are fulfilled or that system performance is sufficiently high. In this paper, we provide a preliminary analysis of some types of norm revision: relaxation and strengthening. Furthermore, with the help of some illustrative scenarios, we show the usefulness of norm revision for better satisfying the overall system objectives.	cs.MA	cs	Reasoning about Norms Revision Davide Dell'Anna, Mehdi Dastani, and Fabiano Dalpiaz Department of Information and Computing Sciences, Utrecht University, Princetonplein 5, 3584 CC Utrecht, The Netherlands Abstract. Norms with sanctions have been widely employed as a mech- anism for controlling and coordinating the behavior of agents without limiting their autonomy. The norms enforced in a multi-agent system can be revised in order to increase the likelihood that desirable system properties are fulfilled or that system performance is sufficiently high. In this paper, we provide a preliminary analysis of some types of norm revi- sion: relaxation and strengthening. Furthermore, with the help of some illustrative scenarios, we show the usefulness of norm revision for better satisfying the overall system objectives. 1 Introduction Modern software systems execute in highly dynamic environments [1]. Typi- cal dynamic settings are open environments, where many (heterogeneous) au- tonomous software components coexist, interact and can join and leave as they please. Dynamicity and uncertainty can be caused by many different (often un- predictable) factors: events from the operating environment, behaviors emerging from the effects of software executions and their interactions, the change of ex- ternal norms [2] and the impact of this on software behaviour, etc. [3]. In order to guarantee desirable overall system level properties, in the multi- agent systems research field, norms with sanctions have been proposed as a means to control and coordinate the behavior of autonomous agents without limiting their autonomy [4, 5]. In many cases, however, it is infeasible for a system designer to anticipate all the possible states that the software systems and its operating environment can reach during execution [3], and to define adequate norms for each of them. Furthermore, system objectives are themselves in constant motion [6, 7]: new goals arise while others are dropped, the desired qualities vary, and the relative priority of the objectives evolves. A static predefined set of norms may often result at runtime inadequate to guarantee the overall system objectives in various contexts [8, 9]. Dynamic update of norms at runtime is therefore one of the key factors to build a versatile normative multi-agent system, able to accommodate an hetero- geneous population of agents and capable of ensuring the overall system objec- tives within a dynamic and uncertain environment [10]. In this paper we provide a preliminary analysis of some main types of norm revision that can be applied to a running system with the goal of better satisfying the overall system objectives. We define the concept of norms revision as the act of replacing a set of norms with a new set. We specialize norms revision in three different types: relaxation, strengthening and generic alteration, which differ for the type of relationship that the revised set has with the original one. Finally, we provide a series of practical examples to clarify the different types of revisions and their possible applications. This paper should be considered as the initial step of a research on runtime supervision of normative systems. We do not report on technical nor experimen- tal results, but rather lay down the foundations of our future research. The remainder of this paper is structured as follows: Section 2 frames our work within the existing literature, Section 3 discusses the problem of norms revision, Section 4 provides a series of examples of revisions of norms and possible applicative scenarios. Section 5 ends the paper with concluding remarks and future work. 2 Background Numerous papers over the years have focused on deciding and proving the cor- rectness of normative systems by model checking formulas describing desired properties such as liveness or safety properties [11, 12, 13]. Despite their ele- gance, these approaches do not fully cope properly with the dynamicity of to- day's complex systems, where the behavior of the system may change over time due to changes in the participating agents, their behaviors, or the active norms. As of today, there is no generally agreed formal methodology to reason about the dynamics of norms and their impact on system specification. However some formal frameworks emerged in recent years to cope with these problems. In the rest of this section we discuss these related works on norm dynamics. Knobbout et al. [10] propose a dynamic logic to formally characterize the dynamics of state-based and action-based norms. Both in Knobbout's works [10, 14] and in Alechina et al.'s [13], norm change is intended as norm addition. Taking these approaches as our baseline, we aim to investigate further types of norms update in order to extend the existing framework for dynamic normative systems. Aucher et al. [15] introduce a dynamic context logic in order to describe the operations of contraction and expansion of theories by introducing or removing new rules. Governatori et al. [16] investigate from a legal point of view the application of theory revision to reason about legal abrogations and annulments. The types of revision presented in this paper can be related to theory revision, but we take a multi-agent systems standpoint, in which norms revision should be studied in terms of its impact on agents autonomy and we leave for future work the study of the impact of a revision on the existing normative system. Norms update has also been studied from the perspective of approximation [17], where an approximated version of a norm is formally obtained to cope with imperfect monitors for the original norm. The concept of approximation is similar to our notion of relaxation, however it is defined with respect to a specific monitor: an approximated norm is synthesized from the original one in order to maximize the violations detectable by the available imperfect monitor. Here we assume perfect monitors and we focus instead on the relationship between different norms in terms of behaviors they allow or prohibit. Whittle and colleagues [3] present early studies on relaxation of requirements of a software system. They define a requirements language for self-adaptive sys- tems that allows to specify, with opportune operators, relaxed versions of a requirement during the requirement elicitation phase. While they focus on a language point of view, useful for the acquisition and specification of norms, in this paper we assume the norms are already specified and we provide an analysis of revision from a normative point of view based on a formal model of multi-agent system that abstracts from the language used to define the norms. 3 Norms Revision The class of models that we use to study and describe normative multi-agent systems are transition systems. These models consist of states and transitions between states. The idea is that such model describes all possible transitions of states that can occur within the system as a result of the actions that the agents perform. Definition 1 (Transition System). Given a set of atomic propositions Props, a transition system is a tuple M = (S, T, s0, V ) where S is a set of states, T ⊆ S × S is a transition relation, s0 ∈ S is a distinguished initial state and V is a valuation function V : S → Props associating propositions with states (i.e., defining which propositions hold in a state). Given a transition system M = (S, T, s0, V ), a path r through M is a sequence s0, s1, s2, . . . of states such that siT si+1 for i = 0, 1, . . .. Each path can be seen as a possible behavior of the system. We consider some of the behaviors as desired and others as undesired and we use the concept of norm to identify these behaviors. In this paper we consider conditional norms with sanctions and deadlines, as they are commonly used in the existing literature on normative multi-agent systems [18, 13] and they have proven to be a reasonable compromise between expressiveness and ease of reasoning [17]. Definition 2 (Conditional Norms). Given a propositional language LN , let cond, φ and d be boolean combinations of propositional variables from LN and let san be a propositional atom. A conditional norm n is represented by a tuple (cond; Z(φ); d; san) where Z can be either O (obligation) or F (prohibition). Given a conditional norm, cond represents the condition that must be satisfied in a state of M in order to detach the norm. A detached norm persists as long as it is not obeyed or violated or the deadline d is not reached (a state where d holds is encountered). Conditional norms are evaluated on paths of the transition system. In this paper we omit a formal definition of a violation, which is strictly dependent on the language used to express them. In the following we denote a violation formula for a conditional norm n by v(n) (see previous publication [17] for a formal definition for PLTL formulae). A norm n is violated on a path r if, and only if, v(n) holds in some state on r. We denote by Viol(M, n) the set of paths through M that violate the norm n. Fig. 1 sketches the main components of a runtime supervision framework that continuously monitors the execution of a multi-agent system (MAS), evaluates its behavior against the currently en- forced norms, and intervenes by deciding which norms should be revised. Norm violation is monitored (and sanctioned) by the Monitoring and sanctioning component. In this paper we ignore the techniques of monitoring and evaluation of norms, for which many works can be found in literature (e.g. [19]), and we assume there exists a perfect monitor for each norm. The aim of this paper is to provide a study of the possible revisions of norms and of the possible conditions when a revision may be useful. The study is a starting point to build a Norm update component able to revise at runtime the currently enforced norms in order to ensure the achievement of the overall objectives. We assume that the Monitoring and sanctioning component stores at runtime information about the obedience or violation of the norms and about the operating contexts in which they are evaluated (e.g. the hour of the day, etc.). This information, to- gether with information about the satisfaction of the overall systems objectives, is used by the Norms update component in order to decide if and how to revise the currently enforced norms. Fig. 1. The main components of the proposed runtime supervision framework aimed at revising norms enforced into a normative MAS. Dotted lines represent the scope of the components. Norm enforcement in multi-agent systems can be done in two possible ways: regimentation and sanctioning [13]. Regimentation makes bad behaviors impos- sible, i.e., it makes some of the paths of M inaccessible. With regimentation, violations of the norms are not possible, however the autonomy of the agents is reduced. Sanctioning norms violation is instead a means to discourage the violation. Sanctions are essentially treated like fines. They penalize agents when Norms with sanctionsS0S1S2.........MASMonitoring and sanctioningNorms updatethey bring the system on an undesired path, however they leave autonomy to the agents, even if their resources are reduced if they violate the norms. A regimented norm n can be described by Viol(M, n) (i.e., by the set of paths of M that violate the norm), while a norm n with a sanction is described by both Viol(M, n) and a propositional sanction atom that is asserted in case of violation of n. We assume in the following to be able to compare two sanctions (e.g., consider the case of numeric sanctions such as money). Given a set of norms N = {n1, ..., nk}, we denote by Viol(M, N ) the set of all paths through M each violating at least one norm in N (i.e., Viol(M, N ) = Viol(M, n1) ∪ ... ∪ Viol(M, nk)), and by San a conjunction of all sanctions asso- ciated with the norms in N (i.e., SanN = san1 ∧ ... ∧ sank). Given a pair of sets of norms, N1 and N2, we denote by -- N1 < N2: the fact that Viol(M, N1) ⊃ Viol(M, N2). In this case, we say that N1 is a set of norms more restrictive than N2, or that N2 is more relaxed than N1. -- N1 ≡ N2: the fact that Viol(M, N1) = Viol(M, N2). We say that the sets of norms N1 and N2 are equally restrictive. Note that the relationships between norms above reported can only be sat- isfied by pairs of sets of norms such that Viol(M, N1) ⊆ (⊇)Viol(M, N n2). Definition 3 (Norms Revision). Let N be a set of norms, a revision of N is a replacement of N with a new set N R. N R may be a more/equivalently/less restrictive set of norms or an alternative set where no relationship can be stated. We distinguish two main types of revision: relaxation and strengthening, and we consider all the other types of revisions as regular alteration. Definition 4 (Revision, Relaxation, Strengthening). A relaxation of a set of norms N is a revision of N with a new set N R such that N R > N . A strengthening of a set of norms N is a revision of N with a new set N R such that N R < N . Any other revision of N with a new set N R such that either N R ≡ N or no relationship can be stated are regular alterations of N . Given a pair of norms, n1 and n2 such that n1 = (cond1; Z(φ1); d1; san1) and n2 = (cond2; Z(φ2); d2; san2), 1. n2 is a relaxation of n1 if and only if at least one of the following holds (all else being equal): (a) cond2 is a stricter formula cond1. Note that with stricter (less strict) formula we mean a formula that is satisfied in strictly less (more) paths of M . A stricter formula for a condition of a conditional norm makes therefore the norm applicable in fewer paths than the original. (b) φ2 is a less strict formula than φ1 if Z = O (obligation), or φ2 is a stricter formula than φ1 if Z = F (prohibition). A less strict obligation (or a stricter prohibition) makes the norm violated in fewer paths, which means that more agents' behaviors are allowed. (c) d2 is a less strict formula than d1. Note that this is valid since we only accept propositional formulae, therefore a less strict deadline for a con- ditional norm means that it can be withdrawn in more states. 2. n2 is a strengthening of n1 if and only if n1 is a relaxation of n2. 3. n2 is a regular alteration of n1 if it is neither a relaxation nor a strengthening Note that while a revision of a single norm can be analyzed in terms of all its components, revision of a set of norms involving more than one norm can only be compared in terms of the set of paths of M and the sanctions. A revision N R of a set N obtained by relaxing some of the norms and strengthening some others can be compared to N only in terms of the resulting set of behaviors allowed in M , since Viol(M, N R) can be either a subset or a superset of Viol(M, N ) or even a disjoint set. However if the revised norms are all relaxed (strengthened) the resulting revised set is a relaxed (strengthened) set of N . Notably, increasing or decreasing sanctions associated to norms, under the assumption that such sanctions are comparable, is an alternative means to in- fluence the behavior of agents that goes beyond relaxation or strengthening of norms. Thus, increasing the sanctions SanN associated with a set of norms N is a way to bring the system to the desired paths defined by N , while decreasing the sanctions is a way to accept undesired paths. In the next section we provide an intuitive explanation of this concept with the help of illustrative examples. 4 Illustrative Examples In this section we provide some illustrative examples of the different types of revision of norms and some illustrative scenarios where to revise norms may be useful to ensure the overall system objectives. Note that in this paper we are not interested in how the norms that we use are acquired. Moreover, to consider only relevant norms, we assume that we are provided with information about the rationale behind norms, that relates them to the overall system objectives. Finally, we assume every parameter of the norms as perfectly monitorable. 4.1 Relaxation Consider norm n1: "cars speed in road r shall be kept below 15km/h, other- wise e10k fine", formally (for brevity assume car and road as implicit) n1 = (inRoad, F (speedAbove15),(cid:62), 10keuros), where (cid:62) provides "always" interpreta- tion. Possible examples of relaxations of n1 are the following (possibly combined): -- r1 = ((inRoad ∧ trafficHigh), F (speedAbove15),(cid:62), 10keuros). r1 differs from n1 only for the condition: (inRoad ∧ trafficHigh) is a stricter formula than inRoad. Since the condition is more specific and the rest of parameters is analogous (case 1a of Definition 4 holds), r1 is applicable in fewer paths of M , therefore fewer paths are violations and more behaviors are allowed in the system. -- r2 = (inRoad, F (speedAbove20),(cid:62), 10keuros). r2 differs from n1 only for the prohibition (case 1b): speedAbove20 is stricter than speedAbove15 because fewer paths are considered prohibited. -- r3 = (inRoad, F (speedAbove15), firstHalfCompleted, 10keuros). r3 differs from n1 only for the deadline (case 1c): firstHalfCompleted is less strict than (cid:62) ("always", e.g. firstHalfCompleted ∧ secondHalfCompleted) since the norm remains valid only when the car is in the first half of the road. Again more behaviors are allowed. Example. Consider a simple scenario where a population of autonomous vehi- cles can take a road r where their speed and the traffic level can be perfectly monitored. The designer of the system M wants to be sure that "no cars stay in road r (10km long) for more than 35 minutes". Norm n1 is currently enforced and at runtime it appears to be too strict for achieving the overall objective: if it is obeyed it prevents the fulfilment of the objective. Since n1 is too strict, one (or a combination) of the following relaxations may be useful to improve the performance of the system: -- replace n1 with r1, r2 or r3. In case of r1, choosing a stricter condition al- lows agents to perform more actions without incurring in sanctions (agents are allowed to drive above 15km/h unless the traffic level is high). Since obeying to n1 prevents the fulfillment of the objective, reducing the con- ditions where it applies means allowing behaviors previously forbidden. An analogous explanation can be provided also for r2 and r3. Notice that an analogous effect can be obtained by replacing n1 with another norm s1 = (inRoad, F (speedAbove15),(cid:62), 5euros), differing from n1 only for the sanction: 5euros < 10keuros. Reducing the fine incentives more agents to violate n1. Since the objective is achieved when n1 is violated we increase the chances the objective is achieved. 4.2 Strengthening Consider a norm n2 = (inRoad, O(speedbelow50), outOfRoad, 5euros). Possible examples of strengthening of n2 are the following (possibly combined): -- r5 = ((inRoad ∨ aroundTheRoad), O(speedbelow50), (outOfRoad ∧ 1kmFarAway), 5euros). r5 differs from n2 for both the condition and the deadline: (inRoad ∨ aroundTheRoad) is less strict than inRoad, while (outOfRoad ∧ 1kmFarAway) is stricter than outOfRoad. Since the condition is less specific, r5 is detached in more paths, and since the deadline is more specific, r5 remains valid in more cases, therefore more paths violate it and less behaviors are allowed. -- r6 = (inRoad, O(typeScooter), outOfRoad, 5euros). O(typeScooter) is stricter than O(speedbelow50): while n2 accepts all execu- tions involving vehicles (scooters or not) that keep the speed below 50km/h, r6 accepts only executions involving scooters (assumed here to have maxi- mum speed of 50km/h). Again less behaviors are allowed. Example. Consider a scenario where a population of autonomous vehicles can take a road r where their speed, type of vehicle and noise they produce can be perfectly monitored. There is an overall system objective "vehicle's noise in the neighborhood is below x db". Norm n2 is currently enforced and it is proven that violation of n2 prevents the achievement of the overall objective. Norm n2 appears at runtime to be too weak to ensure the overall system objective (e.g. it is often violated). One (or a combination) of the following strengthening of n2 may therefore be useful to improve the performance of the system: -- replace n2 with r5. Choosing a less strict condition makes the norm valid in more cases. Deincentivating agents to drive faster than 50km/h also around the road may prevent vehicles to accelerate immediately out of the road, which may be harmful for the achievement of the overall objective of keeping the noise in the neighborhood low. -- replace n2 with r6. Assuming scooters have maximum speed of 50km/h, if the new norm is obeyed by agents, the goal is ensured. Notice one more time that an analogous effect can be obtained by replac- ing n2 with another norm s2 = (inRoad, O(speedbelow50), outOfRoad, 10keuros), differing from n2 only for the sanction: 10keuros > 5euros. Increasing the fine incentives more agents to obey n2. Since the objective is achieved when n2 is obeyed we increase the chances to achieve it. 4.3 Alteration Consider norm n3: "if two cars c1 and c2 are at the opposite ends of the narrowed road, c1 shall move and c2 shall wait, otherwise 10k euros fine to c2", formally n3 = ((firstEnd(c1) ∧ secEnd(c2)), O(move(c1) ∧ wait(c2)), (cid:62), c2 10keuros). Possible examples of alteration of n3 are the following: -- r8 = ((firstEnd(c1) ∧ secEnd(c2)), O(wait(c1) ∧ move(c2)), (cid:62), c1 10keuros). -- r9 = (inRoad(c1), O(speedbelow50), (cid:62), c1 10keuros). Both r8 and r9 define a set of underused paths that is neither a subset or a superset of the paths defined by n3. Therefore they cannot be considered neither a relaxation nor a strengthening. Notice that r9 is reported as an example of a generic alteration, which reflects the definition we provided, even if it is not strictly related to n3. Example. Consider a scenario of a narrowed road and cars coming from both directions. The overall objective is "keep cars queue size in both directions be- low a threshold t". Norm n3 is currently enforced and it proved to be useful to achieve the objective when traffic is low in both directions. However, when traffic is high in the direction of car c1 the queue in the opposite direction grows too much. Norm n3 appears therefore to be not good to ensure the overall system objective in such case. An alteration such as replacing n3 with r9 may be use- ful to improve the performance of the system. Notice that a generic alteration involves a new norm that either is not strictly related to the existing one or it is equally restrictive, therefore it is hard to predict the outcome of the enforce- ment of the new norm. However, assuming the considered norms are related to the same overall system objective, enforcing an alteration in the system may lead to different (hopefully better) performances. 5 Conclusion and Future Work In this paper we proposed a preliminary study of some types of revision of norms that can be enacted on a running multi-agent system in order to enhance its per- formance. We defined the concept of norms revision as a replacement of a set of norms with a new set and we formally defined the notions of relaxation and strengthening, which introduce supersets and subsets of the allowed behaviors, respectively. We provided some examples of revisions and some application sce- narios where a revision of the current set of enforced norms may be useful in order to achieve the overall system objectives. The analysis of different types of norms update provided in this paper is a first step in order to extend existing works for dynamic normative systems (see [14, 10]) with revision of norms. As future work we plan to provide a formal semantics for the concepts here introduced and to develop a syntactically sound and complete reasoning system. We also want to consider the effects of a revision in terms of update of the normative system and to analyse the relation of the revisions here presented with standard operators from the literature (e.g. from AGM framework). Finally we plan to provide a formal discussion about the correlation between the enforced norms and the fulfilment of the overall system objectives and to develop techniques to automatically reason at runtime about this correlation and to automatically suggest and perform a revision. These developments are meant to be part of an adaptive runtime supervision framework that continuously monitors the execution of a normative multi-agent system and intervenes on it in order to improve its performance. References [1] Sommerville, I., Cliff, D., Calinescu, R., Keen, J., Kelly, T., Kwiatkowska, M.Z., McDermid, J.A., Paige, R.F.: Large-scale complex IT systems. Commun. ACM 55(7) (2012) 71 -- 77 [2] Rosemann, M., Recker, J., Flender, C.: Contextualisation of business processes. IJBPIM 3(1) (2008) 47 -- 60 [3] Whittle, J., Sawyer, P., Bencomo, N., Cheng, B.H.C., Bruel, J.: RELAX: a lan- guage to address uncertainty in self-adaptive systems requirement. Requir. Eng. 15(2) (2010) 177 -- 196 [4] Boella, G., van der Torre, L.W.N.: Regulative and constitutive norms in normative multiagent systems. In: Principles of Knowledge Representation and Reasoning: Proceedings of the Ninth International Conference (KR2004), Whistler, Canada, June 2-5, 2004. (2004) 255 -- 266 [5] Bulling, N., Dastani, M.: Norm-based mechanism design. Artif. Intell. 239 (2016) 97 -- 142 [6] Zowghi, D., Gervasi, V.: On the interplay between consistency, completeness, and correctness in requirements evolution. Information & Software Technology 45(14) (2003) 993 -- 1009 [7] Ernst, N.A., Borgida, A., Jureta, I., Mylopoulos, J.: An overview of requirements evolution. In: Evolving Software Systems. (2014) 3 -- 32 [8] Ali, R., Dalpiaz, F., Giorgini, P., Souza, V.E.S.: Requirements evolution: From assumptions to reality. In: Enterprise, Business-Process and Information Systems Modeling - 12th International Conference, BPMDS 2011, and 16th International Conference, EMMSAD 2011, held at CAiSE 2011, London, UK, June 20-21, 2011. Proceedings. (2011) 372 -- 382 [9] Letier, E., van Lamsweerde, A.: Reasoning about partial goal satisfaction for requirements and design engineering. In: Proceedings of the 12th ACM SIGSOFT International Symposium on Foundations of Software Engineering, 2004, Newport Beach, CA, USA, October 31 - November 6, 2004. (2004) 53 -- 62 [10] Knobbout, M., Dastani, M., Meyer, J.C.: A dynamic logic of norm change. In: ECAI 2016 - 22nd European Conference on Artificial Intelligence, 29 August-2 September 2016, The Hague, The Netherlands - Including Prestigious Applica- tions of Artificial Intelligence (PAIS 2016). (2016) 886 -- 894 [11] Dastani, M., Grossi, D., Meyer, J.C., Tinnemeier, N.A.M.: Normative multi- In: Normative Multi-Agent Systems, 15.03. - agent programs and their logics. 20.03.2009. (2009) [12] Knobbout, M., Dastani, M.: Reasoning under compliance assumptions in norma- tive multiagent systems. In: International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2012, Valencia, Spain, June 4-8, 2012 (3 Volumes). (2012) 331 -- 340 [13] Alechina, N., Dastani, M., Logan, B.: Reasoning about normative update. In: IJCAI 2013, Proceedings of the 23rd International Joint Conference on Artificial Intelligence, Beijing, China, August 3-9, 2013. (2013) 20 -- 26 [14] Knobbout, M., Dastani, M., Meyer, J.C.: Reasoning about dynamic normative systems. In: Logics in Artificial Intelligence - 14th European Conference, JELIA 2014, Funchal, Madeira, Portugal, September 24-26, 2014. Proceedings. (2014) 628 -- 636 [15] Aucher, G., Grossi, D., Herzig, A., Lorini, E.: Dynamic context logic. In: Logic, Rationality, and Interaction, Second International Workshop, LORI 2009, Chongqing, China, October 8-11, 2009. Proceedings. (2009) 15 -- 26 [16] Governatori, G., Rotolo, A.: Changing legal systems: legal abrogations and an- nulments in defeasible logic. Logic Journal of the IGPL 18(1) (2010) 157 -- 194 [17] Alechina, N., Dastani, M., Logan, B.: Norm approximation for imperfect monitors. In: International conference on Autonomous Agents and Multi-Agent Systems, AAMAS '14, Paris, France, May 5-9, 2014. (2014) 117 -- 124 [18] Tinnemeier, N.A.M., Dastani, M., Meyer, J.C., van der Torre, L.W.N.: Pro- gramming normative artifacts with declarative obligations and prohibitions. In: Proceedings of the 2009 IEEE/WIC/ACM International Conference on Intelligent Agent Technology, IAT 2009, Milan, Italy, 15-18 September 2009. (2009) 145 -- 152 [19] Bulling, N., Dastani, M., Knobbout, M.: Monitoring norm violations in multi- agent systems. In: International conference on Autonomous Agents and Multi- Agent Systems, AAMAS '13, Saint Paul, MN, USA, May 6-10, 2013. (2013) 491 -- 498
1404.2367	2	1404	2015-02-15T19:41:45	Detecting Possible Manipulators in Elections	[ "cs.MA", "cs.GT" ]	Manipulation is a problem of fundamental importance in the context of voting in which the voters exercise their votes strategically instead of voting honestly to prevent selection of an alternative that is less preferred. The Gibbard-Satterthwaite theorem shows that there is no strategy-proof voting rule that simultaneously satisfies certain combinations of desirable properties. Researchers have attempted to get around the impossibility results in several ways such as domain restriction and computational hardness of manipulation. However these approaches have been shown to have limitations. Since prevention of manipulation seems to be elusive, an interesting research direction therefore is detection of manipulation. Motivated by this, we initiate the study of detection of possible manipulators in an election. We formulate two pertinent computational problems - Coalitional Possible Manipulators (CPM) and Coalitional Possible Manipulators given Winner (CPMW), where a suspect group of voters is provided as input to compute whether they can be a potential coalition of possible manipulators. In the absence of any suspect group, we formulate two more computational problems namely Coalitional Possible Manipulators Search (CPMS), and Coalitional Possible Manipulators Search given Winner (CPMSW). We provide polynomial time algorithms for these problems, for several popular voting rules. For a few other voting rules, we show that these problems are in NP-complete. We observe that detecting manipulation maybe easy even when manipulation is hard, as seen for example, in the case of the Borda voting rule.	cs.MA	cs	Detecting Possible Manipulators in Elections Palash Dey, Neeldhara Misra, and Y. Narahari [email protected], [email protected], [email protected] Department of Computer Science and Automation Indian Institute of Science - Bangalore, India. Date: September 25, 2018 Abstract. Manipulation is a problem of fundamental importance in the context of voting in which the voters exercise their votes strategically in- stead of voting honestly to prevent selection of an alternative that is less preferred. The Gibbard-Satterthwaite theorem shows that there is no strategy-proof voting rule that simultaneously satisfies certain combina- tions of desirable properties. Researchers have attempted to get around the impossibility results in several ways such as domain restriction and computational hardness of manipulation. However these approaches have been shown to have limitations. Since prevention of manipulation seems to be elusive, an interesting research direction therefore is detection of manipulation. Motivated by this, we initiate the study of detection of possible manipulators in an election. We formulate two pertinent computational problems - Coalitional Pos- sible Manipulators (CPM) and Coalitional Possible Manipulators given Winner (CPMW), where a suspect group of voters is provided as input to compute whether they can be a potential coalition of possible manipu- lators. In the absence of any suspect group, we formulate two more com- putational problems namely Coalitional Possible Manipulators Search (CPMS), and Coalitional Possible Manipulators Search given Winner (CPMSW). We provide polynomial time algorithms for these problems, for several popular voting rules. For a few other voting rules, we show that these problems are in NP-complete. We observe that detecting manipu- lation maybe easy even when manipulation is hard, as seen for example, in the case of the Borda voting rule. Keywords: Computational social choice, voting, manipulation, detection 5 1 0 2 b e F 5 1 ] A M . s c [ 2 v 7 6 3 2 . 4 0 4 1 : v i X r a 1. INTRODUCTION 1 Introduction On many occasions, agents need to agree upon a common decision although they have different preferences over the available alternatives. A natural approach used in these situations is voting. Some classic examples of the use of voting rules in the context of multiagent systems are in collaborative filtering [Pennock et al., 2000], rank aggregation in web [Dwork et al., 2001] etc. In a typical voting scenario, we have a set of m candidates and a set of n vot- ers reporting their rankings of the candidates called their preferences or votes. A voting rule selects one candidate as the winner once all the voters provide their votes. A set of votes over a set of candidates along with a voting rule is called an election. A basic problem with voting rules is that the voters may vote strategically instead of voting honestly, leading to the selection of a candidate which is not the actual winner. We call a candidate actual winner if, it wins the election when every voter votes truthfully. This phenomenon of strategic vot- ing is called manipulation in the context of voting. The Gibbard-Satterthwaite (G-S) theorem [Gibbard, 1973, Satterthwaite, 1975] says that manipulation is unavoidable for any unanimous and non-dictatorial voting rule if we have at least three candidates. A voting rule is called unanimous if whenever any can- didate is most preferred by all the voters, such a candidate is the winner. A voting rule is called non-dictatorial if there does not exist any voter whose most preferred candidate is always the winner irrespective of the votes of other vot- ers. The problem of manipulation is particularly relevant for multiagent systems since agents have computational power to determine strategic votes. There have been several attempts to bypass the impossibility result of the G-S theorem. Economists have proposed domain restriction as a way out of the impossibil- ity implications of the G-S theorem. The G-S theorem assumes all possible pref- erence profiles as the domain of voting rules. In a restricted domain, it has been shown that we can have voting rules that are not vulnerable to manipulation. A prominent restricted domain is the domain of single peaked preferences, in which the median voting rule provides a satisfactory solution [Mas-Collel et al., 1995]. To know more about other domain restrictions, we refer to [Gaertner, 2001, Mas-Collel et al., 1995]. This approach of restricting the domain, however, suffers from the requirement that the social planner needs to know the domain of preference profiles of the voters, which is often impractical. 1.1 Related Work Researchers in computational social choice theory have proposed invoking com- putational intractability of manipulation as a possible work around for the G-S theorem. Bartholdi et al. [Bartholdi and Orlin, 1991, Bartholdi et al., 1989] first proposed the idea of using computational hardness as a barrier against manip- ulation. Bartholdi et al. defined and studied the computational problem called manipulation where a set of manipulators have to compute their votes that make their preferred candidate win the election. The manipulators know the votes of the truthful voters and the voting rule that will be used to compute the winner. 2 1. INTRODUCTION Following this, a large body of research [Conitzer et al., 2007, Davies et al., 2011, Dey and Narahari, 2014, Elkind and Erd´elyi, 2012, Elkind and Lipmaa, 2005, Faliszewski et al., 2010, 2013, 2014, Gaspers et al., 2013, Narodytska and Walsh, 2013, Narodytska et al., 2011, Obraztsova and Elkind, 2012, Obraztsova et al., 2011, Xia et al., 2010, 2009, Zuckerman et al., 2011] shows that the manipula- tion problem is in NP-complete (NPC) for many voting rules. However, Procaccia et al. [Procaccia and Rosenschein, 2006, 2007] showed average case easiness of manipulation assuming junta distribution over the voting profiles. Friedgut et al. [Friedgut et al., 2008] showed that any neutral voting rule which is suffi- ciently far from being dictatorial is manipulable with non-negligible probabil- ity at any uniformly random preference profile by a uniformly random prefer- ence. The above result holds for elections with three candidates only. A voting rule is called neutral if the names of the candidates are immaterial. Isaksson et al. [Isaksson et al., 2012] generalize the above result to any number of can- didates. Walsh [Walsh, 2010] empirically shows ease of manipulating an STV (single transferable vote) election -- one of the very few voting rules where ma- nipulation even by one voter is in NPC [Bartholdi and Orlin, 1991]. In addition to the results mentioned above, there exist many other results in the litera- ture that emphasize the weakness of considering computational complexity as a barrier [Conitzer and Sandholm, 2006, Faliszewski and Procaccia, 2010, Walsh, 2011, Xia and Conitzer, 2008b,c]. Hence, the barrier of computational hardness is ineffective against manipulation in many settings. 1.2 Motivation In a situation where multiple attempts for prevention of manipulation fail to provide a fully satisfactory solution, detection of manipulation is a natural next step of research. There have been scenarios where establishing the occurrence of manipulation is straightforward, by observation or hindsight. For example, in sport, there have been occasions where the very structure of the rules of the game have encouraged teams to deliberately lose their matches. Observing such occurrences in, for example, football (the 1982 FIFA World Cup football match played between West Germany and Austria) and badminton (the quarter-final match between South Korea and China in the London 2012 Olympics), the relevant authorities have subsequently either changed the rules of the game (as with football) or disqualified the teams in question (as with the badminiton example). The importance of detecting manipulation lies in the potential for implementing corrective measures in the future. For reasons that will be evident soon, it is not easy to formally define the notion of manipulation detection. Assume that we have the votes from an election that has already happened. A voter is potentially a manipulator if there exists a preference ≻, different from the voter's reported preference, which is such that the voter had an "incentive to deviate" from the former. Specifically, suppose the candidate who wins with respect to this voter's reported preference is preferred (in ≻) over the candidate who wins with respect to ≻. In such a situation, ≻ could potentially be the voter's truthful preference, and the voter could be refraining from being truthful because 3 1. INTRODUCTION an untruthful vote leads to a more favorable outcome with respect to ≻. Note that we do not (and indeed, cannot) conclusively suggest that a particular voter has manipulated an election. This is because the said voter can always claim that she voted truthfully; since her actual preference is only known to her, there is no way to prove or disprove such a claim. Therefore, we are inevitably limited to asking only whether or not a voter has possibly manipulated an election. Despite this technical caveat, it is clear that efficient detection of manipula- tion, even if it is only possible manipulation, is potentially of interest in practice. We believe that, the information whether a certain group of voters have possi- bly manipulated an election or not would be very useful to social planners. For example, the organizers of an event, say London 2012 Olympics, maybe very interested to have this information. Also, in settings where data from many past elections (roughly over a fixed set of voters) is readily available, it is con- ceivable that possible manipulation could serve as suggestive evidence of real manipulation. Aggregate data about possible manipulations, although formally inconclusive, could serve as an important evidence of real manipulation. We re- mark that having a rich history is typically not a problem, particularly for AI related applications, since the data generated from an election is normally kept for future requirements (for instance, for data mining or learning). For exam- ple, several past affirmatives for possible manipulation is one possible way of formalizing the notion of erratic past behavior. Also, applications where benefit of doubt maybe important, for example, elections in judiciary systems, possi- ble manipulation detection seems useful. Thus the computational problem of detecting possible manipulation is of definite interest in this setting. 1.3 Contributions The novelty of this paper is in initiating research on detection of possible ma- nipulators in elections. We formulate four pertinent computational problems in this context: -- CPM: In the coalitional possible manipulators problem, we are interested in whether or not a given subset of voters is a possible coalition of manipulators [Definition 4]. -- CPMW: The coalitional possible manipulators given winner is the CPM problem with the additional information about who the winner would have been if the possible manipulators had all voted truthfully [Definition 2]. -- CPMS, CPMSW: In CPMS (Coalitional Possible Manipulators Search), we want to know, whether there exists any coalition of possible manipulators of a size at most k [Definition 6]. Similarly, we define CPMSW (Coalitional Possible Manipulators Search given Winner ) [Definition 5]. Our specific findings are as follows. -- We show that all the four problems above, for scoring rules and the max- imin voting rule, are in P when the coalition size is one [Theorem 1 and Theorem 4]. 4 2. PRELIMINARIES -- We show that all the four problems, for any coalition size, are in P for a wide class of scoring rules which include the Borda voting rule [Theorem 2, Theorem 3 and Corollary 1]. -- We show that, for the Bucklin voting rule [Theorem 6], both the CPM and CPMW problems are in P. The CPMS and CPMSW problems for the Buck- lin voting rule are also in P, when we have maximum possible coalition size k = O(1). -- We show that both the CPM and the CPMW problems are in NPC for the STV voting rule [Theorem 7 and Corollary 2]. We also prove that the CPMW problem is in NPC for maximin voting rule [Theorem 5]. We observe that all the four problems are computationally easy for many voting rules that we study in this paper. This can be taken as a positive result. The results for the CPM and the CPMW problems are summarized in Table 1. Voting Rule CPM,c = 1 CPM CPMW,c = 1 CPMW Scoring Rules Borda k-approval Maximin Bucklin STV P P P P P ? P P ? P P P P P P ? P P NPC P NPC NPC NPC NPC Table 1. Results for CPM and CPMW (c denotes coalition size). The '?' mark means that the problem is open. This paper is a significant extension of the conference version of this work Dey et al. [2015]: this extended version includes all the proofs. Organization The rest of the paper is organized as follows. We describe the necessary preliminaries in Section 2; we formally define the computational prob- lems in Section 3; we present the results in Section 4 and finally we conclude in Section 5. 2 Preliminaries Let V = {v1, . . . , vn} be the set of all voters and C = {c1, . . . , cm} the set of all candidates. Each voter vi's vote is a preference ≻i over the candidates which is a linear order over C. For example, for two candidates a and b, a ≻i b means that the voter vi prefers a to b. We will use a >i b to denote the fact that 5 3. PROBLEM FORMULATION a ≻i b, a 6= b. We denote the set of all linear orders over C by L(C). Hence, L(C)n denotes the set of all n-voters' preference profile (≻1, · · · , ≻n). We denote the (n − 1)-voters' preference profile by (≻1, · · · , ≻i−1, ≻i+1, · · · , ≻n) by ≻−i. We denote the set {1, 2, 3, . . . } by N+. The power set of C is denoted by 2C, and ∅ denotes the empty set. A map rc : ∪n,C∈N+L(C)n −→ 2C \ {∅} is called a voting correspondence. A map t : ∪C∈N+2C \ {∅} −→ C is called a tie breaking rule. Commonly used tie breaking rules are lexicographic tie breaking rules where ties are broken according to a predetermined preference ≻t∈ L(C). A voting rule is r = t ◦ rc, where ◦ denotes composition of mappings. Given an election E, we can construct a weighted graph GE called weighted majority graph from E. The set of vertices in GE is the set of candidates in E. For any two candidates x and y, the weight on the edge (x, y) is DE(x, y) = NE(x, y) − NE(y, x), where NE(x, y)(respectively NE(y, x)) is the number of voters who prefer x to y (re- spectively y to x). Some examples of common voting correspondences are as follows. -- Positional scoring rules: Given an m-dimensional vector α = (α1, α2, . . . , αm) ∈ Rm with α1 ≥ α2 ≥ · · · ≥ αm and α1 > αm, we can naturally define a voting rule - a candidate gets score αi from a vote if it is placed at the ith position, and the score of a candidate is the sum of the scores it receives from all the votes. The winners are the candidates with maximum score. A scoring rule is called a strict scoring rule if α1 > α2 > · · · > αm. For α = (m − 1, m − 2, . . . , 1, 0), we get the Borda voting rule. With αi = 1 ∀i ≤ k and 0 else, the voting rule we get is known as the k-approval voting rule. The plurality voting rule is the 1-approval voting rule and the veto voting rule is the (m − 1)-approval voting rule. -- Maximin: The maximin score of a candidate x is miny6=xD(x, y). The win- ners are the candidates with maximum maximin score. -- Bucklin: A candidate x's Bucklin score is the minimum number l such that at least half of the voters rank x in their top l positions. The winners are the candidates with lowest Bucklin score. This voting rule is also sometimes referred as the simplified Bucklin voting rule. -- Single Transferable Vote: In Single Transferable Vote (STV), a candidate with least plurality score is dropped out of the election and its votes are transferred to the next preferred candidate. If two or more candidates receive least plurality score, then some predetermined tie breaking rule is used. The candidate that remains after (m − 1) rounds is the winner. 3 Problem Formulation Consider an election that has already happened in which all the votes are known and thus the winner x ∈ C is also known. We call the candidate x the current winner of the election. The authority may suspect that the voters belonging to M ⊂ V have formed a coalition among themselves and manipulated the election by voting non-truthfully. The authority believes that other voters who do not belong to M , have voted truthfully. We denote the coalition size M by c. 6 3. PROBLEM FORMULATION Suppose the authority has auxiliary information, maybe from some other sources, which says that the actual winner should have been some candidate y ∈ C other than x. We call a candidate actual winner if it wins the election where all the voters vote truthfully. This means that the authority thinks that, had the voters in M voted truthfully, the candidate y would have been the winner. We remark that there are practical situations, for example, 1982 FIFA World cup or 2012 London Olympics, where the authority knows the actual winner. This situation is formalized below. Definition 1. Let r be a voting rule, and (≻i)i∈V be a voting profile of a set V of n voters. Let x be the winning candidate with respect to r for this profile. For a candidate y 6= x, M ⊂ V is said to be a coalition of possible manipulators j)j∈M ∈ L(C)M against y with respect to r if there exists a M -voters' profile (≻′ such that x ≻′ j y, ∀j ∈ M , and further, r((≻j )j∈V\M , (≻′ i)i∈M ) = y. Using the notion of coalition of possible manipulators, we formulate a computational problem called Coalitional Possible Manipulators given Winner (CPMW) as follows. Definition 2. (CPMW Problem) Given a voting rule r, a preference profile (≻i)i∈V of a set of voters V over a set of candidates C, a subset of voters M ⊂ V, and a candidate y, determine if M is a coalition of possible manipulators against y with respect to r. In the CPMW problem, the actual winner is given in the input. However, it may very well happen that the authority does not have any other information to guess the actual winner - the candidate who would have won the election had the voters in M voted truthfully. In this situation, the authority is interested in knowing whether there is a M -voter profile which along with the votes in V \ M makes some candidate y ∈ C the winner who is different from the current winner x ∈ C and all the preferences in the M -voters' profile prefer x to y. If such a M -voter profile exists for the subset of voters M , then we call M a coalition of possible manipulators and the corresponding computational problem is called Coalitional Possible Manipulators (CPM). These notions are formalized below. Definition 3. Let r be a voting rule, and (≻i)i∈V be a voting profile of a set V of n voters. A subset of voters M ⊂ V is called a coalition of possible manipula- tors with respect to r if M is a coalition of possible manipulators against some candidate y with respect to r. Definition 4. (CPM Problem) Given a voting rule r, a preference profile (≻i)i∈V of a set of voters V over a set of candidates C, and a subset of voters M ⊂ V, determine if M is a coalition of possible manipulators with respect to r. In both the CPMW and CPM problems, a subset of voters which the au- thority suspect to be a coalition of manipulators, is given in the input. However, there can be situations where there is no specific subset of voters to suspect. In 7 3. PROBLEM FORMULATION those scenarios, it may still be useful to know, what are the possible coalition of manipulators of size less than some number k. Towards that end, we extend the CPMW and CPM problems to search for a coalition of potential possible ma- nipulators and call them Coalitional Possible Manipulators Search given Winner (CPMSW) and Coalitional Possible Manipulators Search (CPMS) respective. Definition 5. (CPMSW Problem) Given a voting rule r, a preference profile (≻i)i∈V of a set of voters V over a set of candidates C, a candidate y, and an integer k, determine whether there exists any M ⊂ V with M ≤ k such that M is a coalition of possible manipulators against y. Definition 6. (CPMS Problem) Given a voting rule r, a preference profile (≻i)i∈V of a set of voters V over a set of candidates C, and an integer k, determine whether there exists any M ⊂ V with M ≤ k such that M is a coalition of possible manipulators. 3.1 Discussion The CPMW problem may look very similar to the manipulation prob- lem [Bartholdi et al., 1989, Conitzer et al., 2007]- in both the problems a set of voters try to make a candidate winner. However, in the CPMW problem, the actual winner must be less preferred to the current winner. Although it may look like a subtle difference, it changes the nature and complexity theoretic behavior of the problem completely. For example, we show that all the four problems have an efficient algorithm for a large class of voting rules that includes the Borda voting rule for which the manipulation problem is in NPC, even when we have at least two manipulators [Betzler et al., 2011, Davies et al., 2011]. Another im- portant difference is that the manipulation problem, in contrast to the problems studied in this paper, does not take care of manipulators' preferences. We believe that there does not exist any formal reduction between the CPMW problem and the manipulation problem. On the other hand, the CPMS problem is similar to the margin of victory problem defined by Xia [Xia, 2012], where also we are looking for changing the current winner by changing at most some k number of votes, which in turn identical to the destructive bribery problem [Faliszewski et al., 2009]. Whereas, in the CPMS problem, the vote changes can occur in a restricted fashion. Here also, the margin of victory problem has the hereditary property which the CPMS problem does not possess. These two problems do not seem to have any obvious complexity theoretic implications. Now, we explore the connections among the four problems that we study here. Notice that, a polynomial time algorithm for the CPM and the CPMW problems gives us a polynomial time algorithm for the CPMS and the CPMSW problems for any maximum possible coalition size k = O(1). Also, observe that, a polynomial time algorithm for the CPMW (respectively CPMSW) problem implies a polynomial time algorithm for the CPM (respectively CPMS) problem. Hence, we have the following propositions. 8 4. RESULTS Proposition 1. For every voting rule, if the maximum possible coalition size k = O(1), then, CP M W ∈ P ⇒ CP M, CP M SW, CP M S ∈ P Proposition 2. For every voting rule, CP M SW ∈ P ⇒ CP M S ∈ P 4 Results In this section, we present our algorithmic results for the CPMW, CPM, CPMSW, and CPMS problems for various voting rules. 4.1 Scoring Rules Below we have certain lemmas which form a crucial ingredient of our algorithms. To begin with, we define the notion of a manipulated preference. Let r be a scoring rule and ≻:= (≻i, ≻−i) be a voting profile of n voters. Let ≻′ i be a preference such that r(≻) >′ i r(≻′ i, ≻−i). Then, we say that ≻′ omit the reference to r if it is clear from the context. i is a (≻, i)-manipulated preference with respect to r. We Lemma 1. Let r be a scoring rule and ≻:= (≻i, ≻−i) be a voting profile of n voters. Let a and b be two candidates such that score≻−i (a) > score≻−i (b), and let ≻′ i be (≻, i)-manipulated preference where a precedes b: ≻′ i:= · · · > a > · · · > b > · · · If a, b are not winners with respect to either (≻′ ≻′′ i obtained from ≻′ i by interchanging a and b is also (≻, i)-manipulated. i, ≻−i) or ≻, then the preference Proof. Let x := r(≻′ proposed profile (≻′′ b with respect to x. First, consider the score of b in the new profile: i, ≻−i). If suffices to show that x continues to win in the i , ≻−i). To this end, it is enough to argue the scores of a and score(≻′′ i ,≻−i)(b) = score≻′′ < score≻′ = score(≻′ ≤ score(≻′ = score(≻′′ i (b) + score≻−i(b) i(a) + score≻−i (a) i,≻−i)(a) i,≻−i)(x) i ,≻−i)(x) (1) i (a) and score≻−i (b) < The second line uses the fact that score≻′′ score≻−i (a). The fourth line comes from the fact that x is the winner and the last line follows from the fact that the position of x is same in both profiles. i (b) = score≻′ 9 4. RESULTS Similarly, we have the following argument for the score of a in the new profile (the second line below simply follows from the definition of scoring rules). score(≻′′ i ,≻−i)(a) = score≻′′ ≤ score≻′ = score(≻′ ≤ score(≻′ = score(≻′′ i (a) + score≻−i (a) i(a) + score≻−i (a) i,≻−i)(a) i,≻−i)(x) i ,≻−i)(x) (2) Since the tie breaking rule is according to some predefined fixed order ≻t∈ L(C) and the candidates tied with winner in (≻′′ i , ≻−i) also tied with winner in (≻′ , ≻−i), we have the following, i r(≻) >′′ i r(≻′′ i , ≻−i) ⊓⊔ We now show that, if there is some (≻, i)-manipulated preference with respect to a scoring rule r, then there exists a (≻, i)-manipulated preference with a specific structure. Lemma 2. Let r be a scoring rule and ≻:= (≻i, ≻−i) be a voting profile of n voters. If there is some (≻, i)-manipulated preference with respect to r, then there also exists a (≻, i)-manipulated preference ≻′ i where the actual winner y immediately follows the current winner x: ≻′ i:= · · · > x > y > · · · and the remaining candidates are in nondecreasing ordered of their scores from ≻−i. Proof. Let ≻′′ be a (≻, i)-manipulated preference with respect to r. Let x := r(≻ ), y := r(≻′′, ≻−i). From Lemma 1, without loss of generality, we may assume that, all candidates except x, y are in nondecreasing order of score≻−i (.) in the preference ≻′′. If ≻′′ i:= · · · ≻ x ≻ y ≻ · · · ≻ · · · from ≻′′ i where y is moved to the position following x and the position of the candidates in between x and y in ≻′′ is deteriorated by one i position each. The position of the rest of the candidates remain same in both ≻′′ i := · · · ≻ x ≻ · · · ≻ y ≻ · · · ≻ · · · , we define ≻′ i. Now we have following, i and ≻′ score(≻′ i,≻−i)(y) = score≻′ ≥ score≻′′ = score(≻′′ i (y) + score≻−i (y) i (y) + score≻−i (y) i ,≻−i)(y) We also have, score(≻′ i,≻−i)(a) ≤ score(≻′′ i ,≻−i)(a), ∀a ∈ C \ {y} (1) (2) 10 4. RESULTS Since the tie breaking rule is according to some predefined order ≻t∈ L(C), we have the following, r(≻) >′ i r(≻′, ≻−i) ⊓⊔ Using Lemmas 1 and 2, we now present the results for the scoring rules. Theorem 1. When c=1, the CPMW, CPM, CPMSW, and CPMS problems for scoring rules are in P. Proof. From Proposition 1, it is enough to give a polynomial time algorithm for the CPMW problem. So consider the CPMW problem. We are given the actual winner y and we compute the current winner x with respect to r. Let ≻[j] be a preference where x and y are in positions j and (j + 1) respectively, and the rest of the candidates are in nondecreasing order of the score that they receive from ≻−i. For j ∈ {1, 2, . . . , m − 1}, we check if y wins with the profile (≻−i, ≻[j]). If we are successful with at least one j we report YES, otherwise we say NO. The correctness follows from Lemma 2. Thus we have a polynomial time algorithm ⊓⊔ for CPMW when c = 1 and the theorem follows from Proposition 1. Now, we study the CPMW and the CPM problems when c > 1. If m = O(1), then both the CPMW and the CPM problems for any anonymous and efficient voting rule r can be solved in polynomial time by iterating over all possible (cid:0)m!+c−1 (cid:1) ways the manipulators can have actual preferences. A voting rule is called efficient if winner determination under it is in P. m! Theorem 2. For scoring rules with α1 − α2 ≤ αi − αi+1, ∀i, the CPMW and the CPM problems are in P. Proof. We provide a polynomial time algorithm for the CPMW problem in this setting. Let x be the current winner and y be the given actual winner. Let M be the given subset of voters. Let ((≻i)i∈M , (≻j)j∈V \M ) be the reported preference profile. Without loss of generality, we assume that x is the most preferred candidate in every ≻i, i ∈ M . Let us define ≻′ i, i ∈ M, by moving y to the second position in the preference ≻i. In the profile ((≻′ i)i∈M , (≻j)j∈V \M ), the winner is either x or y since only y's score has increased. We claim that M is a coalition of possible manipulators with respect to y if and only if y is the winner in preference profile ((≻′ i)i∈M , (≻j)j∈V \M ). This can be seen as follows. Suppose there exist preferences ≻′′ i , with x ≻′′ i y, i ∈ M, for which y wins in the profile ((≻′′ i )i∈M , (≻j)j∈V \M ). Now without loss of generality, we can assume that y immediately follows x in all ≻′′ i , i ∈ M, and α1 − α2 ≤ αi − αi+1, ∀i implies that we can also assume that x and y are in the first and second positions respectively in all ≻′′ i)i∈M , (≻j)j∈V \M ) and ((≻′′ i )i∈M , (≻j)j∈V \M ), the score of x and y are same. But in the first profile ⊓⊔ x wins and in the second profile y wins, which is a contradiction. i , i ∈ M . Now in both the profiles, ((≻′ Theorem 3. For scoring rules with α1 − α2 ≤ αi − αi+1, ∀i, the CPMSW and the CPMS problems are in P. 11 4. RESULTS Proof. From Proposition 2, it is enough to prove that CP M SW ∈ P. Let x be the current winner, y be the given actual winner and s(x) and s(y) be their current respective scores. For each vote v ∈ V, we compute a number ∆(v) = α2 − αj − α1 + αi, where x and y are receiving scores αi and αj respectively from the vote v. Now, we output yes iff there are k votes vi, 1 ≤ i ≤ k such that, Pk i=1 ∆(vi) ≥ s(x) − s(y), which can be checked easily by sorting the ∆(v)'s in nonincreasing order and checking the condition for the first k ∆(v)'s, where k is the maximum possible coalition size specified in the input. The proof of correctness follows by exactly in the same line of argument as the proof of ⊓⊔ Theorem 2. For the plurality voting rule, we can solve all the problems easily using max flow. Hence, from Theorem 2 and Theorem 3, we have the following result. Corollary 1. The CPMW, CPM, CPMSW, and CPMS problems for the Borda and k-approval voting rules are in P. 4.2 Maximin Voting Rule For the maximin voting rule, we show that all the four problems are in P, when we have a coalition size of one to check for. Theorem 4. The CPMW, CPM, CPMSW, and CPMS problems for maximin voting rule are in P for coalition size c = 1 or maximum possible coalition size k = 1. Proof. Given a n-voters' profile ≻∈ L(C)n and a voter vi, let the current winner be x := r(≻) and the given actual winner be y. We will construct ≻′= (≻′ i, ≻−i), if it exists, such that r(≻) >′ i r(≻′) = y, thus deciding whether vi is a possible manipulator or not. Now, the maximin score of x and y in the profile ≻′ can take one of values from the set {score≻−i (x) ± 1} and {score≻−i (y) ± 1}. The algorithm is as follows. We first guess the maximin score of x and y in the profile ≻′. There are only four possible guesses. Suppose, we guessed that x's score will decrease by one and y's score will decrease by one assuming that this guess makes y win. Now notice that, without loss of generality, we can assume that y immediately follows x in the preference ≻′ i since y is the winner in the profile ≻′. This implies that there are only O(m) many possible positions for x and y in ≻′ i. Let B(x) and B(y) be the sets of candidates with whom x and respectively y performs worst. Now since, x's score will decrease and y's score will decrease, we have the following constraint on ≻′ i. There must be a candidate each from B(y) and B(x) that will precede x. We do not know a-priori if there is one candidate that will serve as a witness for both B(x) and B(y), or if there separate witnesses. In the latter situation, we also do not know what order they appear in. Therefore we guess if there is a common candidate, and if not, we guess the relative ordering of the distinct candidates from B(x) and B(y). Now we place any candidate at the top position of ≻′ i if this action does not make y lose the election. If there i. We guess the position of x and thus the position of y in ≻′ 12 4. RESULTS are many choices, we prioritize in favor of candidates from B(x) and B(y) -- in particular, we focus on the candidates common to B(x) and B(y) if we expect to have a common witness, otherwise, we favor a candidate from one of the sets according to the guess we start with. If still there are multiple choices, we pick arbitrarily. After that we move on to the next position, and do the same thing (except we stop prioritizing explicitly for B(x) and B(y) once we have at least one witness from each set). The other situations can be handled similarly with minor modifications. In this way, if it is able to get a complete preference, then it checks whether vi is a possible manipulator or not using this preference. If yes, then it returns YES. Otherwise, it tries other positions for x and y and other possible scores of x and y. After trying all possible guesses, if it cannot find the desired preference, then it outputs NO. Since there are only polynomial many possible guesses, this algorithm runs in a polynomial amount of time. The proof of correctness follows from the proof of Theorem 1 in [Bartholdi et al., 1989]. ⊓⊔ We now show that the CPMW problem for maximin voting rule is in NPC when we have c > 1. Towards that, we use the fact that the unweighted coalitional manipulation (UCM) problem for maximin voting rule is in NPC [Xia et al., 2009], when we have c > 1. The UCM problem is as follows. Definition 7. (UCM Problem) Given a voting rule r, a set of manipulators M ⊂ V, a profile of non- manipulators' vote (≻i)i∈V\M , and a candidate z ∈ C, we are asked whether j)j∈M such that r((≻i)i∈V\M , (≻′ there exists a profile of manipulators' votes (≻′ )j∈M ) = z. Assume that ties are broken in favor of z. j We define a restricted version of the UCM problem called R-UCM as follows. Definition 8. (R-UCM Problem) This problem is the same as the UCM problem with a given guarantee - let c := M . The candidate z loses pairwise election with every other candidate by 4c votes. For any two candidates a, b ∈ C, either a and b ties or one wins pairwise election against the other one by margin of either 2c + 2 or of 4c or of 8c. We denote the margin by which a candidate a defeats b, by d(a, b). The R-UCM problem for maximin voting rule is in NPC [Xia et al., 2009], when we have c > 1. We will need the following lemma to manipulate the pairwise difference scores in the reduction. The lemma has been used before [McGarvey, 1953, Xia and Conitzer, 2008a]. Lemma 3. For any function f : C × C −→ Z, such that 1. ∀a, b ∈ C, f (a, b) = −f (b, a). 2. ∀a, b ∈ C, f (a, b) is even, there exists a n voters' profile such that for all a, b ∈ C, a defeats b with a margin of f (a, b). Moreover, n = O   X {a,b}∈C×C f (a, b)  13 4. RESULTS Theorem 5. The CPMW problem for maximin voting rule is in NPC, for c > 1. Proof. Clearly the CPMW problem for maximin voting rule is in NP. We provide a many-one reduction from the R-UCM problem for the maximin voting rule to it. Given a R-UCM problem instance, we define a CPMW problem instance Γ = (C′, (≻′ i)i∈V ′ , M ′) as follows. C′ := C ∪ {w, d1, d2, d3} We define V ′ such that d(a, b) is the same as the R-UCM instance, for all a, b ∈ C and d(d1, w) = 2c+ 2, d(d1, d2) = 8c, d(d2, d3) = 8c, d(d3, d1) = 8c. The existence of such a V ′ is guaranteed from Lemma 3. Moreover, Lemma 3 also ensures that V ′ is O(mc). The votes of the voters in M is w ≻ . . . . Thus the current winner is w. The actual winner is defined to be z. The tie breaking rule is ≻t= w ≻ z ≻ . . . , where z is the candidate whom the manipulators in M want to make winner in the R-UCM problem instance. Clearly this reduction takes polynomial amount of time. Now we show that, M is a coalition of possible manipulators iff z can be made a winner. The if part is as follows. Let ≻i, i ∈ M be the votes that make z win. We can assume that z is the most preferred candidate in all the preferences ≻i, i ∈ M . Now consider the preferences for the voters in M is follows. ≻′ i:= d1 ≻ d2 ≻ d3 ≻ w ≻i, i ∈ M The score of every candidate in C is not more than z. The score of z is −3c. The score of w is −3c − 2 and the scores of d1, d2, and d3 are less than −3c. Hence, M is a coalition of possible manipulators with the actual preferences ≻′ i:= d1 ≻ d2 ≻ d3 ≻ w ≻i, i ∈ M . i, i ∈ M . Consider the preferences ≻′ The only if part is as follows. Suppose M is a coalition of possible manipula- tors with actual preferences ≻′ i, i ∈ M , but restricted to the set C only. Call them ≻i, i ∈ M . We claim that ≻i, i ∈ M with the votes from V makes z win the election. If not then, there exists a candidate, say a ∈ C, whose score is strictly more than the score of z - this is so because the tie breaking rule is in favor of z. But this contradicts the fact that z wins the election when the voters in M vote ≻′ i, i ∈ M along with the votes from V ′. ⊓⊔ 4.3 Bucklin Voting Rule In this subsection, we design polynomial time algorithms for both the CPMW and the CPM problem for the Bucklin voting rule. Again, we begin by showing that if there are profiles witnessing manipulation, then there exist profiles that do so with some additional structure, which will subsequently be exploited by our algorithm. Lemma 4. Consider a preference profile (≻i)i∈V , where x is the winner with respect to the Bucklin voting rule. Suppose a subset of voters M ⊂ V form a coalition of possible manipulators. Let y be the actual winner. Then there exist preferences (≻′ i)i∈M ), and further: i)i∈M such that y is a Bucklin winner in ((≻i)i∈V\M, (≻′ 14 4. RESULTS 1. y immediately follows x in each ≻′ i. 2. The rank of x in each ≻′ i is in one of the following - first, b(y) − 1, b(y), b(y) + 1, where b(y) be the Bucklin score of y in ((≻i)i∈V\M, (≻′ i)i∈M ). Proof. From Definition 3, y's rank must be worse than x's rank in each ≻′ i. We now exchange the position of y with the candidate which immediately follows x in ≻′ i. This process does not decrease Bucklin score of any candidate except possibly y's, and x's score does not increase. Hence y will continue to win and thus ≻′ i satisfies the first condition. Now to begin with, we assume that ≻′ i satisfies the first condition. If the position of x in ≻′ i is b(y) − 1 or b(y), we do not change it. If x is above b(y) − 1 in ≻′ i, then move x and y at the first and second positions respectively. Similarly if x is below b(y) + 1 in ≻′ i, then move x and y at the b(y) + 1 and b(y) + 2 positions respectively. This process does not decrease score of any candidate except y because the Bucklin score of x is at least b(y). The transformation cannot increase the score y since its position has only been improved. Hence y continues to win and thus ≻′ ⊓⊔ i satisfies the second condition. Lemma 4 leads us to the following theorem. Theorem 6. The CPMW problem and the CPM problems for Bucklin voting rule are in P. Therefore, by Proposition 1, the CPMSW and the CPMS problems are in P when the maximum coalition size k = O(1). Proof. Proposition 1 says that it is enough to prove that the CPMW problem is in P. Let x be the current winner and y be the given actual winner. For any final Bucklin score b(y) of y, there are polynomially many possibilities for the positions of x and y in the profile of ≻i, i ∈ M , since Bucklin voting rule is anonymous. Once the positions of x and y is fixed, we try to fill the top b(y) positions of each ≻′ i - place a candidate in an empty position above b(y) in any ≻′ i if doing so does not make y lose the election. If we are able to successfully fill the top b(y) positions of all ≻′ i for all i ∈ M , then M is a coalition of possible manipulators. If the above process fails for all possible above mentioned positions of x and y and all possible guesses of b(y), then M is not a coalition of possible manipulators. Clearly the above algorithm runs in poly(m,n) time. The proof of correctness is as follows. If the algorithm outputs that M is a coalition of possible manipulators, then it actually has constructed ≻′ i for all i ∈ M with respect to which they form a coalition of possible manipulators. On the other hand, if they form a coalition of possible manipulators, then Lemma 4 ensures that our algorithm explores all the sufficient positions of x and y in ≻′ i for all i ∈ M . Now if M is a possible coalition of manipulators, then the corre- sponding positions for x and y have also been searched. Our greedy algorithm must find it since permuting the candidates except x and z which are ranked above b(y) in ≻′ i cannot stop y to win the election since the Bucklin score of ⊓⊔ other candidates except y is at least b(y). 15 4. RESULTS 4.4 STV Voting Rule Next, we prove that the CPMW and the CPM problems for STV rule is in NPC. To this end, we reduce from the Exact Cover by 3-Sets Problem (X3C), which is known to be in NPC [Garey and Johnson, 1979]. The X3C problem is as follows. Definition 9. (X3C Problem) Given a set S of cardinality n and m subsets S1, S2, . . . , Sm ⊂ S with Si = 3, ∀i = 1, . . . , m, does there exist an index set I ⊆ {1, . . . , m} with I = S 3 such that ∪i∈I Si = S. Theorem 7. The CPM problem for STV rule is in NPC. Proof sketch: It is enough to show the theorem for the case c = 1. Clearly the problem is in NP. To show NP hardness, we show a many-one reduction from the X3C problem to it. The reduction is analogous to the reduction given in [Bartholdi and Orlin, 1991]. Hence, we give a proof sketch only. Given an X3C instance, we construct an election as follows. The unspecified positions can be filled in any arbitrary way. The candidate set is as follows. C = {x, y} ∪ {a1, . . . , am} ∪ {a1, . . . , am} ∪ {b1, . . . , bm} ∪ {b1, . . . , bm} ∪ {d0, . . . , dn} ∪ {g1, . . . , gm} The votes are as follows. 3 votes for d0 ≻ x ≻ y ≻ . . . -- 12m votes for y ≻ x ≻ . . . -- 12m − 1 votes for x ≻ y ≻ . . . -- 10m + 2n -- 12m − 2 votes for di ≻ x ≻ y ≻ . . . , ∀i ∈ [n] -- 12m votes for gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 6m + 4i − 5 votes for bi ≻ bi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 2 votes for bi ≻ dj ≻ x ≻ y ≻ . . . , ∀i ∈ [m], ∀j ∈ Si -- 6m + 4i − 1 votes for bi ≻ bi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 2 votes for bi ≻ d0 ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 6m + 4i − 3 votes for ai ≻ gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 1 vote for ai ≻ bi ≻ gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 2 votes for ai ≻ ai ≻ gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 6m + 4i − 3 votes for ai ≻ gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 1 vote for ai ≻ bi ≻ gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] -- 2 votes for ai ≻ ai ≻ gi ≻ x ≻ y ≻ . . . , ∀i ∈ [m] The tie breaking rule is ≻t= · · · ≻ x. The vote of v is x ≻ · · · . We claim that v is a possible manipulator iff the X3C is a yes instance. Notice that, of the first 3m candidates to be eliminated, 2m of them are a1, . . . , am and a1, . . . , am. Also exactly one of bi and bi will be eliminated among the first 3m candidates to be eliminated because if one of bi, bi then the other's score exceeds 12m. We show that the winner is either x or y irrespective of the vote of one more candidate. 16 5. CONCLUSION Let J := {j : bj is eliminated before bj}. If J is an index of set cover then the winner is y. This can be seen as follows. Consider the situation after the first 3m eliminations. Let i ∈ Sj for some j ∈ J. Then bj has been eliminated and thus the score of di is at least 12m. Since J is an index of a set cover, every di's score is at least 12m. Notice that bj has been eliminated for all j /∈ J. Thus the revised score of d0 is at least 12m. After the first 3m eliminations, the remaining candidates are x, y, {di : i ∈ [n]}, {gi : i ∈ [m]}, {bj : j /∈ J}, {bj : j ∈ J}. All the remaining candidates except x has score at least 12m and x's score is 12m − 1. Hence x will be eliminated next which makes y's score at least 24m − 1. Next di's will get eliminated which will in turn make y's score (12n + 36)m − 1. At this point gi's score is at most 32m. Also all the remaining bi and bi's score is at most 32m. Since each of the remaining candidate's scores gets transferred to y once they are eliminated, y is the winner. Now we show that, if J is not an index of set cover then the winner is x. This can be seen as follows. If J > n 3 , then the number of bj that gets eliminated in the first 3m iterations is less than m − n 3 . This makes the score of d0 at most 12m − 2. Hence d0 gets eliminated before x and all its scores gets transferred to x. This makes the elimination of x impossible before y and makes x the winner of the election. If J ≤ n 3 and there exists an i ∈ S that is not covered by the corresponding set cover, then di gets eliminated before x with a score of 12m − 2 and its score gets transferred to x. This makes x win the election. Hence y can win iff X3C is a yes instance. Also notice that if y can win the election, then it can do so with the voter v voting a preference like · · · ≻ x ≻ y ≻ · · · . (cid:3) From the proof of the above theorem, we have the following corollary by specifying y as the actual winner for the CPMW problem. Corollary 2. The CPMW problem for STV rule is in NPC. 5 Conclusion In this work, we have initiated a promising research direction for detecting ma- nipulation in elections. We have proposed the notion of possible manipulation and explored several concrete computational problems, which we believe to be important in the context of voting theory. These problems involve identifying if a given set of voters are possible manipulators (with or without a specified candidate winner). We have also studied the search versions of these problems, where the goal is to simply detect the presence of possible manipulation with the maximum coalition size. We believe there is theoretical as well as practical interest in studying the proposed problems. We have provided algorithms and hardness results for many common voting rules. In this work, we considered elections with unweighted voters only. An im- mediate future research direction is to study the complexity of these problems in weighted elections. Further, verifying the number of false manipulators that 17 6. ACKNOWLEDGEMENT this model catches in a real or synthetic data set, where, we already have some knowledge about the manipulators, would be interesting. It is our conviction that both the problems that we have studied here have initiated an interesting research direction with significant promise and potential for future work. 6 Acknowledgement We acknowledge the anonymous reviewers of AAMAS 2014 for providing con- structive comments to improve the paper. 18 Bibliography Bartholdi, J. and Orlin, J. (1991). Single transferable vote resists strategic vot- ing. Social Choice and Welfare (SCW), 8(4):341 -- 354. Bartholdi, J., Tovey, C., and Trick, M. (1989). The computational difficulty of manipulating an election. Social Choice and Welfare (SCW), 6(3):227 -- 241. Betzler, N., Niedermeier, R., and Woeginger, G. J. (2011). Unweighted coali- tional manipulation under the borda rule is NP-hard. In IJCAI, volume 11, pages 55 -- 60. Conitzer, V. and Sandholm, T. (2006). Nonexistence of voting rules that are usu- ally hard to manipulate. In International Conference on Artificial Intelligence (AAAI), volume 6, pages 627 -- 634. Conitzer, V., Sandholm, T., and Lang, J. (2007). When are elections with few candidates hard to manipulate? Journal of the ACM (JACM), 54(3):14. Davies, J., Katsirelos, G., Narodytska, N., and Walsh, T. (2011). Complexity of and algorithms for borda manipulation. In Proceedings of the International Conference on Artificial Intelligence (AAAI), pages 657 -- 662. Dey, P., Misra, N., and Narahari, Y. (2015). Detecting possible manipulators in elections. In Proceeding of the 14th International Conference on Autonomous Systems and Multiagent Systems (AAMAS-15). Dey, P. and Narahari, Y. (2014). Asymptotic collusion-proofness of voting rules: the case of large number of candidates. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 1419 -- 1420. International Foundation for Autonomous Agents and Multiagent Systems. Dwork, C., Kumar, R., Naor, M., and Sivakumar, D. (2001). Rank aggregation methods for the web. In Proceedings of the 10th International Conference on World Wide Web (WWW), pages 613 -- 622. ACM. Elkind, E. and Erd´elyi, G. (2012). Manipulation under voting rule uncertainty. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 627 -- 634. International Foundation for Autonomous Agents and Multiagent Systems. Elkind, E. and Lipmaa, H. (2005). Hybrid voting protocols and hardness of manipulation. In Algorithms and Computation, pages 206 -- 215. Springer. Faliszewski, P., Hemaspaandra, E., and Hemaspaandra, L. A. (2009). How hard is bribery in elections? Journal of Artificial Intelligence Research (JAIR), 35(2):485. Faliszewski, P., Hemaspaandra, E., and Hemaspaandra, L. A. (2010). Using complexity to protect elections. Communications of the ACM, 53(11):74 -- 82. Faliszewski, P., Hemaspaandra, E., and Hemaspaandra, L. A. (2013). Weighted electoral control. In Proceedings of the 12th International Conference on Au- tonomous Agents and Multiagent Systems (AAMAS), pages 367 -- 374. Interna- tional Foundation for Autonomous Agents and Multiagent Systems. BIBLIOGRAPHY BIBLIOGRAPHY Faliszewski, P. and Procaccia, A. D. (2010). Ai's war on manipulation: Are we winning? AI Magazine, 31(4):53 -- 64. Faliszewski, P., Reisch, Y., Rothe, J., and Schend, L. (2014). Complexity of ma- nipulation, bribery, and campaign management in bucklin and fallback voting. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 1357 -- 1358. International Founda- tion for Autonomous Agents and Multiagent Systems. Friedgut, E., Kalai, G., and Nisan, N. (2008). Elections can be manipulated often. In IEEE 49th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 243 -- 249. IEEE. Gaertner, W. (2001). Domain Conditions in Social Choice Theory. Cambridge University Press. Garey, M. R. and Johnson, D. S. (1979). Computers and Intractability, volume 174. freeman New York. Gaspers, S., Kalinowski, T., Narodytska, N., and Walsh, T. (2013). Coali- tional manipulation for schulze's rule. In Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 431 -- 438. International Foundation for Autonomous Agents and Multiagent Systems. Gibbard, A. (1973). Manipulation of voting schemes: a general result. Econo- metrica: Journal of the Econometric Society, pages 587 -- 601. Isaksson, M., Kindler, G., and Mossel, E. (2012). The geometry of manipulation - a quantitative proof of the gibbard-satterthwaite theorem. Combinatorica, 32(2):221 -- 250. Mas-Collel, A., Whinston, M. D., and Green, J. (1995). Microeconomic theory. McGarvey, D. C. (1953). A theorem on the construction of voting paradoxes. Econometrica: Journal of the Econometric Society, pages 608 -- 610. Narodytska, N. and Walsh, T. (2013). Manipulating two stage voting rules. In Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 423 -- 430. International Foundation for Autonomous Agents and Multiagent Systems. Narodytska, N., Walsh, T., and Xia, L. (2011). Manipulation of nanson's and baldwin's rules. In Proceedings of the International Conference on Artificial Intelligence (AAAI), pages 713 -- 718. Obraztsova, S. and Elkind, E. (2012). Optimal manipulation of voting rules. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 619 -- 626. International Foundation for Autonomous Agents and Multiagent Systems. Obraztsova, S., Elkind, E., and Hazon, N. (2011). Ties matter: Complexity of voting manipulation revisited. In The 10th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 71 -- 78. Inter- national Foundation for Autonomous Agents and Multiagent Systems. Pennock, D. M., Horvitz, E., and Giles, C. L. (2000). Social choice theory and recommender systems: Analysis of the axiomatic foundations of collaborative filtering. In Proceedings of the 17th International Conference on Artificial Intelligence (AAAI). 20 BIBLIOGRAPHY BIBLIOGRAPHY Procaccia, A. and Rosenschein, J. (2006). Junta distributions and the average- case complexity of manipulating elections. In Proceedings of the Fifth Interna- tional Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 497 -- 504. ACM. Procaccia, A. and Rosenschein, J. (2007). Average-case tractability of manipu- lation in voting via the fraction of manipulators. In Proceedings of the 6th In- ternational Conference on Autonomous Agents and Multiagent Systems (AA- MAS), volume 7. Satterthwaite, M. (1975). Strategy-proofness and arrow's conditions: Existence and correspondence theorems for voting procedures and social welfare func- tions. Journal of Economic Theory (JET), 10(2):187 -- 217. Walsh, T. (2010). An empirical study of the manipulability of single trans- ferable voting. In Proceedings of the 19th European Conference on Artificial Intelligence (ECAI), pages 257 -- 262. Walsh, T. (2011). Is computational complexity a barrier to manipulation? Annals of Mathematics and Artificial Intelligence, 62(1-2):7 -- 26. Xia, L. (2012). Computing the margin of victory for various voting rules. In Proceedings of the 13th ACM Conference on Electronic Commerce (EC), pages 982 -- 999. ACM. Xia, L. and Conitzer, V. (2008a). Determining possible and necessary winners under common voting rules given partial orders. In International Conference on Artificial Intelligence (AAAI), volume 8, pages 196 -- 201. Xia, L. and Conitzer, V. (2008b). Generalized scoring rules and the frequency of coalitional manipulability. In Proceedings of the 9th ACM conference on Electronic Commerce (EC), pages 109 -- 118. ACM. Xia, L. and Conitzer, V. (2008c). A sufficient condition for voting rules to In Proceedings of the 9th ACM conference on be frequently manipulable. Electronic Commerce (EC), pages 99 -- 108. ACM. Xia, L., Conitzer, V., and Procaccia, A. D. (2010). A scheduling approach to coalitional manipulation. In Proceedings of the 11th ACM conference on Electronic Commerce (EC), pages 275 -- 284. ACM. Xia, L., Zuckerman, M., Procaccia, A. D., Conitzer, V., and Rosenschein, J. S. (2009). Complexity of unweighted coalitional manipulation under some com- mon voting rules. In Proceedings of the 21st International Joint Conference on Artificial Intelligence (IJCAI), volume 9, pages 348 -- 353. Zuckerman, M., Lev, O., and Rosenschein, J. S. (2011). An algorithm for the coalitional manipulation problem under maximin. In The 10th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 845 -- 852. International Foundation for Autonomous Agents and Multiagent Systems. 21
0801.0249	1	0801	2007-12-31T23:22:50	A mathematical formalism for agent-based modeling	[ "cs.MA", "cs.DM", "math.CO" ]	Many complex systems can be modeled as multiagent systems in which the constituent entities (agents) interact with each other. The global dynamics of such a system is determined by the nature of the local interactions among the agents. Since it is difficult to formally analyze complex multiagent systems, they are often studied through computer simulations. While computer simulations can be very useful, results obtained through simulations do not formally validate the observed behavior. Thus, there is a need for a mathematical framework which one can use to represent multiagent systems and formally establish their properties. This work contains a brief exposition of some known mathematical frameworks that can model multiagent systems. The focus is on one such framework, namely that of finite dynamical systems. Both, deterministic and stochastic versions of this framework are discussed. The paper contains a sampling of the mathematical results from the literature to show how finite dynamical systems can be used to carry out a rigorous study of the properties of multiagent systems and it is shown how the framework can also serve as a universal model for computation.	cs.MA	cs	A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING REINHARD LAUBENBACHER, ABDUL S. JARRAH, HENNING MORTVEIT, AND S. S. RAVI Abstract. Many complex systems can be modeled as multiagent systems in which the constituent entities (agents) interact with each other. The global dynamics of such a system is determined by the nature of the local interactions among the agents. Since it is difficult to formally analyze complex multiagent systems, they are often studied through computer simulations. While computer simulations can be very useful, results obtained through simulations do not formally validate the observed behavior. Thus, there is a need for a mathematical framework which one can use to represent multiagent systems and formally establish their properties. This work contains a brief exposition of some known mathematical frameworks that can model multiagent systems. The focus is on one such framework, namely that of finite dynamical systems. Both, deterministic and stochastic versions of this framework are discussed. The paper contains a sampling of the mathematical results from the literature to show how finite dynamical systems can be used to carry out a rigorous study of the properties of multiagent systems and it is shown how the framework can also serve as a universal model for computation. Contents Glossary 1. Definition of the sub ject and its importance 2. Introduction 2.1. Examples of Agent-Based Simulations 3. Existing mathematical frameworks 3.1. Cellular Automata 3.2. Hopfield Networks 3.3. Communicating Finite State Machines 4. Finite Dynamical Systems 4.1. Definitions, Background, and Examples 4.2. Stochastic Finite Dynamical Systems 4.3. Agent-based Simulations as Finite Dynamical Systems 5. Finite dynamical systems as theoretical and computational tools 5.1. A Computational View of Finite Dynamical Systems: Definitions 5.2. Configuration Reachability Problem for SDSs 5.3. Turing Machines: A Brief Overview 5.4. How SDSs Can Mimic Turing Machines 5.5. TRANSIMS related questions 6. Mathematical results on finite dynamical systems 6.1. Parallel update systems 6.2. Sequential update systems 6.3. The category of sequential dynamical systems 7. Future directions 8. Bibliography Primary Litreture Books and Reviews 1 2 2 3 4 6 6 7 7 8 8 10 11 12 12 12 13 13 14 14 15 16 18 19 19 19 23 2 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI Glossary Agent-based simulation. An agent-based simulation of a complex system is a computer model that consists of a collection of agents/variables that can take on a typically finite collection of states. The state of an agent at a given point in time is determined through a collection of rules that describe the agent’s interaction with other agents. These rules may be deterministic or stochastic. The agent’s state depends on the agent’s previous state and the state of a collection of other agents with whom it interacts. Mathematical framework. A mathematical framework for agent-based simulation consists of a collection of mathematical ob jects that are considered mathematical abstractions of agent-based simulations. This collection of ob jects should be general enough to capture the key features of most simulations, yet specific enough to allow the development of a mathematical theory with meaningful results and algorithms. Finite dynamical system. A finite dynamical system is a time-discrete dynamical system on a finite state set. That is, it is a mapping from a cartesian product of finitely many copies of a finite set to itself. This finite set is often considered to be a field. Dynamics is generated by iteration of the mapping. 1. Definition of the subject and its importance Agent-based simulations are generative or computational approaches used for analyzing “complex systems.” What is a “system?” Examples of systems include a collection of molecules in a container, the population in an urban area, and the brokers in a stock market. The entities or agents in these three systems would be molecules, individuals and stock brokers, respectively. The agents in such systems interact in the sense that molecules collide, individuals come into contact with other individuals and brokers trade shares. Such systems, often called multiagent systems, are not necessarily complex. The label “complex” is typically attached to a system if the number of agents is large, if the agent interactions are involved, or if there is a large degree of heterogeneity in agent character or their interactions. This is of course not an attempt to define a complex system. Currently there is no generally agreed upon definition of complex systems. It is not the goal of this article to provide such a definition – for our purposes it will be sufficient to think of a complex system as a collection of agents interacting in some manner that involves one or more of the complexity components just mentioned, that is, with a large number of agents, heterogeneity in agent character and interactions and possibly stochastic aspects to all these parts. The global properties of complex systems, such as their global dynamics, emerge from the totality of local interactions between individual agents over time. While these local interactions are well understood in many cases, little is known about the emerging global behavior. Thus, it is typically difficult to construct global mathematical models such as systems of ordinary or partial differential equations, whose properties one could then analyze. Agent-based simulations are one way to create computational models of complex systems that take their place. An agent-based simulation, sometimes also called an individual-based or interaction-based sim- ulation (which we prefer), of a complex system is in essence a computer program that realizes some (possibly approximate) model of the system to be studied, incorporating the agents and their rules of interaction. The simulation might be deterministic (i.e., the evolution of agent-states is governed by deterministic rules) or stochastic. The typical way in which such simulations are used is to initialize the computer program with a particular assignment of agent states and to run it for some time. The output is a temporal sequence of states for all agents, which is then used to draw conclusions about the complex system one is trying to understand. In other words, the computer A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 3 program is the model of the complex system, and by running the program repeatedly, one expects to obtain an understanding of the characteristics of the complex system. There are two main drawbacks to this approach. First, it is difficult to validate the model. Simulations for most systems involve quite complex software constructs that pose challenges to code validation. Second, there are essentially no rigorous tools available for an analysis of model properties and dynamics. There is also no widely applicable formalism for the comparison of models. For instance, if one agent-based simulation is a simplification of another, then one would like to be able to relate their dynamics in a rigorous fashion. We are currently lacking a mathematically rich formal framework that models agent-based simulations. This framework should have at its core a class of mathematical ob jects to which one can map agent-based simulations. The ob jects should have a sufficiently general mathematical structure to capture key features of agent-based simulations and, at the same time, should be rich enough to allow the derivation of substantial mathematical results. This chapter presents one such framework, namely the class of time-discrete dynamical systems over finite state sets. The building blocks of these systems consist of a collection of variables (mapping to agents), a dependency graph that captures the dependence relations of agents on other agents, a local update function for each agent that encapsulates the rules by which the state of each agent evolves over time, and an update discipline for the variables (e.g. parallel or sequential). We will show that this class of mathematical ob jects is appropriate for the representation of agent-based simulations and, therefore, complex systems, and is rich enough to pose and answer relevant mathematical questions. This class is sufficiently rich to be of mathematical interest in its own right and much work remains to be done in studying it. We also remark that many other frameworks such as probabilistic Boolean networks [80] fit inside the framework described here. 2. Introduction Computer simulations have become an integral part of today’s research and analysis methodolo- gies. The ever-increasing demands arising from the complexity and sheer size of the phenomena studied constantly push computational boundaries, challenge existing computational methodolo- gies, and motivate the development of new theories to improve our understanding of the potential and limitations of computer simulation. Interaction-based simulations are being used to simulate a variety of biological systems such as ecosystems and the immune system, social systems such as urban populations and markets, and infrastructure systems such as communication networks and power grids. To model or describe a given system, one typically has several choices in the construction and design of agent-based models and representations. When agents are chosen to be simple, the simulation may not capture the behavior of the real system. On the other hand, the use of highly sophisticated agents can quickly lead to complex behavior and dynamics. Also, use of sophisticated agents may lead to a system that scales poorly. That is, a linear increase in the number of agents in the system may require a non-linear (e.g., quadratic, cubic, or exponential) increase in the computational resources needed for the simulation. Two common methods, namely discrete event simulation and time-stepped simulation, are often used to implement agent-based models [1; 45; 67]. In the discrete event simulation method, each event that occurs in the system is assigned a time of occurrence. The collection of events is kept in increasing order of their occurrence times. (Note that an event occurring at a certain time may give rise to events which occur later.) When all the events that occur at a particular time instant have been carried out, the simulation clock is advanced to the next time instant in the order. Thus, the differences between successive values of the simulation clock may not be uniform. Discrete event simulation is typically used in contexts such as queuing systems [58]. In the time-stepped method of simulation, the simulation clock is always advanced by the same amount. For each value of the simulation clock, the states of the system components are computed using equations that 4 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI model the system. This method of simulation is commonly used for studying, e.g., fluid flows or chemical reactions. The choice of model (discrete event versus time-stepped) is typically guided by an analysis of the computational speed they can offer, which in turn depends on the nature of the system being modeled, see, e.g., [37]. Tool-kits for general purpose agent-based simulations include Swarm [29; 57] and Repast [68]. Such tool-kits allow one to specify more complex agents and interactions than would be possible using, e.g., ordinary differential equations models. In general, it is difficult to develop a software package that is capable of supporting the simulation of a wide variety of physical, biological, and social systems. Standard or classical approaches to modeling are often based on continuous techniques and frame- works such as ordinary differential equations (ODEs) and partial differential equations (PDEs). For example, there are PDE based models for studying traffic flow [38; 47; 85]. These can accurately model the emergence of traffic jams for simple road/intersection configurations through, for exam- ple, the formation of shocks. However, these models fail to scale to the size and the specifications required to accurately represent large urban areas. Even if they hypothetically were to scale to the required size, the answers they provide (e.g. car density on a road as a function of position and time) cannot answer questions pertaining to specific travelers or cars. Questions of this type can be naturally described and answered through agent-based models. An example of such a system is TRANSIMS (see Section 2.1.1), where an agent-based simulation scheme is implemented through a cellular automaton model. Another well-known example of the change in modeling paradigms from continuous to discrete is given by lattice gas automata [32] in the context of fluid dynamics. Stochastic elements are inherent in many systems, and this usually is reflected in the resulting models used to describe them. A stochastic framework is a natural approach in the modeling of, for example, noise over a channel in a simulation of telecommunication networks [8]. In an economic market or a game theoretic setting with competing players, a player may sometimes decide to provide incorrect information. The state of such a player may therefore be viewed and modeled by a random variable. A player may make certain probabilistic assumptions about other players’ environment. In biological systems, certain features and properties may only be known up to the level of probability distributions. It is only natural to incorporate this stochasticity into models of such systems. Since applications of stochastic discrete models are common, it is desirable to obtain a better understanding of these simulations both from an application point of view (reliability, validation) and from a mathematical point of view. However, an important disadvantage of agent-based models is that there are few mathematical tools available at this time for the analysis of their dynamics. 2.1. Examples of Agent-Based Simulations. In order to provide the reader with some concrete examples that can also be used later on to illustrate theoretical concepts we describe here three examples of agent-based descriptions of complex systems, ranging from traffic networks to the immune system and voting schemes. 2.1.1. TRANSIMS ( TRansportation ANalysis SIMulation System). TRANSIMS is a large-scale computer simulation of traffic on a road network [63; 66; 76]. The simulation works at the resolution level of individual travelers, and has been used to study large US metropolitan areas such as Portland,OR, Washington D.C. and Dallas/Fort Worth. A TRANSIMS-based analysis of an urban area requires (i) a population, (ii) a location-based activity plan for each individual for the duration of the analysis period and (iii) a network representation of all transportation pathways of the given area. The data required for (i) and (ii) are generated based on, e.g., extensive surveys and other information sources. The network representation is typically very close to a complete description of the real transportation network of the given urban area. TRANSIMS consists of two main modules: the router and the cellular automaton based micro- simulator. The router maps each activity plan for each individual (obtained typically from activity A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 5 surveys) into a travel route. The micro-simulator executes the travel routes and sends each indi- vidual through the transportation network so that its activity plan is carried out. This is done in such a way that all constraints imposed on individuals from traffic driving rules, road signals, fellow travelers, and public transportation schedules are respected. The time scale is typically 1 second. The micro-simulator is the part of TRANSIMS responsible for the detailed traffic dynamics. Its implementation is based on cel lular automata which are described in more detail in Section 3.1. Here, for simplicity, we focus on the situation where each individual travels by car. The road network representation is in terms of links (e.g. road segments) and nodes (e.g. intersections). The network description is turned into a cell-network description by discretizing each lane of every link into cells. A cell corresponds to a 7.5 meter lane segment, and can have up to four neighbor cells (front, back, left and right). The vehicle dynamics is specified as follows. Vehicles travel with discrete velocities 0, 1, 2, 3, 4 or 5 which are constant between time steps. Each update time step brings the simulation one time unit forward. If the time unit is one second then the maximum speed of vmax = 5 cells per time unit corresponds to an actual speed of 5 × 7.5 m/s = 37.5 m/s which is 135 kmh or approximately 83.9 mph. Ignoring intersection dynamics, the micro-simulator executes three functions for each vehicle in every update: (a) lane-changing, (b) acceleration and (c) movement. These functions can be implemented through four cellular automata, one each for lane change decision and execution, one for acceleration and one for movement. For instance, the acceleration automaton works as follows. A vehicle in TRANSIMS can increase its speed by at most 1 cell per second, but if the road ahead is blocked, the vehicle can come to a complete stop in the same time. The function that is applied to each cell that has a car in it uses the gap ahead and the maximal speed to determine if the car will increase or decrease its velocity. Additionally, a car may have its velocity decreased one unit as determined by a certain deceleration probability. The random deceleration is an important element of producing realistic traffic flow. A ma jor advantage of this representation is that it leads to very light-weight agents, a feature that is critical for achieving efficient scaling. 2.1.2. CImmSim. Next we discuss an interaction-based simulation that models certain aspects of the human immune system. Comprised of a large number of interacting cells whose motion is constrained by the body’s anatomy, the immune system lends itself very well to agent-based sim- ulation. In particular, these models can take into account three-dimensional anatomical variations as well as small-scale variability in cell distributions. For instance, while the number of T-cells in the human body is astronomical, the number of antigen-specific T-cells, for a specific antigen, can be quite small, thereby creating many spatial inhomogeneities. Also, little is known about the global structure of the system to be modeled. The first discrete model to incorporate a useful level of complexity was ImmSim [22; 23], devel- oped by Seiden and Celada as a stochastic cellular automaton. The system includes B-cells, T-cells, antigen presenting cells (APCs), antibodies, antigens, and antibody-antigen complexes. Receptors on cells are represented by bit strings, and antibodies use bit strings to represent their epitopes and peptides. Specificity and affinity are defined by using bit string similarity. The bit string approach was initially introduced in [31]. The model is implemented on a regular two-dimensional grid, which can be thought of as a slice of a lymph node, for instance. It has been used to study var- ious phenomena, including the optimal number of human leukocyte antigens in human beings [22], the autoimmunity and T-lymphocyte selection in the thymus [60], antibody selection and hyper- mutation [24], and the dependence of the selection and maturation of the immune response on the antigen-to-receptor’s affinity [15]. The computational limitations of the Seiden-Celada model have been overcome by a modified model, CimmSim [20], implemented on a parallel architecture. Its complexity is several orders of magnitude larger than that of its predecessor. It has been used to model hypersensitivity to chemotherapy [19] and the selection of escape mutants from immune 6 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI recognition during HIV infection [14]. In [21] the CimmSim framework was applied to the study of mechanisms that contribute to the persistence of infection with the Epstein-Barr virus. 2.1.3. A Voting Game. The following example describes a hypothetical voting scheme. The voting system is constructed from a collection of voters. For simplicity, it is assumed that only two candidates, represented by 0 and 1, contest in the election. There are N voters represented by the set {v1 , v2 , . . . , vN }. Each voter has a candidate preference or a state. We denote the state of voter vi by xi . Moreover, each voter knows the preferences or states of some of his or her friends (fellow voters). This friendship relation is captured by the dependency graph which we describe later in Section 4.1. Starting from an initial configuration of preferences, the voters cast their votes in some order. The candidate that receives the most votes is the winner. A number of rules can be formulated to decide how each voter chooses a candidate. We will provide examples of such rules later, and as will be seen, the outcome of the election is governed by the order in which the votes are cast as well as the structure of the dependency graph. 3. Existing mathematical frameworks The field of agent-based simulation currently places heavy emphasis on implementation and computation rather than on the derivation of formal results. Computation is no doubt a very useful way to discover potentially interesting behavior and phenomena. However, unless the simulation has been set up very carefully, its outcome does not formally validate or guarantee the observed phenomenon. It could simply be caused by an artifact of the system model, an implementation error, or some other uncertainty. A first step in a theory of agent-based simulation is the introduction of a formal framework that on the one hand is precise and computationally powerful, and, on the other hand, is natural in the sense that it can be used to effectively describe large classes of both deterministic and stochastic systems. Apart from providing a common basis and a language for describing the model using a sound formalism, such a framework has many advantages. At a first level, it helps to clarify the key structure in a system. Domain specific knowledge is crucial to deriving good models of complex systems, but domain specificity is often confounded by domain conventions and terminology that eventually obfuscate the real structure. A formal, context independent framework also makes it easier to take advantage of existing general theory and algorithms. Having a model formulated in such a framework also makes it easier to establish results. Additionally, expressing the model using a general framework is more likely to produce results that are widely applicable. This type of framework also supports implementation and validation. Modeling languages like UML [16] serve a similar purpose, but tend to focus solely on software implementation and validation issues, and very little on mathematical or computational analysis. 3.1. Cellular Automata. In this section we discuss several existing frameworks for describing agent-based simulations. Cellular automata (CA) were introduced by Ulam and von Neumann [84] as biologically motivated models of computation. Early research addressed questions about the computational power of these devices. Since then their properties have been studied in the context of dynamical systems [40], language theory [52], and ergodic theory [51] to mention just a few areas. Cellular automata were popularized by Conway [35] (Game of Life) and by Wolfram [55; 86; 88]. Cellular automata (both deterministic and stochastic) have been used as models for phenomena ranging from lattice gases [32] and flows in porous media [77] to traffic analysis [33; 64; 65]. A cellular automaton is typically defined over a regular grid. An example is a two-dimensional grid such as Z2 . Each grid point (i, j ) is referred to as a site or node. Each site has a state xi,j (t) which is often taken to be binary. Here t denotes the time step. Furthermore, there is a notion of a neighborhood for each site. The neighborhood N of a site is the collection of sites that can A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 7 influence the future state of the given site. Based on its current state xi,j (t) and the current states of the sites in its neighborhood N , a function fi,j is used to compute the next state xi,j (t + 1) of the site (i, j ). Specifically, we have (3.1) xi,j (t + 1) = fi,j ( ¯xi,j (t)) , where ¯xi,j (t) denotes the tuple consisting of all the states xi′ ,j ′ (t) with (i′ , j ′ ) ∈ N . The tuple consisting of the states of all the sites is the CA configuration and is denoted x(t) = (xi,j (t))i,j . Equation (3.1) is used to map the configuration x(t) to x(t + 1). The cellular automaton map or dynamical system, is the map Φ that sends x(t) to x(t + 1). A central part of CA research is to understand how configurations evolve under iteration of the map Φ and what types of dynamical behavior can be generated. A general introduction to CA can be found in [43]. 3.2. Hopfield Networks. Hopfield networks were proposed as a simple model of associative mem- ories [42]. A discrete Hopfield neural network consists of an undirected graph Y (V , E ). At any time t, each node vi ∈ V has a state xi (t) ∈ {+1, −1}. Further, each node vi ∈ V has an associated threshold τi ∈ R. Each edge {vi , vj } ∈ E has an associated weight wi,j ∈ R. For each node vi , the neighborhood Ni of vi includes vi and the set of nodes that are adjacent to vi in Y . Formally, Ni = {vi} ∪ {vj ∈ V : {vi , vj } ∈ E }. States of nodes are updated as follows. At time t, node vi computes the function fi defined by wi,j xj (t) fi (t) = sgn  −τi + Xvj ∈Ni  , where sgn is the map from R to {+1, −1}, defined by sgn(x) = (cid:26) 1, if x ≥ 0 and −1 otherwise. Now, the state of vi at time t + 1 is xi (t + 1) = fi (t). Many references on Hopfield networks (see for example [42; 78]) assume that the underlying undirected graph is complete; that is, there is an edge between every pair of nodes. In the definition presented above, the graph need not be complete. However, this does not cause any difficulties since the missing edges can be assigned weight 0. As a consequence, such edges will not play any role in determining the dynamics of the system. Both synchronous and asynchronous update models of Hopfield neural networks have been considered in the literature. For theoretical results concerning Hopfield networks see [69; 70] and the references cited therein. Reference [78] presents a number of applications of neural networks. In [54], a Hopfield model is used to study polarization in dynamic networks. 3.3. Communicating Finite State Machines. The model of communicating finite state ma- chines (CFSM) was proposed to analyze protocols used in computer networks. In some of the literature, this model is also referred to as “concurrent transition systems” [36]. In the CFSM model, each agent is a process executing on some node of a distributed computing system. Although there are minor differences among the various CFSM models proposed in the literature [17; 36], the basic set-up models each process as a finite state machine (FSM). Thus, each agent is in a certain state at any time instant t. For each pair of agents, there is a bidirectional channel through which they can communicate. The state of an agent at time t+1 is a function of the current state and the input (if any) on one or more of the channels. When an agent (FSM) undergoes a transition from one state to another, it may also choose to send a message to another agent or receive a message from an agent. In general, such systems can be synchronous or asynchronous. 8 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI As can be seen, CFSMs are a natural formalism for studying protocols used in computer networks. The CFSM model has been used extensively to prove properties (e.g. deadlock freedom, fairness) of a number of protocols used in practice (see [17; 36] and the references cited therein). Other frameworks include interacting particle systems [50], and Petri nets [59]. There is a vast literature on both, but space limitations prevent a discussion here. 4. Finite Dynamical Systems Another, quite general, modeling framework that has been proposed is that of finite dynamical systems, both synchronous and asynchronous. Here the proposed mathematical ob ject representing an agent-based simulation is a time-discrete dynamical system on a finite state set. The description of the systems is modeled after the key components of an agent-based simulation, namely agents, the dependency graph, local update functions, and an update order. This makes a mapping to agent- based simulations natural. In the remainder of this chapter we will show that finite dynamical systems satisfy our criteria for a good mathematical framework in that they are general enough to serve as a broad computing tool and mathematically rich enough to allow the derivation of formal results. 4.1. Definitions, Background, and Examples. Let x1 , . . . , xn be a collection of variables, which take values in a finite set X . (As will be seen, the variables represent the entities in the system being modeled and the elements of X represent their states.) Each variable xi has associated to it a “local update function” fi : X n −→ X , where “local” refers to the fact that fi takes inputs from the variables in the “neighborhood” of xi , in a sense to be made precise below. By abuse of notation we also let fi denote the function X n −→ X n which changes the i-th coordinate and leaves the other coordinates unchanged. This allows for the sequential composition of the local update functions. These functions assemble to a dynamical system Φ = (f1 , . . . , fn ) : X n −→ X n , with the dynamics generated by iteration of Φ. As an example, if X = {0, 1} with the standard Boolean operators AND and OR, then Φ is a Boolean network. The assembly of Φ from the local functions fi can be done in one of several ways. One can update each of the variables simultaneously, that is, Φ(x1 , . . . , xn ) = (f1 (x1 , . . . , xn ), . . . , fn(x1 , . . . , xn )) . In this case one obtains a paral lel dynamical system. Alternatively, one can choose to update the states of the variables according to some fixed update order, for example, a permutation or a word on the set {1, . . . , n}. That is, let π = (π1 , . . . , πt ) be a word using the alphabet {1, . . . , n}. The function (4.1) Φπ = fπt ◦ fπn−1 ◦ · · · ◦ fπ1 , is called a sequential dynamical system (SDS) and, as before, the dynamics of Φπ is generated by iteration. The case when π is a permutation on {1, . . . , n} has been studied extensively [9; 10; 11; 12]. It is clear that using a different permutation γ may result in a different dynamical system Φγ . Remark 4.1. Notice that for a fixed π , the function Φπ is a parallel dynamical system: once the update order π is chosen and the local update functions are composed according to π , that is, the function Φπ has been computed, then Φπ (x1 , . . . , xn ) = g(x1 , . . . , xn ) where g is a parallel update dynamical system. However, the maps gi are not local functions. The dynamics of Φ is usually represented as a directed graph on the vertex set X n , called the phase space of Φ. There is a directed edge from v ∈ X n to w ∈ X n if and only if Φ(v) = w. A second graph that is usually associated with a finite dynamical system is its dependency graph Y (V , E ). In general, this is a directed graph, and its vertex set is V = {1, . . . , n}. There is a A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 9 directed edge from i to j if and only if xi appears in the function fj . In many situations, the interaction relationship between pairs of variables is symmetric; that is, variable xi appears in fj if and only if xj appears in fi . In such cases, the dependency graph can be thought of as an undirected graph. We recall that the dependency graphs mentioned in the context of the voting game (Section 2.1.3) and Hopfield networks (Section 3.2) are undirected graphs. The dependency graph plays an important role in the study of finite dynamical systems and is sometimes listed explicitly as part of the specification of Φ. Example 4.2. Let X = {0, 1}. Suppose we have four variables and the local Boolean update functions are f1 = x1 + x2 + x3 + x4 , f2 = x1 + x2 , f3 = x1 + x3 , f4 = x1 + x4 , where “+” represents XOR, the exclusive OR function. The dynamics of the function Φ = (f1 , . . . , f4 ) : X 4 −→ X 4 is the directed graph on the left in Figure 1 while the dependency graph is on the right. 0 1 1 1 1 0 0 0 0 0 0 1 1 1 1 0 0 0 1 0 1 1 0 1 0 1 0 0 1 0 1 1 1 1 1 1 1 0 0 1 1 0 1 0 1 1 0 0 0 0 0 0 0 1 1 0 0 1 0 1 0 0 1 1 x1 x3 x2 x4 Figure 1. The phase space of the parallel system Φ is on the left and dependency graph of the Boolean function from Example 4.2 is on the right. Example 4.3. Consider the local functions in the Example 4.2 above and let π = (2, 1, 3, 4). Then Φπ = f4 ◦ f3 ◦ f1 ◦ f2 : X 4 −→ X 4 . The phase space of Φπ is the directed graph on the left in Figure 2, while the phase space of Φγ , where γ = id is on the right in Figure 2. 0 0 0 0 0 0 0 1 0 0 1 0 0 0 1 1 0 1 0 0 0 1 0 1 0 1 1 0 0 1 1 1 0 0 0 0 0 0 0 1 0 0 1 0 0 0 1 1 0 1 0 0 0 1 0 1 0 1 1 0 0 1 1 1 1 0 1 0 1 0 0 1 1 1 1 1 1 1 0 0 1 1 1 0 1 1 0 1 1 0 1 1 1 0 0 0 1 1 0 1 1 1 1 0 1 0 0 0 1 0 1 1 1 0 0 1 1 0 1 0 1 1 0 0 1 1 1 1 Figure 2. The phase spaces from Example 4.3: Φπ (left) and Φid (right) Notice that the phase space of any function is a directed graph in which every vertex has out- degree one; this is a characteristic property of deterministic functions. Making use of Boolean arithmetic is a powerful tool in studying Boolean networks, which is not available in general. In order to have available an enhanced set of tools it is often natural to make an additional assumption regarding X , namely that it is a finite number system, a finite field [49]. This amounts to the assumption that there are “addition” and “multiplication” operations defined 10 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI on X that satisfy the same rules as ordinary addition and multiplication of numbers. Examples include Zp , the integers modulo a prime p. This assumption can be thought of as the discrete analog of imposing a coordinate system on an affine space. When the set X is a finite field, it is easy to show that for any local function g , there exists a polynomial h such that g(x1 , . . . , xn ) = h(x1 , . . . , xn ) for all (x1 , . . . , xn ) ∈ X n . To be precise, suppose X is a finite field with q elements. Then n Yi=1 g(x1 , . . . , xn ) = X(c1 ,...,cn )∈X n (1 − (xi − ci )q−1 ). This observation has many useful consequences, since polynomial functions have been studied extensively and many analytical tools are available. Notice that cellular automata and Boolean networks, both parallel and sequential, are special classes of polynomial dynamical systems. In fact, it is straightforward to see that g(c1 , . . . , cn ) (4.2) (4.3) x ∧ y = x · y , x ∨ y = x + y + xy and ¬x = x + 1. Therefore, the modeling framework of finite dynamical systems includes that of cellular automata, discussed earlier. Also, since a Hopfield network is a function X n −→ X n , which can be represented through its local constituent functions, it follows that Hopfield networks also are special cases of finite dynamical systems. 4.2. Stochastic Finite Dynamical Systems. The deterministic framework of finite dynamical systems can be made stochastic in several different ways, making one or more of the system’s defining data stochastic. For example, one could use one or both of the following criteria. Assume that each variable has a nonempty set of local functions assigned to it, together with a probability distribution on this set, and each time a variable is updated, one of these local functions is chosen at random to update its state. We call such systems probabilistic finite dynamical systems (PFDS), a generalization of probabilistic Boolean networks [81]. Fix a subset of permutations T ⊆ Sn together with a probability distribution. When it is time for the system to update its state, a permutation π ∈ T is chosen at random and the agents are updated sequentially using π . We call such systems stochastic finite dynamical systems (SFDS). Remark 4.4. By Remark 4.1, each system Φπ is a parallel system. Hence a SFDS is nothing but a set of parallel dynamical systems {Φπ : π ∈ T }, together with a probability distribution. When it is time for the system to update its state, a system Φπ is chosen at random and used for the next iteration. To describe the phase space of a stochastic finite dynamical system, a general method is as follows. Let Ω be a finite collection of systems Φ1 , . . . , Φt , where Φi : X n −→ X n for all i, and consider the probabilities p1 , . . . , pt which sum to 1. We obtain the stochastic phase space (4.4) ΓΩ = p1Γ1 + p2Γ2 + · · · + ptΓt , where Γi is the phase space of Φi . The associated probability space is F = (Ω, 2Ω , µ), where the probability measure µ is induced by the probabilities pi . It is clear that the stochastic phase space can be viewed as a Markov chain over the state space X n . The adjacency matrix of ΓΩ directly encodes the Markov transition matrix. This is of course not new, and has been done in, e.g.,[28 ; 81; 83]. But we emphasize the point that even though SFDS give rise to Markov chains our study of SFDS is greatly facilitated by the rich additional structure available in these systems. To understand the effect of structural components such as the topology of the dependency graph or the stochastic nature of the update, it is important to study them not as Markov chains but as SFDS. A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 11 Example 4.5. Consider Φπ and Φγ from Example 4.3 and let Γπ and Γγ be their phases spaces as 1 1 1 Γγ of the stochastic sequential Γπ + . The phase space shown in Figure 2. Let p1 = p2 = 2 2 2 dynamical system obtained from Φπ and Φγ (with equal probabilities) is presented in Figure 3. 0001 1101 1010 1110 1001 0010 0110 0101 0011 0000 0100 1000 1111 1011 1100 0111 Figure 3. The stochastic phase space for Example 4.5 induced by the two deter- ministic phase spaces of Φπ and Φγ from Figure 2. For simplicity the weights of the edges have been omitted. 4.3. Agent-based Simulations as Finite Dynamical Systems. In the following we describe the generic structure of the systems typically modeled and studied through agent-based simulations. The central notion is naturally that of an agent. Each agent carries a state that may encode its preferences, internal configuration, perception of its environment, and so on. In the case of TRANSIMS, for instance, the agents are the cells making up the road network. The cell state contains information about whether or not the cell is occupied by a vehicle as well as the velocity of the vehicle. One may assume that each cell takes on states from the same set of possible states, which may be chosen to support the structure of a finite field. The agents interact with each other, but typically an agent only interacts with a small subset of agents, its neighbors. Through such an interaction an agent may decide to change its state based on the states (or other aspects) of the agents with which it interacts. We will refer to the process where an agent modifies its state through interaction as an agent update. The precise way in which an agent modifies its state is governed by the nature of the particular agent. In TRANSIMS the neighbors of a cell are the adjacent road network cells. From this adjacency relation one obtains a dependency graph of the agents. The local update function for a given agent can be obtained from the rules governing traffic flow between cells. The updates of all the agents may be scheduled in different ways. Classical approaches include synchronous, asynchronous and event-driven schemes. The choice will depend on system properties or particular considerations about the simulation implementation. In the case of CImmSim, the situation is somewhat more complicated. Here the agents are also the spatial units of the system, each representing a small volume of lymph tissue. The total volume is represented as a 2-dimensional CA, in which every agent has 4 neighbors, so that the dependency graph is a regular 2-dimensional grid. The state of each agent is a collection of counts for the various immune cells and pathogens that are present in this particular agent (volume). Movement between cells is implemented as diffusion. Immune cells can interact with each other and with pathogens while they reside in the same volume. Thus, the local update function for a given cell of the simulation is made up of the two components of movement between cells and interactions within a cell. For instance, a B cell could interact with the Epstein-Barr virus in a given volume and transition from uninfected to infected by the next time step. Interactions as well as movement are stochastic, resulting in a stochastic finite dynamical system. The update order is parallel. Example 4.6 (The voting game (Section 2.1.3)). The following scenario is constructed to illustrate how implementation choices for the system components have a clear and direct bearing on the dynamics and simulation outcomes. 12 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI Let the voter dependency graph be the star graph on 5 vertices with center vertex a and sur- rounding vertices b, c, d and e. Furthermore, assume that everybody votes opportunistically using the ma jority rule: the vote cast by an individual is equal to the preference of the ma jority of his/her friends with the person’s own preference included. For simplicity, assume candidate 1 is preferred in the case of a tie. If the initial preference is xa = 1 and xb = xc = xd = xe = 0 then if voter a goes first he/she will vote for candidate 0 since that is the choice of the ma jority of the neighbor voters. However, if b and c go first they only know a’s preference. Voter b therefore casts his/her vote for candidate 1 as does c. Note that this is a tie situation with an equal number of preferences for candidate 1 (a) and for candidate 2 (b). If voter a goes next, then the situation has changed: the preference of b and c has already changed to 1. Consequently, voter a picks candidate 1. At the end of the day candidate 1 is the election winner, and the choice of update order has tipped the election! This example is of course constructed to illustrate our point. However, in real systems it can be much more difficult to detect the presence of such sensitivities and their implications. A solid mathematical framework can be very helpful in detecting such effects. 5. Finite dynamical systems as theoretical and computational tools If finite dynamical systems are to be useful as a modeling paradigm for agent-based simulations it is necessary that they can serve as a fairly universal model of computation. We discuss here how such dynamical systems can mimic Turing Machines (TMs), a standard universal model for computation. For a more thorough exposition, we refer the reader to the series of papers by Barrett et al. [2; 3; 4; 5; 6]. To make the discussion reasonably self-contained, we provide a brief and informal discussion of the TM model. Additional information on TMs can be found in any standard text on the theory of computation (e.g. [82]). 5.1. A Computational View of Finite Dynamical Systems: Definitions. In order to un- derstand the relationship of finite dynamical systems to TMs, it is important to view such systems from a computational stand point. Recall that a finite dynamical system Φ : X n −→ X n , where X is a finite set, has an underlying dependency graph Y (V , E ). From a computational point of view, the nodes of the dependency graph (the agents in the simulation) are thought of as devices that compute appropriate functions. For simplicity, we will assume in this section that the dependency graph is undirected, that is, all dependency relations are symmetric. At any time, the state value of each node vi ∈ V is from the specified domain X . The inputs to fi are the current state of vi and the states of the neighbors of vi as specified by Y . The output of fi , which is also a member of X , becomes the state of vi at the next time instant. The discussion in this section will focus on sequentially updated systems (SDS), but all results discussed apply to parallel systems as well. Each step of the computation carried out by an SDS can be thought as consisting of n “mini steps;” in each mini step, the value of the local transition function at a node is computed and the state of the node is changed to the computed value. Given an SDS Φ, a configuration C of Φ is a vector (c1 , c2 , . . . , cn ) ∈ X n . It can be seen that each computational step of an SDS causes a transition from one configuration to another. 5.2. Configuration Reachability Problem for SDSs. Based on the computational view, a number of different problems can be defined for SDSs (see for example, [4; 6; 7]). To illustrate how SDSs can model TM computations, we will focus on one such problem, namely the configuration reachability (CR) problem: Given an SDS Φ, an initial configuration C and another configuration C ′ , will Φ, starting from C , ever reach configuration C ′ ? The problem can also be expressed in terms of the phase space of Φ. Since configurations such as C and C ′ are represented by nodes in the phase space, the CR problem boils down to the question of whether there is a directed path in the phase space from C to C ′ . This abstract problem can be mapped to several problems in the simulation of multiagent systems. Consider for example the TRANSIMS context. Here, the initial A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 13 configuration C may represent the state of the system early in the day (when the traffic is very light) and C ′ may represent an “undesirable” state of the system (such as heavy traffic congestion). Similarly, in the context of modeling an infectious disease, C may represent the initial onset of the disease (when only a small number of people are infected) and C ′ may represent a situation where a large percentage of the population is infected. The purpose of studying computational problems such as CR is to determine whether one can efficiently predict the occurrence of certain events in the system from a description of the system. If computational analysis shows that the system can indeed reach undesirable configurations as it evolves, then one can try to identify steps needed to deal with such situations. 5.3. Turing Machines: A Brief Overview. A Turing machine (TM) is a simple and commonly used model for general purpose computational devices. Since our goal is to point out how SDSs can also serve as computational devices, we will present an informal overview of the TM model. Readers interested in a more formal description may consult [82]. A TM consists of a set Q of states, a one-way infinite input tape and a read/write head that can read and modify symbols on the input tape. The input tape is divided into cells and each cell contains a symbol from a special finite alphabet. An input consisting of n symbols is written on the leftmost n cells of the input tape. (The other cells are assumed to contain a special symbol called blank.) One of the states in Q, denoted by qs , is the designated start state. Q also includes two other special states, denoted by qa (the accepting state) and qr (the rejecting state). At any time, the machine is in one of the states in Q. The transition function for the TM specifies for each combination of the current state and the current symbol under the head, a new state, a new symbol for the current cell (which is under the head) and a movement (i.e., left or right by one cell) for the head. The machine starts in state qs with the head on the first cell of the input tape. Each step of the machine is carried out in accordance with the transition function. If the machine ever reaches either the accepting or the rejecting state, it halts with the corresponding decision; otherwise, the machine runs forever. A configuration of a TM consists of its current state, the current tape contents and the position of the head. Note that the transition function of a TM specifies how a new configuration is obtained from the current configuration. The above description is for the basic TM model (also called single tape TM model). For convenience in describing some computations, several variants of the above basic model have been proposed. For example, in a multi-tape TM, there are one or more work tapes in addition to the input tape. The work tapes can be used to store intermediate results. Each work tape has its own read/write head and the definitions of configuration and transition function can be suitably modified to accommodate work tapes. While such an enhancement to the basic TM model makes it easier to carry out certain computations, it does not add to the machine’s computational power. In other words, any computation that can be carried out using the enhanced model can also be carried out using the basic model. As in the case of dynamical systems, one can define a configuration reachability (CR) problem for TMs: Given a TM M , an initial configuration I M and another configuration CM , will the TM starting from IM ever reach CM ? We refer to the CR problem in the context of TMs as CR-TM. In fact, it is this problem for TMs that captures the essence of what can be effectively computed. In particular, by choosing the state component of CM to be one of the halting states (qa or qr ), the problem of determining whether a function is computable is transformed into an appropriate CR-TM problem. By imposing appropriate restrictions on the resources used by a TM (e.g. the number of steps, the number of cells on the work tapes), one obtains different versions of the CR-TM problem which characterize different computational complexity classes [82]. 5.4. How SDSs Can Mimic Turing Machines. The above discussion points out an important similarity between SDSs and TMs. Under both of these models, each computational step causes 14 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI a transition from one configuration to another. It is this similarity that allows one to construct a discrete dynamical system Φ that can simulate a TM. Typically, each step of a TM is simulated by a short sequence of successive iterations Φ. As part of the construction, one also identifies a suitable mapping between the configurations of the TM being simulated and those of the dynamical system. This is done in such a way that the answer to CR-TM problem is “yes” if and only if the answer to the CR problem for the dynamical system is also “yes.” To illustrate the basic ideas, we will informally sketch a construction from [4]. This construction produces an SDS Φ that simulates a restricted version of TMs; the restriction being that for any input containing n symbols, the number of work tape cells that the machine may use is bounded by a linear function of n. Such a TM is called a linear bounded automaton (LBA) [82]. Let M denote the given LBA and let n denote the length of the input to M . The domain X for the SDS Φ is chosen to be a finite set based on the allowed symbols in the input to the TM. The dependency graph is chosen to be a simple path on n nodes, where each node serves as a representative for a cell on the input tape. The initial and final configurations C and C ′ for Φ are constructed from the corresponding configurations of M . The local transition function for each node of the SDS is constructed from the given transition function for M in such a way that each step of M corresponds to exactly one step of Φ. Thus, there is a simple bijection between the sequence of configurations that M goes through during its computation and the sequence of states that Φ goes through as it evolves. Using this bijection, it is shown in [4] that the answer to the CR-TM problem is “yes” if and only if Φ reaches C ′ starting from C . Reference [4] also presents a number of sophisticated constructions where the resulting dynamical system is very simple; for example, it is shown that an LBA can be simulated by an SDS in which X is the Boolean field, the dependency graph is d-regular for some constant d and the local transition functions at all the nodes are identical. Such results point out that one does not need complicated dynamical systems to model TM computations. Barrett et al. [5] have also considered stochastic SDS (SSDS), where the local transition function at each node is stochastic. For each combination of inputs, a stochastic local transition function specifies a probability distribution over the domain of state values. It is shown in [5] that SSDSs can effectively simulate computations carried out by probabilistic TMs (i.e., TMs whose transition functions are stochastic). 5.5. TRANSIMS related questions. Section 2.1.1 gave an overview of some aspects of the TRANSIMS model. The micro-simulator module is specified as a functional composition of four cellular automata of the form ∆4 ◦ ∆3 ◦ ∆2 ◦ ∆1 . (We only described ∆3 which corresponds to velocity updates.) Such a formulation has several advantages. First, it has created an abstraction of the essence of the system in a precise mathematical way without burying it in contextual domain details. An immediate advantage of this specification is that it makes the whole implementation process more straightforward and transparent. While the local update functions for the cells are typically quite simple, for any realistic study of an urban area the problem size would typically require a sequential implementation, raising a number of issues that are best addressed within a mathematical framework like the one considered here. 6. Mathematical results on finite dynamical systems In this section we outline a collection of mathematical results about finite dynamical systems that is representative of the available knowledge. The ma jority of these results are about deterministic systems, as the theory of stochastic systems of this type is still in its infancy. We will first consider synchronously updated systems. Throughout this section, we make the assumption that the state set X carries the algebraic structure of a finite field. Accordingly, we use the notation k instead of X . It is a standard result that in this case the number q of elements in k has the form q = pt for some prime p and t ≥ 1. The A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 15 reader may keep the simplest case k = {0, 1} in mind, in which case we are considering Boolean networks. Recall Equation (4.2). That is, any function g : kn −→ k can be represented by a multivariate polynomial with coefficients in k . If we require that the exponent of each variable be less than q , then this representation is unique. In particular Equation (4.3) implies that every Boolean function can be represented uniquely as a polynomial function. 6.1. Parallel update systems. Certain classes of finite dynamical systems have been studied extensively, in particular cellular automata and Boolean networks. Many of these studies have been experimental in nature, however. Some more general mathematical results about cellular automata can be found in the papers of Wolfram and collaborators [87]. The results there focus primarily on one-dimensional Boolean cellular automata with a particular fixed initial state. Here we collect a sampling of more general results. We first consider linear and affine systems Φ = (f1 , . . . , fn ) : kn −→ kn That is, we consider systems for which the coordinate functions fi are linear, resp. affine, polyno- mials. (In the Boolean case this includes functions constructed from XOR and negation.) When each fi is a linear polynomial of the form fi (x1 , . . . , xn ) = ai1x1 + · · · + ainxn , the map Φ is nothing but a linear transformation on kn over k , and, by using the standard basis, Φ has the matrix representation x1 ... xn x1 ... xn , a11 ... an1 · · · a1n ... . . . · · · ann    =     Φ          where aij ∈ k for all i, j . Linear finite dynamical systems were first studied by Elspas [30]. His motivation came from studying feedback shift register networks and their applications to radar and communication sys- tems and automatic error correction circuits. For linear systems over finite fields of prime cardi- nality, that is, q is a prime number, Elspas showed that the exact number and length of each limit cycle can be determined from the elementary divisors of the matrix A. Recently, Hernandez [41] re-discovered Elspas’ results and generalized them to arbitrary finite fields. Furthermore, he showed that the structure of the tree of transients at each node of each limit cycle is the same, and can be completely determined from the nilpotent elementary divisors of the form xa . For affine Boolean networks (that is, finite dynamical systems over the Boolean field with two elements, whose local functions are linear polynomials which might have constant terms), a method to analyze their cycle length has been developed in [56]. After embedding the matrix of the transition function, which is of dimension n × (n + 1), into a square matrix of dimension n + 1, the problem is then reduced to the linear case. A fast algorithm based on [41] has been implemented in [44], using the symbolic computation package Macaulay2. It is not surprising that the phase space structure of Φ should depend on invariants of the matrix A = (aij ). The rational canonical form of A is a block-diagonal matrix and one can recover the structure of the phase space of A from that of the blocks in the rational form of A. Each block represents either an invertible or a nilpotent linear transformation. Consider an invertible block B . If µ(x) is the minimal polynomial of B , then there exists s such that µ(x) divides xs − 1. Hence B s − I = 0 which implies that B sv = v. That is, every state vector v in the phase space of B is in a cycle whose length is a divisor of s. Definition 6.1. For any polynomial λ(x) in k [x], the order of λ(x) is the least integer s such that λ(x) divides xs − 1. 16 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI The cycle structure of the phase space of Φ can be completely determined from the orders of the irreducible factors of the minimal polynomial of Φ. The computation of these orders involves in particular the factorization of numbers of the form q r − 1, which makes the computation of the order of a polynomial potentially quite costly. The nilpotent blocks in the decomposition of A determine the tree structure at the nodes of the limit cycles. It turns out that all trees at all periodic nodes are identical. This generalizes a result in [55] for additive cellular automata over the field with two elements. While the fact that the structure of the phase space of a linear system can be determined from the invariants associated with its matrix may not be unexpected, it is a beautiful example of how the right mathematical viewpoint provides powerful tools to completely solve the problem of relating the structure of the local functions to the resulting (or emerging) dynamics. Linear and affine systems have been studied extensively in several different contexts and from several different points of view, in particular the case of cellular automata. For instance, additive cellular automata over more general rings as state sets have been studied, e.g., in [25]. Further results on additive CAs can also be found there. One important focus in [25] is on the problem of finding CAs with limit cycles of maximal length for the purpose of constructing pseudo random number generators. Unfortunately, the situation is more complicated for nonlinear systems. For the special class of Boolean synchronous systems whose local update functions consist of monomials, there is a polynomial time algorithm that determines whether a given monomial system has only fixed points as periodic points [27]. This question was motivated by applications to the modeling of biochemical networks. The criterion is given in terms of invariants of the dependency graph Y . For a strongly connected directed graph Y (i.e., there is a directed path between any pair of vertices), its loop number is the greatest common divisor of all directed loops at a particular vertex. (This number is independent of the vertex chosen.) Theorem 6.2 ([27]). A Boolean monomial system has only fixed points as periodic points if and only if the loop number of every strongly connected component of its dependency graph is equal to 1. In [26] it is shown that the problem for general finite fields can be reduced to that of a Boolean system and a linear system over rings of the form Z/prZ, p prime. Boolean monomial systems have been studied before in the cellular automaton context [13]. 6.2. Sequential update systems. The update order in a sequential dynamical system has been studied using combinatorial and algebraic techniques. A natural question to ask here is how the system depends on the update schedule. In [9; 10; 61; 72] this was answered on several levels for the special case where the update schedule is a permutation. We describe these results in some detail. Results about the more general case of update orders described by words on the indices of the local update functions can be found in [34]. Given local update functions fi : kn −→ k and permutation update orders σ, π , a natural question is when the two resulting SDS Φσ and Φπ are identical and, more generally, how many different systems one obtains by varying the update order over all permutations. Both questions can be answered in great generality. The answer involves invariants of two graphs, namely the acyclic orientations of the dependency graph Y of the local update functions and the update graph of Y . The update graph U (Y ) of Y is the graph whose vertex set consists of all permutations of the vertex set of Y [72]. There is an (undirected) edge between two permutations σ = (σ1 , . . . , σn ) and τ = (τ1 , . . . , τn ) if they differ by a transposition of two adjacent entries σi and σi+1 such that there is no edge in Y between σi and σi+1 . The update graph encodes the fact that one can commute two local update functions fi and fj without affecting the end result Φ if i and j are not connected by an edge in Y . That is, · · · fi ◦ fj · · · = · · · fj ◦ fi · · · if and only if i and j are not connected by an edge in Y . A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 17 All permutations belonging to the same connected component in U (Y ) give identical SDS maps. The number of (connected) components in U (Y ) is therefore an upper bound for the number of functionally inequivalent SDS that can be generated by just changing the update order. This can also be characterized in terms of acyclic orientations of the graph Y : each component in the update graph induces a unique acyclic orientation of the graph Y . Moreover, we have the following result: Proposition 6.3 ([72]). There is a bijection FY : SY / ∼Y −→ Acyc(Y ) , where SY / ∼Y denotes the set of connected components of U (Y ) and Acyc(Y ) denotes the set of acyclic orientations of Y . This upper bound on the number of functionally different systems has been shown in [72] to be sharp for Boolean systems, in the sense that for a given Y one construct this number of different systems, using appropriate combinations of NOR functions. For two permutations σ and τ it is easy to determine if they give identical SDS maps: one can just compare their induced acyclic orientations. The number of acyclic orientations of the graph Y tells how many functionally different SDS maps one can obtain for a fixed graph and fixed vertex functions. The work of Cartier and Foata [18] on partially commutative monoids studies a similar question, but their work is not concerned with finite dynamical systems. Note that permutation update orders have been studied sporadically in the context of cellular automata on circle graphs [71] but not in a systematic way, typically using the order (1, 2, . . . , n) or the even-odd/odd-even orders. As a side note, we remark that this work also confirms our findings that switching from a parallel update order to a sequential order turns the complex behavior found in Wolfram’s “class III and IV” automata into much more regular or mundane dynamics, see e.g. [79]. The work on functional equivalence was extended to dynamical equivalence (topological conju- gation) in [10; 61]. The automorphism group of the graph Y can be made to act on the components of the update graph U (Y ) and therefore also on the acyclic orientations of Y . All permutations contained in components of an orbit under Aut(Y ) give rise to dynamically equivalent sequential dynamical systems, that is, to isomorphic phase spaces. However, here one needs some more tech- nical assumptions, i.e., the local functions must be symmetric and induced, see [11]. This of course also leads to a bound for the number of dynamically inequivalent systems that can be generated by varying the update order alone. Again, this was first done for permutation update orders. The theory was extended to words over the vertex set of Y in [34; 75]. The structure of the graph Y influences the dynamics of the system. As an example, graph invariants such as the independent sets of Y turn out to be in a bijective correspondence with the invariant set of sequential systems over the Boolean field k that have nort : k t → k2 given by nort (x1 , . . . , xt ) = (1 + x1 ) · · · (1 + xt) as local functions [73]. This can be extended to other classes such as those with order independent invariant sets as in [39]. We have already seen how the automorphisms of a graph give rise to equivalence [61]. Also, if the graph Y has non-trivial covering maps we can derive simplified or reduced (in an appropriate sense) versions of the original SDS over the image graphs of Y , see e.g. [62; 74]. Parallel and sequential dynamical systems differ when it comes to invertibility. Whereas it is generally computationally intractable to determine if a CA over Zd is invertible for d ≥ 2 [46], it is straightforward to determine this for a sequential dynamical system [61]. For example, it turns out that the only invertible Boolean sequential dynamical systems with symmetric local functions are the ones where the local functions are either the parity function or the logical complement of the parity function [10]. Some classes of sequential dynamical systems such as the ones induced by the nor-function have desirable stability properties [39]. These systems have minimal invariant sets (i.e. periodic states) 18 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI that do not depend on the update order. Additionally, these invariant sets are stable with respect to configuration perturbations. If a state c is perturbed to a state c′ that is not periodic this state will evolve to a periodic state c′′ in one step; that is, the system will quickly return to the invariant set. However, the states c and c′′ may not necessarily be in the same periodic orbit. 6.3. The category of sequential dynamical systems. As a general principle, in order to study a given class of mathematical ob jects it is useful to study transformations between them. In order to provide a good basis for a mathematical analysis the ob jects and transformations together should form a category, that is, the class of transformations between two ob jects should satisfy certain reasonable properties (see, e.g., [53]). Several proposed definitions of a transformation of SDS have been published, notably in [48] and [74]. One possible interpretation of a transformation of SDS from the point of view of agent-based simulation is that the transformation represents the approximation of one simulation by another or the embedding/pro jection of one simulation into/onto another. These concepts have obvious utility when considering different simulations of the same complex system. One can take different points of view in defining a transformation of SDS. One approach is to require that a transformation is compatible with the defining structural elements of an SDS, that is, with the dependency graph, the local update functions, and the update schedule. If this is done properly, then one should expect to be able to prove that the resulting transformation induces a transformation at the level of phase spaces. That is, transformations between SDS should preserve the local and global dynamic behavior. This implies that transformations between SDS lead to transformations between the associated global update functions. Since the point of view of SDS is that global dynamics emerges from system properties that are defined locally, the notion of SDS transformation should focus on the local structure. This is the point of view taken in [48]. The definition given there is rather technical and the details are beyond the scope of this article. The basic idea is as follows. Let Φπ = fπ(n) ◦ · · · ◦ fπ(1) and Φσ = gσ(m) ◦ · · · ◦ gσ(1) with the dependency graphs Yπ and Yγ , respectively. A transformation F : Φπ −→ Φσ is determined by a graph mapping ϕ : Yπ −→ Yγ (reverse direction); a family of maps kφ(v) −→ kv , v ∈ Yπ ; an order preserving map σ 7−→ π of update schedules. These maps are required to satisfy the property that they “locally” assemble to a coherent trans- formation. Using this definition of transformation, it is shown [48, Theorem 2.6] that the class of SDS forms a category. One of the requirements, for instance, is that the composition of two transformations is again a transformation. Furthermore, it is shown [48, Theorem 3.2] that a trans- formation of SDS induces a map of directed graphs on the phase spaces of the two systems. That is, a transformation of the local structural elements of SDS induces a transformation of global dynamics. One of the results proven in [48] is that every SDS can be decomposed uniquely into a direct product (in the categorical sense) of indecomposable SDS. Another possible point of view is that a transformation F : (Φπ : kn → kn ) −→ (Φγ : km → km ) is a function F : kn −→ km such that F ◦ Φπ = Φγ ◦ F , without requiring specific structural properties. This is the approach in [74]. This definition also results in a category, and a collection of mathematical results. Whatever definition chosen, much work remains to be done in studying these categories and their properties. A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 19 7. Future directions Agent-based computer simulation is an important method for modeling many complex systems, whose global dynamics emerges from the interaction of many local entities. Sometimes this is the only feasible approach, especially when available information is not enough to construct global dynamic models. The size of many realistic systems, however, leads to computer models that are themselves highly complex, even if they are constructed from simple software entities. As a result it becomes challenging to carry out verification, validation, and analysis of the models, since these consist in essence of complex computer programs. This chapter argues that the appropriate approach is to provide a formal mathematical foundation by introducing a class of mathematical ob jects to which one can map agent-based simulations. These ob jects should capture the key features of an agent-based simulation and should be mathematically rich enough to allow the derivation of general results and techniques. The mathematical setting of dynamical systems is a natural choice for this purpose. The class of finite dynamical systems over a state set X which carries the structure of a finite field satisfies all these criteria. Parallel, sequential, and stochastic versions of these are rich enough to serve as the mathematical basis for models of a broad range of complex systems. While finite dynamical systems have been studied extensively from an experimental point of view, their math- ematical theory should be considered to be in its infancy, providing a fruitful area of research at the interface of mathematics, computer science, and complex systems theory. 8. Bibliography Primary Litreture [3] [6] [4] [1] R. L. Bagrodia, Paral lel languages for discrete-event simulation models, IEEE Computa- tional Science and Engineering, 5 (1998), pp. 27–38. [2] C. L. Barrett, H. B. Hunt III, M. V. Marathe, S. S. Ravi, D. J. Rosenkrantz, and R. E. Stearns, On some special classes of sequential dynamical systems., Annals of Combinatorics, 7 (2003), pp. 381–408. , Reachability problems for sequential dynamical systems with threshold functions., Theor. Comput. Sci., 295 (2003), pp. 41–64. , Complexity of reachability problems for finite discrete sequential dynamical systems., J. Computer and System Sciences, 72 (2006), pp. 1317–1345. [5] C. L. Barrett, H. B. Hunt III, M. V. Marathe, S. S. Ravi, D. J. Rosenkrantz, R. E. Stearns, and M. Thakur, Computational aspects of analyzing social network dynamics., in Proc. International Joint Conference on Artificial Intellligence (IJCAI 2007), 2007, pp. 2268– 2273. , Predecessor existence problems for finite discrete dynamical systems, 2007. To appear in Theoretical Computer Science. [7] C. L. Barrett, H. B. Hunt III, M. V. Marathe, S. S. Ravi, D. J. Rosenkrantz, R. E. Stearns, and P. Tosic, Garden of eden and fixed point configurations in sequential dynamical systems, in Proc. International Conference on Combinatorics, Computation and Geometry (DM-CCG’2001), 2001, pp. 95–110. [8] C. L. Barrett, M. V. Marathe, J. P. Smith, and S. S. Ravi, A mobility and traffic generation framework for modeling and simulating ad hoc communication networks, in SAC’02 Madrid, Spain, ACM, 2002, pp. 122–126. [9] C. L. Barrett, H. S. Mortveit, and C. M. Reidys, Elements of a theory of simulation II: Sequential dynamical systems, Appl. Math. and Comput., 107 (2000), pp. 121–136. , Elements of a theory of simulation III: Equivalence of SDS, Appl. Math. and Comput., 122 (2001), pp. 325–340. [10] 20 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI [11] , Elements of a theory of computer simulation. IV. sequential dynamical systems : fixed points, invertibility and equivalence, Appl. Math. Comput., 134 (2003), pp. 153–171. [12] C. L. Barrett and C. M. Reidys, Elements of a theory of simulation I: Sequential CA over random graphs, Appl. Math. and Comput., 98 (1999), pp. 241–259. [13] R. Bartlett and M. Garzon, Monomial cel lular automata, Complex Systems, 7 (1993), pp. 367–388. [14] M. Bernaschi and F. Castiglione, Selection of escape mutants from immune recognition during hiv infection, Immunology and Cell Biology, 80 (2002), pp. 307–313. [15] M. Bernaschi, S. Succi, and F. Castiglione, Large-scale cel lular automata simulations of the immune system response, Phys. Rev. E, 61 (2000). [16] G. Booch, J. Rumbaugh, and I. Jacobson, Unified Modeling Language User Guide, Addison-Wesley Professional, 2nd ed., 2005. [17] D. Brand and P. Zafiropulo, On communicating finite-state machines, J. ACM, 30 (1983), pp. 323–342. [18] P. Cartier and D. Foata, Problemes combinatoires de commutation et re´arrangements, vol. 85 of Lecture Notes in Mathematics, Springer Verlag, 1969. [19] F. Castiglione and Z. Agur, Analyzing hypersensitivity to chemotherapy in a cel lular automata model of the immune system, in Cancer modeling and simulation, L. Preziosi, ed., Chapman and Hall/CRC Press, London, 2003. [20] F. Castiglione, M. Bernaschi, and S. Succi, Simulating the Immune Response on a Distributed Paral lel Computer, International Journal of Modern Physics C, 8 (1997), pp. 527– 545. [21] F. Castiglione, K. Duca, A. Jarrah, R. Laubenbacher, D. Hochberg, and D. Thorley-Lawson, Simulating Epstein-Barr virus infection with C-ImmSim, Bioinfor- matics, 23 (2007), pp. 1371–1377. [22] F. Celada and P. Seiden, A computer model of cel lular interactions in the immune syste, Immunology Today, 13 (1992), pp. 56–62. , A model for simulating cognate recognition and response in the immune system, J Theor Biol, 158 (1992), pp. 235–270. , Affinity maturation and hypermutation in a simulation of the humoral immune response, Eur J Immunol, 26 (1996), pp. 1350–1358. [25] P. P. Chaudhuri, Additive Cel lular Automata. Theory and Applications, vol. 1, IEEE Com- puter Society Press, 1997. [26] O. Col´on-Reyes, A. Jarrah, R. Laubenbacher, and B. Sturmfels, Monomial dynam- ical systems over finite fields, Complex Systems, 16 (2006), pp. 333–342. [27] O. Col´on-Reyes, R. Laubenbacher, and B. Pareigis, Boolean monomial dynamical systems, Annals of Combinatorics, 8 (2004), pp. 425–439. [28] D. Dawson, Synchronous and asynchronous reversible markov systems, Canad. Math. Bull., 17 (1974), pp. 633–649. [29] W. Ebeling and F. Schweitzer, Swarms of particle agents with harmonic interactions, Theory in Biosciences, 120-3/4 (2001), pp. 207–224. [30] B. Elspas, The theory of autonomous linear sequential networks, IRE Transaction on Circuit Theory, (1959), pp. 45–60. [31] J. Farmer, N. Packard, and A. Perelson, The immune system, adaptation, and machine learning, Phys. D, 2 (1986), pp. 187–204. [32] U. Frish, B. Hasslacher, and P. Y., Lattice-gas automata for the Navier-Stokes equations, Physical Review Letters, 56 (1986), pp. 1505–1508. [33] H. Fuk´s, Probabilistic cel lular automata with conserved quantities, Nonlinearity, 17 (2004), pp. 159–173. [23] [24] A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 21 [34] L. D. Garcia, A. S. Jarrah, and R. Laubenbacher, Sequential dynamical systems over words, Appl. Math. Comput., 174 (2006), pp. 500–510. [35] M. Gardner, The fantastic combinations of John Conway’s new solitaire game ”life”, Scien- tific American, 223 (1970), pp. 120–123. [36] M. Gouda and C. Chang, Proving liveness for networks of communicating finite-state ma- chines, ACM Transactions on Programming Languages and Systems, 8 (1986), pp. 154–182. [37] Y. Guo, W. Gong, and D. Towsley, Time-stepped hybrid simulation (tshs) for large scale networks, in INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 2, 2000, pp. 441–450. [38] A. Gupta and V. Katiyar, Analyses of shock waves and jams in traffic flow, Journal of Physics A, 38 (2005), pp. 4069–4083. [39] A. A. Hansson, H. S. Mortveit, and C. M. Reidys, On asynchronous cel lular automata, Advances in Complex Systems, (2005). submitted. [40] G. Hedlund, Endomorphisms and automorphisms of the shift dynamical system, Math. Syst. Theory, 3 (1969), pp. 320– 375. [41] A. Hern´andez-Toledo, Linear finite dynamical systems, Communications in Algebra, 33 (2005), pp. 2977–2989. [42] J. Hopfield, Neural networks and physical systems with emergent col lective computational properties, Proc. National Academy of Sciences of the USA, 79 (1982), pp. 2554–2588. [43] A. Ilichinsky, Cel lular Automata: A Discrete Universe, World Scientific Publishing Com- pany, July 2001. [44] A. Jarrah, R. Laubenbacher, M. Stillman, and P. Vera-Licona, An efficient algo- rithm for the phase space structure of linear dynamical systems over finite fields. Submitted, 2007. [45] D. R. Jefferson, Virtual time, ACM Transactions on Programming Languages and Systems, 7 (1985), pp. 404–425. [46] J. Kari, Theory of cel lular automata: A survey, Theoretical Computer Science, 334 (2005), pp. 3–33. [47] B. L. Keyfitz, Hold that light! modeling of traffic flow by differential equations., Stud. Math. Libr., 26 (2004), pp. 127–153. [48] R. Laubenbacher and B. Pareigis, Decomposition and simulation of sequential dynamical systems, Adv. Appl. Math., 30 (2003), pp. 655–678. [49] R. Lidl and H. Niederreiter, Finite Fields, Cambridge University Press, New York, 1997. [50] T. M. Liggett, Interacting particle systems, Classics in Mathematics, Springer-Verlag, Berlin, 2005. Reprint of the 1985 original. [51] D. A. Lind, Applications of ergodic theory and sofic systems to cel lular automata, Physica D, 10D (1984), pp. 36–44. [52] K. Lindgren, C. Moore, and M. Nordahl, Complexity of two-dimensional patterns, Jour- nal of Statistical Physics, 91 (1998), pp. 909–951. [53] S. Mac Lane, Category Theory for the Working Mathematician, no. 5 in GTM, Springer Verlag, 2nd ed., 1998. [54] M. W. Macy, J. A. Kitts, and A. Flache, Dynamic social network nodeling and analysis, The National Academies Press, 2003, ch. Polarization in Dynamic Networks: A Hopfield Model of Emergent Structure, pp. 162–173. [55] O. Martin, A. Odlyzko, and S. Wolfram, Algebraic properties of cel lular automata, Commun. Math. Phys., 93 (1984), pp. 219–258. [56] D. Milligan and M. Wilson, The behavior of affine boolean sequential networks, Connection Science, 5 (1993), pp. 153–167. [57] N. Minar, R. Burkhart, C. Langton, and A. Manor, The swarm simulation sys- tem: A toolkit for building multi-agent simulations, Santa Fe Institute preprint series, (1996). 22 R. LAUBENBACHER, A. S. JARRAH, H. MORTVEIT, AND S. S. RAVI [62] [74] http://www.santafe.edu/sfi/publications/96wplist.html. [58] J. Misra, Distributed discrete-event simulation, ACM Computing Surveys, 18 (1986), pp. 39– 65. [59] T. Moncion, G. Hutzler, and P. Amar, Verification of biochemical agent-based mod- els using petri nets, in International Symposium on Agent Based Modeling and Simulation, ABModSim’06, T. Robert, ed., Austrian Society for Cybernetics Studies, 2006, pp. 695–700. [60] D. Morpurgo, R. Serentha, P. Seiden, and F. Celada, Model ling thymic functions in a cel lular automaton, International Immunology, 7 (1995), pp. 505–516. [61] H. S. Mortveit and C. M. Reidys, Discrete, sequential dynamical systems, Discrete Math- ematics, 226 (2001), pp. 281–295. , Reduction of discrete dynamical systems over graphs, Advances in Complex Systems, 7 (2004), pp. 1–20. [63] K. Nagel, M. Rickert, and C. L. Barrett, Large-scale traffic simulation, Lecture notes in computer science, 1215 (1997), pp. 380–402. [64] K. Nagel and M. Schreckenberg, A cel lular automaton model for freeway traffic, Journal de physique I, 2 (1992), pp. 2221–2229. [65] K. Nagel, M. Schreckenberg, A. Schadschneider, and N. Ito, Discrete stochastic models for traffic flow, Physical Review E, 51 (1995), pp. 2939–2949. [66] K. Nagel and P. Wagner, Traffic Flow: Approaches to Model ling and Control, John Wiley & Sons, 2006. [67] R. E. Nance, A history of discrete event simulation programming languages, ACM SIGPLAN Notices, 28 (1993), pp. 149–175. [68] M. J. North, N. T. Collier, and J. R. Vos, Experiences creating three implementations of the repast agent modeling toolkit., ACM Trans. Modeling and Computer Simulation, 16 (2006), pp. 1–25. [69] P. Orponen, Computational complexity of neural networks: A survey, Nordic Journal of Computing, 1 (1994), pp. 94–110. , The computational power of discrete hopfield networks with hidden units, Neural Com- putation, 8 (1996), pp. 403–415. [71] J. K. Park, K. Steiglitz, and W. P. Thruston, Soliton-like behavior in automata, Phys- ica D, 19D (1986), pp. 423–432. [72] C. Reidys, Acyclic Orientations of Random Graphs, Advances in Applied Mathematics, 21 (1998), pp. 181–192. [73] C. M. Reidys, On acyclic orientations and sequential dynamical systems, Adv. Appl. Math., 27 (2001), pp. 790–804. , On Certain Morphisms of Sequential Dynamical Systems, Discrete Mathematics, 296 (2005), pp. 245–257. , Sequential dynamical systems over words, Annals of Combinatorics, 10 (2006). [75] [76] M. Rickert, K. Nagel, M. Schreckenberg, and A. Latour, Two lane traffic simula- tions using cel lular automata, Physica A, 231 (1996), pp. 534–550. [77] D. H. Rothman, Cel lular-automaton fluids: A model for flow in porous media, Geophysics, 53 (1988), pp. 509–518. [78] S. Russell and P. Norwig, Artificial Intel ligence: A Modern Approach, Prentice-Hall, 2003. [79] B. Schonfisch and A. de Roos, Synchronous and asynchronous updating in cel lular au- tomata, BioSystems, 51 (1999), pp. 123–143. [80] I. Shmulevich, E. R. Dougherty, S. Kim, and W. Zhang, Probabilistic boolean net- works: A rule-based uncertainty model for gene regulatory networks, Bioinformatics, 18 (2002), pp. 261–274. [81] I. Shmulevich, E. R. Dougherty, and W. Zhang, From boolean to probabilistic boolean networks as models of genetic regulatory networks, Proceedings of the IEEE, 90 (2002), [70] A MATHEMATICAL FORMALISM FOR AGENT-BASED MODELING 23 pp. 1778–1792. [82] M. Sipser, Introduction to the Theory of Computation, PWS Publishing Co., 1997. [83] L. Vasershtein, Markov processes over denumerable products of spaces describing large sys- tem of automata, Problemy Peredachi Informatsii, 5 (1969), pp. 64–72. [84] J. von Neumann, Theory of Self-Reproducing Automata, University of Illinois Press, 1966. Edited and completed by Arthur W. Burks. [85] G. Whitham, Linear and Nonlinear Waves, Pure and Applied Mathematics: A Wiley- Interscience Series of Texts, Monographs and Tracts, Wiley-Interscience, reprint edition ed., 1999. [86] S. Wolfram, Statistical mechanics of cel lular automata, Rev. Mod. Phys., 55 (1983), pp. 601– 644. [87] [88] , Theory and Applications of Cel lular Automata, vol. 1 of Advanced Series on Complex Systems, World Scientific Publishing Company, 1986. , A New Kind of Science, Wolfram Media, 2002. Books and Reviews [1] Hopcroft, J. E., R. Motwani, and J. D. Ullman (2000). Automata Theory, Languages and Computation. Reading: Addison Wesley. [2] Kozen, D. C. (1997). Automata and Computability. New York: Springer-Verlag. [3] Wooldridge, M. (2002). Introduction to Multiagent Systems. Chichester, UK: John Wiley & Sons. Virginia Bioinformatics InstituteVirginia Polytechnic Institute and State University E-mail address, Reinhard Laubenbacher: [email protected] E-mail address, Abdul S. Jarrah: [email protected] E-mail address, Henning Mortveit: [email protected] Department of Computer Science, University at Albany–State University of New York E-mail address, S. S. Ravi: [email protected]
1909.02475	2	1909	2019-09-11T14:12:40	Lower bound performances for average consensus in open multi-agent systems (extended version)	[ "cs.MA", "math.DS" ]	We derive fundamental limitations on the performances of intrinsic averaging algorithms in open multi-agent systems, which are systems subject to random arrivals and departures of agents. Each agent holds a value, and their goal is to estimate the average of the values of the agents presently in the system. We provide a lower bound on the expected Mean Square Error for any estimation algorithm, assuming that the number of agents remains constant and that communications are random and pairwise. Our derivation is based on the expected error obtained with an optimal algorithm under conditions more favorable than those the actual problem allows, and relies on an analysis of the constraints on the information spreading mechanisms in the system, and relaxations of these.	cs.MA	cs	Lower bound performances for average consensus in open multi-agent systems (extended version) Charles Monnoyer de Galland and Julien M. Hendrickx 9 1 0 2 p e S 1 1 ] A M . s c [ 2 v 5 7 4 2 0 . 9 0 9 1 : v i X r a Abstract -- We derive fundamental limitations on the perfor- mances of intrinsic averaging algorithms in open multi-agent systems, which are systems subject to random arrivals and departures of agents. Each agent holds a value, and their goal is to estimate the average of the values of the agents presently in the system. We provide a lower bound on the expected Mean Square Error for any estimation algorithm, assuming that the number of agents remains constant and that communications are random and pairwise. Our derivation is based on the expected error obtained with an optimal algorithm under conditions more favorable than those the actual problem allows, and relies on an analysis of the constraints on the information spreading mechanisms in the system, and relaxations of these. I. INTRODUCTION Multi-agent systems show great benefits for modeling and solving problems in various domains including sensor net- works [1], [2], vehicle coordination [3], or social phenomena [4], [5]. Among the most cited properties of multi-agent systems are their flexibility, their scalability, or their robust- ness. Yet, most results around multi-agent systems stand for asymptotic properties under the convenient assumption that their composition remains unchanged. The increasing size of the systems challenges this assumption as it implies slower processes and higher probabilities of arrivals and departures, making those non-negligible. It is also challenged by the chaotic nature of some systems, where communications can be difficult, or happen at a time-scale highly comparable to that of the arrivals and departures: vehicles can for instance share a stretch of road before heading to different destinations in collaborative multi-vehicles systems. All those reasons bring up the emergence of the study of Open Multi-Agent Systems. Results about closed systems do not easily extend to open ones: repeated arrivals and departures imply important differences in the design and analysis of such systems, and result in several challenges, see e.g. [6], [7]. First, with the frequent arrivals and departures of agents, the size of an open system changes with time, making the analysis of its state challenging. Moreover, incessant perturbations impact the state, but also in some cases the objective pursued by the agents of the system: algorithms then have to adapt to that variable objective, making their design challenging, and the usual convergence cannot be achieved anymore. ICTEAM institute, UCLouvain (Belgium). J.H. is also with the CISE, Boston University (USA). This work was supported by "Communaut´e franc¸aise de Belgique - Actions de Recherche Concert´ees". C.M. is a FRIA fellow (F.R.S.-FNRS), and J.H. holds a WBI.World excellence fellowship. Email adresses: [email protected], [email protected]. A. State of the art There is little theoretical analysis of open multi-agent systems, as most multi-agent results rely on the assumption that systems are closed. However, some results considered arrivals and departures, such as simulation-based analyses of social phenomena performed in [8]. Also, it has been shown in [6] that systems subject to gossip interactions in open features can be analyzed through size-independent descriptors, whose evolution described by a dynamical sys- tem is shown to asymptotically converge to some steady states. Moreover, algorithm design has been explored for MAX-consensus problems with arrivals and departures in [7] through additional variables, and where performance was measured by the probability for the estimate to eventually converge if the system closes. Similarly, THOMAS architec- ture was designed to maintain connectivity into P2P networks [9]. Openness was also considered in applications such as VTL for autonomous cars to deal with cross-sections [10]. Nevertheless, up to now, efficiently studying performances of algorithms and analyzing open systems remains a chal- lenge in general. As agents cannot instantaneously react to perturbations, obtaining an exact result is often out of range. Hence, the goal should be to remain near a time- varying objective, and usual convergence is no more relevant to be studied as it cannot be achieved. A step towards understanding open systems is the derivation of fundamental performance limitations: lower bounds on the performances that are possible to achieve. Through those limitations, one can obtain some quality criterion for algorithm design in open systems, but also get a better understanding of the possible bottlenecks that could arise. B. Contribution We establish fundamental performance limitations in open multi-agent systems for intrinsic averaging consensus prob- lems (i.e. estimating the average of all intern values owned by agents present in the system at that time), where information is exchanged between agents through random pairwise com- munications. In this analysis, we focus on systems of fixed size: we assume that each departure of an agent is instanta- neously followed by an arrival, so that only replacements occur, and the system size remains constant, see Section II for a formal definition. The approach used for deriving the fundamental limitations is detailed in Section III, and consists in evaluating the performances of an algorithm that is provably optimal under favorable settings where agents have access to more resources than decentralized problems typically allow. These performances depend on some compli- cated distribution related to information propagation. Bounds on this distribution were obtained by considering relaxed conditions for the information exchange, namely the Ping model in Section IV and the Infection model in Section V, to properly define the performance limitations. Since consensus problems are commonly a building block for various multi-agent complex applications (such as de- centralized optimization [11] or several control applications [3]), we expect the techniques we use for deriving those limitations to be extendable to more advanced tasks on open multi-agent systems. II. PROBLEM STATEMENT A. System description 2 We consider a set of N agents, labelled from 1 to N. Each agent i ∈ {1 . . .N} holds an i.i.d. intrinsic value xi(t) ∈ R ran- domly selected from a distribution which we assume without loss of generality to be zero mean and of variance Ex2 i = σ 2. This value remains constant except at a replacement, which we model as the complete erasure of the agent's memory and the attribution of a new value xi(t) drawn from the same distribution. Those replacements happen according to a Poisson clock defined by an individual replacement rate λr, so that on average Nλr replacements happen in the whole system per unit of time. Agents interact via pairwise communications: every two agents interact with each other at random times determined by a Poisson clock with rate λc. There are thus on average N(N−1) λc communications taking place in the system per unit of time. Agents are deterministic and have unlimited memory and computation power and there is no restriction on the size or nature of the messages they can exchange. They have a unique identifier that they can use and they have access to a common universal time. The way we model replacements as a memory erasure also means agents know the correspondence between the agents having left and those having replaced them since they share the same label. We assume moreover that they know the number of agents the parameters λr, λc, and the distribution of the xi N, (although we will see that the results would be the same if they knew only the expected value of xi, assumed here to be 0). Agents have thus access to significantly more information than in many works on multi-agent systems, but our lower bounds on the algorithm performance will of course also apply to these more usual situations, since our setting allows implementing any algorithm that could be implemented under more restrictive conditions. B. Objective In intrinsic averaging, agents try to estimate the average of the intern values of the agents present in the system at that time, denoted ¯x(t). For that purpose, every agent i maintains its own estimate yi(t) of that average based on its knowledge about the system, and updates it in continuous time. Under ideal conditions, one would expect that estimate to become yi(t) = ¯x(t) = 1 j=1 x j(t). N ∑N This is not achievable in open systems because of the variable objective and the delays to transmit information in the system, leading to the absence of usual convergence. We thus need a quantitative measure of the performance of an algorithm solving this problem in open systems, and choose the classical Mean Square Error (MSE) of the estimation of all the agents at time t, presented in equation (1). j=1( ¯x(t)− y j(t))2 C(t) := 1 (1) In this analysis, we focus on the steady state-error: we assume the system has been running since −∞, so that the effect of initial conditions has disappeared, and E [C(t)] can be considered independent of the time (one can verify that our processes remain well defined). We thus derive fundamental performance limitations for all algorithms that can be implemented in our setting as a (time-invariant) lower bound on (2). E [C(t)] = E(cid:2)( ¯x(t)− yi(t))2(cid:3) N ∑N (2) C. Preliminary notions Before stating our first result, we need to introduce certain notions related to the information available to an agent when computing its estimate, and to the age of its information. We first formalize in the next definition all the information about events and values x j to which an agent i could possibly access, and thus influence its estimate yi(t). Definition 1: At its arrival at time t, an agent i whose value is set to xi owns a knowledge set ωi(t) such that ωi(t+) = {i,xi,t}. If agents i and j interact at time t, then ωi(t+) = ω j(t+) = ωi(t−) ∪ ω j(t−) ∪ {i - j; t,xi(t),x j(t)}. The last expression of the union defining an interaction denotes that i, j have interacted at time t and confirms their values at that time.1 In other words, the knowledge sets after interaction consist of the union of the knowledge sets, to which is added the information about the interaction and confirmation of the values. Standard results in distributed computation show that any estimate yi(t) that an agent can compute in our settings can actually be computed based only on ωi(t) and the time t. Observe that ωi(t) may contain various values x j(t(cid:48)) for different t(cid:48) < t. However, we will see later that only the most recent known values about the other agents are needed to build the estimate yi(t). Definition 2: The most recent value known by the agent is denoted x(i) j at time t i about j (t) := max{s : x j(s) ∈ ωi(t)} is the time at which that most t(i) recent value was first obtained. We denote the age of that information by T (i) j (t) := t − t(i) j (t) = x j j (t). (cid:0)t(i) j (t)(cid:1), where By convention, if no value x j lies in ωi(t), we set x j to 0 and the corresponding age to +∞. Moreover, the estimate of an agent about itself is always correct, and the corresponding age of the information is 0. In steady state, and due to the symmetry between the j the distribution of T (i) j (t) is independent of i, agents, 1This confirmation is included for simplicity, but is actually redundant, as the values at that time could be obtained from ωi(t−) and ω j(t−). and t. Hence, all agents share a common cdf and pdf for that variable, respectively denoted F(s) and f (s). Moreover, with a small abuse of language, we say that another pdf f ∗(s) bounds f (s) when the corresponding cdf F∗(s) satisfies F∗(s) ≥ F(s) ∀s, which is always satisfied for a random variable T∗ ≤st T , in the usual stochastic order [12]. III. BOUND IN TERMS OF THE INFORMATION SPREADING MECHANISM We present in Theorem 1 a general lower bound on (2) depending on the way information spreads in the system. To properly define a bound, we will thus need to find relaxations of that information spreading to instantiate the expression. Theorem 1: The performances of any algorithm in the setting defined in Section II are bounded from below as E [C(t)] ≥ N − 1 N2 σ 2dt where f ∗(t) bounds f (t) as defined in Section II-C. 0 f ∗(t) 1− e−2λrt(cid:17) (cid:16) (cid:90) ∞ (3) The bound derived in the above theorem is obtained by studying an optimal algorithm in the setting of Section II (Section III-A). It relies on the decomposition of the contri- butions of all the agents for the estimation of the average to obtain the MSE for some knowledge (Sections III-B to III- D), and then on the derivation of a steady state through the analysis of the way information spreads in the system (Sections III-E and III-F). A. Optimal algorithm definition We first build an algorithm that is optimal for solving intrinsic averaging based on a knowledge set. Proposition 1: In the setting described in Section II, the d j=1 following estimate is optimal in the sense of criterion (2). N ∑N E [x j(t)ωi(t)] yi(t) = E [ ¯x(t)ωi(t)] = 1 rithm (4) is optimal since it minimizes the MSE: (4) Proof: Conditional to the knowledge set ωi(t), algo- dyi(t) Hence, since any other algorithm only depends on ωi(t) and t, the result of any of them y∗ E(cid:2)( ¯x(t)− yi(t))2ωi(t)(cid:3) = 2yi(t)− 2E [ ¯x(t)ωi(t)] = 0. i (t))2ωi(t)(cid:3) ≥ E(cid:2)( ¯x(t)− yi(t))2ωi(t)(cid:3) . E(cid:2)( ¯x(t)− y∗ i (t))2ωi(t)(cid:3)(cid:3) ≥ E(cid:2)E(cid:2)( ¯x(t)− yi(t))2ωi(t)(cid:3)(cid:3) . E(cid:2)E(cid:2)( ¯x(t)− y∗ Considering the expected value on all ωi(t), it follows: i (t) is such that (cid:105)(cid:105) Since the algorithm (4) is optimal in the sense of criterion (2), its performance E [C(t)] = E(cid:104)E(cid:104) ( ¯x(t)− E [ ¯x(t)ωi(t)])2ωi(t) (5) provides a lower bound on the performance of all algorithms that can be deployed in our settings. B. Individual contribution We show in the following proposition that expression (5) conveniently reduces to the analysis of a single agent. Proposition 2: The criterion (2) for the algorithm (4) conditional to the knowledge set ωi(t) reduces to ( ¯x(t)− yi(t))2ωi(t) = 1 N2 ∑N j=1 (x j(t)− x j(t))2ωi(t) (6) E(cid:104) E(cid:104) (cid:105) (cid:105) where x j(t) := E [x j(t)ωi(t)] (we lighten the notation by dropping the dependence of x(i) j (t) on i). Proof: With the optimal algorithm (4), one has yi(t) = 1 E [x j(t)ωi(t)] (cid:20)(cid:16) It follows that the error given ωi(t) is written N ∑N (cid:105) j=1 ( ¯x(t)− yi(t))2ωi(t) = E(cid:104) (cid:0)x j(t)− x j(t)(cid:1)(cid:17)2(cid:12)(cid:12)(cid:12)(cid:12)ωi(t) (cid:21) . E 1 N2 ∑N j=1 The absence of correlation between the agents values xi(t) finally allows to nullify the crossed-product terms of the squared sum, to obtain the final expression. C. Single agent estimate We can then explicitly write an expression for the estimate of a single agent x j(t). Proposition 3: There holds, j x j(t − T (i) j ). x j(t) = E [x j(t)ωi(t)] = e−λrT (i) (7) Hence, the estimate x j(t) only depends on the most recent information about the agent j (note that the time-dependence of T (i) j (t) is removed to lighten the notations). Proof: The most recent information we know about j is given by Definition 2, and we have no information about whether it was replaced since then. Hence, denoting R the event that j has been replaced, and ¯R that it has not, x j(t) = E [x j(t) ¯R]· P( ¯R) + E [x j(t)R]· (1− P( ¯R)) . By definition, E [x j(t)R] = 0. Then, from Poisson properties: (cid:16) (cid:17) x j(t) = 0 + x(i) j (t) e −λr t−t(i) j (t) and the conclusion follows from the definition of the age of the most recent information. The result above also stands when an agent estimates its own value or that of an agent for which it has no information, leading respectively to x(i) D. Individual error i (t) = xi(t) and x(i) j (t) = 0. For concision matters, we denote j (t) := (x j(t)− x j(t))2 C(i) (8) We can then write the MSE of the estimation of a single agent by injecting (7) into (8). E(cid:104) Proposition 4: The MSE when estimating a single agent for algorithm (4) conditional to the knowledge set ωi(t) is (9) and is thus entirely characterized by the age of the most recent information about it. (x j(t)− x j(t))2ωi(t) 1− e−2λrT (i) (cid:17) (cid:16) (cid:105) σ 2 = j Proof: Denoting R the event of at least one replacement and by ¯R the event of no replace- of the agent j during T (i) j ment, we can develop (8) with a case-by-case analysis: E(cid:104) (cid:105) j (t)ωi(t) C(i) = E(cid:104) j (t) ¯R C(i) We develop each term to obtain the final result: j (t)R C(i) (cid:105)· P(R) +E(cid:104) (cid:16) (cid:105) (cid:105) (cid:16) j (t)R C(i) j (t) ¯R C(i) = = E(cid:104) E(cid:104) j 1 + e−2λrT (i) 1− e−λrT (i) j (cid:105)· P( ¯R). (cid:17) (cid:17)2 σ 2 σ 2 P(R) = 1− e−λrT (i) P( ¯R) = e−λrT (i) j j This last result allows to write the following: 1− e−2λrT (i) j N2 ∑N j=1 E [C(t)ωi(t)] = 1 E. Global expected value (cid:16) (cid:17) σ 2. (10) 0 (cid:90) ∞ 1− e−2λrt(cid:17) (cid:16) We have developed an expression for the error in terms of the age of the most recent information about the other agents in (10). We can now obtain an expression for criterion (2) by computing the expected value of that result. N − 1 N2 Proposition 5: The MSE defined in (2) is given by E [ ¯x(t)− yi(t)]2 = σ 2 dt, (11) where we remind that Exi = 0 and Ex2 i = σ 2. It is thus entirely characterized by the distribution of the age of an information, and knowing f (t) leads to a proper bound. f (t) Proof: By definition, the global MSE is Using the result (10), it becomes E E(cid:2)( ¯x(t)− yi(t))2(cid:3) = E(cid:2)E(cid:2)( ¯x(t)− yi(t))2ωi(t)(cid:3)(cid:3) . (cid:19)(cid:21) (cid:20)(cid:18) E(cid:2)( ¯x(t)− yi(t))2(cid:3) = 1 N2 ∑N (cid:19)(cid:21) 1− e−2λrt(cid:17) (cid:16) (cid:90) ∞ follow f (t) from Section II-C, Finally, since all T (i) j 1− e−2λrT (i) 1− e−2λrT (i) (cid:20)(cid:18) f (t) dt, j=1 E σ 2. = j j 0 and the conclusion is then direct, where we remind that T (i) i = 0 by definition (cfr Section II-C). F. Relaxation of the communication process The previous result is actually that of Theorem 1 where the pdf is exactly f (s). Since this pdf reveals to be hard to f ∗(s) compute, we show in the next proposition that a pdf bounding f (s) as defined in Section II-C leads to a proper lower bound on (11), which concludes the development of Theorem 1. In the next then present two possible relaxations of the information spreading mechanism that will both lead to an appropriate f ∗(s) to instantiate the bound. 0 F(cid:48)(t) · err(t)dt where F(t) is a cdf, and where err(t) is a positive non- decreasing function, then for any other cdf F∗(t) ≥ F(t) ∀t, (12) Proposition 6: Given some value E =(cid:82) ∞ two Sections, we will (cid:90) ∞ E∗ = Proof: By defining F∗ = F + ∆ with ∆ ≥ 0, one has (F∗)(cid:48)(t)· err(t)dt ≤ E (cid:90) ∞ 0 E∗ = E − ∆· err(cid:48)(t)dt thanks to cdf properties. Using then the positivity of ∆ and err(cid:48)(t), the conclusion is direct. 0 IV. STRONG ASSUMPTION: PING MODEL A. Assumption description The first relaxation we consider relies on a strong simpli- fication of the communication mechanism. We assume that each time it communicates, an agent acquires all the infor- mation about all the agents presently in the system, and never forgets it, even at replacements. We refer to this assumption as the Ping model, in reference to the ping software used j (t) = x j (cid:0)t(i) j (t)(cid:1) where t(i) to test the reachability of machines in a network. The most recent value known by the agent i about j at time t is then given by x(i) j (t) is the time of the very last communication in which agent i was involved, and the age of that information T (i) j (t) is the time spent since then. Proposition 7: In steady state, ∀i (cid:54)= j, the age of the most j,Ping(t) follows (13) recent information about j held by i, noted T (i) f Ping(s) = (N − 1)λce−(N−1)λcs, which bounds f (s) as defined in Section II-C. j,Ping(t) = t − t(i) Proof: The random variable T (i) j,Ping(t) defines the time spent since the last communication involving the agent i, and thus follows by definition of the communications a Poisson process of rate (N − 1)λc. Its cdf is then given by = 1− e−(N−1)λcs. FPing(s) = P j,Ping(t) ≤ s T (i) (cid:104) (cid:105) The pdf is then obtained with f Ping(s) = dFPing(s) . ds T (i) j,Ping(t) is actually the minimal value that can take T (i) j (t). Hence T (i) j,Ping(t) ≤st T (i) j (t), and f Ping(s) bounds f (s). B. Results Theorem 2: The performances of any algorithm in the setting defined in Section II are bounded from below as (cid:32) (cid:33) E [C(t)] ≥ N − 1 N2 (cid:90) ∞ 1 1 + N−1 2 λc λr σ 2 (14) f Ping(t), the Proof: Applying Theorem 1 with the pdf (N − 1)λce−(N−1)λcs(cid:16) bound is given by performing the integration E [C(t)] ≥ N−1 N2 A few algebraic operations lead to the conclusion. 1− e−2λrs(cid:17) 0 σ 2ds. Fig. 1: Bound derived with the Ping model (14) in terms of the rate ratio between pairwise communications and replacements. Each curve stands for a different number of agents N, with a unit variance σ 2 = 1. Several observations on the evolution of the error in terms of the ratio of the rates λc/λr and N arise from Fig. 1. When communications are rather rare (λc/λr → 0), the error tends to N−1 N2 σ 2, i.e. the error that an agent would obtain only considering itself in the estimation. As the communications become more frequent, the second factor of (14) makes the bound decrease, to ultimately bring it to 0 as λc/λr → +∞, which is the error to which we converge in a close system. Observe that ¯λc = (N − 1)λc is the communication rate for one agent. Hence, the bound is actually separated into two factors: N−1 N2 σ 2 and the second one that only depends on ¯λc/λr which characterizes the expected number of inter- actions involving one agent before it is replaced. V. WEAKER ASSUMPTION: INFECTION MODEL A. Assumption description The random variable T (i) j (t) denotes the age of the most j available to i at time t, i.e. the recent information about time since this information was emitted by j. Since the distribution of T (i) j (t) is independent of the time t, its cdf F(s) corresponds to the probability that an information x j(τ) (with 0 ≤ τ ≤ s) is available to i at time s. This probability is lower than what it would be if the agents were never replaced (as information is erased with replacements). Hence we use this latter replacement-free situation as a relaxation to obtain a bound on the cdf. One can verify that when agents are never replaced, the information propagation follows a simple infection process: j is infected at time 0, and agents get infected as soon as they communicate with an infected agent. The probability of i having that information about j at time s is then equal to the probability of i being infected at time s, denoted by P[T (i) j,In f is the time of its first infection. Based on this, one can prove that the cdf of T (i) j,In f bounds F(s) as defined in Section II-C. j,In f (t) ≤ s] where T (i) Proposition 8: In steady state, the pdf of T (i) f In f (s), bounds f (s) as defined in Section II-C. j,In f (t), noted j (t). Proof: Considering one realization of communications, there exists a sequence of events where T (i) j (t) is such that it did not suffer replacements. Since that sequence always also exists with the Infection model, T (i) B. Result j,In f (t) ≤st T (i) Proposition 9: In steady state, ∀i (cid:54)= j, the age of the most recent information agent i has about agent j with the Infection model, denoted T (i) FIn f (s) = ∑N k−1 N−1 Pk(s) where Pk(s) is defined by an ODE system: Pk(s) = (k− 1)(N − k + 1)λcPk−1(s)− k(N − k)λcPk(s) (16) such that P1(0) = 1 and Pk(cid:54)=1(0) = 0. j,In f (t), follows the cdf Proof: Let I(s) be the set containing the labels of its size, and denote time s, NI(s) the infected agents at Pk(s) = P[NI(s) = k]. Then, one has FIn f (s) = P[i ∈ I(s)] = ∑N k=1 Pk(s)· P[i ∈ I(s)NI(s) = k]. From the symmetry between the agents, and since they (15) k=1 are all interchangeable, one has P[i ∈ I(s)NI(s) = k] = k−1 N−1 . Finally, the ODE system defining Pk(s) is a standard result from continuous time Markov chains theory. Theorem 3: The performances of any algorithm in the (cid:0)1− wT A(2λrI − A)−1e1 (cid:1)σ 2 setting defined in Section II are bounded from below as E [C(t)] ≥ N−1 N2 (17) where A is the bidiagonal matrix defining the ODE system (16) from Proposition 9 i.e., Aii = − i(N − i)λc and Ai+1,i = i(N − i)λc, and wT = (cid:2)0 P(s) = AP(s) with P(s) =(cid:2)P1(s) P2(s) with λr (cid:54)= 0. Proof: Equation (16) defines 1 N−1 a whose solution is given by 2 N−1 . . . 1(cid:3), . . . PN(s)(cid:3)T system , linear P(s) = eAsP(0) with P(0) = e1. It follows from (15) that the cdf and pdf are given by FIn f (s) = wT eAse1 and f In f (s) = wT AeAse1. 0 0 0 ds = dse1. f In f (s) (cid:90) ∞ e(A−2λrI)sds. (cid:16) eAse−2λrs(cid:17) Injecting that result in the integral from Theorem 1, which Since λr (cid:54)= 0, then (A− 2λrI) is invertible and has strictly (cid:90) ∞ (cid:90) ∞ Since A and −2λrI commute, the integral reduces to we will see is well defined, one has ds = 1− wT A (cid:90) ∞ e(A−2λrI)s(cid:105)∞ 1− e−2λrs(cid:17) (cid:16) (cid:16) eAse−2λrIs(cid:17) (cid:90) ∞ e(A−2λrI)sds = (A−2λrI)−1(cid:104) (cid:16) 1− e−2λrs(cid:17) (cid:90) ∞ (cid:32) Corollary 1: The expression below is equivalent to (17). Hence, the integral from Theorem 1 is given by and the conclusion is direct from Theorem 1. ds = 1− wT A(2λrI − A)−1e1, negative eigenvalues, leading to = (2λrI−A)−1. f In f (s) 0 0 0 0 (cid:16) (cid:17)(cid:33) E [C(t)] ≥ N − 1 N2 (cid:32) h(N,L) = N ∑ k=2 k− 1 N − 1 k−1 ∏ j=1 1− 1 1 + N−1 2 h N, λc λr λc λr j(N − j)L 2 + ( j + 1)(N − j− 1)L σ 2 (cid:33) (18) (19) with the Corollary 1 presents an algebraic expression for from Theorem 1, and is proven in Ap- bound (17) that pendix A. Detailed empirical explorations suggest h(N,L) ≤ L 2kL 2+2kL . Using results on bounds in numerical integration, we can then find a compact bound on that expression, and derive the following empirical ex- (cid:32) pression for bounding E [C(t)] from below. 2 ∑N−2 (cid:33)(cid:32) (cid:32) 1 + (N − 2) λc 2 + L (cid:33)(cid:33) k=1 1 + 1 2 log 1 + λc λr λr σ 2 (20) N − 1 N2 1 1 + N−1 2 λc λr C. Discussion Fig. 2 compares the bounds derived with the Ping and Infection models, and the expression (20) based on the Infection bound, in logarithmic scale. The first observation is the improvement of both bounds from the Infection model, which relies on a weaker relaxation, in comparison with the Ping model. That improvement is even more apparent when the rate ratio λc/λr increases, since the impact of the corresponding factor gets more important. In particular, the expression (20) is actually exactly that of the bound from the Ping model multiplied by a logarithmic factor that scales to logN when λc/λr is large. This could relate to the fact that information propagates in one hop in the Ping model, whereas it needs time with the Infection model. Observe also in Fig. 2 that the bound follows two distinct regimes: when λc/λr is very small, it has a very low impact We focused on the analysis of the performances of intrinsic averaging algorithms for open systems of constant size, for which we derived fundamental performance limitations as lower bounds on the Mean Square Error of estimation in steady state. This was done by studying the performances of an algorithm optimal in a context more favorable to the agents than what is usually assumed in multi-agent systems, and relied then on the relaxation of the information spreading mechanism in the system. Two of such relaxations were studied, leading to a conservative but readable bound, and to a tighter but more complex bound. Possible extensions include the study of more complicated problems (e.g. decentralized optimization [11]), more struc- tured constraints on the inter-agent communications, and the consideration of variable-size systems. More fundamentally, our bounds purely rely on limitations on the information propagation, and are valid even if the agents have some strong knowledge on the system and use identifiers or heterogeneous algorithms. Stronger lower bounds can be ob- tained under more restrictive assumptions, e.g. by improving the Infection model from Section V with healing. Yet, it would be interesting to investigate what factors significantly impact the bound: in particular, an anonymous algorithm was already shown in Section V-C to exhibit performances not too different from our bound, questioning the impact of anonymity of the agents. REFERENCES [1] J. Predd, S. Kulkarni, and H. V. Poor, "Distributed learning in wireless sensor networks," Signal Processing Magazine, IEEE, vol. 23, pp. 56 -- 69, 07 2006. [2] R. Olfati-Saber and N. F. Sandell, "Distributed tracking in sensor networks with limited sensing range," pp. 3157 -- 3162, 07 2008. [3] Y. Cao, W. Yu, W. Ren, and G. Chen, "An overview of recent progress in the study of distributed multi-agent coordination," IEEE Transactions on Industrial Informatics, vol. 9, 07 2012. [4] V. Blondel, J. M. Hendrickx, and J. Tsitsiklis, "Continuous-time average-preserving opinion dynamics with opinion-dependent com- munications," SIAM Journal on Control and Optimization, vol. 48, 07 2009. [5] R. Hegselmann and U. Krause, "Opinion dynamics and bounded confidence models, analysis and simulation," Journal of Artificial Societies and Social Simulation, vol. 5, 07 2002. [6] J. M. Hendrickx and S. Martin, "Open multi-agent systems: Gossiping with deterministic arrivals and departures," in Proceedings of the 56th IEEE CDC, pp. 1094 -- 1101, 09 2016. [7] M. Abdelrahim, J. M. Hendrickx, and W. M. Heemels, "Max- consensus in open multi-agent systems with gossip interactions," in Proceedings of the 56th IEEE CDC, pp. 4753 -- 4758, 12 2017. [8] P. Sen and B. K. Chakrabarti, Sociophysics: an introduction. Oxford University Press, 2013. [9] A. Giret, V. Julin, M. Rebollo, E. Argente, C. Carrascosa, and V. Botti, "An open architecture for service-oriented virtual organizations," pp. 118 -- 132, 09 2010. [10] O. K. Tonguz, "Red light, green lightno light: Tomorrow's commu- nicative cars could take turns at intersections," IEEE Spectrum, vol. 55, pp. 24 -- 29, Oct 2018. [11] J. C. Duchi, A. Agarwal, and M. J. Wainwright, "Dual averaging for distributed optimization: Convergence analysis and network scaling," IEEE Transactions on Automatic Control, vol. 57, 05 2010. [12] A. Muller and D. Stoyan, Comparison methods for stochastic models and risks. Chichester ; New York, NY : Wiley, 2002. [13] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip algorithms," IEEE/ACM Trans. Netw., vol. 14, pp. 2508 -- 2530, June 2006. Fig. 2: Bounds derived with the Ping model (dashed), the Infection model (plain), and the expression (20) (dotted) in terms of the rate ratio λc/λr, for different numbers of agents N, and with a unit variance σ 2 = 1. Fig. 3: Performances comparison for solving an intrinsic consensus problem with 10 agents between the bound from the Infection model (plain) and a simulation using a simple Gossip algorithm (dotted), both in logarithmic scale. The simulations are performed with 10000 iterations of the simulation over 200 events, which is empirically shown to be enough for the system to be considered in steady state in that setting. yi(t−)+y j(t−) on the bound, which stays around N−1 N2 σ 2; whereas after some threshold, the bound decays according to 1/(λc/λr). Finally, we have compared our bound with the results of the simplest version of average gossip [13]: agents take their own value xi as initial estimate yi(t), and whenever two agents i, j communicate, their estimates become . 2 yi(t+) = y j(t+) = (21) This algorithm computes the average in closed system. Observe that it is consistent with our setting presented in Section II, and that we did not make any adaptation to the open character. Fig. 3 compares the performances of the algorithm (21) with the bound from the Infection model (17). It is interesting to notice that even though there is a gap between the curves, the behavior of the error in terms of λc/λr is well captured by the bound, especially when that ratio gets large enough. Interestingly, the algorithm relies on only one variable, and does not make use of identifiers nor of any other information provided to the agents as described in Section II, while our bound is valid for algorithms potentially using this information. Hence one could wonder about the precise impact of that information and the efficiency gains it allows. VI. CONCLUSIONS We considered in this paper the possibility of arrivals and departures of agents in the study of multi-agent systems, and highlighted several challenges arising from this property. In particular, it prevents algorithms to converge for many problems, making their analysis challenging. APPENDIX A. Proof of Corollary 1 (cid:32) We prove in this Section that the expression (17) is equivalent to the following algebraic expression with 1− 1 1 + N−1 2 E [C(t)] ≥ N − 1 (cid:32) N2 h(N,L) = N ∑ k=2 k− 1 N − 1 k−1 ∏ j=1 (cid:16) (cid:17)(cid:33) h N, λc λr λc λr j(N − j)L σ 2 (cid:33) 2 + ( j + 1)(N − j− 1)L . Then, since P−1e1 = P−1 wT Q−1P−12λre1 = 11 e1 and P−1 11 = 2λr+(N−1)λc 2λr From the definition of Q−1 and thus of ak, one has then wT A(2λrI − A)−1e1 = . Algebraic manipulations of this result within equation (17) 2λr+(N−1)λc ∑N k=1 2λr from Theorem 3 allow to conclude. 1 , 2λr+(N−1)λc wT Q−1e1. (cid:16) k−1 N−1 ∏k−1 j=0 a j (cid:17) of bidiagonal matrices. We will use the following standard result on the inverse Lemma 1: The inverse A−1 of a bidiagonal matrix with a unit diagonal  1 −a1 ... 0 0 A =  0 0 ... 0 1 . . . 0 0 ... ... . . . −aN−1 1 0 1 ... 0 i−1 ∏ k= j is lower triangular matrix with unit diagonal and A−1 i, j = ak (∀i > j) We also need the following technical lemma. Lemma 2: For all matrix A, and ∀β ∈ R such that (β I − A)−1 is well defined, there holds A(β I − A)−1 = (β I − A)−1β − I Proof: Let A := A− β I, then A(β I − A)−1 = −( A + β I) A−1 = −I − β A−1. (22) The conclusion follows that − A−1 = (β I − A)−1. We start from the result of Theorem 3: (cid:0)1− wT A(2λrI − A)−1e1 (cid:1)σ 2. E [C(t)] ≥ N−1 N2 Since (2λrI − A) is bidiagonal and invertible when λr (cid:54)= 0, one writes (2λrI − A) = PQ with • P a diagonal matrix with Pii = i(N − i)λc + 2λr; • Q a bidiagonal matrix of unit diagonal with −i(N−i)λc Qi+1,i = 2λr+(i+1)(N−i−1)λc From this definition, one has . (2λr − A)−1 = Q−1P−1 , and where Q−1 is obtained using where P−1 Lemma 1 with ii = 1 2λr+i(N−i)λc ak = k(N−k)λc 2λr+(k+1)(N−k−1)λc . From Lemma 2 and the previous result, there holds wT A(2λrI − A)−1e1 = wT [Q−1P−12λr − I]e1 with wT Ie1 = wT e1 = 0.
1705.10259	1	1705	2017-05-29T15:34:27	Distributed Communication-aware Motion Planning for Multi-agent Systems from STL and SpaTeL Specifications	[ "cs.MA", "cs.LO", "cs.RO", "eess.SY" ]	In future intelligent transportation systems, networked vehicles coordinate with each other to achieve safe operations based on an assumption that communications among vehicles and infrastructure are reliable. Traditional methods usually deal with the design of control systems and communication networks in a separated manner. However, control and communication systems are tightly coupled as the motions of vehicles will affect the overall communication quality. Hence, we are motivated to study the co-design of both control and communication systems. In particular, we propose a control theoretical framework for distributed motion planning for multi-agent systems which satisfies complex and high-level spatial and temporal specifications while accounting for communication quality at the same time. Towards this end, desired motion specifications and communication performances are formulated as signal temporal logic (STL) and spatial-temporal logic (SpaTeL) formulas, respectively. The specifications are encoded as constraints on system and environment state variables of mixed integer linear programs (MILP), and upon which control strategies satisfying both STL and SpaTeL specifications are generated for each agent by employing a distributed model predictive control (MPC) framework. Effectiveness of the proposed framework is validated by a simulation of distributed communication-aware motion planning for multi-agent systems.	cs.MA	cs	Distributed Communication-aware Motion Planning for Multi-agent Systems from STL and SpaTeL Specifications Zhiyu Liu, Bo Wu, Jin Dai, and Hai Lin 7 1 0 2 y a M 9 2 ] A M . s c [ 1 v 9 5 2 0 1 . 5 0 7 1 : v i X r a Abstract -- In future intelligent transportation systems, net- worked vehicles coordinate with each other to achieve safe operations based on an assumption that communications among vehicles and infrastructure are reliable. Traditional methods usually deal with the design of control systems and communi- cation networks in a separated manner. However, control and communication systems are tightly coupled as the motions of vehicles will affect the overall communication quality. Hence, we are motivated to study the co-design of both control and communication systems. In particular, we propose a control the- oretical framework for distributed motion planning for multi- agent systems which satisfies complex and high-level spatial and temporal specifications while accounting for communication quality at the same time. Towards this end, desired motion specifications and communication performances are formulated as signal logic (SpaTeL) formulas, respectively. The specifications are encoded as constraints on system and environment state variables of mixed integer linear programs (MILP), and upon which control strategies satisfying both STL and SpaTeL specifications are generated for each agent by employing a distributed model predictive control (MPC) framework. Effectiveness of the pro- posed framework is validated by a simulation of distributed communication-aware motion planning for multi-agent systems. logic (STL) and spatial-temporal temporal I. INTRODUCTION Multi-agent systems like intelligent transportation systems and smart cities have emerged as a hot research topic in the interdisciplinary study of control theory, robotics and computer science due to their wide applications in both academic and industrial domains. Being one of the funda- mental research problems in this context, motion planning and control of multi-agent systems has drawn a consider- able amount of research interest in recent years. Motion and control strategies are developed for various global/local coordination purposes, see e.g. [1] and the references therein. Recent theoretical and technological developments have enhanced the application of formal methods to address complex and high-level motion planning objectives. Formal languages, such as linear temporal logic (LTL) and compu- tation tree logic (CTL) [2], provide a concise formalism for specifying and verifying desired logic behavior of dynamical systems [3]. To pursue satisfaction of temporal specifica- tions, a vast majority of research efforts has been devoted to abstraction-based approaches (see e.g. [4] -- [6] and the references therein), where the temporal logic formula spec- ification is translated into an automata representation whose This work was supported by the National Science Foundation (NSF-CNS- 1239222 and NSF- EECS-1253488) Z. Liu, B. Wu, J. Dai, and H. Lin are with the Department of Electrical Engineering, University of Notre Dame, Notre Dame, IN 46556, USA (e- mail: [email protected]; [email protected]; [email protected]; [email protected]). accepting runs correspond to satisfaction of the formula, while the environment of the system is also abstracted into a finite transition diagram. Based on these abstractions, algo- rithms are derived for verification and synthesis of discrete controllers that drive the system to satisfy the specification. Despite their success in the correct-by-construction design of controllers, abstraction-based approaches suffer from the "state explosion" issues as both the synthesis and abstraction algorithms scale at least exponentially with the dimension of the discretized configuration space [3], rendering such approaches impractical for large-scale multi-agent systems. Furthermore, LTL and/or CTL specifications require only the order of events that should be executed by the system, whereas temporal distance between them is often neglected. Many attempts have been made to multi-agent systems to ease the computational burden while obeying temporal logic specifications. On the one hand, approaches that do not require abstractions are proposed, such as sampling- based approach [7] and mixed integer linear programming techniques [8] -- [10]. On the other hand, model predictive control (MPC) schemes are introduced to reduce the formal synthesis problem into a series of smaller optimization prob- lems of a shorter horizon. Wongpiromsarn et al. [11] utilized MPC to accomplish motion planning for LTL specifications, whereas Tumova and Dimarogonas [12] applied MPC for coordination of multi-agent systems. Kuwata and How [13] integrated MILP formalism with MPC techniques to pursue distributed trajectory optimization; nevertheless, satisfaction of more complex specifications were not studied. It is worth pointing out that aforementioned results as- sumed perfect inter-agent communication, which turned out to be oversimplified in practice, since communication quality of service (QoS) is crucial in maintaining multi-agent coordi- nation. In intelligent transportation systems, communications among vehicles and infrastructure are not always reliable. To pursue the co-optimization of motion and communica- tion, Grancharova et al. [14] proposed a trajectory planning scheme for agents subject to communication capacity con- straints by applying distributed MPC. The same approach is utilized in [15] for motion planning of multi-agent systems under radio communication constraints while agents are treated as relay nodes where communication channels are modeled as disk model which assumes communication is off beyond a certain threshold. Yan and Mostofi [16] modeled the communication channel between a robot and a base station as a Gaussian process with fading and shadowing effects; however, the optimization was performed only with respect to the robot's motion velocity, transmission rate and stop time, while the robot was assumed to travel along a pre-defined trajectory. Nevertheless, communication models presented in the previous work are either too complicated for multi-agent systems or oversimplified in practice. Motivated by the aforementioned concerns, we focus on construction of local controller for multi-agent systems such that certain motion requirements are fulfilled in the presence of communication capacity constraints in this paper. Given multiple agents moving in a shared environment with com- munication base stations, our design objective is to drive each agent to satisfy its own motion specifications, while unsafe zone, collision and poor communication QoS are avoided. Towards this end, we use signal temporal logic (STL) [17], [18] formulas to describe local motion and safety require- ments for each agent, and spatial-temporal logic (SpaTel) [19] formulas to represent global safety and communication QoS requirements, based on which an MILP formalism is established to solve not only the joint motion-communication co-optimization problem, but synthesizes collision-avoiding motion controllers as well. Note that different from our previ- ous work [20], we present a distributed synthesis framework by employing MPC such that appropriate controllers can be synthesized locally. The rest of this paper is organized as follows. In Section II, we briefly introduce necessary preliminaries of STL and SpaTel. We then formally present the communication-aware motion planning problem from STL-SpaTel specifications in Section III. In Section IV, we propose the MILP encoding schemes for STL and SpaTel specifications. Based on these MILP constraints, we exploit a distributed MPC strategy in Section V to synthesize local controllers for each agent. Sim- ulation examples are presented in Section VI for validating our proposed framework. Section VII concludes the paper. II. PRELIMINARIES A. Agent Models The multi-agent system under consideration in this pa- per consists of P agents with unique identities P = {1, 2, . . . , P} which perform their missions in a shared 2- D environment. For each i ∈ P, the agent dynamics is dominated by the linear dynamics of the following form xi(t) = Axi(t) + Bui(t), (1) i where xi ∈ R4 is the state of agent i with xi = [pT i ]T , vT where pi, vi ∈ R2 are the position and velocity of the agent, respectively; ui = [ui,1 ui,2]T ∈ U ⊆ R2 is the local admissible control inputs, and xi(0) = xi,0 ∈ R4 is the initial state. (A, B) is a controllable pair with proper dimensions. The environment X is given by a large convex polygonal subset of the 2-D Euclidean space R2. Let Xobs ⊆ X be the regions in the environment occupied by polygon obstacles. Xf ree = X \ Xobs denotes the obstacle-free working space for the multi-agent system. To run the distributed communication-aware motion plan- ning in an online manner, we borrow the idea from [21] and assume that the agent dynamics (1) admits a discrete-time approximation of the following form, given an appropriate sampling time ∆t > 0: xi(tk+1) = Adxi(tk) + Bdui(tk), (2) where k ∈ N is the sampling index and ∆t is selected such that (Ad, Bd) is controllable. The sampling is uniformly performed, i.e., for each k > 0, tk+1 − tk = ∆t and we use [a, b] as an abbreviation for the set {a, a + 1, . . . , b}. Given xi,k ∈ R2 and uH i ∈ U ω, i ∈ P, a (state) run i = xi,kxi,k+1xi,k+2 . . . xi,k+H−1 generated by agent i's xH dynamics (2) with control input uH is a finite sequence ob- tained from agent i's state trajectory, where xi,k = xi(tk) ∈ i R4 is the state of the system at time index t, and for each k ∈ N, there exists a control input ui,k = ui(tk) ∈ U such that xi(tk+1) = Adxi(tk) + Bdui(tk). Under MPC framework with planning horizon H (cf. Section III), given a local state xi,k and a sequence of local control inputs uH i = ui,kui,k+1ui,k+2 . . . ui,k+H−1, the resulting horizon-H run of agent i, xi(xi,k, uH i ) = xi,kxi,k+1xi,k+2 . . . xi,k+H−1 is unique. B. Communication Typically, agents within the team may be assigned with different roles and responsibilities, and inter-agent commu- nication is required to ensure proper coordination between them for safety and efficient mission execution. Furthermore, reliable communication is also needed between the agents and base stations to collect the sensed environment data and let base stations provide global information for agents, which is essential for distributed algorithms. Therefore in this paper we explicitly consider communication as an optimization objective. As shown in Fig. 1, we assume the environment is a gridded square world. There are four base stations, one at the center of each quadrant. To reduce the interference and increase the number of agents that can be served, similar to cellular communication, each base station is equipped with four 90◦ sector antenna systems to fully cover each quadrant. The quality of service (QoS) of inter-agent communication and between the individual agent and the base station is assumed to be subject to the path loss and the shadowing effect [22] due to the mountains. Fig. 1: The gridded environment C. Signal Temporal Logic We consider STL formulas which are defined recursively as follows. cursively as: Definition 1 (STL Syntax): STL formulas are defined re- ϕ ::= Trueπµ¬πµϕ ∧ ψϕ ∨ ψ2[a,b]ψϕ (cid:116)[a,b] ψ where πµ is an atomic predicate Rn → {0, 1} whose truth value is determined by the sign of a function µ : Rn → R, i.e., πµ is true if and only if µ(x) > 0; and ψ is an STL formula. The "eventually" operator 3 can also be defined here by setting 3[a,b]ϕ = True (cid:116)[a,b] ϕ. The semantics of STL with respect to a discrete-time signal x are introduced as follows, where (x, tk) = ϕ denotes for which signal values and at what time index the formula ϕ holds true. Definition 2 (STL Semantics): The validity of an STL for- time tk is defined to signal x at mula ϕ with respect inductively as follows: (x, tk) = ψ; 1) (x, tk) = µ, if and only if µ(xk) > 0; 2) (x, tk) = ¬µ, if and only if ¬((x, tk) = µ); 3) (x, tk) = ϕ ∧ ψ, if and only if (x, tk) = ϕ and 4) (x, tk) = ϕ∨ψ, if and only if (x, tk) = ϕ or (x, tk) = ψ; 5) (x, tk) = 2[a,b]ϕ, if and only if ∀tk(cid:48) ∈ [tk + a, tk + b], (x, tk(cid:48)) = ϕ; 6) (x, tk) = ϕ(cid:116)[a,b] ψ, if and only if ∃tk(cid:48) ∈ [tk +a, tk +b] such that (x, tk(cid:48)) = ψ and ∀tk(cid:48)(cid:48) ∈ [tk, tk(cid:48)], (x, tk(cid:48)(cid:48) ) = ϕ. A signal x = x0x1x2 . . . satisfies ϕ, denoted by x = ϕ, if (x, t0) = ϕ. Intuitively, x = 2[a,b]ϕ if ϕ holds at every time step between a and b, x = ϕ (cid:116)[a,b] ψ if ϕ holds at every time step before ψ holds, and ψ holds at some time step between a and b, and x = 3[a,b]ϕ if ϕ holds at some time step between a and b. An STL formula ϕ is bounded-time if it contains no unbounded operators. The bound of ϕ can be interpreted as the horizon of future predicted signals x that is needed to calculate the satisfaction of ϕ. D. Spatial Temporal Logic SpaTeL is defined by the combination of STL and Tree Spatial Superposition Logic (TSSL) where STL is respon- sible for describing the temporal properties and TSSL is related to spatial properties [23]. Prior to define the syntax and semantics of SpaTeL, we first introduce quad transition systems (QTS) with a quad tree data structure to model spatial characteristics. A QTS [23] is defined through a quad tree structure by dividing environment into four quadrants recursively. It is defined as a tuple Q(t) = (V,E, v0, Vf , µ,L, l), where V is the set of nodes vi. E ⊂ V × V is the set of directed transitions. v2 is a child of v1 if (v1, v2) ∈ E. v0 is the root of the tree (the only node which is not a child of another node). Vf is the set of leaves (nodes without children). µ : 6 4 3 2 2 3 4 6 4 4 4 3 3 4 4 4 3 4 4 3 0 0 4 3 2 3 2 2 0 2 3 2 2 3 0 1 0 3 3 2 3 4 0 2 3 4 4 3 3 6 4 3 3 4 6 3 2 3 3 2 2 3 3 2 Fig. 2: Matrix C for the environment v0 NW NE SW SE v1 v2 v3 v4 NW NE SW SE v5 v6 v7 v8 NW NE SW SE vf vf vf vf Fig. 3: Partial QTS for Matrix C V×R≥0 → R+ is the valuation function assigning each node a real positive number. L is a finite set of labels. l : E → L is the labeling function which maps each edge to a label. A labeled path of a QTS is defined as a function which maps a node to a set of infinite sequences of nodes: λB :=(cid:8)(v0, v1, v2, . . . , ¯vf )(vi, vi+1) ∈ E, l(vi, vi+1) ∈ B(cid:9) , (cid:8)0, . . . , T(cid:9) → Q. (3) where B ⊆ L; ¯vf denotes infinite repetitions of leaf node vf ; i ∈ N≥0. A trace corresponding to a trajectory traveling in the divided environment is defined as a function q : Fig. 2 and Fig. 3 show an example of formulating QTS. Let matrix C ∈ R2D×2D be the representation of the environment in Fig. 1 divided by 2D × 2D grids and elements in it represent a valuation function µ for corresponding area, where D ∈ N+ is the resolution of the matrix (or the depth of the tree), which is 3 in this case. A QTS can be constructed from C by choosing it as the root node v0 then dividing C into four 2D−1 × 2D−1 sub-matrices. Each of them is a child of v0 and the directed transitions are stored in E with directional labels from the set L =(cid:8)N W, N E, SE, SW(cid:9). One can expand each child with the same method applied on the root node until we have 2D × 2D leaf nodes. Valuation function µ can be defined for each leaf nodes by using the elements in C and for other nodes by the mean of the valuations of its children recursively: µ(v) = 1 4 µ(vc),∀v ∈ V \ Vf , (4) (cid:88) (v,vc)∈E Based on the definition of QTS, the syntax and semantics of TSSL are well defined in [24]. With STL and TSSL, one can define SpaTeL as follow [23]. Definition 3 (SpaTeL Syntax): The spatial formulas are defined recursively as: ψ ::=Truem ∼ d¬ψψ1 ∧ ψ2∃B (cid:13) ψ∀B (cid:13) ψ ∃Bψ1Ukψ2∀Bψ1Ukψ2 where ∼∈(cid:8)≥,≤(cid:9), d ∈ [0, b], b ∈ R+, k ∈ N+, B ⊆ L with B (cid:54)= ∅, and m ∈ µ. Uk and (cid:13) are until and next operators respectively. The syntax of SpaTeL formulas are defined as follow: φ ::= ψ¬φφ1 ∧ φ22[a,b]φφ1 (cid:116)[a,b] φ2. As we noticed, SpaTeL is a combination of STL and TSSL by replacing the STL predicates with TSSL formulas. Definition 4 (SpaTeL Semantics): The validity of an Spa- TeL formula φ with a trace q at time tk is defined inductively as follows: 1) (q, tk) = ψ, if and only if (q, v0, tk) = ψ; 2) (q, tk) = ¬φ, if and only if ¬((q, tk) = φ); 3) (q, tk) = φ1 ∧ φ2, if and only if (q, tk) = φ1 and (q, tk) = φ2; 4) (q, tk) = 2[a,b]φ, if and only if ∀tk(cid:48) ∈ [tk + a, tk + b], (q, tk(cid:48)) = φ; 5) (q, tk) = φ1 (cid:116)[a,b] φ2, if and only if ∃tk(cid:48) ∈ [tk + a, tk + b] such that (q, tk(cid:48)) = φ2 and ∀tk(cid:48)(cid:48) ∈ [tk, tk(cid:48)], (q, tk(cid:48)(cid:48) ) = φ1; where ψ is a TSSL formula. E. Distributed MPC Instead of solving optimization problems for the whole time horizon Tf , MPC solves problems within a finite horizon H < Tf starting from current states and provides a finite control inputs uH k . Only the first control input will be implemented and agents' states will be sampled again. Then the new optimization problem will be computed based on the new states within the finite horizon H. MPC has a better performance comparing to other methods in terms of handling model uncertainty and external disturbance [25]. It also reduces the size of problem by only considering H steps ahead. For distributed MPC, each agent has its own MPC based optimization problem which only considers its own neighbor [13] and therefore makes it much smaller than the centralized problem. By applying distributed MPC, one can deal with larger scale cases [26]. III. PROBLEM STATEMENT A. STL Motion Planning Specifications We now proceed to the communication-aware motion planning problems for multi-agent systems. Let us consider a team of P agents conducting motion behavior in the shared environment X , each of which is governed by the discretized dynamics (2). We assign a goal region Xi,goal for agent i, i ∈ P that is characterized by a polytope [3] in Xf ree, i.e., there exist M ≥ 3 and ai,j ∈ R2, bi,j ∈ R, j = 1, 2, . . . , Mi such that Xi,goal = {p ∈ R2aT i,jp + bi,j ≤ 0, j = 1, 2, . . . , Mi}. (5) In other words, Xi,goal = {x ∈ XaT i,j[I2 O2]x+bi,j ≤ 0, j = 1, 2, . . . , Mi}. (6) where I2, O2 ∈ R2×2 denote the 2-dimensional identity and zero matrices, respectively. We assume that all agents share a synchronized clock. The terminal time of multi-agent motion is upper-bounded by tf = Tf ∆t with Tf ∈ N+, and the planning horizon is then given by [0, Tf ]. Accomplishment of individually-assigned specifications is of practical importance, for instance search and rescue missions or coverage tasks are often specified to mobile robots individually. In this paper, local motion planning tasks for agent i are summarized as the following STL formula: for i ∈ P, require: ϕi = ϕi,p ∧ ϕi,s, where 1) the motion performance property (cid:0)aT i,j[I2 O2]xi + bi,j ≤ 0(cid:1) Mi(cid:94) j=1 ϕi,p = 3[0,Tf ] (7) (8) (9) requires that agent i enter the goal region within Tf time steps; 2) the safety property (cid:94) [(pi,1 − pj,1 ≥ d1) ϕi,s = 2[0,Tf ] j∈Ni,j(cid:54)=i ∧ (pi,2 − pj,2 ≥ d2)] ensures that agent i shall never encounter obstacle re- gions. Here d1 and d2 are pre-defined safety distances between two agents in the two dimensions. Ni ⊆ P denotes the set for agent i's neighbor which will be described in Section V. B. SpaTeL Specifications of Communication and Safety The shared environment X is divided into 2D × 2D same sized grids such that the environment can be represented by a quad tree, where D is the depth of the tree. Without loss of generality, we also assume that the obstacle region Xobs, like the mountains in Fig. 1, is a rectangular subset of X which is also represented by grids. Fig. 2 and Fig. 3 illustrate such setup with corresponding QTS, where D = 3. The number in each grid is the valuation function for each leaf node. We use SpaTeL specifications to specify obstacles avoidance, optimal communication with base stations and preventing traffic congestion. To this regard, we assign each leaf node with a number listed in Fig.2 representing the maximum amount of agents where each grid can have at the same time. For the obstacles represented with red grids, the number of agents allowed in it is zero. We place four communication base stations shown in Fig. 1 located at the center of v1 to v4. Considering the path loss and the shadowing effect in wireless communication, we assume grids that are far away from base stations and are blocked by obstacles will suffer poor communication quality leading to the maximum amount of agents in those grids are relatively small. The grids with number 6 are the initial and terminal locations for all agents which have a higher capacity for accommodating agents. To specify the spatial temporal specifications mentioned (10) (11) above, we formulate the SpaTeL formula as bellow: φ = 2[0,Tf ](ψ1) ∧ 2[0,Tf ](ψ2 ∧ ψ3 ∧ ψ4 ∧ ψ5) where ψi are TSSL formulas: ψ1 =∀SW (cid:13) ∀N E (cid:13) ∀(cid:8)N W, N E, SW(cid:9) (cid:13) (µ = 0)∧ ψ1 specifies the safety property of avoiding obstacles. ψ2 =∀N W (cid:13) ∀N W (cid:13) ∀N W (cid:13) (µ ≤ 6)∧ ∀SE (cid:13) ∀N W (cid:13) ∀N W (cid:13) (µ = 0)∧ ∀N E (cid:13) ∀SW (cid:13) ∀(cid:8)N W, N E(cid:9) (cid:13) (µ = 0) ∀N W (cid:13) ∀N W (cid:13) ∀(cid:8)N E, SW, SE(cid:9) (cid:13) (µ ≤ 4)∧ ∀N W (cid:13) ∀N E (cid:13) ∀(cid:8)N W, SE(cid:9) (cid:13) (µ ≤ 3)∧ ∀N W (cid:13) ∀SW (cid:13) ∀(cid:8)N W, SE(cid:9) (cid:13) (µ ≤ 3)∧ ∀N W (cid:13) ∀SE (cid:13) ∀(cid:8)N E, SE(cid:9) (cid:13) (µ ≤ 2)∧ ∀N W (cid:13) ∀SW (cid:13) ∀N E (cid:13) (µ ≤ 4)∧ ∀N W (cid:13) ∀SW (cid:13) ∀SW (cid:13) (µ ≤ 2)∧ ∀N W (cid:13) ∀SE (cid:13) ∀N W (cid:13) (µ ≤ 4)∧ ∀N W (cid:13) ∀N E (cid:13) ∀SW (cid:13) (µ ≤ 4)∧ ∀N W (cid:13) ∀N E (cid:13) ∀N E (cid:13) (µ ≤ 2)∧ ∀N W (cid:13) ∀SE (cid:13) ∀SW (cid:13) (µ ≤ 3) (12) ψ2 specifies the desired pattern in terms of communication quality and avoiding traffic congestion for the upper left quadrant. ψ3 to ψ5 specify the rest of environment following the same procedure. Comparing SpaTeL formulas (10) and STL formulas (7), it is worth pointing out that STL formulas define the require- ment for each agent while SpaTeL formulas define global specifications for all agents. By using both STL and SpaTeL formulas in our design, we are able to consider both local and global properties in a decentralized method. C. MPC based Co-optimization Problem We wish to achieve a co-optimization for motion planning and communication QoS. To this regard, the cost function Ji,1 in the following linear quadratic form represents the energy consumption for agent i. k(cid:48)+H−1(cid:88) k=k(cid:48) k(cid:48)+H−1(cid:88) k=k(cid:48) di,k Ji,1 = (qTxi,tk + rTui,tk) + h(k(cid:48)) (13) where H is the planning horizon, q and r are non-negative weighting vectors and . denotes the element-wise absolute value. Second term is time penalty multiplying goal penalty such that each agent can move towards its goal. We define goal penalty as di,k = pT i,goal and time penalty as h(k) = λk2, where pi,goal ∈ R2 denotes the geometric center of goal region Xi,goal for agent i and λ is a user defined parameter. i,k − pT As for communication QoS, we consider both base station- to-agent and agent-to-agent scenarios. For inter-agent com- munication, we assume in each planning period, for agent i, agents within its neighbor are static where the commu- nication channel among them are affected by path loss. For agent j ∈ Ni, one can defines a matrix C(cid:48) j similar to C which defines its communication QoS pattern in the environment. Using similar idea from our previous work [20], the communication cost is formulated as follow. k(cid:48)+H−1(cid:88) 2D(cid:88) 2D(cid:88) Ni(cid:88) n=1 m=1 k=k(cid:48) Ji,2 = (Cm,n + j,m,n)(Om,n,tk − 1) C(cid:48) (14) where C is defined in Fig.2 representing communication quality with base stations. Ot is a binary matrix for capturing the occupancy of the grids where Om,n,t is zero if and only if the agent is in m-th row and n-th column. j∈Ni Given the aforementioned preliminaries and cost func- tions, we formally formulate the distributed communication- aware motion planning from STL-SpaTeL specifications as below: Problem 1: Given a multi-agent system with P agents whose dynamic behaviors are determined by (2) with initial states xi,0, a planning horizon H, a local STL formula ϕi in (7) and a global formula φ in (10), we find local control inputs ui(tk) for all agents such that the following cost function is optimized: i )) = αJi,1 + (1 − α)Ji,2 uH min i ,i∈P Ji(xi(xi,0, uH s.t. ∀i ∈ P, (15) i ) = ϕi, xi(tk+1) = Adxi(tk) + Bdui(tk), xi(xi,k, uH (q, tk) = φ, ui ∈ U = [−umax, umax] × [−umax, umax], vi < vmax, ui mivi ≤ umax mivmax ωi = where α ∈ [0, 1] is a user defined parameter. IV. MILP ENCODING OF COMMUNICATION-AWARE MOTION PLANNING A. MILP Encoding of Agent Dynamics In this section, we replace tk with t and denote xit and uit as the state and control inputs of agent i at time step t. To encode the motion planning cost (13) as linear programming, we employ Manhattan distance for di,k and introduce slack vectors αit, βit, γit and additional constraints [27] such that Ji,1 can be transformed to linear cost function. k(cid:48)+H−1(cid:88) t=k(cid:48) Ji,1 = (qT αit + rT βit) + h(k(cid:48)) k(cid:48)+H−1(cid:88) 2(cid:88) t=k(cid:48) k=1 γitk (16) s.t. ∀t ∈ [k(cid:48), k(cid:48)+H − 1],∀j ∈ [1, 4],∀k ∈ [1, 2] xitj ≤ αitj,−xitj ≤ αitj uitk ≤ βitk,−uitk ≤ βitk and and xitk − pi,goal,k ≤ γitk,−xitk + pi,goal,k ≤ γitk and xi(t + 1) = Adxi(t) + Bdui(t) (17) To transfer the nonlinear velocity constraints, we use the method in [8] by introducing an arbitrary number L of linear constraints leading to the 2-D velocities approximated by a regular L-sided polygon. ∀l ∈ [1, L], i ∈ [1, P ], t ∈ [k(cid:48), k(cid:48) + H − 1] ) ≤ vmax ) + vit2cos( vit1sin( 2πl L 2πl L (18) B. Boolean Encoding of STL Constraints For the MILP encoding of STL specifications in (7), we denote two Boolean variables zϕi,p whose value depends on the satisfaction of ϕi,p and ϕi,s respectively [18], [20]. Then the satisfaction of ϕi at time step t can be represented by Boolean variable zϕi t which is determined by zϕi,p and zϕi,s and zϕi,s as follow. t t t t t = zϕi,p zϕi t ∧ zϕi,s t with ∀i ∈ P : t ≤ zϕi,p zϕi t ≥ zϕi,p zϕi t ≤ zϕi,s , zϕi t − 1 t + zϕi,s t t (19) (20) where zϕi,p ing specifications are satisfied. and zϕi,s t t are one if and only if their correspond- C. Boolean Encoding of SpaTeL Constraints Similar to encoding STL as MILP, we use the method in [19] to encode SpaTeL formulas (10) as MILP recursively. We denote SpaTeL formulas φ1 = 2[0,Tf ](ψ1) and φ2 = 2[0,Tf ](ψ2∧ψ3∧ψ4∧ψ5). We define three Boolean variables zφ v,t, zφ1 v,t whose truth values correspond to the satisfaction of φ, φ1 and φ2 respectively and are determined by the following constraints. v,t and zφ2 with v,k v,t, zφ v,k, zφ2 ∀i ∈ P,∀k ∈ [t, Tf ] : v,t ≤ zφ2 v,t ≤ zφ1 zφ v,t, zφ v,t ≤ zψ2−5 v,t ≤ zψ1 Tf(cid:88) zφ1 v,t ≥ v,k − Tf + t, zφ2 zφ1 zψ1 v,k ≥ 5(cid:88) v,k ≤ zψ3 v,k ≤ zψ2 v,k, zψ2−5 zψ2−5 v,k − 3 zψ2−5 zψj k=t v,t + zφ2 v,t − 1 v,t ≥ zφ1 Tf(cid:88) v,k − Tf + t zψ2−5 v,k ≤ zψ4 v,k, zψ2−5 v,k, zψ2−5 v,t ≥ k=t v,k ≤ zψ5 v,k j=2 (22) v,k, i ∈ [1, . . . , 5] represent the where Boolean variables zψi satisfaction of TSSL formulas ψ1 to ψ5. We assume the sampling period ∆t for the discretized system is 1. V. DISTRIBUTED MODEL PREDICTIVE CONTROL SYNTHESIS We wish to construct a distributed and online framework for the communication-aware motion planning. Towards this end, we employ MPC as the basic framework such that the sub-problem for each agent can be solved online and the system is robust enough to deal with model uncertainty and external disturbance. Another strategy we employed for achieving distributed and online manner is only considering its neighbor for each agent during planning so that the size of each sub-problem will reduce significantly for the size of the formulated MILP problem in previous section depends on the number of agents directly. We assume that global information like time, synchronization, states of each agent are available for each agent through communication base stations. The algorithm is summarized in Algorithm 1. Since it is reasonable to omit the agents that are far away, we define the neighbor of each agent by choosing the agents whose distance to the agent are shorter than a certain threshold. We assign each agent a unique priority in each planning period randomly. Agents can only plan after all the agents in its neighbor with higher priority have planned. Fig. 4 illustrates our idea. Agents connected with edges are considered in the same neighbor. For the agent with priority 2, it plans first in each period since it has the highest priority in its neighbor and the agent with priority 4 will plan right after it. In each planning period, agents formulate their own optimization problem in (15) encoded as MILP and run commercial MILP solver to find the control inputs within the planning horizon. After all agents planned, they implement the first step of the control inputs and move into next period. The algorithm stops when all agents reach their goals or it reaches the time limit. zφ v,t = zφ1 v,t ∧ zφ2 Tf(cid:94) v,t Tf(cid:94) zψ2−5 v,t = v,k v,k ∧ zψ5 v,k ∧ zψ4 k=t v,k zψ1 v,k, zφ2 v,k ∧ zψ3 zφ1 v,t = k=t zψ2−5 v,k = zψ2 VI. SIMULATION RESULTS (21) To justify our distributed co-optimization strategy, we implemented our MPC based framework with MATLAB. The encoded MILP problem was modeled by AMPL, an are marked as black rectangular. Communication quality is considered in this case. Due to the shadowing effect caused by obstacles, the communication quality is the worst in the space between two obstacles. Fig. 2 shows the communica- tion quality pattern in terms of base stations and obstacles. The outputs of the simulation are agents' states and control inputs at all steps. Algorithm 1 Distributed Communication-aware Motion Planning with MPC Input: Intial states and goal positions of each agent Output: Return states and control inputs of each agent Initialization while AgentSet (cid:54)= ∅ or t ≤ Tf do Randomly set a unique priority for each agent AgentSet for Agent i, at time t do j , j ∈ P Update the list of all agents' states xH inputs uH if Agent j, j ∈ P is close enough to agent i then Add agent j into agent i's neighbour Ni j and control in end Wait until all agents in Ni with higher priority than i planned, then for t ∈ [t, t + H] do Minimize the cost function Ji in (15) subjecting to corresponding constraints and uH i to the whole end Broadcast the results of xH i group end Implement the first step of control inputs uH i Global time t = t + 1 if Any agent i ∈ P reaches its target then Remove it from AgentSet end end Fig. 5: Distributed Communication-aware Motion Planning 3 8 5 1 2 6 7 4 Fig. 4: Neighbor algebraic modeling language for large scale mathematical programming [28], and solved by Gurobi, a commercial solver for MILP [29]. Fig. 1 illustrates our basic setup. Fig. 5 shows the simu- lation result of the MPC based distributed co-optimization. Fig. 6 demonstrates agents distribution at different time steps. Twelve agents with initial states and target positions marked as red cross are given. The agent dynamics ruled by (2) is given by setting matrices Ad and Bd as: 1 0 0 1 0 0 0 0  , Bd = 0.5 0 1 0 1 0 1 0 0 1 0 1  . 0 0.5 0 1 Ad = (23) We choose H = 5, L = 8, ∆t = 1, Tf = 50, λ = 0.005, d1 = d2 = 1, α = 0.5. The working space is set as a 160m × 160m square which is represented by QTS with D = 3 shown in yellow grids. Four communication base stations are marked as red stars in the graph. Obstacles Fig. 6: Agents distribution As we can see from the result, using the proposed distributed co-optimization strategy, all agents are able to satisfy the specifications. Several agents avoid the poor communication quality area to reach their goals, compared to Fig. 7 without considering communication. The simulation was run on a PC with Intel core i7- 4710MQ 2.50 GHz processor and 8GB RAM. The algorithm was run distributively among agents. Each agent can solve its own MILP problem around 0.1 second. VII. CONCLUSION As an extension of our previous work [20], we develop a distributed MPC based algorithm for communication-aware [3] M. Kloetzer and C. Belta, "A fully automated framework for control of linear systems from temporal logic specifications," IEEE Transactions on Automatic Control, vol. 53, no. 1, pp. 287 -- 297, 2008. [10] E. M. Wolff, U. Topcu, and R. M. Murray, "Optimization-based trajectory generation with linear temporal logic specifications," in 2014 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2014, pp. 5319 -- 5325. [11] T. Wongpiromsarn, U. Topcu, and R. M. Murray, "Receding horizon temporal logic planning," IEEE Transactions on Automatic Control, vol. 57, no. 11, pp. 2817 -- 2830, 2012. [12] J. Tumova and D. V. Dimarogonas, "Multi-agent planning under local ltl specifications and event-based synchronization," Automatica, vol. 70, pp. 239 -- 248, 2016. [13] Y. Kuwata and J. P. How, "Cooperative distributed robust trajectory optimization using receding horizon milp," IEEE Transactions on Control Systems Technology, vol. 19, no. 2, pp. 423 -- 431, 2011. [14] A. Grancharova, E. I. Grøtli, D.-T. Ho, and T. A. Johansen, "Uavs trajectory planning by distributed mpc under radio communication path loss constraints," Journal of Intelligent & Robotic Systems, vol. 79, no. 1, pp. 115 -- 134, 2015. [15] E. I. Grøtli and T. A. Johansen, "Path planning for uavs under com- munication constraints using splat! and milp," Journal of Intelligent & Robotic Systems, vol. 65, no. 1-4, pp. 265 -- 282, 2012. [16] Y. Yan and Y. Mostofi, "Co-optimization of communication and motion planning of a robotic operation under resource constraints and in fading environments," IEEE Transactions on Wireless Communica- tions, vol. 12, no. 4, pp. 1562 -- 1572, 2013. [17] O. Maler and D. Nickovic, "Monitoring temporal properties of contin- uous signals," in Formal Techniques, Modelling and Analysis of Timed and Fault-Tolerant Systems. Springer, 2004, pp. 152 -- 166. logic specifications," in Proceedings of [18] V. Raman, A. Donz´e, M. Maasoumy, R. M. Murray, A. Sangiovanni- Vincentelli, and S. A. Seshia, "Model predictive control with signal the 53rd IEEE temporal Conference on Decision and Control (CDC). IEEE, 2014, pp. 81 -- 87. [19] I. Haghighi, S. Sadraddini, and C. Belta, "Robotic swarm control from spatio-temporal specifications," in Decision and Control (CDC), 2016 IEEE 55th Conference on. IEEE, 2016, pp. 5708 -- 5713. [20] Z. Liu, J. Dai, B. Wu, and H. Lin, "Communication-aware motion planning for multi-agent systems from signal temporal logic specifi- cations," in American Control Conference (ACC), Proceedings of the 2017. [21] V. Raman, A. Donz´e, D. Sadigh, R. M. Murray, and S. A. Seshia, "Reactive synthesis from signal logic specifications," in Proceedings of the 18th International Conference on Hybrid Systems: Computation and Control (HSCC). ACM, 2015, pp. 239 -- 248. temporal [22] A. F. Molisch, Wireless communications. vol. 34. John Wiley & Sons, 2012, [23] I. Haghighi, A. Jones, Z. Kong, E. Bartocci, R. Gros, and C. Belta, "Spatel: a novel spatial-temporal logic and its applications to net- worked systems," in Proceedings of the 18th International Conference on Hybrid Systems: Computation and Control. ACM, 2015, pp. 189 -- 198. [24] E. Bartocci, E. A. Gol, I. Haghighi, and C. Belta, "A formal methods approach to pattern recognition and synthesis in reaction diffusion networks," IEEE Transactions on Control of Network Systems, 2016. [25] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. Scokaert, "Constrained model predictive control: Stability and optimality," Au- tomatica, vol. 36, no. 6, pp. 789 -- 814, 2000. [26] Y. Kuwata, A. Richards, T. Schouwenaars, and J. P. How, "Distributed robust receding horizon control for multivehicle guidance," IEEE Transactions on Control Systems Technology, vol. 15, no. 4, pp. 627 -- 641, 2007. [27] M. Athans and P. L. Falb, Optimal control: an introduction to the theory and its applications. Courier Corporation, 2013. [28] R. Fourer, D. Gay, and B. Kernighan, Ampl. Boyd & Fraser Danvers, MA, 1993, vol. 117. [29] Z. Gu, E. Rothberg, and R. Bixby, "Gurobi optimizer reference manual," URL: http://www. gurobi. com, vol. 2, pp. 1 -- 3, 2012. Fig. 7: Distributed Motion Planning motion planning. By using STL-SpaTeL formulas to specify motion planning and communication requirements and en- coding them into MILP under distributed MPC, the proposed algorithm is able to find online control inputs for each agent distributively such that desired specifications and patterns can be satisfied and hence demonstrates the ability of dealing with large scale systems. The algorithm is validated by a co-optimization simulation for multi-agent system. REFERENCES [1] Y. Cao, W. Yu, W. Ren, and G. Chen, "An overview of recent progress in the study of distributed multi-agent coordination," IEEE Transactions on Industrial informatics, vol. 9, no. 1, pp. 427 -- 438, 2013. [2] C. Baier, J.-P. Katoen, and K. G. Larsen, Principles of model checking. Boston: MIT Press, 2008. [4] C. Belta, A. Bicchi, M. Egerstedt, E. Frazzoli, E. Klavins, and G. J. Pappas, "Symbolic planning and control of robot motion [grand challenges of robotics]," IEEE Robotics & Automation Magazine, vol. 14, no. 1, pp. 61 -- 70, 2007. [5] M. Kloetzer and C. Belta, "Automatic deployment of distributed teams of robots from temporal logic motion specifications," IEEE Transactions on Robotics, vol. 26, no. 1, pp. 48 -- 61, 2010. [6] G. E. Fainekos, A. Girard, H. Kress-Gazit, and G. J. Pappas, "Temporal logic motion planning for dynamic robots," Automatica, vol. 45, no. 2, pp. 343 -- 352, 2009. [7] S. Karaman and E. Frazzoli, "Sampling-based motion planning with deterministic µ-calculus specifications," in Decision and Control, 2009 held jointly with the 2009 28th Chinese Control Conference. CDC/CCC 2009. Proceedings of the 48th IEEE Conference on. IEEE, 2009, pp. 2222 -- 2229. [8] A. Richards and J. P. How, "Aircraft trajectory planning with collision avoidance using mixed integer linear programming," in Proceedings of the 2002 American Control Conference, vol. 3. IEEE, 2002, pp. 1936 -- 1941. [9] S. Karaman and E. Frazzoli, "Linear temporal logic vehicle routing with applications to multi-uav mission planning," International Jour- nal of Robust and Nonlinear Control, vol. 21, no. 12, pp. 1372 -- 1395, 2011.
1505.07257	1	1505	2015-05-27T10:33:31	A First Step Towards Dynamic Hybrid Traffic Modeling	[ "cs.MA" ]	Hybrid traffic modeling and simulation provide an important way to represent and evaluate large-scale traffic networks at different levels of details. The first level, called "microscopic" allows the description of individual vehicles and their interactions as well as the study of driver's individual behavior. The second, based on the analogy with fluidic dynamic, is the "macroscopic" one and provides an efficient way to represent traffic flow behavior in large traffic infrastructures, using three aggregated variables: traffic density, mean speed and traffic volume. An intermediate level called "mesoscopic" considers a group of vehicles sharing common properties such as a same origin and destination. The work conducted in this paper presents a first step allowing simulation of wide area traffic network on the basis of dynamic hybrid modeling, where the representation associated to a network section can change at runtime. The proposed approach is implemented in a simulation platform, called Jam-free.	cs.MA	cs	A FIRST STEP TOWARDS DYNAMIC HYBRID TRAFFIC MODELING Najia Bouha(cid:63)(cid:63) , Gildas Morvan(cid:63), Hassane Abouaıssa(cid:63), Yoann Kubera(cid:63) (cid:63) Univ. Lille Nord France, 59000 Lille, France, U-Artois, LGI2A (EA 3926), Technoparc Futura, 62400 B´ethune, France Email: (hassane.abouaissa, gildas.morvan, yoann.kubera)@univ-artois.fr (cid:63)(cid:63) Universit´e Ibn Zohr, Facult´e des Sciences, Agadir, Morocco, Email: [email protected] 5 1 0 2 y a M 7 2 ] A M . s c [ 1 v 7 5 2 7 0 . 5 0 5 1 : v i X r a KEYWORDS Agent-based modeling; Multi-level modeling; Intelligent transportation systems; Simulation; Traffic flow ABSTRACT Hybrid traffic modeling and simulation provide an important way to represent and evaluate large-scale traffic networks at different levels of details. The first level, called "microscopic" allows the description of individual vehicles and their inter- actions as well as the study of driver's individual behavior. The second, based on the analogy with fluidic dynamic, is the "macroscopic" one and provides an efficient way to represent traffic flow behavior in large traffic infrastructures, using three aggregated variables: traffic density, mean speed and traffic volume. An intermediate level called "mesoscopic" considers a group of vehicles sharing common properties such as a same origin and destination. The work conducted in this paper presents a first step allowing simulation of wide area traffic network on the basis of dynamic hybrid modeling, where the representation associated to a network section can change at runtime. The proposed approach is implemented in a simulation platform, called JAM-FREE. INTRODUCTION Severe congestion is a daily problem which leads to a con- tinuously growth of direct and indirect cost. Traffic congestion which can be recurrent typical to rush hours or non recurrent due to accidents, works, . . . represents a major preoccupation of many transportation institutions and practitioners, calls for an efficient and intelligent dynamic management. Accordingly, several works were undertaken to study the traffic phenomena, to implement effective strategies for an optimal use of the existing infrastructure and to minimize the congestion effects as well as a high quality of service. Nevertheless, the im- plementation of traffic measurement and control algorithms calls for a deep understanding of the traffic phenomena. In this context, traffic flow modeling and simulation play an important role and constitute efficient tools to perform tasks such as traffic prediction and monitoring, traffic control and forecasting, the repercussion of the construction of new parts on infrastructure onto the global behavior of the traffic flow, . . . According to the defined objective, several models have been developed and can be classified into microscopic, mesoscopic and macroscopic models. However, to simulate large-scale road networks, it can be interesting to integrate different representations in the same framework which leads to the so-called "hybrid modeling" as shown on fig. 1. Note that the concept of hybrid modeling has different meanings according to the studied domain. Here, hybrid modeling means the coupling of different models. In this paper, the first step is devoted to the integration of microscopic and macroscopic models into a single framework. The concept of hybrid modeling has been developed by several authors. Hence, some existing hybrid micro-macro traffic models are shown in Table I. simulated entities control strategies macro flow of vehicles ramp metering meso group of vehicles variable-message panels micro vehicle vehicle instrumentation Fig. 1. Hybrid traffic simulation and control approach The models presented in Table I share the same limitation: connections between levels are fixed a priori and cannot be changed at runtime. Therefore, to be able to observe some emerging phenomena such as congestion formation or to find the exact location of a jam in a large macro section, a dynamic hybrid modeling approach is needed. The work presented in this paper is devoted to overcome these shortcomings and proposes the first step towards the development and implementation of dynamic hybrid models. Such a new approach provides an efficient way to change the model Magne et al. (2000) Poschinger et al. (2002) Bourrel and Lesort (2003) Mammar and Haj-Salem (2006) Espi´e et al. (2006) El hmam (2006) Joueiai et al. (2013) micro model SITRA-B+ IDM optimal velocity ARCHISM generic ABM IDM macro model SIMRES Payne LWR ARZ SSMT LWR LWR, ARZ, Payne STATIC MICRO-MACRO TRAFFIC FLOW MODELS PROPOSED IN THE TABLE I LITERATURE level of representation dynamically. As a result of these devel- opment, a platform called JAM-FREE has been implemented. In the following sections, we identify the main issues related to dynamic hybrid traffic modeling and then propose some generic solutions to them. Then, the JAM-FREE simulator is presented, as well as some simulation results. Finally, we conclude and suggest some research perspectives. DYNAMIC HYBRID TRAFFIC MODELING: MOTIVATIONS AND EXISTING SOLUTIONS Motivations Traffic simulation is generally used to: 1) simulate the road traffic flow on very large and complex infrastructures including both city and highways while 2) being able to test the effects of different traffic signs, for example, or traffic assignment and dynamic routing strategies. It provides also a simple way to evaluate the impact of dynamic traffic control (ramp-metering, dynamic speed limit, . . . ) on the whole behavior of the traffic, 3) assess their precise influence on the traffic flow and 4) understand how and why traffic perturbations leading to appearance of congestions do occur. The first goal is easily achieved using macroscopic simula- tion models at the expense of the fourth goal. Conversely, the fourth goal is easily carried-out using microscopic simulation models at the expense of the first goal. To deal with this paradox, and benefit from both approaches, a dynamic hybrid model can be used. The advantages of this hybrid approach include the ability: • To obtain both quantitative and qualitative information about the road traffic, using respectively macroscopic and microscopic representations in the same simulation using the clusters principle. • to switch between these different levels of representation locally depending: -- on the simulation needs; for instance understanding the source of a traffic congestion, -- on the computation constraints; for instance manag- ing the CPU load, • to experiment both macroscopic and microscopic dy- namic routing strategies, i.e. road load balancing strate- gies, and others dynamic trafic control. Use cases examples In order to assess the relevance of our approach for dynamic hybrid simulation, use cases demonstrating the limits of the existing solutions depicted in Table I are presented in this section. Reduce the CPU load during peak hours: The hybrid model can change the traffic representation of a portion of the road network dynamically. This feature can be used to switch from a microscopic representation to a macroscopic one, when the CPU is overused in order to reduce its load. Such situations include cases where the number of simulated vehicles becomes too high to be managed satisfyingly by the CPU (for instance during peak hours). This case is illustrated in figure 2. Balance the CPU load between clusters: The hybrid model can change the traffic representation of a portion of the road network dynamically. This feature can be used to switch from: • a microscopic representation to a macroscopic one when the CPU is overused in order to reduce its load, • inversely (from macroscopic to microscopic), in order to detect the reason for the occurrence of traffic congestion. These features can be used jointly to balance the CPU load between two clusters. Such situations include cases where a traffic jam (or con- gestion) appears in a cluster using a macroscopic model and where the number of simulated vehicles in another cluster is significantly reduced (for instance during off-peak hours). Such situations also include cases where a traffic jam appears in a cluster using a microscopic model and where the number of simulated vehicles in another cluster is significantly reduced (for instance during fluidic period of the flow) as illustrated in figure 3. Find out the reason of the occurence of the traffic jam: Fig- ure 4 shows an example of the occurence of traffic congestions. It demonstrates the relevance of dynamic hybrid simulations to switch between two levels of representation in order to find out the causes of the appearance of these phenomena. Follow moving macroscopic phenomena: The static division of the road network into clusters that can be found in most of the existing approaches causes precision losses. Indeed, a traffic flow simulator has to be able to reproduce macroscopic phenomena such as shockwaves, capacity drop, . . . Yet, the static decomposition found in static hybrid models will cause eventually a loss of quantitative information. To avoid this issue, a dynamic hybrid model gives the ability to move, resize, split and merge the clusters of the road network (see figure 5). Limits of the existing solutions To the best of our knowledge, the only other work describ- ing a dynamic hybrid model is Sewall et al. (2011). However, authors focus on the visualization of data rather than accurate simulation of traffic flow: zooming in a part of the network triggers a micro representation on this area. Thus, there is no bi-directional information exchange between the micro and macro representations. Fig. 2. Illustration of the use case "Reduce the CPU load during peak hours". Fig. 3. Illustration of the use case "Balance the CPU load between clusters". Fig. 4. Illustration of the use case "Find out the reason of a traffic jam". Fig. 5. Dynamic hybrid model of a shockwave phenomenon. DYNAMIC HYBRID TRAFFIC MODELING: ISSUES AND PROPOSED SOLUTIONS Main issues First, we identify the main issues to address in order to simulate dynamic hybrid traffic models accurately. 1) How can the simulated network be structured to switch between different models at runtime? 2) How to switch between different models (particularly from macro to micro)? 3) How the temporalities of the different interacting models can be managed to avoid bias? In the next section, we present possible generic solutions to these issues. Proposed solutions Dynamic decomposition of the network into autonomous clusters: The network can be dynamically decomposed into autonomous clusters. Each cluster operates autonomously and simulates the traffic flow using a model of its own. However, cutting the road network is not arbitrary, to avoid incoherent situation like having an area of a few square meters. To ensure the integrity of the simulation, a minimum cut is deducted from sensors: the entry and exit points of a cluster are necessarily sensors. Thus, the road network can be viewed as a directed graph of interconnected sensors. Each arc of the graph represents a path joining two sensors directly, without going through an intermediate sensor. A cluster is then characterized by input and output sensors and eventually sensors situated inside the cluster. The mini- mum cutting of the road network is then defined by: • a cluster representing the outside of the simulated net- work, • clusters representing minimal subsets in the road network; a cluster is minimal if there is no sensor inside the cluster. Clusters of a simulation are necessarily disjoint combination of minimum contiguous clusters. An example of minimal cutting is depicted in figure 6. Each cluster transmits mean speed and flow information from its bounding sensors to the upstream and downstream clusters. In the case of a macro cluster, this information is integrated using flow conservation and prediction formulas to compute the flow at next time. In the case of a micro cluster, it consists in changing the frequency of generation of vehicles and their speed based on data from the sensors, and measuring the number of vehicles passing the sensors and their speed, for a given time interval to deduct the sensor data. Aggregation and disaggregation of traffic variables: When running the simulation, an agent manages the decomposition of the cluster network. This agent encapsulates rules, determining whether the model used in a cluster must be changed (for example from micro to macro) and determining whether the area of a cluster must be enlarged or reduced according to the needs of the simulation (see the previous section). Switching from micro to macro is straightforward: the mean speed and flow inside the cluster can be directly computed from sensors. On the other hand, switching from macro to micro requires to perform a local warm-up so the speed, density and flow of vehicles can be accurate. Formal multi-level agent-based meta-model: Regular multi- agent based simulation meta-models lack the structure to manage such hybrid approaches: their representation of the agents, the environment and the temporal dynamics of the system is designed to support a single viewpoint. Multi-level agent-based modeling is an interesting approach to simulate such systems. Indeed it offers a large range of techniques to dynamically adapt the level of detail of simulations, couple heterogenous models or detect and reify emergent phenom- ena (Gaud et al., 2008; Gil-Quijano et al., 2012; Picault and Mathieu, 2011). Many meta-models and simulation engines dedicated to multi-level agent-based modeling have been proposed in the literature. See e.g. Morvan (2013) for a complete review. Man- aging multiple viewpoints on the same phenomenon induces the use of heterogeneous time models, thus raising issues related to time and consistency. However, simulation bias can be controlled by rules that constraint perceptions, influence production and reaction computation, according to causality and coherence principles (Morvan et al., 2011). To deal with this issue, we developed a generic approach called SIMILAR to design simulations (see Morvan and Kubera (2014) for more detailed information about the main principles of SIMILAR). Indeed, this approach relies on a multi-level, influence-reaction and agent-based knowledge representation, more fitting to multiple viewpoints (Ferber and Muller, 1996; Michel, 2007; Soyez et al., 2013). THE JAM-FREE SIMULATOR Architecture JAM-FREE is based on a multi-level architecture (see e.g., figure 7): • the infrastructure level models the components of the road network; the naming and identification of road network components follow the S´etra (Service d'´etudes sur les transports, les routes et leurs am'enagements, is a French technical service of the Minist`ere de l' ´Ecologie, du D´eveloppement Durable et de l' ´Energie dedicated to transportation issues) specification (S´etra, 2010). • the control level is the core of the dynamic hybrid approach: it manages the dynamic decomposition of the network into clusters. • the simulation dependent levels model the traffic dynamic using a dedicated representation; by default, three traffic levels are considered: microscopic, macroscopic and data- driven (i.e., using real data). Models used in JAM-FREE Road network structure: We consider that the road network can be divided dynamically into subsets called clusters. The traffic is simulated on each cluster using various heterogeneous models depending on the situation: either a microscopic model or a macroscopic model. For computation time efficiency Fig. 6. Minimal cutting of a network • Highway extraction node Microscopic level models: The behavior of a vehicle can be seen as the sum of three different parts: • A navigation behavior when the current lane of the vehicle does not lead to the desired destination of the vehicle. This behavior consists in changing the lane of the vehicle until the current lane of the vehicle leads to the desired destination. • An overtaking behavior when the vehicle either wants to increase its speed by changing its lane, or when the vehicle has to swerve because a faster vehicle is tailing it. • A behavior when no lane change is required (acceleration model). Fig. 7. Multi-level architecture of JAM-FREE reasons, we choose not to model the road network at a physical level. Indeed, such a model requires the vehicles to interpret a large continuous zone of asphalt with blank lines as separated lanes. Such computations are unnecessary complex, since in our use case (France) rational drivers usually never drive over the blank lines separating the lanes if they are not overtaking. To keep the model simple, the road network is only modeled at a semantic level. The road network is modeled as a set of interconnected roads. Each road contains a set of lanes (usually from 1 to 5). Roads also contain vertical signs (e.g. stop sign, speed limit sign) that are bond either to all the lanes or to specific lanes (e.g. extraction lane in the highway, slow vehicles lane). In our model, we support a specific subset of the french vertical signs. The connection points of the roads are called nodes. We distinguish different types of nodes: • Crossroads nodes • Roundabout nodes • Highway insertion node JAM-FREE implements the highly used Intelligent-Driver Model (IDM) to model the acceleration behavior of vehi- cles (Kesting et al., 2010). IDM is a microscopic traffic flow model, i.e., each vehicle- driver combination constitutes an active "particle" in the sim- ulation. Such models characterize the traffic state at any given time by the positions and speeds of all simulated vehicles. In case of multi-lane traffic, the lane index complements the state description. More specifically, IDM is a car-following model. In such models, the decision of any driver to accelerate or to brake depends only on his or her own speed, and on the position and speed of the "leading vehicle" immediately ahead. The model structure of IDM can be described as follows: • the influencing factors (model input) are the own speed v, the bumper-to-bumper gap s to the leading vehicle, and the relative speed (speed difference) of the two vehicles (positive when approaching), • the model output is the acceleration chosen by the driver for this situation, • the model parameters describe the driving style, i.e., whether the simulated driver drives slow or fast, careful or reckless. Lane changes takes place, if: • the potential new target lane is more attractive, i.e., the "incentive criterion" is satisfied, • and the change can be performed safely, i.e., the "safety criterion" is satisfied. JAM-FREE implements the lane changing model MOBIL (Kest- ing et al., 2007). Chosen acceleration and lane changing models are presented in detail in the JAM-FREE documentation. Macroscopic level model: Any macroscopic traffic flow model can be used in JAM-FREE since it defines generic connectors to convert micro/macro representations. By default, a generic implementation of the METANET model is pro- vided (Messner and Papageorgiou, 1990). METANET is based on a second order macroscopic model and uses the following traffic flow equations: ρi(k + 1) = ρi(k) + [qi−1(k)− qi(k)] Ts Li (1) where, ρi, defines the traffic density in (veh/km/lane). Li, k and Ts represent, the segment length (km), the simulation step and its duration, respectively. qi in (vehicles/h) defines the traffic flow in the segment i: qi = ρivi (2) The speed vi in (km/h) represents the velocity of vehicles in segment i at time kTs. vi = Ve (ρi) (3) where Ve (ρi) is the static speed-density characteristic called a fundamental diagram. The momentum equation is obtained by introducing two small constant parameters in (3). Such equation reads: (cid:124) (cid:125) Ts τ [Ve(ρi)− vi(k)] vi(k)[vi−1(k)− vi+1(k)] vi(k + 1) = Tsη Li (cid:123)(cid:122) (cid:125) convection relaxation (cid:123)(cid:122) (cid:124) − Tsν τLi + (cid:124) (cid:123)(cid:122) ρi+1(k)− ρi(k) ρi(k) + κ (cid:125) anticipation (4) The first term of equation (4) represents relaxation to the equilibrium. This is the most dominant term in the equation since the other terms manifest their effects only when there are (occasional) fluctuations in the traffic speed and density upstream or downstream. This relaxation term describes the fact that the drivers adjust their speeds to the equilibrium speed-density characteristic Ve (ρi) according to the reaction time τ. The second term, represents the convection, i.e., the influence of the upstream traffic mean speed. The third term, called "anticipation", translates the effect of drivers reacting to downstream traffic density variations. Drivers tend to decelerate if the downstream density is higher and accelerate if it is lower. ν, κ are a model parameters. Traffic generation: In our model, each input point is a special connector that will generate traffic on the road it is attached to. Note that such a connector has to be created and put on each lane where the traffic appears. The creation of vehicles is managed by a traffic input point agent. Various implementation of this agent currently exist: • "Flow-mass traffic input point" agents, generating the traffic flow using a flow-mass parameter • "Scripted traffic input point" agents, generating the traffic flow using user-defined events. The created vehicles and the creation dates are manually specified by the users. A first experimentation To validate the models implemented in JAM-FREE, as well as the flow (dis)aggregation algorithms, we performed a first hybrid simulation using real data from of the A25 highway in France. We simulate a 4780m portion of the highway, de- composed into 11 clusters (see fig. 8). The clusters containing insertion or extraction lanes are simulated with a microscopic model, the others with a macroscopic model. Fig. 8. Description of the simulated network The obtained results permit to validate the performance of JAM-FREE as well as its microscopic and macroscopic models. We cannot present in this paper for an evident question of space. The figure 9 shows some simulation results in cluster R8 (i.e., after (dis)aggregation processes), using typical model parameter values (i.e., without specific calibration). them all CONCLUSION AND PERSPECTIVES In this paper a first step towards the implementation of a dynamic hybrid model was proposed. The main objectives of this step were to identify use-cases, propose generic solutions to the dynamic hybrid simulation problem and study the core aspects of the multi-level simulation approach. These ideas have then been implemented in a simulator called JAM-FREE. The first experiments using real data allow us to validate several issues addressed in the paper and the implemented models. Further works will focus on the introduction of several 0.8 0.7 0.6 0.5 0.4 0.3 0.2 32 30 28 26 24 real simulated 2000 4000 6000 8000 (a) density (vehicle/m) real simulated 2000 4000 time (s) (b) speed (m/s) 6000 8000 Fig. 9. Simulation results in cluster R8 traffic models allowing to the practitioners multiple choices in function of their simulation objective and the validation of the identified use-cases. Moreover, a more accurate macro/micro switching algorithm is being developed using a sophisticated continuous warm-up process. In addition, some control strate- gies will be added to JAM-FREE in order to evaluate their impact on the global behavior of the networks. REFERENCES Bourrel, E. and Lesort, J. (2003), Mixing micro and macro representations of traffic flow: a hybrid model based on the LWR theory, 82th Annual Meeting of the Transportation Research Board . El hmam, M. (2006), Contribution `a la mod´elisation et `a la simulation hybride du flux de trafic, PhD thesis, Universit´e d'Artois. Espi´e, S., Gattuso, D. and Galante, F. (2006), Hybrid traffic model coupling macro- and behavioral microsimulation, 85th Annual Meeting of Transportation Research Board. Ferber, J. and Muller, J.-P. (1996), Influences and reaction: a model of situated multiagent systems, 2nd International Conference on Multi-agent systems (ICMAS'96), pp. 72 -- 79. Gaud, N., Galland, S., Gechter, F., Hilaire, V. and Koukam, A. (2008), Holonic multilevel simulation of complex sys- tems : Application to real-time pedestrians simulation in virtual urban environment, Simulation Modelling Practice and Theory 16, 1659 -- 1676. Gil-Quijano, J., Louail, T. and Hutzler, G. (2012), From biological to urban cells: Lessons from three multilevel agent-based models, Principles and Practice of Multi-Agent Systems, Vol. 7057 of LNCS, Springer, pp. 620 -- 635. Joueiai, M., van Lint, H. and Hoogendoorn, S. (2013), Generic solutions for consistency problems in multi-scale traffic flow models - analysis and preliminary results, 16th Int. IEEE Conf. on Intelligent Transportation Systems (ITSC), pp. 310 -- 315. Kesting, A., Treiber, M. and Helbing, D. (2007), General lane- changing model MOBIL for car-following models, Trans- portation Research Record: Journal of the Transportation Research Board 1999(1), 86 -- 94. Kesting, A., Treiber, M. and Helbing, D. (2010), Enhanced intelligent driver model to access the impact of driving strategies on traffic capacity, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineer- ing Sciences 368(1928), 4585 -- 4605. Magne, L., Rabut, S. and Gabard, J. (2000), Towards an hybrid macro-micro traffic flow simulation model, Proceedings of the INFORMS Salt Lake City String 2000 Conference . Mammar, S.and Lebacque, J. and Haj-Salem, H. (2006), Hybrid model based on second-order traffic model, 85th Annual Meeting of the Transportation Research Board. Messner, A. and Papageorgiou, M. (1990), Metanet: A macro- scopic simulation program for motorway networks, Traffic Engineering & Control 31(8-9), 466 -- 470. Michel, F. (2007), The IRM4S model: the influence/reaction principle for multiagent based simulation, Proc. of 6th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS), pp. 1 -- 3. Morvan, G. (2013), Multi-level agent-based modeling - a literature survey, CoRR abs/1205.0561. Morvan, G. and Kubera, Y. (2014), SIMILAR documentation. http://www.lgi2a.univ-artois.fr/∼morvan/similar.html. Morvan, G., Veremme, A. and Dupont, D. (2011), IRM4MLS: the influence reaction model for multi-level simulation, Multi-Agent-Based Simulation XI, Vol. 6532 of LNCS, Springer, pp. 16 -- 27. Picault, S. and Mathieu, P. (2011), An interaction-oriented model for multi-scale simulation, Proc of the 22nd Int. Joint Conf. on Artificial Intelligence, AAAI Press, pp. 332 -- 337. Poschinger, A., Kates, R. and Keller, H. (2002), Coupling of concurrent macroscopic and microscopic traffic flow models using hybrid stochastic and deterministic disaggregation, Transportation and Traffic Theory for the 21st century . S´etra (2010), Identification et localisation sur le reseau routier national, Technical report, S´etra, 110 rue de Paris 77171 Sourdun - France. http://dtrf.setra.fr/pdf/pj/Dtrf/0005/Dtrf- 0005792/DT5792.pdf Sewall, J., Wilkie, D. and Lin, M. C. (2011), Interactive hybrid simulation of large-scale traffic, ACM Transaction on Graphics (Proceedings of SIGGRAPH Asia) 30(6). Soyez, J.-B., Morvan, G., Dupont, D. and Merzouki, R. (2013), A methodology to engineer and validate dynamic multi- level multi-agent based simulations, Multi-Agent-Based Simulation XIII, Vol. 7838 of LNCS, Springer, pp. 130 -- 142.
1003.1810	1	1003	2010-03-09T07:20:49	Reconfigurable Parallel Data Flow Architecture	[ "cs.MA" ]	This paper presents a reconfigurable parallel data flow architecture. This architecture uses the concepts of multi-agent paradigm in reconfigurable hardware systems. The utilization of this new paradigm has the potential to greatly increase the flexibility, efficiency, expandability of data flow systems and to provide an attractive alternative to the current set of disjoint approaches that are currently applied to this problem domain. The ability of methodology to implement data flow type processing with different models is presented in this paper.	cs.MA	cs	Reconfigurable Parallel Data Flow Architecture Hamid Reza Naji International Center for Science and High Technology & Environmental Sciences implementing hardware agents for data/control flow type environments in four models: two models that describe deterministic hardware agents in fine and coarse grain modes; one model of hardware agents handling both control flow and data flow; and one intelligent, non- deterministic model that hints at some of the more advanced possibilities of hardware agent use. Section 4 presents implementation and results. The results of implementing dataflow operations with hardware agents show the high processing speed of input tokens and producing results in compare to software agents implementation. Section 5 provides conclusions. II. Reconfigurable Hardware Agents The current state of Field Programmable Gate Array (FPGA) technology and other reconfigurable hardware [13] makes it possible for hardware implementations to enjoy much of the flexibility that was formally only associated with software. Unlike conventional fixed hardware systems, reconfigurable hardware has the potential to be configured in a manner that matches the structure of the application. In some cases, this can be done statically before execution begins where the structure of the hardware will remain unchanged as the system operates. In other cases it is possible to re- configure the hardware dynamically as the systems is operating to allow it to adapt to changes in the environment or the state of the system itself. In other words, the design of the hardware may actually change in response to the demands placed upon the system throughout the scope of the application. The system could be a hybrid of both low-level hardware based agents and higher-level software implementations of agents which cooperate to achieve the desired results. Implementation of agent techniques in re-configurable hardware[11,12] allows for creation of high-speed systems that can exploit a much finer grained parallelism than is possible with distributed software based systems. It is assumed that an embedded system will be created that utilizes adaptable (reconfigurable) hardware which can be created using FPGA, System on a Chip (SOC)[14], or custom technology. In such an architecture, the functionality of the reconfigurable hardware is controlled by placing design information directly into the configuration memory. In this way, the external environment has the capability to either change the hardware’s functionality dynamically or at the time that the application is created by introducing agents into the appropriate area of configuration memory that controls Abstract- This paper presents a reconfigurable parallel data flow architecture. This architecture uses the concepts of multi-agent paradigm in reconfigurable hardware systems. The utilization of this new paradigm has the potential to greatly increase the flexibility, efficiency, expandability of data flow systems and to provide an attractive alternative to the current set of disjoint approaches that are currently applied to this problem domain. The ability of methodology to implement data flow type processing with different models is presented in this paper. Key Words: Dataflow, Reconfigurable Systems, Multi- agents I. Introduction The focus of this paper is to illustrate how multi-agent concept can be employed within today's reconfigurable hardware design environments for data flow processing. We call these new agents that run inside reconfigurable logic “Hardware Agents [10],” as opposed to the more traditional software agents that normally reside in computer program memory (RAM) and execute using commercially available microprocessors. Such design environments often utilize hardware description languages such as VHDL to capture the design and use synthesis tools to translate this high level description of the system into a low level bit stream that can be used to program the reconfigurable devices. We will utilize and adapt a reduced form of the Belief, Desire and Intention (BDI) architecture [9] for our agents. In this architecture, the term, beliefs, represent the set of working assumptions that the agent has about itself and the environment in which it functions. This forms informational state of a BDI agent -- where such information may be incomplete and inaccurate but often can be modified in a local manner by the agent as a byproduct of the agent's interactions with other agents and the environment. The term, desires, represent the high- level set of objectives and goals that the agent is trying to achieve. The agent’s desires must be realistic and must not conflict with each other. Intentions represent the deliberative state of the BDI agent. It is here that the detailed sequences of actions, called plans, made to the environment and other cooperating agents through actuators are maintained. Section 2 introduces the basic concepts associated with the hardware multi-agent paradigm and reconfigurable computing environment. In section 3, paper illustrates the . (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 244http://sites.google.com/site/ijcsis/ ISSN 1947-5500 the interconnectivity of and functionality the this architecture, reconfigurable hardware. the In reconfigurable to support partial is assumed logic reconfiguration in that it is assumed that segments of its logic can be changed without affecting other segments (for example the Xilinx Virtex-II architecture supports powerful new configuration modes, including partial reconfiguration. Partial reconfiguration can be performed with and without shutting down the device)[13]. Interaction with the external environment is supported by I/O connections made directly to the reconfigurable logic. This allows high speed sensor and actuator operations to be controlled directly by the reconfigurable logic without processor intervention. Figure 1 illustrates a generic dynamically adaptable embedded system environments that support the hardware agent model that is proposed in this paper. In Figure 1 an embedded processor/controller the to is connected reconfigurable hardware in a manner that allows it to alter the configuration memory associated with one or more segments of the partially reconfigurable logic. In this model it is the responsibility of the embedded processor to initiate the configuration of each segment of the reconfigurable hardware by transferring configuration data from processor-controlled memory spaces to the configuration memory using memory mapped or DMA type operations. It should also be noted the that configuration memories are shown as if they were spatially separated from the logic elements that they control but this is usually not the case. In general configuration memory the throughout is dispersed reconfigurable hardware. Figure 1. A Processor-Controlled Dynamically Reconfigurable Embedded System Environment The processing time of a hardware agent can be one or two orders of magnitude greater than an equivalent software agent due to the speed of hardware compared to the speed of microprocessor-based software. This speed could be especially desirable in a real-time system and real-time processing requiring high-speed signal conditioning. In special cases, if the beliefs and the inputs to the agent are expressed as Boolean values, then the function that maps the current set of beliefs and the input data to a new set of beliefs can be implemented as . combinatorial logic. The speed of this implementation would be much faster than performing the comparable operation in software. Likewise, if desires and intentions are both expressed as Boolean values, the function that maps desires into intentions can also be implemented in combinatorial logic; again, at very high speed. III. Design of Multi Hardware Agent Systems to Implement Data Flow Operations The ability of hardware agents to implement data flow type synchronization with different models is presented in this section. This type of synchronization is often employed when to hardware creating modern communicate between asynchronous hardware elements. In a data flow operation, the execution of each operation is driven by the data that is available to that operation. The behavior of data flow operations can be shown by data the data represent flow graphs(DFG) which dependencies between a number of operations. A data flow graph is made up of operators (actors) connected by arcs that convey data. An operator has input and output arcs that carry tokens bearing values to and from the actor. When tokens are present on each input arc and there are no tokens on any output arc, actors are enabled (fired). This means removing one token from each input arc, applying the specified operation to the values associated with those tokens, and placing tokens labeled with the result value on the output arcs. We will present four models to show the ability of hardware agents to implement data flow operations in different scenarios. A. Deterministic Fine Grain Hardware Agents Consider using the dataflow graph shown in figure 2 to 1O . In this dataflow graph there are four find the output 4I ) and 5 nodes(operations). 2I , inputs( 1I , 3I , 1I 2I 4I 3I 5I 6I 2op 4op 1op 3op 5op 1O Figure 2. A sample dataflow graph If we use a multi-agent system to perform this operation, we can implement each of the nodes (operations) with a single agent if we define them at a fine grain level. In this example we use five different agents and each of them runs one single operation as is shown in Figure 3. The agents act in parallel on isolated operations, get information (data) from the environment, and send the results back to the environment. (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 245http://sites.google.com/site/ijcsis/ ISSN 1947-5500 Agent4 4op 2I 3I 4I 5I 6I Agent3 3op 1I According to Figure 4, Agent1,Agent2,and Agent3 sends intermediate results through their TR_Agent port and Agent4 and Agent5 receive this information through Agent1 Agent2 2op their RS_Agent port. Agent1, Agent2, Agent3 can use 1op their Request and Acknowledge signals to interact with the environment (send a request which means the agent is free and ready to receive new tokens, and send an acknowledge which means the agent has received the input tokens). Agent5 Agent5, after processing the final operation, sends the result through its Output port to the environment and 1O informs the environment by setting its done signal. As we mentioned before, the most important advantage of Figure 3. A multi-agent system to implement the data hardware agents for implementing data flow operations is flow operations of Figure 2 the the high speed for processing data inputs and By implementing this data flow graph with hardware producing the outputs. Thus, the speed of information agents we benefit from the speed of the specialized flow can be several times the speed of the flow of hardware in processing the input tokens and producing information when the same flow graph is implemented in the results. Five hardware agents cooperate together to software. form a multi-agent architecture for this data flow graph. Using the reconfigurability of hardware agents we can As Figure 4 shows, in this structure, Agent1, Agent2, and reconfigure the agents in the same multi-agent system to Agent3 receive the input tokens from the enviroment, implement a different data flow graph. For example, process them and send the results to Agent4 and Agent5. different data flow graphs can be implemented using the Finally Agent5 sends the overall result to the same multi-agent system of Figure 4 as we will see later. environment. A signal from the environment activates this In this model agents are small, simple and easy to multi-agent system and when each agent completes the implement for simple operations, but to implement a operation on its input tokens it will set its done signal and complex system we need many agents and a lot of send the value of that result to the agent in the next level. communication between agents with high latency. So, It will inform that agent by sending its done signal to the deterministic fine grain hardware agents system is strobe signal of its successive agent. In this model, suitable for simple deterministic systems. hardware agents use done and strobe signals for handshaking. Multi Hardware Agent System 1I 2I t t n n e e m m n n o o r r i i v v n n Sel Done (1) E E Str (1) m m Str (2) o o r r f f HW Agent5 t t u u p p n RS (1) Out (1) n I I RS (2) 5I 6I activate multi-agent system Figure 4. Multi hardware agent system of figure3 Sel Done (1) Str (1) Req (1) Str (2) Ack (1) Ack (2) HW Agent4 RS (1) TR (1) RS (2) Sel Done(1) Str(1) Req(1) Ack(1) HW Agent1 In (1) TR(1) In (2) Sel Done(1) Str(1) Req(1) Ack(1) HW Agent2 In (1) TR(1) In (2) Sel Done(1) Str(1) Req(1) Ack(1) HW Agent3 In (1) TR(1) In (2) O u t p u t t o E n v i r o n m e n t 1O 5op 3I 4I . (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 246http://sites.google.com/site/ijcsis/ ISSN 1947-5500 B. Deterministic Fine/Coarse Grain Hardware Agents To show the power gained by reconfiguring hardware agents and also to demonstrate how hardware agents can provide support for both fine grain and coarse grain abstractions[15], we implement the data flow graph of Figure 5 using the same multi-agent system of Figure 3. 1I 2I 4I 3I 5I 6I 7I 8I 1op 9I 2op 3op 4op 5op Agent2 Agent3 Agent4 6op Agent1 7op 9op Agent5 8op 11op 10op 3O 2O Figure 5. A sample fine grain/coarse grain dataflow graph As we see in Figure 5, Agent1 and Agent5 are coarse grained agents, with several data flow operations implemented in the same agent,while Agent3, Agent4, and Agent5 are fine grained agents, with only a single data flow operation implemented per agent. They act in parallel in those operations which are not dependent on the other operations, and also cooperate together to find the outputs. The major reason for combining some operations in a single agent is to reduce the hardware interface and the amount of inter-agent communication. This is analogous to the grain packing problem associated with traditional parallel processing problems[16]. We can implement the data flow graph of Figure 5 using the same multi hardware agent system of Figure 4, just agents are reconfigured. In this model with combination of fine and coarse grain agents a trade of between the agent simplicity and communication time depending to the complexity of system can be provided. So, a combination of fine grain and coarse grain hardware agents is suitable for systems consisting both simple and complex deterministic operations. C. Control /Data Flow Hardware Agents This model demonstrates how control flow as well as data flow can be implemented using hardware agents.We consider the following graph (Figure 6) which contains both control flow and data flow. In this system, according to the events in the environment, the control part will . 1op & 3op 6I choose the time that the data flow operations 4op should be activated. 2op & or Events 2I 5I 1I Sensor Agent1 Condition (threshold) 3I 4I 2op 1op 3op 4op Agent2 Agent3 Actuator Fig. 6. A sample dataflow graph We use a multi-agent system with three agents to implement this control and dataflow graph, as shown in Figure 6. In this system, Agent1 is used to implement control flow while Agent2 and Agent3 are used to implement data flow part. The implementation of this control/data flow graph is with a multi hardware agent system similar to the system shown in Figure 4 but with three hardware agents. In this model agents can run both data processing and control on the processing in the same multi-agent system at the same time. So, control/data flow hardware agents provides more independent powerful multi-agent systems. D. Non-Deterministic Hardware Agents This model demonstrates how hardware agents can be used in a non-deterministic and intelligent manner. We consider the control and data flow graph shown in Figure 7 that consists of three hardware agents. Agent1 is non- deterministic and intelligent. It receives input information, and saves its current state in memory. Its new state is a combination of the old state, what it has learned from the environment, and what it has calculated itself. In this system, according to the events in the environment, Agent1 will choose to activate Agent2 or Agent3, separately or in tandem, or will choose not to activate them at all. If Agent1 doesn’t receive any information within a certain period of time it will timeout and take appropriate action relative to the environment, according to its current state. We can define the learning and (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 247http://sites.google.com/site/ijcsis/ ISSN 1947-5500 decision making capability of Agent1 by the following function description: Function HW-Agent (percept) returns action Static: memory ; the agent’s memory memory ← update-mem(memory, percept) ;Learning by perception from environment action ← take-decision(memory) ;Decision-making by its knowledge memory ← update-mem(memory, action) ;Learning by last action return action According to this function a hardware agent can learn and update its memory using its current knowledge and percepts (set of perceptions or inputs) from the environment, its current state and calculations based both on state and environmental input. The agent makes decisions using its total knowledge, environmental, state, and current calculations on both environmental and state. As we see in Figure 7, the possible actions (plans) of this multi-agent system, which its non- illustrate determinism are: 1op → 3op → action Plan1: by cooperation of Agent1 & Agent2 2op → 3op → action Plan2: by cooperation of Agent1 & Agent3 3op → action memory → Plan3: agent 1 3op ) by Agent1(its knowledge & running memory → action Plan4: agent 1 by Agent1(its knowlwdge) With such a non-deterministic structure, the fault tolerant capability of hardware agents can be easily demonstrated. Suppose Agent1 has a timer, which times out after an input token is not received for a period of time. In this case, a value based on previous state or previous outputs can be presented as the output of the system. The implementation of this data flow graph is with a multi hardware agent system similar to the system implemented for Figure 6 but with three reconfigured hardware agents. With considering some situations which are not predefined or cannot predicted mainly in real-time systems then having agents with non-deterministic behavior in the multi-agent system will be useful. So non- deterministic hardware agents provide multi-agent systems with the high capability of responding to the non- deterministic real-time situations. . Events 3I sensor percept 1I 2I Agent1 Agent2 Agent3 Memory actuator 2op 3op 1op Actions Figure 7. A sample non-deterministic dataflow graph IV. Implementation and results Suppose that we are using dataflow operations of Figure 2, multi-agent system of Figure 3 and multi hardware agents of Figure 4 for a data fusion system as 1s and 2s are the sensory inputs to shown in Figure 8. the system. 2s 1s 1s correlation 2s 1s closeness 2s average Agent1 Agent2 Agent3 confidence Agent4 Agent5 Figure 8. Data flow operations of Figure 3 for data fusion In the first level of fusion process, Hardware Agent1 computes the correlation between sensors 1s and 2s by i ba , ( ) using pairs of observation of these sensors using i equation 1. fusion k k k ∑∑∑ b baN a − i i i i 1 1 1 k k k k ∑∑ ∑∑ 2 2 a b aN bN [ [ ( ]) ( 2/12 − − i i i i 1 1 1 1 (1) ]) 2/12 Correlation = (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 248http://sites.google.com/site/ijcsis/ ISSN 1947-5500 ) (3) γ)). ban ,( (2) Thus Correlation=1 means correlation, perfect Correlation= inverse correlation, and -1 indicates Correlation=0 indicates no correlation between the data. In this application it is assumed that inverse correlation is unlikely so that we can use the square of the Correlation, (Correlation)2 to obtain a metric that can be used by advanced stages of the fusion process. Another metric that is computed by Hardware Agent2 is the closeness between two sensors which is defined by equation 2. Closeness Coefficient: Sensor abs Sensor ( − 1 b a −=γ threshold Clossness _ Sensor are the sensors’ values and Sensor and Here, b a Closeness_threshold is the maximum difference between two sensors. If two sensors have the same value then 1=γ , otherwise γ is less than one. Hardware Agent3 is calculating the average of sensors' value (a simple fusion method). Hardware Agent4 determines the confidence factor between two sensors which is defined by equation 3. Factor Min Confidence ba Correlatio ),( ( . = Finally Hardware Agent5 determines the fusion value using the outputs of Agent3 and Agent4. The code for this model of multi-hardware agent data flow fusion system is written using VHDL. In this model the hardware agents set their initial belief set with the first value of the sensors and thresholds for the correlation and closeness of the sensors’ data. They will update their belief set with a new value of its sensor (interaction with the environment) and interaction with the other agents (the main agent will change the correlation and closeness threshold if there is not a high enough degree of confidence of the fusion result). There are several intentions (plans) for this multi-hardware agent system to reach its desire and we assigned each plan to a separate agent to use the collaboration of agents for achieving the global goal or desire which is fusion. The first plan is to determine the correlation between the sensors’ data, the second plan is to find the closeness of the sensors’ data, the third plan is computing the average of sensors' data, and the forth plan is to find the confidence of system. The desire of this multi hardware agent system is to find the fusion value. The similar multi software agent system implemented for this model has beliefs, intentions (plans), and desires exactly the same as the hardware agents. It should be noted that if the software agents were implemented in Aglets [17] or a similar software agent framework, that the use of Java and other overhead would make the software agent slower than the version of the software agent that was implemented in C++. This means that our software agents implemented in C++ are more efficient than most traditional software agent implementations. . This would imply that the speed comparison between hardware and software agents that follows is a more strict speed comparison for hardware agents. The run time of the BDI software agents implemented for this fusion system on a 2.6 GHZ Pentium is 2 us. The run time and speedup of the hardware agent implemented for this fusion system as compared to the equivalent software agent for 8 bit,16 bit, and 32 bit agents for several types of FPGA are shown in Table 1 – Table 3. In 8 and 16 bit modes agents are implemented on Xilinx Virtex II-2v40fg256 and Xilinx- VirtexII-2v10000ff1517 and in 32 bit mode agents are implemented on Xilinx Virtex II-2v500fg456 and Xilinx Virtex II-2v10000ff517. In each table, the first type of FPGA is the minimum size of the FPGA for each agent and the second type is a common large size FPGA. As the results of these tables show, hardware agents are much faster than the similar software agents for the same application. For example, according to the results of Table 1 and Table 3 the speed of an eight bit hardware agent is 80 times and a thirty two bit hardware agent is 19 times that of a similar software fusion agent, using a Xilinx-Virtex 2v10000ff1517. Of course if the software is, for instance, coded and optimized directly in assembler (a software abstraction level similar than the hardware abstraction level managed by FPGAs synthesis tools), all the software layers present in a general purpose computer such as operating systems procedures removed, and the FPGA re/configuration time is taken into account, then the speed up should be a little bit lower. Table 1. 8 bits Hardware Agents 8 bits FPGA Agent speedup run time 26ns 77 25 ns 80 Xilinx- VirtexII 2v40fg256 Xilinx- VirtexII 2v10000ff1517 Table 2. 16 bits Hardware Agents 16 bits FPGA Agent speedup run time 64 ns 31 Xilinx- VirtexII 2v80fg256 Xilinx- VirtexII 2v10000ff1517 Table 3. 32 bits Hardware Agents 51 ns 39 FPGA Agent Xilinx- VirtexII 2v500fg456 Xilinx- VirtexII 2v10000ff1517 32 bits speedup run time 113 ns 17 106 ns 19 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 249http://sites.google.com/site/ijcsis/ ISSN 1947-5500 Table 4 – Table 6 show the number of logic gates used in each device for the implementation and utilization of hardware agents. Device selection varies according to the policy of the design (distributed or concentrated) and the size and the number of agents that we need to build our hardware agent system. For example according to the results of Table 4 and Table 6 we can implement up to 18 eight bit hardware agents and up to 4 thirty two bit hardware agents similar to the data flow operation system of Figure 6 in each Xilinx-Virtex 2v10000ff1517. Table 4. Device Utilization (8 bits HW Agents) Xilinx- VirtexII Resource FPGA 2v10000ff1517 Utilization 21.00 % 4.43 % 4.41 % 2.00 % Used 225 5314 2658 192 Ios Function Generators CLB Slices DFFs or Latches Table 5. Device Utilization (16 bits HW Agents) Xilinx- VirtexII Resource FPGA 2v10000ff1517 Utilization Used 113 10.54 % 2048 1.70 % 1.70 % 1024 160 1.66 % Ios Function Generators CLB Slices DFFs or Latches Table 6. Device Utilization (32 bits HW Agents) Resource FPGA Xilinx- VirtexII 2v10000ff1517 Utilization Used 5.32 % 57 939 0.78 % 0.77 % 471 1.54 % 148 Ios Function Generators CLB Slices DFFs or Latches V. Conclusion In this paper, a general architectural framework for implementing agents in reconfigurable hardware has been presented. Hardware implementations have always been known to be faster than software implementations, but at the cost of great loss in flexibility. The use of reconfigurable hardware added flexibility to hardware, while still retaining most of the speed of hardware. Now the use of hardware agents can greatly expanded this flexibility of reconfigurable hardware. In the future, such improvements to reconfigurable hardware such as faster programming independently and more times, reconfigurable sections in the reconfigurable hardware will make hardware agents even more flexible while . coming even closer to retaining the speed that makes hardware-based implementations desirable. The hardware agents developed for data flow application display many of the features associated with more traditional agents implemented in software. The results of hardware agents implementation in this paper show that the speed of hardware agents can be over an order of magnitude greater than an equivalent software agent the implementation. The parallel nature of reconfigurable hardware would cause this speedup to be than one agent were increased further if more implemented in the reconfigurable hardware. It is believed that the use of hardware agents may prove useful in a number of application domains, where speed, flexibility, and evolutionary design goals are important issues. References [1] Walter B, Zarnekow R. Intelligent Software Agents, Springer-Verlag, Berlin Heidelberg, New York, NY, 1998. [2] Jennings N, Wooldrige M. Agent Technology, Springer-Verlag, New York, NY, 1998. [3] Jennings N, Wooldridge M. Intelligent Agents: Theory and Practice, The Knowledge Engineering Review, 1995; 10(2):115-152. [4 Brooks R. Intelligence Without Reason, Massachusetts Institute of Technology, Artificial Intelligence Laboratory, A.I. Memo, 1991. [5] Ambrosio J, Darr T. Hierarchical Concurrent Engineering in a Multi-agent Framework, Concurrent Engineering Research and Application Journal, 1996;4:47-57. [6] Weiss G. Multiagent Systems-A Modern Approach to Distributed Artificial intelligence, Cambridge: MIT Press 1999. [7] Flores-Mendez R. Towards a Standardization of Multi-agent System Frameworks, ACM Crossroads Special Issue on Intelligence Agents, 1999;5(4):18-24. [8] Jennings N, Sycara K, Wooldridge M. A Roadmap of Agent Research and Development, Autonomous Agents and Multi-Agent Systems Journal, Kluwer Publishers, 1998;1(1):7-38. [9] Rao A. BDI Agents: From Theory to Practice, ICMAS ’95 First International Conference on Multi-agent System, 1995. [10] Naji H. R., Wells B. E., Aborizka M., Hardware Agents, Proceedings of the ISCA 11th International Conference on Intelligent Systems on Emerging Technologies (ICIS-2002), Boston, MA, 2002. [11] Naji H. R., Implementing data flow operations with multi hardware agent systems, Proceedings of the IEEE 2003 Southeastern Symposium on System Theory, Morgantown, WV, March 2003. [12] Naji H. R., Wells B.E., Etzkorn L., Creating an Adaptive Embedded System by Applying Multi-agent (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 250http://sites.google.com/site/ijcsis/ ISSN 1947-5500 Techniques to Reconfigurable Hardware, Future Generation Computer Systems, 2004, (20) 1055-1081. [13] Guccione S. Reconfigurable Computing at Xilinx, Proceedings of Euromicro Symposium on Digital Systems Design, 2001. [14] Becker J, Pionteck T, Glesner M. Adaptive Systems- on-chip: Architectures, Technologies and Applications, 14th Symposium on Integrated Circuits and Systems Design, 2001. [15] Srinivasan V, Govindarajan S, Vemuri R. Fine- Grained and Coarse-grained behavioral partitioning with effective utilization of memory and design space exploratin for multi-FPGA architecture, IEEE Transactions on very large scale integration (VLSI) systems, 2001. [16] Kruatrachue B, Lewis T. Grain Size Determination for Parallel Processing, IEEE trans. On Software, 1998;5(1):23-32. [17] G. Karjoth, D.B. Lange, A security Model for Aglets, Internet Comput, IEEE,1997;1(4):68-77. Hamid Reza Naji is an assistant professor in the International Center for Science and High Technology & Environmental Sciences in Kerman, Iran. His research interests include embedded, reconfigurable, and multiagent systems, networks, and security. Naji has a PhD in computer engineering from the University of Alabama in Huntsville, USA. He is a professional member of the IEEE. Contact him at [email protected] . (IJCSIS) International Journal of Computer Science and Information Security, Vol. 7, No. 2, February 2010 251http://sites.google.com/site/ijcsis/ ISSN 1947-5500
cs/0602056	1	0602	2006-02-15T14:46:05	Building Scenarios for Environmental Management and Planning: An IT-Based Approach	[ "cs.MA" ]	Oftentimes, the need to build multidiscipline knowledge bases, oriented to policy scenarios, entails the involvement of stakeholders in manifold domains, with a juxtaposition of different languages whose semantics can hardly allow inter-domain transfers. A useful support for planning is the building up of durable IT based interactive platforms, where it is possible to modify initial positions toward a semantic convergence. The present paper shows an area-based application of these tools, for the integrated distance-management of different forms of knowledge expressed by selected stakeholders about environmental planning issues, in order to build alternative development scenarios. Keywords: Environmental planning, Scenario building, Multi-source knowledge, IT-based	cs.MA	cs	BUILDING SCENARIOS FOR ENVIRONMENTAL MANAGEMENT AND PLANNING: AN IT-BASED APPROACH1 by Dino Borri and Domenico Camarda2 Dipartimento di Architettura e Urbanistica, Politecnico di Bari, Italy. Abstract Oftentimes, the need to build multidiscipline knowledge bases, oriented to policy scenarios, entails the involvement of stakeholders in manifold domains, with a juxtaposition of different languages whose semantics can hardly allow inter-domain transfers. A useful support for planning is the building up of durable IT-based interactive platforms, where it is possible to modify initial positions toward a semantic convergence. The present paper shows an area-based application of these tools, for the integrated distance- management of different forms of knowledge expressed by selected stakeholders about environmental planning issues, in order to build alternative development scenarios. Keywords: Environmental planning, Scenario building, Multi-source knowledge, IT-based participation 1. Introduction Contemporary spatial planning tries to face the emerging social-environmental crisis by using more sophisticated and participated policies and forms of knowledge, that can be collected through friendly technologies to facilitate interactions. 1 The paper is the result of a common work carried out by the two authors jointly: nonetheless, chapters 1 and 3 have been written by D.Borri, whereas chapter 2 has been written by D.Camarda. 2 Proofs and reprints should be sent to Domenico Camarda, Dipartimento di Architettura e Urbanistica, Politecnico di Bari, Via Orabona n.4, 70125 – Bari (Italy). Email: [email protected]. 1 New government arenas and new forums are opened, in which bottom-up adversary movements tend to be replaced by new institutional configurations of governance, including third parties between state and market, variably organized and associated. Typical interactions, “discourses” of this new type of planning –called either strategic or interactive, communicative, collaborative- call for committed and complex organizations, durable and able to give listening and symbolic answers, more than practical ones. There is the need for learning organizations, new “communities of practice”, new understanding of interaction and linguistic mediation, new –also technologic- abilities of recording knowledge and wills, both expert (concentrated) and common (diffused), new abilities of mediation, negotiation, bargain and field agreement. The important social-environmental objectives of contemporary spatial planning are helped by technological innovation. Noticeable aspects are the growing diffusion of geographical information systems, their increasingly intelligent and interoperable configurations, architectures and functions more and more hybrid, the diffusion of interactive knowledge-acquisition procedures in multi-agent and remote-interaction systems, the return of interest for the procedures of decision and value appraisal-attribution in new and more complex uncertain and multi-agent contexts. In this situation, it is important to develop GISs increasingly dealing with both quantitative and qualitative information, to work upon intelligent systems for monitoring, recognition, classification, decision and management, and to develop organizations and knowledge in complex and often remote interaction networks. Yet, it is important to reflect on the adequacy of technologies to compete with social-environmental problems, that could even be burdened with difficulties to access knowledge and procedures developed in specialized environments, with technologies not sufficiently tested yet. There are some unresolved questions about the collection of knowledge. It is increasingly assumed that “communities of discourse” and “communities of practice” build the space of planning. But the organization of discourse, its rules, opportunities, evolutions, the plurality of discourse agents are still all points to be explored. The same can be said about the communities of practice, whereas 2 today practices in cities are innovated by increasing symbolic transactions and the dematerialization of technologies. Which tools can be set up to make formal/informal knowledge explicit? The production of new forms of knowledge both at individual levels (single groups, organizations, individuals, communities) and at the level of comprehensive, changing, dynamic organizations still represents a barely known field. Learning-skilled organizations represent the contexts in which processes of production of new knowledge are drawn out, knowledge in use is recorded or external knowledge is acquired. What is then the role of planners in those production processes: activating or exploring “communities of practice and discourse”, mediating or settling down differences, facilitating or surveying processes? Some opportunities are offered to the creation of virtual spaces of interaction by new technologies, that allow the large diffusion of spaces for knowledge interaction, transfer and production. However, there are questions that still remain problematic, such as the ones connected with the forms of representation of knowledge, the methods of exploration, the methods of creating possible syntheses, the forms of interaction. Is such a deep penetration into communities really effective, in the processes of explicitation, collection and production of knowledge for planning? Virtual spaces still need to be studied with reference to the mechanisms of cognitive interaction activated/activable in multiple environments, with the aim of identifying the transitions among the different levels of the same environment and among different environments -considering such transitions as processes modifying environments themselves and involved relationships, with reference to cognitive and organizational dimensions. Remarkable problems still remain, such as cognitive scales and hierarchies, and the need to navigate among abstraction levels [20]. They are general problems in intelligent environments but even specific problems in intelligent spatial environments [17], essential for the management of complexity in spatial reasoning [1]. In such virtual spaces it is necessary to support processes of 3 knowledge acquisition, recording, use and transfer with reference to cognitive contexts that are fragmented, conflicting, incomplete, changing and not uniform. How to represent different types of knowledge in a form that does not alter the mutual mechanisms of relationship which characterize the different forms of interaction? Few agents are more easily referable to coordinate and cooperative behaviours, but they are little representative of complex environments where coordination and cooperation are unstable conditions, that can be achieved through a flow of “critical states”. How to choose the number of possible intelligent agents to be set up in order to support interactions and the preservation/creation of relationships? What are the roles to be assigned to such agents? It is still unclear what kinds of cognitive developments do take place in collaborative multi-agent planning environments, and what are their structures and dynamics. An important discussion is currently focusing on the real cognitive transformations taking place in multi-agent environments in front of the extreme organizational, procedural and linguistic complexity of such pluri-logic procedures as compared to the traditional mono-logic ones. This literary question appears to be even more remarkable and perhaps more explorable in planning environments, representing organized institutional environments for spatial transformations. In it, the potentials of multi-agency are conditioned by the forms of power distribution, of deliberative practices, of the strong orientation to solution. In some cases, a useful support for planning is represented by the building up of a shared evaluative and interpretative platform, where the interaction is long lasting, and is commonly defined by scientific literature as a ‘double ring’. Within this interaction, it is possible to modify and fine-tune initial positions, in order to allow the convergence toward a unique strategic approach. In complex regional contexts, and for particularly delicate issues, the ‘ring’ needs to embrace a number of very skilled and rare knowledge sources, represented by scholars scattered around the world and greatly distant from one another. In such difficult cases, the possibility to reach the 4 desired semantic convergence can be enhanced by using IT-based tools, networks, and dedicated software. The work presented in this paper is an area-based application of IT-based tools, made up by using MeetingWorks©, a groupware tool for electronic brainstorming and idea organization, originally designed to facilitating corporate decision making. After this introduction, the first chapter describes the activity of scenario building in the Mediterranean context and its suitability to be aided by the particular IT tool, used for the management of the different forms of knowledge. The chapter also deals with the architecture of the system and the process flow, showing some critical points, both positive and negative, of carrying out a computer-aided participatory session, especially when compared to a traditional face-to-face session. The last chapter draws out some final comments, pointing in particular to some interesting perspectives of IT-based multiple-knowledge management in decision making processes. 2. IT and environmental policies 2.1. Building future scenarios Under a project financed by the European Union, oriented to building sustainable development scenarios in the Mediterranean region, some experiences have been carried out since 1999. A Concerted Action of the EU INCO-DC Commission (Dg XII), the project aims at enabling policies for sustainable development in the Mediterranean region, particularly focused on soil and water. The first activity was concentrated on the Tunis case, and its topic was the interplay between agriculture and urbanization; the activity in Izmir dealt with coastal zone management; the third activity in Rabat dealt with globalization vs. local resources, under the prespective of the emerging Euro-Med free trade zone (see the official website http://www.iamb.it/incosusw for details). The scenario-building approach first used in Izmir, Turkey (2001), then being amended and fine-tuned in the case of Rabat, Morocco (2002), is based on a variant to the strategic choice approach by Friend and Hickling [7], as developed by the recent evolution of futures studies. 5 In general, the whole set of future studies methodologies is really multifarious, ranging during its long history from quantitative to qualitative and hybrid methods. Quantitative methods, classically linked to the Delphi iterative method of obtaining shared convergence on issues, do have peculiar military origins. Rigid forecasts and rational predictions were major aims, and their outcomes ended up being particularly appealing also to private entrepreneural management [23]. Because of the increasing awareness of the importance of multiple knowledge in planning processes, together with a larger acquaintance with the underlying presence of uncertainty, then methods tended to evolve towards softer and more hybrid methodologies. Scenario-building approaches begun to combine expert and local-based forms of knowledge, either complementing them with Delphi or moving toward structured but more qualitative knowledge interaction processes [22]. One of the most advanced results has been a pronounced openness to problem-structuring, more than problem-solving approaches. Hence, a higher stress on participatory arenas, both in knowledge sharing and knowledge enhancing interactions, particularly aiming at turning planning and decisionmaking procedures into democratic processes [13]. Participatory arenas have been frequently indicated as essential in the building up of shared future scenarios, with their consensus-enhancing character. In some particular contexts, where problems and expectations are remarkable, scenarios can be gathered through the so-called future workshops [12]. The Mediterranean context, with its complexity and current criticality, proved to be particularly suitable to such methodology, and in fact that was the methodology selected by the project steering group. The future workshop methodology is outlined in Table 1. It was selected among other scenario- building methodologies after analyzing the particular implementation context under several points of view. As a result, several reasons were fundamental for the final decision on its utilization, ranging from a strong commitment to action, to a significant involvement of non-experts in 6 designing futures, an enhancement of participants’ creativity while building shared visions, a strong orientation to interactive learning, and a useful process flexibility [14]. During the Critique phase, stakeholders are asked to describe major structural changes occurred in the area in the past and major problems of the present situation. As Ziegler points out, this involves a basic hypothesis at the very beginning, i.e., that a broad discontent with the current situation does exist, that creates an intense aspiration toward a development change [24]. In fact, contrasting dissatisfactions and desires are supposed to boost the realization of significant future visions. As operational tool, a modified version of Delphi method is applied in order to acquire a common knowledge base on structural changes and current problems for the subsequent phases3. In the subsequent Imagination phase, the process is carried out in the form of computer-mediated brainstorming session, with the effort to let stakeholders get rid of the current urgencies and to project themselves to the distant future4. Traditionally, this phase is carried out as an oral brainstorming session, helped by notes taken throughout the interaction: computer-mediated interactions should then replace informal contacts and notes with a written exchange of ideas. Although consensus on few visions is a desirable goal, computing should not reduce the richness of images and discourses, in order to prevent limitations on future strategies: Delphi method is then at times replaced or hybridized by more argumentative tools [21]. The Implementation phase is devoted to the transformation of visions into operational strategies. In this phase, participants are solicited first to reflect on possible actions aimed at implementing their future visions, then to single out policies and related resources in order to enforce those actions. As a result, complex strategies are built up, whose aim is to provide implementation alternatives that should support public decisionmaking for the given time span. Although not so creative as the fantasy phase, but rather routinary and structurable, this phase is relationally complex and nested and hard to be tackled by computerized interactions. The whole 7 architecture of the process does then show several elements of complexity and hybridation, with particular concentration in the fantasy and implementation phases. 2.2. Scenario building in the Mediterranean Throughout the INCOSUSW project, future workshops have been set up either as a vis-à-vis, computer-free session or as a computer-aided session, with mixed but encouraging results [14]. In the case of scenario building in Izmir, two parallel sessions were developed in the same time: a vis-à-vis session and a computer-aided session. The two sessions were intended as different methods to implement a similar approach in building scenarios, where stakeholders were involved, coming from institutional boards, NGOs and University. The reasons of exploring the possibility to use a computer program in some phases of the scenario- building activity in Izmir were several. Above all, a major reason was a better manageability of the preliminary Delphi tour for the identification of problematic areas in local coastal zone issues. It was thought that at least some routinary stages, such as the setting up and the exchange of questionnaires among stakeholders, could be accelerated and could facilitate the convergence of outcomes. After a research on software tools available in the market, useful to assist decisionmaking, to build questionnaires and/or to manage multi-stakeholder meetings, it was decided to choose MeetingWorks©, by meetingworks.com. A LAN-based system, MeetingWorks© provides a Chauffeur station for use by the meeting facilitator or an assistant to create a meeting agenda and to run the meeting. Participants have access to Participant stations, where they can enter their ideas, votes, comments and other inputs, all anonymously. The software was conceived as a tool for executive, staff or personnel meetings within firms or societies, where the main aim is to ease the achievement of agreement on projects and issues. In the case of scenario building, instead, the major aim is to enhance the interaction and the share of knowledge from different sources, namely 3 In the case of Rabat, the last twenty years were the time span analyzed. 8 stakeholders, so finally easing the drawing out of alternative community scenarios, development strategies and policies [21]. This difference is important in understanding the pros and cons, as well as the real potential of the tool. In the last case-study of Rabat, an attempt was made to set up a computer-based interaction throughout the whole process. An initial purpose was to start the interactive session basing on a framework of scenarios previously built up by 'expert' stakeholders. They were to be arranged on the basis of data available from local, regional and national government, as well as from knowledge obtained from scientific and other sources. For each group of experts (Public associations, Scholars and academicians, NGOs), the process leading to scenarios can be summarized and divided into 7 main stages (Figure 1): 1) Reflecting on the concerns raised by an upcoming globalization in local economy, society, environment, as well as on issues and problems generated from that (Critique); 2) Generation of desired visions concerning the consequences of globalization, after forgetting the constraints of the present reality (Fantasy); 3) Identification of obstacles and limits to the implementation of visions; 4) Definition of possible actions aimed at overcoming obstacles and achieving the desired visions; 5) Definition of resources to be mobilized in order to implement actions; 6) Aggregation of outcomes into two or three alternative scenarios. 7) Grouping homologous scenarios drawn out by groups, in order to have three or four scenarios at the end of the process. After the completion of the above sequence, scenarios were supposed to be evaluated by a more numerous and assorted group of stakeholders, both experts and non-experts, in order to assess the validity of assumptions, plausibility of development and interpretation of the local context. As a matter of facts, time, organizational and communication constraints among Countries made it impossible to gather the three expert groups: even 'non-expert' (common) stakeholders were 4 In Rabat, the time horizon of scenarios was the year 2030. 9 difficult to be found in sound number. As a result the three parallel processes were unified and carried out by a single group of assorted stakeholders from institutional, academic and NGO sectors together. With reference to figure 1, the actual process architecture can be referred to the 'Group 1' column only, the last stage being the sixth one. 2.3. The interaction process architecture: limits and potentials The Rabat workshop is still under way, so it seems insignificant to discuss deeply the whole experience thoroughly, and to critically compare outcomes with previous case-studies, at this stage5. Nonetheless, an overall description of the system architecture for the defining process of scenarios is useful, in order to explore particularly the relational aspects of multi-source knowledge interactions, as supported by computer technology. The description will be carried out with reference to the routine for the first stage (the critique phase, fig.2), under both technical and substantial point of views. The routine pattern is basically similar for each stage, except for some significant adaptations, due to the different themes dealt with. This induces a problem of imperfect repeatability of routines that is addressed with ad-hoc approaches, described thereafter. 2.3.1. The basic routine. The routine flow is made up of 7 consecutive steps, plus some collateral steps (3a to 6a) aiming at monitoring qualitatively some characters of the learning process, both single and collective. The first step is used to write down the concerns over globalization as well as over productive and urban changes as intended by each stakeholder. Lists are prepared singularly and then merged by the software: interactions are anonymous, but participants' responses may be traced back by an ex-post analysis. In Rabat, stakeholders' interaction has been remarkable in this step: it confirms that free and anonymous computer-based interactions allow a more democratic expression of ideas [21], so enhancing responses (63 out of 11 participants). On the other side, this may induce the 10 unmanageability of outcomes, in that it increases the repetition of similar issues, that voice-based interactions might instead filter and prevent. Therefore, a second step is just needed for the elimination of redundancies, contradictions, inconsistencies among responses, and for the grouping of similar issues under the same heading, in order to have a manageable list for the next steps. In a traditional session this step is typically carried out by means of a general discussion, coordinated by a facilitator and heuristically simplified by informal human interactions. The software does not allow an effective simplification and, therefore, a compromise is currently being carried out, by means of an ex-post simplification of the list, manually grouped and made consistent by the facilitator and his staff. By making this, biases and distortions are not avoided, and the hexogenous synthesis may even create substantial divergences from original ideas. However, this solution seems preferable to the traditional voice discussion because of its better manageability in terms of time and suitability, particularly as regards possible remote, internet-based sessions. In the actual Rabat case, the ex-post grouping was operated by complying with a strict safeguard of the complexity and variety of ideas, so inducing a reduced list of 40 items, with only a 30% reduction rate. Step 3 is aimed at further reducing the list, but basing on more substantial criteria, able to involve the knowledge-based and experiential values of each single stakeholder. As a matter of facts, they are asked to rate (0-5) the list of issues in terms of importance. Basically, this is the first evaluative step of an iteration, where different rating methods are suggested for each step (steps 4, 5a, 6a, 7), aiming both at singling out the top issues and at controlling the consistency of evaluations. The final purpose is to obtain a manageable list of issues based on a shared consensus, as a result of a mutual learning and awareness-enhancing process, not of a merely reductionist procedure. The produced list is not important per se, but in that it involves a learning and knowledge-sharing process, that can be a useful support for the subsequent generation of scenarios [13]. Interestingly, in the Rabat case step 3 did not produce a useful consensus on indicating the most important issues: in fact, three 5 A proper official INCOSUSW report is being prepared on the Rabat workshop, and a further comparative paper will 11 quarters of all concerns top ranked 4.5 to 3. Admittedly, it is often remarked that pure numbers are not able to properly qualify judgement values, and that such kinds of methodologies are therefore intrinsically erroneous [23]. However, recalling the overall process-oriented aim, this first evaluative step may arguably enhance mutual knowledge exchange and awareness on issues, and this should produce useful effects with subsequent iterations. A further evaluative step is then step 4, where stakeholders are asked to scroll the list resulted from the preceding step and to choose the 10 most important issues according to their own views, ranking them (1-10). In the Rabat experience, the agreement on a reduced list proved to be difficult, and only 5 out of 40 items where eliminated from the list by the interactive evaluation. As a matter of facts, classical Delphi literature report a formal impossibility of reaching convergence on ideas at an early stage, and the Rabat occurrence confirmed such position [22] [23]. This induces the need of further iterating the evaluation but, in order to overcome associated time constraints, the possibility of disregarding some low ranking issues is envisioned by the process flow. The dropping-off of the least ranked issues is located in step 5. This phase of selection is manually carried out by the facilitator's staff as a hard cut-off of the top items, basing on the top average rankings. Of course, this operation produces the elimination of ideas that a proper Delphi iteration would instead, probably, enhanced. Being manual and undisclosed, it may jeopardize the original representativeness of issues, especially if least issues remain numerous or there is a high disagreement on them. This last case did occur in the Rabat session, where the 17 selected issues out of 40 were not really founded on a substantial agreement of expressed ideas. This risky operation was then made possible only by the experimental nature of the project, but it can be said that normally such manual simplification may well involve only few low-ranking agreed-upon issues. In step 6, the iterative process started in step 4 is carried on. In fact, after the first ranking, participants are asked to interact and comment with one another by using an embedded tool of the be forthcoming, following previous ones [2] [14]. 12 computer program: the 'chat'. This unstructured interaction allows a certain relaxation in communication schemes, so emulating a voice discussion that is able to enhance the exchange –not necessarily the production- of ideas. It typically safeguards the democratic expression of ideas, since no precedence is given to fluent catalizing leaders but the random rapidity in digitizing. Further, the classical criticality implicitly included in any discussion, i.e., the difficulty in taking memory of all issues and the richness of their complex grey shades, is easily overcome by means of the chat log, a detailed record of all exchanges. In the Rabat case, the chat tool, intrinsically less structured than other evaluative tools, was enthusiastically accepted by participants, who found themselves more comfortable in exchanging informal comments and ideas. However, outcomes were not significantly effective in enhancing the final consensus in that case, so confirming that when disagreement is high the evaluating cycle should be probably reiterated again and again [15]. In that case a foreseeable risk is casting an excessive time burden over the process flow, and then predetermined time constraints may be useful when using the 'chat' tool, although they may raise doubts of mortifying creativity. After chatting, if the disagreement on the outcoming list of concerns is low or fairly non-existent, then the last ranked list is the final one of the whole critique phase, and the process can move to the next phase of the scenario-building activity. On the contrary, if the desired convergence of outcomes is not reached at the end of step 6, then an iteration is started up in step 7, carried out as Delphi tours, in order to reach a possible convergence on a consistent list of items at the very end of the stage. Here a question of time is strongly persistent, and in fact, in the Rabat case, at the end of the secondly iterated step 4, the realization of a further disagreement on the reduction of issues suggested to give up with further attempts and to proceed to subsequent stages with the whole list of 17 resulted issues (concerns). The above main routine is aimed at structuring the proper scenario-building process throughout all the constituting stages of the future workshop. However, a frequently upcoming critique over the communicative approach behind the process is that not the relevance but rather the argumentative 13 ability of each issue can enhance the achievement of the desired convergence [6]. It is clear that such a critique is able to challenge the whole conceptual framework of building scenarios basing on the mutual exchange and enhancement of single knowledge levels of participants. Should it be possible to have an evidence of actual increases of stakeholders' knowledge levels, then the effectiveness of scenario-building approaches could seem less questionable. In order to try to explore such possibility, some collateral steps are then carried out together with the mainstream of the main routine. In particular, where iterative cycles are set up, the ratings of each stakeholder are taken out after each evaluation step, as ex-post by-products of resulting data (steps 3a and 4a). The longer the iteration, the larger the time series on which to carry out the behavioural analysis. In order to have a more complete set of behavioural data during the whole process, other evaluative steps, less structured than proper rating and ranking steps, are complemented with analyzing tools. In particular, steps 5a and 6a are again aimed at looking into stakeholders’ behaviours, even if sympathetically and qualitatively. In these steps, participants are invited to reflect on the impact of the interaction process on their informational level, and to self-rank the level of knowledge gained from the interaction. However, as a whole, these delicate steps are not yet completely defined at present, and the way how to carry out them effectively is still being explored. As a matter of facts, in the case of Rabat both steps 3a to 6a were not performed at all, because of contingent problems, particularly technical and time. 2.3.2. Notes on ad-hoc approaches in subsequent stages. Basically, the above description reveals the intrinsic impossibility of carrying out a perfectly integrated process, fully automated and properly routinary. When moving from the first stage (critique phase) to subsequent phases (fantasy and implementation), the issues that characterize each phase change substantially, since aims themselves are different from one phase to another. The 14 process is therefore sensibly different from the basic one, with the need of amendments and deviations that induce a substantial hybridization of the whole system architecture. Far from being useless burdens, such ad-hoc approaches aim at increasing the interaction effectiveness, in that they enhance the possibility of letting mutual knowledge exchanges cope with the increased complexity of issues at hand. A typical case is the generation of visions connected with future images of a desired future, dealt with in the fantasy phase. In that phase, creativity needs to be largely enhanced, even to the possible detriment of the optimal manageability of process routines. In fact, experience shows the difficulty of letting people freely express their desires if communication tools act as practical limitation and obstacles of the present reality, so threatening the whole scope of work [14]. In some cases, for example, the imagination phase is carried out by relying almost completely on brainstorming and discussion phases, rather than on Delphi tours. Admittedly, even if computer-aided, discussions are in this case difficult to be recorded properly, and only final agreements and collective outcomes become usable data for subsequent phases, so loosing memory for more detailed information. However, generated images and future visions often prove to be more effective and representative, and more useful to hinge subsequent strategic scenarios on. Particularly complex contexts seem to suggest such a more hybrid approach, unlike communities with more standard and regular everyday problems6. In the implementation phase, creativity is substantially less crucial than the previous one, and therefore routinary steps may be more adherently performed. However, since the generation of 'obstacles' toward the realization of future visions is followed by the generation of related 'policies' and then 'resources' aimed at overcoming such obstacles, a possible 'combinatorial explosion' may easily make issues unmanageable by the process flow [14]. Possible further simplification and hybridations may then be needed in order to properly manage the process, to the possible detriment of scenarios richness and representativeness. The final outcomes of the Rabat future workshop will soon provide useful feedback in these fields. 15 2.4. Significant overall features Some positive and negative aspects can be drawn out about the use of such software tools in running computer-aided sessions of scenario-building activity. First, the use of an ad-hoc computer software was helpful in framing scenarios with robust and credible issues: the anonymous interaction among stakeholders enhanced the free democratic expression of opinions, themes and concerns, particularly important in some Developing Countries. Secondly, the computer program could ease the real-time setting up of structured scenario alternatives: the main result of accelerated interactions and feedbacks was then to have a resulting data set easy to be read and managed for future policy actions. A third important feature of a computer-mediated session, when compared to a vis-à-vis session, is that stakeholders can interact without (or at least with a tiny) filter of a human facilitator. This is more likely to happen in a remote interaction, when stakeholders are far from one another, and they are in touch only by means of the computerized session. In this light, MeetingWorks allows the setting up a web-based remote session, either in the same time or in different times, interacting on the basis of an ad-hoc agenda published on an Internet webpage. A last point is the possibility to check, at least in a qualitative way, what level of knowledge can be raised by the whole iterative process. In fact, the above-mentioned increasing criticism is particularly hard to be challenged, also because in traditional vis-à-vis interactions an evaluation of the level of knowledge, reached either by single participants or by the entire group, is in fact difficult to be verified and likely to be influenced by the interaction itself. As a matter of fact, a computer-based interaction is instead mainly mediated by a software tool, then more likely to minimize biases and mutual influences of stakeholders. Therefore, steps of collective evaluation or self-assessment of the level of knowledge acquired can be more likely to be properly effective. 6 In the Rabat case, an hybridation of voice discussions with chat interactions is being explored. 16 The above aspects constitute the potential of the use of computer-mediated interactions: however, there are several important critical aspects that need to be addressed, as well. A first negative point to be underlined is represented by the mediation offered by the computer itself: a cold and non-stimulating medium that is hardly able to make up for warm, informal, sympathetic interactions among humans, as often claimed [3][21]. This occurrence determined several negative by-products, especially after the first intriguing impact: monotony and tedium in replying, scarce involvement, low attention. Shortened stages and discussion intervals can be an answer to such shortcomings, but they are not always applicable in the case of long scenario- building activities, especially if carried out remotely. A second important negative feature of the software is related to a character that typically automated routines lack, when compared to manual, traditional face-to-face interaction: the control of process flows. In fact, whereas in a traditional interaction among stakeholders issues come out during a continuous process, attended and controlled by the attention of facilitators and stakeholders, in an automated routine processes are often invisible, and there is the risk of hiding steps and generating false heuristics. However, this shortcoming cannot be easily challenged within a software conceived for specific and often undisclosed company processes, such as MeetingWorks. A possible hybridation with human control, or with some other complementary software could be a possible solution that is currently being explored. A further point is linked to the need of having manageable lists during the process, by eliminating redundancies and grouping similar issues under the same heading. As mentioned before, the software does not allow an effective simplification, and facilitators need to make a selection of issue categories prior to the session (in the Izmir case, these issue areas were five: Political/institutional, Physical/environmental, Social-cultural, Economic and Others -a free field for uncategorizable issues), which is a compromise between a free identification issue areas –leading to a combinatorial explosion- and a stricter classification of ideas. The result is a sequence of several different steps in the process, where all single issues typically need to be individually addressed to each 17 predetermined category. In this case, the high fragmentation of the sequence into micro-steps mutually complementary enhances the complexity of a reduction activity, so involving a long routinary process, able to cause disorientation and loss of attention and interest on stakeholders. Results typically show in the end a high presence of redundant and inconsistent responses, probably stemming from such induced loss of concentration. In this concern, an automated or semi-automated routine, based on genetic algorithms, would be useful, if properly devised [8][11]. In particular, in a first stage, the recognition of issue areas could be derived basing on the analysis of recurrences of concepts by means of GA-funded procedure of text-analysis software tools. In a second stage, the assignments of statements and issues to issue areas could be realized by analyzing the consistency with a representative criteria for each issue. As happening in all classical GA applications, the GA-based routine should then continue until each condition is met, for each statement and toward the completion of each area. 2.5. Outcomes and remarks After some experiences carried with IT-based scenario building, particularly in Developing Countries, some interesting results can be pointed out. First result is that, although not univocally effective in all situations, the computer allowed the carrying out of rapid and remote sessions (especially Delphi) hard to be developed alternatively. Secondly, the bunch of resulted datasets is given in real time, which is a remarkable outcome mainly because –above all- it saves time for the next hard decisionmaking and policymaking processes. Moreover, apart from consolidated results, there are some suggestions that come out particularly in the case-study of Rabat, even if they cannot be claimed as proper results, since the process is not complete. Among a number of intriguing but fuzzy suggestions, one needs to be mentioned in terms of stakeholders’ learning behaviour. In fact, a quick survey of the partial report of the process seems to show that evaluation criteria slightly changed during the various evaluative session. In particular, while mainly 18 financial/economic criteria were used by the group in the very first evaluative step, collective interactions seem to have determined a progressive shift toward social and environmental criteria. Such kind of result, especially if confirmed by the behavioural process of single participants (still under way), can reveal the importance of multi-discipline group interactions in modifying and perhaps socializing the knowledge level of stakeholders participating in the group. Instead of a simple juxtaposition of different languages, peculiar of the cognitive domains that produced them, the interaction may have allowed a certain level of inter-domain transfers, which would be a very interesting perspective in terms of the underlying knowledge-learning process. Of course, such intriguing results are still vaguely supported at present, and need to be argued and confirmed by further stages of the session and by further analyses and research. As a whole, the experience carried out seems to suggest new perspectives in the quest for building multi-actor, multi-discipline knowledge bases, oriented to policy scenarios. Especially the obstacle represented by distance seems now to be tackled by web-based computerized interactions. This occurrence is extremely significant, meaning that in times of globalization, getting multi-discipline knowledge from scattered stakeholders within structured frameworks is not utopian, and its management by local decision makers is not utopian but possible and useful for local communities. 3. Final comments and research perspectives Sustainable planning is currently looking for an equilibrium of interests between the dimensions of “self” and “hethero”. Varieties, differences and tones are indispensable ingredients for the coexistence and the development of lives in the society and the environment. If the “reductio ad unum” induced by globalization is a threat for localisms and biodiversities, globalization involves also the space for new entities: institutions, concepts, information flows, etc. These entities, by introducing new meanings and potentials from different environments, tend to make and keep them crucial for the existence, by making the link between local and global 19 inextricable (an apparent paradox: the search for the “local” derives from the increasing reinforcement of the “global”). “Self” and “hethero” show also more implications in terms of planning, if we look at plans as micro-anticipations and micro-simulations of knowledge-in-practice within evolutionary systems. As planners, we face complex systems, whose values are changing, adaptive and self-organizing. We must structure and solve problems created by those systems. In fact, both in human (individuals, groups, organizations, economies, etc.) and environmental domains, they are dynamic systems involving the cooperation and/or the conflict among multiple agents with multiple cultures. On the other side, structures and organizations of such problems and systems are created by our dynamic representations and “beliefs”. In problem setting and in problem solving a sort of self- organization is involved, due to learning cycles and to the adaptation of the problem through information sharing, conflict mediation and negotiation, reframing. The role of awareness becomes central. Awareness represents the ability of self-organizing answers, operating through rational/emotional and moral cognition including subjective and qualitative experiences. Awareness means experimenting the “singularity” and thinking of “hethero” as stemming from “self”, assuming the raising out of “one from two” and of “two from one”. This is a rather mysterious dialectic, even if materialism explains it through the emerging proprieties of neural system. In the recent mystic vision of problem-making and decision-making, “singularity” is associated with love, and “separateness” with fear. Possibilities of conflict resolution derive from spiritual “singularity” and depend on the ability of different agents to assume their own diversity as different manifestation of “self”. Once again, learning processes become crucial, since actors learn their competitors’ movements and act consequently. In such environments there is an equilibrium that is able to self-confirm itself, more than what Nash equilibrium does [18]: i.e., a situation giving also a practical knowledge to individuals/elements facing each other, rather than a mere conceptual and theoretical support to decision. 20 However, new internet- or intranet-based forms of negotiation still put questions of problem structuring (or pre-structuring), representativeness of actors, comprehension of texts-languages and different cultures, showing the irreducible harshness of every knowledge-in-practice transaction in multi-agent and multi-actor contexts. Information and communication technologies have impacts on organizational-operational decisions, often giving a decisional environment which is familiar to the more traditional game theory and conflict analysis but requires dynamic approaches careful to non-equilibrium. The development of information and communication technologies and information sciences gives new opportunities to negotiations through forums of “knowledge” and “interests”. Negotiations become increasingly multi-cultural and extended in space and time and such expansion creates in turn inter-cultural and intra-cultural contexts in which negotiations take place. When do practical and successful (fair) negotiations take place in the social-environmental domain planners belong to? How is it possible to adopt integrated negotiation or problem-solving approach (perhaps in a win-win process) to negotiation rather than distributive (win/loose) negotiation? How to build different and compatible objectives? Simple answers do not exist, out of general, somehow moral principles, oriented to putting true cooperation and mutual comprehension at the basis of negotiation. Current literature outlines that inter-cultural negotiations are difficult and a positive development depends on the knowledge of each party about else’s culture, on the way how each negotiator perceives else’s behaviour, on implicit objectives emphasized by each culture. The need to avoid the “Babele effect” and the failure (linguistic confusion producing paralysis) caused by generalized incomprehension is still remarkable. During negotiations, cultural differences (global environments) create difficulties not only in understanding words but also in interpreting actions, in giving shape and substance to the given problem and in the selection of the style of negotiation. According to Faure [5], globalization has brought people closer and supports the management of conflicts through negotiation rather than through the destruction of opponents. As regards 21 negotiation and the role of local/global agents, negotiations have some key elements such as actors, games and strategies (organizations): outcomes (pre- and post-negotiation activities) and in general cultural aspects are crucial. The model of collective decision by Condorcet is different [4]: “the performance of a group in reaching a correct judgement basing on a majority rule will be superior to individual performances, provided that some conditions are respected”. Such conditions need to be explored, and they are generally the following: the probability of a correct decision depends on the technical experience of group voters; voters share a common aim; voters are statistically independent. Also in Condorcet’s model the definition of “correctness” is very complex. In real life it is preferable to replace the concept of correctness with the concept of “expected choice”, that is the choice of voters from an infinitely large population within a polling procedure. This is a change that guarantees that decisions are built through a correct process (socially appropriate) and inclusive, which makes use of pertinent data and information. In fact, the process of group consensus represents paths-toward-truth in contexts of incomplete and partial knowledge of discourse-dependent situations. The integration of multiple-source knowledge as a multiple-objective optimization problem is challenged by large spaces of research and by the difficulty of finding optimal solutions. Moreover, such integration needs to face the continuous risk of redundancy, subsuming and contradiction. It is generally guaranteed by the confidence placed in shaping/solving problems in intelligent environments more than in individual knowledge. Recent developments of artificial intelligence oriented to integrating knowledge from multiple sources include the use of knowledge systems, cognitive dictionaries, datasets, data dictionaries. In them, every knowledge set is provided with a knowledge dictionary, recording characters and classes of every knowledge share: moreover, even this scheme requires datasets and dictionaries itself. Although intriguing, proposed methods for knowledge integration seem to be less important: for example, genetic algorithms are proposed for the selection of the “optimal” set of rules on the 22 basis of a performance evaluation [9][10]. However, there is no agreement on the integration of different dictionaries regulating the interaction of multiple cognitive sources. Approaches discussed above operate by using processing technologies of natural language whose aim is to translate the representation of a source into a final representation that can be integrated in other operational systems. These translations require knowledge structuring through levels of conventional linguistic description. Today, the use of multiple cognitive sources in planning requires peculiar technological organizations making use, in turn, of tools for processing knowledge avoiding easy validations of models and outcomes. Today’s research perspectives in the field of IT-based scenario management and planning are actually oriented to these fronts. References [1] A.Barbanente, D.Borri, G.Concilio, S.Macchi, E.Scandurra, Dealing with Environmental Conflicts in Evaluation: Cognitive Complexity and Scale Problems, in: N.Litchfield et al., Eds. (1998) [2] [3] A.Barbanente, M.Puglisi, C.Tedesco, 2002 K.M.Carley, Communication Technologies and their Effect on Cultural Homogeneity, Consensus, and the Diffusion of New Ideas, Sociological Perspectives 38, No. 4 (1995). [4] M.J.Condorcet, Sketch for a Historical Picture of the Progress of the Human Mind (Hyperion Pr, Winnipeg, 1989). [5] J.C.Faure, Development Co-Operation Report 2000: Efforts and Policies of the Members of the Development Assistance Committee (OECD, Paris, 2001) [6] [7] J.Forester, Planning in the Face of Power, (Berkeley University Press, Berkeley, 1989) J.Friend, A. Hickling, Planning under Pressure : The Strategic Choice Approach (Butterworth-Heinemann, Oxford, 1997) 23 [8] F.Herrera, E.Herrera-Viedma, Linguistic Decision Analysis: Steps for Solving Decision Problems under Linguistic Information, Fuzzy Sets and Systems 115, (2000). [9] F.Herrera, E.López, C.Mendaña, M.A.Rodríguez, Solving an Assignment-Selection Problem with Verbal Information and Using Genetic Algorithms, European Journal of Operational Research 119 (1999). [10] J.H.Holland, Adaptation in Natural and Artificial Systems (Univ. Michigan Press, Ann Arbor, 1975) [11] H.Ishibuchi, T.Murata, I.B.Türkşen, Single-Objective and Two-Objective Genetic Algorithms for Selecting Linguistic Rules for Pattern Classification Problems”, Fuzzy Sets and Systems 89 (1997). [12] R.Jungk, N.Mullert, Future Workshop: How to Create Desirable Futures (Institute for Social Inventions, London, 1996). [13] A.Khakee, “Participatory Scenarios for Sustainable Development”, Foresight 1, No 3, pp. 229-240 (1999). [14] A.Khakee, A.Barbanente, DCamarda, M.Puglisi, “With or without? Comparative study of preparing participatory scenarios using computer-aided and traditional brainstorming”, Journal of Future Research 9 (2002). [15] H.A.Linstone, “The Delphi Technique”, in J.Fowles, Ed., Handbook of Futures Research (Greenwood, Westport, 1978), pp. 273-300. [16] N.Litchfield, et alii, Eds., Evaluation in Planning (Kluver, Dordrecht, 1998) [17] P.A.Longley, M.F.Goodchild, D.J.Maguire, D.V.Rhind, Geographical Information Systems: Principles, Techniques, Management, and Applications (Guilford, New York, 1999). [18] J.Nash, Equilibrium Points in n-Person Games, Proceedings, National Academy of Science (1950). [19] Puglisi 2002 Harmattan [20] E.D.Sacerdoti, A Structure for Plans and Behavior (American Elsevier, New York, 1977). 24 [21] J.A.A.Sillince, M.H.Saeedi, Computer-Mediated Communication; Problems and Potentials of Argumentation Structures, Decision Support Systems 26, No. 4 (1999). [22] J.A.M. Van Dijk, “Delphi questionnaires versus individual and group interviews”, Technological Forecasting and Social Change, No. 37 (1990). [23] A. Woudenberg, “An evaluation of Delphi”, Technological Forecasting and Social Charge, No. 40 (1991). [24] W. Ziegler, “Envisioning the future”, Futures, June, pp. 516-527 (1991). 25 Table 1. The future workshop methodology [19]. PHASE 1. Preparation 2. Critique 3. Fantasy 4. Implementation Future Workshops CONTENTS The issue to be analysed is decided and the structure and environment of sessions are prepared. Clarification, on the issue selected, of dissatisfactions and negative experiences of the present situation. Free idea generation (as an answer to the problems) and of desires, dreams, fantasies, opinions concerning the future. The participants are asked to forget practical limitation and obstacles of the present reality. Going back to the present reality, to its power structures and to its real limits to analyse the actual feasibility of the previous phase solutions and ideas. Obstacles and limits to the plan implementation identification and definition of possible ways to overcome them. EXPECTED RESULTS Summary of contributions. Problematic areas for the following discussion definition. Indication of a collection of ideas and choice of some solutions and planning guide lines.. Creation of strategic lines to be followed in order to fulfil the traced goals. Action plan and implementation proposal drawing. 26 Figure 1. The whole process architecture Group 1 Group 2 Group 3 1. Critique phase 1. Critique phase 1. Critique phase 2. Fantasy phase 2. Fantasy phase 2. Fantasy phase 3. Identification of obstacles 3. Identification of obstacles 3. Identification of obstacles 4. Identif.of actions 5. Identif.of resources 4. Identif.of actions 5. Identif.of resources 4. Identif.of actions 5. Identif.of resources 6. Aggregation into scenarios 6. Aggregation into scenarios 6. Aggregation into scenarios 7. Grouping similar scenarios Alternative scenario 1 Alternative scenario 2 Alternative scenario 3 Fig. 2: Critique phase Step 1 - Listing concerns over globalization list 1 list 2 list 3 list 4 list 5 list 6 list ... list n Generated list of concerns Step 2 - Grouping similar and/or redundant issues under synthetic headings Step 3 – Rating grouped issues rate rate 1 2 rate 3 rate 4 rate 5 rate 6 rate rate ... n Rated list Step 3a -Extracting list with single ratings Step 4 – Selecting and ranking top 10 issues in the list rank rank rank rank rank rank rank rank 1 2 3 4 5 6 ... n Ranked list Step 5 - Dropping off the least ranked issues Step 6 - Free discussion on the outcomes of the ranking (on-line ‘chat’) Step 4a -Extracting list with single ratings Step 5a -Self-ranking level of acq. knowledge Step 6a -Self-ranking level of acq. knowledge Is there agreement about the ranked list? YES Final shared list of concerns over globalization NO Step 7 - Resume Delphi Tour
1603.01787	1	1603	2016-03-06T04:17:56	Approaches to Modeling Insurgency	[ "cs.MA", "cs.CY", "cs.SI" ]	This paper begins with an introduction to qualitative theories and models of insurgency, quantitative measures of insurgency, influence diagrams, system dynamics models of insurgency, agent based molding of insurgency, human-in-the-loop wargaming of insurgency, and statistical models of insurgency. The paper then presents a detailed case study of an agent-based model that focuses on the Troubles in Northern Ireland starting in 1968. The model is agent-based and uses a modeling tool called Simulation of Cultural Identities for Prediction of Reactions (SCIPR). The objective in this modeling effort was to predict trends in the degree of population's support to parties in this conflict. The case studies describes in detail the agents, their actions, model initialization and simulation process, and the results of the simulation compared to actual historical results of elections.	cs.MA	cs	Approaches to Modeling Insurgency Alexander Kott, Bruce Skarin Introduction More writings about insurgency appeared in the last few years than in the preceding hundred years (Kilcullen 2008). The explosion of interest in the subject has much to do with international interventions: insurgency is the single most difficult and frightening challenge that an intervention-military or non- military-may face. For the purposes of this paper, we define insurgency as an organized movement that uses armed violence to overthrow a country's government while often hiding within the civilian population and using civilians to perform combat support functions. The use of civilian population differentiates insurgency from the regular warfare where such an exploitation of civilians would constitute a war crime. Similarly, a rebellion where anti-government forces do not disguise themselves as civilians and fight as a regular, identifiable military is different from insurgency. The significant involvement of civilian population also distinguishes insurgency from a purely terrorist movement, which relies primarily on a tight network of professional terrorists. Although our definition, like any other (e.g., US DoD, 2007), leaves room for gray areas, it serves to emphasize the key feature of insurgency-its reliance upon, and exploitation of civilian population. Because the literature on modeling, simulation, and analysis of regular warfare is vast and readily available, and because insurgencies are often associated with interventions, in this paper we limit our discussion to insurgencies. Insurgency forces may include a combination of the following:  an ideology-based movement that fights to overthrow the current form of the country's government and to establish a different regime;  a personality-based movement driven to install its leader as the ruler of the country;  a religious movement that wishes to defend its religious freedoms or to establish a religion- based regime in the country;  an ethnic minority demanding greater rights or independence;  a regional movement demanding secession or a greater share of the country's resources;  an ethnic majority fighting against the rule of an ethnic minority or a colonial power. Counterinsurgency forces also take a variety of forms:  a democratic state that enjoys the support of a majority of the population;  a dictatorship that relies on coercion to maintain its rule;  a colonial government that represents a foreign power;  a state that receives a limited support of a foreign power but is independent in its actions and could conceivably survive on its own;  a state largely reliant on resources and support of a foreign power. Regardless of the forces on the insurgent and counterinsurgent sides, importance and effectiveness of insurgencies have grown since the Second World War, for numerous reasons. These include the reluctance of Western or Western-supported governments to apply the brutal methods common in prior centuries; the effectiveness, low cost and ease of use of modern small arms like the Kalashnikov rifle (Singer ,2006); the easy availability of arms from a range of state and non-state supporters through channels of modern commerce (Anderson, 2007). An international intervention can be a response to an insurgency, either in support of the insurgent side, e.g., African Union peacekeeping in Darfur since 2004 (Lynch, 2007), or in support of the counterinsurgents (e.g., the US support to the Colombian government fighting the FARC insurgents [Marcella, 2003]). On the other hand, an international intervention is often a cause of an insurgency, or a major factor in changing the insurgency's intensity or character. Thus, a change in insurgency can be both a cause and an effect of an intervention. For example, a diplomatic intervention may induce a third party to discontinue its support to an insurgency; or compel a counterinsurgency-fighting government to conciliate with insurgents. An international famine aid or economic development assistance may reduce populations' grievances and its support to insurgents; but it may also increase the resources available to insurgents through protection racket (Baker, 2009). Similarly, an international informational campaign that condemns an oppressive government may fan the flames of insurgency against the government; yet a campaign in support of a government may convince a part of the population that the government is an illegitimate foreign puppet. Finally, a military or law-enforcement intervention is likely to cause popular resentment at foreign meddling, or drive a segment of population to insurgency by depriving them of their prior privileges and wealth. In turn, insurgency affects other phenomena (Kott and Citrenbaum, 2010). Economic development suffers, and illicit economy flourishes. Political dynamics shifts toward the competing positions on the issue of how to fight or to accommodate the insurgency. Information channels become key tools-and casualties-of insurgents and counterinsurgents. Crime and corruption multiply as all sides may resort to bribes, death threats, protection racket, drug revenue, ransom and extortions. Ethnic, social and religious divisions are exploited and magnified in an insurgency. Qualitative Theories and Models Theorists and practitioners of insurgency and counterinsurgency have outlined a number of key factors that affect the strengths of insurgency. Lenin (Osanka, 1962) stressed the importance of economic and social discontent of masses as a precondition to successful insurgency, as well as presence of a well- organized core of revolutionaries able to mobilize and guide the insurgency. Lawrence (1920, 1935) emphasized the need for an insurgent base inaccessible to the counterinsurgents forces, with protective terrain, adequate supplies of munitions, and at least passively supportive population-a safe haven where insurgents can hide and regroup. He also noted that insurgents benefit when counterinsurgent force relies on a vulnerable technology, such as a railroad. Galula (1964) wrote about the critical role of civilian population who tends to be largely neutral in the conflict and shifts its support to insurgents or counterinsurgents depending on perceived benefits and outcomes of such a support. The population's support also depends on the actions, such as assistance or violent reprisals, taken by either side toward the population. Malayan insurgency (Nagl, 2002) offered the evidence that insurgency loses its strengths when population is physically separated and protected against the insurgents; when counterinsurgents offer economic benefits, security from violence, and political conciliation to the population. In addition, counterinsurgents benefit when they are able to attract a large fraction of population by exploiting ethnic and other differences. Indigenous counterinsurgent forces are more effective than foreign counterinsurgency forces in gaining population's loyalty. Leites and Wolf (1970) point out that insurgency declines when deprived of resource inflows (such as munitions, supplies, and finances) and when its organizational structure and competency are disrupted by counterinsurgents. Respect and fear of government and its forces are important to dissuade population from supporting insurgents (Peters, 2006). War-weariness and anti-war sentiments among the counterinsurgent population and government may encourage and strengthen the insurgency (Iyengar and Monten, 2008; Anderson, 2007). Amnesty, financial rewards and offers of government and military positions can induce insurgents to switch sides (Kahl, 2007). While the aforementioned factors are the most common drivers of insurgency, many other phenomena are important in specific situations. For example, a large pool of displaced persons or refugees can become a highly productive recruiting ground for insurgents as well as an opportunity to skim the foreign food aid (Cuny and Hill, 1999). Large-scale international economic aid programs can become the primary financing mechanism for an insurgency, through protection racket (Baker, 2009). In an attempt to integrate a range of theoretical findings and practical observations, the US military produced a counterinsurgency manual (US Army, 2006), which is in part a comprehensive qualitative model of insurgency. The widely-cited manual identifies multiple factors that encourage and discourage insurgency, stresses that application of force can be a major factor in increasing population's resentment of counterinsurgency, and highlights population's security, good governance and essential services as key factors that diminish the population support to insurgency. Unfortunately, empirical support to qualitative theories of insurgency tends to be anecdotal rather than scientifically rigorous. The work by Iyengar and Monten (2008) is a relatively uncommon example of a model-based, quantitative examination of a qualitative theory. These authors test the argument that anti-war sentiments in the USA emboldens the anti-US insurgents in Iraq and influences them to increase the rate of attacks on the US forces. Iyengar and Monten construct a theoretical model that relates the behavior of Iraq insurgents, specifically the rate of attacks on US forces and on Iraqi Government forces, to their perception of anti-war sentiments in the US. In this model, insurgents are rational, strategic actors who attempt to optimize the distribution of their attacks over time in such a manner that the insurgents preserve their resources while maximizing the anti-war opinions in the US. The authors compute the differences in predictions of the model for different areas of Iraq-some with greater access to information about the US public opinions than others-and compare these estimates with the reported insurgent attacks. They find that in periods immediately after the US media reports a spike in anti-war sentiments, the level of insurgent attacks increases. Also unfortunately, interpretation and application of qualitative models to practical decision-making is an imprecise art. When in 2007-2007 the US decision-makers pondered whether to increase or to decrease the number of US troops of Iraq-the so-called Surge decision (Woodward, 2008)-the qualitative theory was hardly in question. Most likely, all participants in the debate agreed that increasing the number of US troops fighting the Iraqi insurgency may improve the security for a fraction of the Iraqi population; it may also increase the population's anger at foreign occupation; it may also give the Iraqi government additional time to strengthen its political and military posture; or, it may also lull the government into complacent reliance on the US protection. However, the decision-makers and consultants disagreed strongly on the relative quantitative magnitudes of these potential qualitative effects, and on the resulting balance. The overwhelming majority of the US senior military leaders believed that on the balance the Surge- rapid temporary injection of additional US troops into the counterinsurgency efforts-would be counter- productive because it would merely encourage the Iraqi government to continue its complacent dependency on the US (Woodward, 2008; pp.224-281). A small group of civilian theoreticians and retired generals believed otherwise and urged President George W. Bush to accept the Surge plan. In the event, President Bush decided to execute the Surge, and a major reduction of insurgency followed a few months later, in the middle of 2008. Opinions still differ on whether the Surge worked as its proponents expected, or whether other, unrelated mechanisms caused the reduction in insurgency (Woodward 2008; pp.380-384; Pierson 2008). Qualitative models are insufficient to answer such questions; they require quantitative models with the corresponding quantitative metrics, variables, and relationships. Quantitative Measures of Insurgency To construct quantitative model of a complex phenomenon, such as insurgency, one needs ways to measure attributes and dynamics pertaining to that phenomenon. Formulating meaningful metrics of insurgency, however, is a significant challenge. Insurgencies are largely about human perceptions, which are contextual. For example, public opinion about the quality of current situation in a country is highly dependent on past historical experiences and availability of alternatives. Thus, interpretation of a metric's magnitude or event its trend is dependent on other, often intangible variables (Campbell et al, 2009). Most commonly used metrics of insurgency measure level of violence, e.g., the number of insurgent attacks per month; quality of government institutions, e.g., public opinion polls regarding the level of corruption; and strengths of security forces, e.g., the number of counterinsurgent troops and their degree of readiness. For example, Brookings Institution offers comprehensive datasets of metrics (O'Hanlon and Campbell; Campbell and Shapiro) for insurgencies in Iraq (since March 2003) and Afghansistan (since October 2001). These datasets include several dozens of metrics such as fatalities and counterinsurgent troops, number of insurgent attacks of different types, strength of counterinsurgent troops, strengths of anti-insurgent militia, unemployment, electricity generation, inflation, GDP, and public opinion polls. Others begin to explore more comprehensive processes of measuring insurgency, with a special focus on the insurgency's less-tangible aspects like population attitudes and perceptions. For example, the MPICE program (Dziedzic, 2008) has developed a broad-ranging recommendation for gathering a variety of in- depth metrics with computer tools. These would include semi-automated analysis of a country's media content to gauge popular and elite impressions of insurgency-related issues; creation of a panel of experts to assess issues of interest (e.g. the capacity of law enforcement agencies to perform essential functions); and specially constructed public opinion surveys. There is no shortage of complaints about metrics being potentially meaningless and even misleading. For example, Clancy and Crosett (2007) describe history of several insurgencies and find that metrics used in those insurgencies were highly misleading. Still, critics of metrics agree that analysts and decision-makers must look for insightful metrics, and for better means to interpret their meaning. Quantitative models can help do exactly that. Influence Diagrams Also called the causal loop diagram, influence diagram occupies the middle ground between a qualitative model and a quantitative model. Like a qualitative model, an influence diagram describes key aspects of insurgency phenomena and the influences between them. In addition, however, an influence diagram offers features that make it a stepping-stone toward a quantitative model: the diagram names specific quantitative variables, identifies dependent variables for each variable, and specifies whether an increase in a variable causes an increase or decrease in its dependent variable. Figure 1. Key variables and their relations for the insurgency of Anglo-Irish War 1916-1923 In Figure 1, the influence diagram shows key variables and their relations that describe the insurgency of Anglo-Irish War (also known as the Irish War of Independence) of 1916-1923 (Anderson, 2007). Let us begin with the variable called Number of Insurgents, which reflects the number of active anti-British insurgents operating in Ireland. An increase in Number of Insurgents leads to increase in Insurgent Attacks-note that the two variables are connected with an arrow and marked with the plus sign (increase leads to increase). The growing Number of Attacks in turn leads to an increase in British Public War-Weariness (again, the plus sign indicates that increase leads to an increase) and prompts the British forces in Ireland to energize their British Security Measures, which leads to greater Irish Population Resentment and also to more arrests that deplete the Number of Insurgents. Note the last arrow (from British Security Measures to Number of insurgents) is marked with a minus sign, because here increase leads to decrease. And so on. The diagram points to potentially very complex non-linear dynamics of insurgency. Even in this simple model, the number of Insurgent Attacks, for example, affects the Number of Insurgents in five different ways-there are five distinct paths from Insurgent Attacks to Number of Insurgents. The benefits of constructing such a diagram include the following:  The modeler or analyst elucidates and formalizes her thinking about the insurgency phenomena;  Specific variables and their relations are identified and documented;  Qualitative nature (e.g., increase in A leads to decrease in B) of the variable dependencies are determined and documented;  Complex feedback loops and side effects become clearly visible;  Subject matter experts and other analysts and modelers can review and confirm or question the visual representation of the model. Ideally, the modeler derives the influence diagram directly from relevant qualitative theories. For example, Pierson faithfully followed a single qualitative model--the Counterinsurgency Manual (US Army, 2006)--to build an influence diagram of Iraq insurgency (started in 2003), with a large number of variables and influence lines (Pierson, 2008). Choucri et al (2006) formulate their influence diagram of insurgency while rigorously documenting social science theoretical literature in support of each of their model's influences. For example, instead of merely asserting as self-evident the influence "More Insurgents Lead to More Regime Opponents," they cite literature that supports existence of such an influence. They also attempt to justify the validity of variables they introduce into their model. For example, they introduce a variable called State Resiliency and justify it by comparing the State Resiliency to the determinants of civil war of Hegre et al (2001). System Dynamics Models of Insurgency The modeler may continue to develop the model of Fig. 1 by specifying equations that relate each variable to the variables that influence it, e.g., the equation that computes the Number of Insurgents as a function of British Security Measures and of Irish Population Resentment. The resulting system of equations (typically a system of coupled non-linear differential equations) can be solved, for example, by numerical simulation. The solution will shows how each variable evolves over time. System Dynamics (Sterman, 2001) is a technique that simplifies specifying and solving such systems of equations. A variable is represented as a "stock" of goods. Inflows and outflows represent temporal changes to the variable. A "valve" that opens and closes as a function of other variables controls the rate of a flow. Figure 2. A fragment of a System Dynamics model Fig. 2 depicts a fragment of a System Dynamics model that elaborates the influence diagram of Fig. 1. Here, the Number of Insurgents is a stock, or a level of liquid in a reservoir. The incoming pipe carries the flow of new recruits; the valve opens wider when the Irish Population Resentment is greater. One outgoing pipe represents the depletion of Number of Insurgents due to arrests. The valve opening on that pipe depends on the British Security Measures. The second outgoing pipe represents the number of insurgents lost in action, and the valve is controlled by the number of Insurgent Attacks. The modeler must specify an equation for each valve. A computerized System Dynamic tool such as (isee systems, 2009) helps to specify the model and then solves it automatically. System Dynamics is arguably the most popular technique of insurgency modeling. For example, Fig. 1 is partially adapted from Anderson (2007) who constructed a system dynamics model of the Anglo-Irish War, possibly the first modern urban insurgency. Anderson used only a few causal loops-closed paths through a set of variables. One loop represents the insurgency suppression and creation: coercive acts of British forces increase interference in civil life, which increases population resentment, which increases number of insurgents and their anti-British attacks, which leads to increase in British coercive acts. Another critical loop reflects the impact of British war weariness: as insurgent violence increases, British public's war-weariness increases, leading to the public pressure to remove British troops from Ireland. If Anderson (2007) exemplifies a high-level model capturing overall dynamics of the entire insurgency, the work of Grynkewich and Reifel (2006) is an example of a detailed model of a particular sub-feature of insurgency. They model the financial operations and organizational behavior of the Salafist Group for Preaching and Combat (known by its French initials, GSPC). The model relates intensity of insurgent combat operations, expenses to support the operations, and influence of combat operations on population's willingness to support insurgents financially. The key stock in this model is the pool of finances available to GSPC; key inflows include extortion from population, voluntary donations, smuggling operations, kidnappings and ransoms. Outflows include organizational overhead and operational costs. Authors use limited published data and educated guesses to derive values of equation parameters such as the fraction of population willing to donate to insurgents and the amount of donations. The authors report that the model's simulations agree qualitatively with the available information regarding the GSPC's operations and finances. While few would claim that a system dynamics model of insurgency provides a reliable prediction of an insurgency's future evolution, there are other significant benefits in constructing and simulating such a model:  The model helps analysts and decision-makers see un-anticipated side effects, particularly those due to feedback loops;  The modelers and analysts document systematically a formal model that includes a rich, integrated set of factors, processes, and quantitative dependencies;  The simulation of the model illustrates the complexity of the nonlinear temporal dynamics of the insurgency system;  Sensitivity analysis aids analysts and decision-makers in forming insights and intuition toward formulation of intervention plans and policies. Agent Based Modeling of Insurgency Let us return to Fig. 1. Even in this simple diagram, we see several distinct actors, each with its own set of goals, actions, culture, resources, and relations: insurgent organization, general Irish public, British forces in Ireland, British public and British government. We may conclude that this set of actors (agents) is too simplistic: after all, Irish insurgents included various movements with different tactics and leadership, the British public included pro-war and anti-war segments, and the British government included parties with different views on the Irish question, and so on. If we wish to model the dynamic relations and mutual influences of all these agents, we should add multiple variables associated with each of the agents, and influence lines between the additional variables. The model becomes unwieldy. An alternative is to use the agent-based modeling paradigm. An agent-based model consists of agents – software representations of individuals or groups of individuals. Groups can be represented at different scales of abstraction: organizations, segments of populations, ethnic or religious groups, social classes, political parties, or movements, even whole countries. A model may also include elements of the agents' environment. For example, we may want to represent significant geographic areas where the insurgency unfolded: Dublin, Southern Ireland, Northern Ireland, etc. An agent has attributes, e.g., an insurgent organization may be characterized by the number of members, amount of munitions, funds available, level of combat training, political objectives, and organizational competency. An agent has relations to other agents, e.g., an insurgent group may be an ally of another insurgent group. An agent has means by which to make decisions about the actions it will take. Computationally, agents can be implemented, for example, as objects. In that case, an agent has methods by which it makes decision, i.e., choices among available actions. Such a method may include rules or decision-making algorithms, stochastic or deterministic. An agent has a set of actions it can take, e.g., bombing the barracks of counterinsurgent forces, moving itself to another county, or adopting a more positive attitude toward a rival insurgent group. When executed, an action affects attributes and relations of this and other agents. For example, a bombing attack by an insurgent group reduces the strength and tolerance level of counterinsurgent agent, reduces the amount of explosives available to the insurgent group, and increases its morale and reputation for effectiveness. To construct a complete model, the modeler identifies the appropriate set of agents (possibly starting with an influence diagram like Fig. 1), assigns attributes and relations for agents, and codes the methods for agent decision-making and for action impacts. Prior to executing the simulation of the model, the modeler also assigns initial values (i.e., the values at the start of the simulation) of the attributes and relations. The simulation of the model usually proceeds in time-steps. A time-step for modeling insurgency is often a month or a week. At the beginning of the simulation, at week-1, each agent uses its decision-making method to select one or more actions. Then each agent executes the selected actions and each action modifies values of appropriate attributes. This completes the first time-step, and the process repeats for the next time-step, week-2, and so on. Because attributes and relations of agents change with time, in each time-step agents may select different actions (or no action). The simulation ends when agents reach the last time-step. Usually, the analyst who uses the model specifies the number of time-steps. For example, if the analyst studies the potential insurgency effects of an intervention campaign that is to last five years, she may specify 60 time-steps, with each step corresponding to a month of time. At the end of the simulation, the analyst reviews the history of the simulated agents: changes in their attributes and relations over time. For example, an analyst may observe that an agent representing an insurgent group rapidly increases its strength between months 1 and 15, then begins to lose support of local population between months 15 and 20, rapidly depletes its strengths between months 20 and 23, and finally merges with another insurgent group at month 27. Depending on the tool or model used, an agent may have a memory; it may accumulate experiences, learn new rules, and change its beliefs. For example, the CORES system (Kowalchuck et al, 2004) models an agent's belief in its own actions. When an action does not succeed, the agent's belief in the worth of the action diminishes. Thus, a counterinsurgency agent may gradually come to conclude that harsh retributions are not effective. Nexus (Duong, 2007) agent-modeling tool pays even greater attention to the cognitive nature of its agents. A Nexus agent has a degree of historical consciousness; it assigns and reassigns blame for past actions of agents, changes beliefs in trustworthiness of other agents, judges their ideology and looks for friendship with its enemy's enemy. Nexus played a major role in a large-scale, real-world study by a US government agency, in a situation that involved potential international intervention and insurgency. Senturion's agents possess a complex decision-making mechanism that comprises a set of algorithms drawn from game theory, decision theory, spatial bargaining, and microeconomics. Together, they model how agents interact in a political process. This tool has produced multiple real-world predictions of insurgency-driven situations, such as those in Iraq, Palestinian Territories, and Darfur in 2004-2009 (Abdollahian, 2005; Sentia, 2008). Other Modeling Methods and Tools Broad overviews of tools relevant to insurgency modeling are found in (Hartley, 2008; Benedict and Simmons, 2007). Virtually all tools able to generate an anticipatory estimate of an intervention's impact on insurgency fall into one of the two categories we already discussed: system dynamics modeling or agent-based modeling. However, one finds a few exceptions that fall into two other categories: human- driven wargaming and statistical correlations. Human-in-the-Loop Wargaming PSOM model (Parkman, 2005) is an example of a wargaming-based approach-a computerized, time- stepped war-game where human players decide the actions and moves of insurgent and counterinsurgent forces. In PSOM, the geographic area of operations (the wargame board) is divided into 50 km squares. Each square has attributes such as its degree of urbanization, nature of terrain, population density, quality of infrastructure, cultural values, population's perception of security and support to the government. Human players operate the insurgency and counterinsurgency forces. At the beginning of the wargame, the players allocate their respective force units to selected squares of the wargame board. Players assign particular missions to these force units: enforce, stabilize, disrupt, and others. During each time- step, the computer determines the outcome of each force unit's mission based on the current condition in the square, and on actions and strengths of the opponents' forces in the square. The outcome then leads to changes in the square's attributes values. For example, if the counterinsurgent force unit deems successful in its security-enhancement mission, the value of the security attribute in the square increases. Then the game proceeds to the next time-step, and so on. One can notice here a certain similarity to the agent-based modeling, except that in PSOM the human players select the agents' actions (missions and moves), while in the agent-based paradigm agents make decisions without human intervention. Statistical Correlations Application of statistical techniques to historical data on insurgencies yields valuable correlations. Some regularity in data is noticeable even without a formal analysis. For example, (Quinlivan, 1997) offers a compelling visual correlation between an intervention's success and the number of security personnel (military plus police) deployed per thousand of the country's inhabitants. Historically, successful suppression of an insurgency requires about ten or more security personnel per thousand of population. Elbadawi and Sambanis (2000) offer a rigorous quantitative analysis of factors affecting the duration of civil wars (including insurgencies). They find that an external intervention tend to prolong a conflict. They also find a strong U-shaped correlation between the duration of an ethnically based conflict and the ethic fractionalization index. Conflict lasts longer in countries with two or few large ethnic groups than in those with many small groups or a single dominant group. Initialization, Calibration and Validation An insurgency model requires initial values of variables and values of constant parameters or coefficients in equations or rules. The modeler also needs a validation process that shows an acceptable degree of agreement between the model's outputs and data or trends observed in the real insurgency. Often, modelers have to make educated guesses based on very limited data. E.g., Grynkewich and Reifel (2006) are compelled to use a single newspaper quote of an unnamed "Hezbollah operative" to assign a cost to an insurgent operation. When no insurgency-specific data are available, modelers resort to the use of data from domains partially similar to insurgency. For example, Robbins (2005) presents a system dynamics model for reconstruction and stabilization. The model includes an insurgency sub-module that accounts for such factors as ethnic fractionalization, effect of unemployment and urbanization. Lacking insurgency-related data, Robbins uses correlations obtained from studies of crime dynamics. Others derive quantitative parameters using rather sophisticated models, extensive collection of real- world data and comprehensive statistical analysis. For example, Iyengar and Monten use such formidable arsenal of tools to quantify the degree of influence that the apparent lack of resolve among the US public has on the intensity of insurgent attacks. In many cases, the modeler obtains a model's parameters by calibration, i.e., by changing the values until the model's outputs match the available data. For example, Leweling and Sieber (2006) calibrate their model of human resources of an insurgent organization against data derived from publically available news reports, such as numbers of insurgents arrested. They adjust both the structure and parameters of the model in order to obtain satisfactory agreement between the model's output and the data. Although hardly the best practice, some modelers consider calibration identical with validation. The modeler calibrates the parameters of his model, shows a reasonable agreement between the model's output and the available real-world data, and then declares the model valid. Ideally, he should calibrate with one set of data and validate with respect to another set data. Often, unfortunately, only one set of data is available. For example, Anderson (2007) validates his system dynamics model by comparing the model's results with the data describing the Anglo-Irish War of 1916-23. He uses this particular civil war for validation purposes because it is the first modern urban insurgency, and because it is a rare case of a well- documented insurgency. The model was able to replicate approximately the dynamic behavior of the Anglo-Irish War, suggesting a degree of model's validity. Anderson lists numerous parameters and values of these parameters without explaining how he obtained the values. One has to presume that he calibrated the values in a way that maximized the agreement between the model results and the real- world data. Case Study: Conflict in Northern Ireland, 1966-1998 The Republic of Ireland occupies about 80% of the island of Ireland. The remaining northeastern area of the island, the Northern Ireland, is a part of the United Kingdom. Between 1966 and 1997, the Northern Ireland experienced an armed conflict known as the Troubles (Tonge, 2006). Several irregular combatant groups rooted in Irish Catholic population, notably the Provisional Irish Republican Army (Provisional IRA) and the Official Irish Republican Army (Official IRA), fought for reunification of the Northern Ireland with the Republic of Ireland. Their opponents, irregular combatant groups of Protestant origins such as the Ulster Volunteer Force (UVF) and Ulster Defense Association (UDA), fought to maintain the Northern Ireland as a part of the United Kingdom. All irregular combatant groups were in conflict with the government of the United Kingdom and its armed forces. Because the irregular combatants disguised themselves as civilians and relied on widespread popular support, the conflict is best characterized as an insurgency. Political parties in the Northern Ireland were strongly polarized along the lines of ethno-religious affiliation. Pro-British Protestant parties included the Ulster Unionist Party (UUP) and the Democratic Unionist Party (DUP). Supporters of these parties were more likely to sympathize with the UVF and UDA combatants. Irish Catholic population tended to support such parties as Social Democratic and Labor Party (SDLP) and Sinn Fein. The last one is often described as the political arm of IRA, and Sinn Fein's supporters were likely to assist IRA combatants (Silke, 1999). Let us review in detail a model (Grier et al, 2008) that focuses on the Troubles in Northern Ireland starting in 1968. The model is agent-based and uses a modeling tool called Simulation of Cultural Identities for Prediction of Reactions (SCIPR). Our objective in this modeling effort is to predict trends in the degree of population's support to parties in this conflict. In effect, we ask the following question: if we were to have a model like this in 1969, could we predict trends in population's sympathies to political movements like DUP and Sinn Fein. Arguably, insurgents on both sides draw their strength from population segments that identify with extremes of the political spectrum. If we can predict trends in extreme political opinions, we are better prepared to anticipate changes in the strength of insurgency. Another important question is how much data we need to construct and simulate such a model. Models that require less data are less expensive to construct, and easier to understand. In this case study, we use a rather simple model that requires little data. We find, encouragingly, that the simple model with a modest amount of readily available data produces potentially useful predictions of trends. Agents Using SCIPR, we construct about 5000 agents that represent the entire population of Northern Ireland. Each agent represents a group of approximately 300 individuals with approximately similar identities, residing near the same locale. An agent has several attributes including the district of residence (one of the 26 districts),   ethno-religious affiliation (Catholic or Protestant),  a number between 0.0 and 1.0 representing the agent's opinion on the issue of Northern Ireland affiliation (0.0 means Northern Ireland must remain British, 1.0 means Northern Ireland must unite with the Republic of Ireland), and the political party that the agent supports.  An agent has social links to an average of ten other agents, of which 90% are of the same religion. These networks are fixed and are formed under the assumption that individuals in Northern Ireland have around ten people that they are in regular contact with to discuss political issues, and that they are generally of a similar identity. Using these links, an agent can communicate its political opinion to other agents. There is no particular theoretical basis for using 5,000 agents, and not 500 or 50,000, but for simulations of less than 1,000 agents we find the statistical variance of the generated agent population from the input distributions is significant. With simulations greater than 5,000 there is no significant difference in either the initialization or outcome. Therefore, the number 5,000 is merely a modeling assumption, and it seems to work for our purposes in this model. Note what we do not attempt to model: we do not model political movements explicitly, nor political leaders, nor other influential countries like the Eire or the United States, nor the rest of the Great Britain. Neither do we represent explicitly the insurgency groups, counterinsurgency forces, economics, and many other potentially significant factors. Agent Actions An agent can perform several actions:     communicate its current political opinion to another agent through an existing link we mentioned earlier, change its political opinion on the question of Northern Ireland affiliation in response to receiving an opinion from another agent, change its political opinion on the question of Northern Ireland affiliation in response to the news of a latest sectarian killing, and change its party affiliation. Figure 3. A model of an agent's change of its political opinion. An agent changes its political opinion in the manner depicted in Fig. 3. An agent's opinion is characterized by the opinion number, e.g., 0.5, and the opinion's confidence bounds, e.g., (0.25; 0.75). When the agent receives an opinion, .e.g., 0.7, from another agent within his social network (i.e., connected by an existing link), the receiver modifies its opinion partially, in the direction of the sender's opinion. When the sender's opinion is outside the receiver's confidence bounds, the receiver ignores the sender's opinion. This model, and the opinion scaling equations, follows largely (Friedkin, 1999; Hegselman and Krause 2002). An agent also changes its political opinion in response to events, in this case, the latest episode of sectarian killings. Responses to the event are specified by identities as a distribution of reactions to an opinion. After the occurrence of an event, agents within the region of the event sample a value from the distribution of reactions for their identities. This value is then used to scale the maximum opinion change parameter and added to an agent's current opinion. This new opinion is then evaluated using the same bounded confidence procedure described above. In this model, when for example a Catholic is killed, agents of the same religion increase the strength of their opinion in favor of the united Ireland. An agent switches to support of another party when the agent finds its opinion more closely aligned with those of the members of another party than with the members of the agent's current party. For the sake of clarity and brevity, we omit a few other relevant details, such as an agent's change of opinion in response to a sectarian murder. Model Initialization We initialize our model to resemble the Northern Ireland of 1968 by assigning each agent the values of its attributes. We pick the values stochastically but our distributions are such that the total numbers of Protestant and Catholic agents, as well as fractions of supporters of each party in each county correspond to the demographic and voting data of 1968 (Fig. 4). Figure 4. Demographic and voting categories. We also create a social network of agents: we assign a link to a randomly chosen pair of agents in such a manner that each agent has on average about 10 links, with about 45% of the links connecting to agents within its immediate vicinity, 30% to other agents residing in the same district, and the rest---to agents in other locales of Northern Ireland. We also ensure that about 90% of the links are between agents of the same religion. These parameters-10, 45%, and 30%--are merely modeling assumptions, without a substantive empirical basis. A more rigorous modeling effort should consider obtaining empirical data to support and improve these assumptions. We also initialize a table of sectarian killings: for each day, how many (if any) Catholics and Protestants perish in inter-communal violence. Our model does not predict such data; they have to come from a separate source. Now we have the initial attributes of agents and their links fully defined; the model is ready for simulation. Northern Ireland Population by ProvinceProtestant CatholicUnionist BlockNationalist BlockModerateStrongly ForStrongly OpposedOpinion toward a united IrelandOther Nationalists SDLP Sinn Fein DUP UUP Other UnionistOtherAPNIMale FemaleSimulation Process The simulation algorithm begins its process on the first day of the year 1969, in simulated time. The algorithm randomly picks a number of pair of linked agents. In each pair, the algorithm randomly designates one of the agents to be the sender and another as the receiver of the political opinion. The number of pairs selected on each day is such that the every agent, on average, acts as a sender approximately every three days. This parameter-three days-is merely a modeling assumption; a more rigorous model should test the empirical validity of this assumption. When selected by the algorithm, the sender communicates its current political opinion to the receiver. The receiver then either ignores the received opinion, or shifts its own opinion partially toward the sender's opinion as we discussed above. The simulation algorithm also notifies each agent of sectarian killings that occurred on that day. Agents adjust their political opinions correspondingly. The simulation algorithm then performs the same process on the next day of the simulated time, and so on. At the end of each year, each agent re-evaluates the alignment of its political opinion with the members of political parties. Depending on the alignment, the agent changes its party support. The algorithm stops at the end of the year 2005, as instructed by the modeler, and outputs the results. Results of the Simulation Fig . 5 and 6 compare the historical results of elections and polls in Northern Ireland with the results of our simulation. Figure 5. Historical results of elections. Historical Elections0%20%40%60%80%100%19691973197419751979198119831985198719891993199619971999200120042005DUPUUPOther U.APNIOther NonOtherOther N.SDLPSF Figure 6. Simulated results of elections. On the one hand, there is a notable similarity in general trends. For example, the simulated growth trends in the levels of support for Sin Fein post-1981 and for UUP in years 1973-1993 are broadly consistent with actual historical data. The relatively constant levels of simulated support for DUP and SDL are also comparable to real history. On the other hand, the simulation clearly arrives to a steady state after1994 and seems unable to project any further changes. In our experience, this is a general limitation of the Bound Confident model which tends to converge to an artificial steady state after a period of simulation. The simulation also strongly overestimates the support to moderate parties, those between UUP and SDLP. Still, it is encouraging to see that an admittedly simple model shows the ability to predict trends in population's sympathies nearly 15 years forward! It is even more remarkable considering that the modelers used very small amount of readily available information: elections results and basic demographics. Simulated Elections0%20%40%60%80%100%19691973197419751979198119831985198719891993199619971999200120042005DUPUUPOther U.APNIOther NonOtherOther N.SDLPSFAs of this writing, extended models of this type, e.g., (Kott and Corpac, 2007; Kott et al, 2007; Kott et al, 2010) are used to study practical problems of modern insurgencies. Practical Tips  Insurgencies are diverse, and no ready-made model will represent all features of the particular insurgency and situation that interest your customer (here, customer is the organization that wishes to use the results of your analysis for applied decision-making). When selecting a model for your analysis, look for one that allows you to make major extensions and modifications.  A number of simple yet insightful models of insurgency have been published by academic researchers such as the Center for Contemporary Conflict of the Naval Postgraduate School (http://www.nps.edu/Academics/centers/ccc/). Consider using one of such models as an initial baseline, then experiment and gradually extend it for your purposes.  The ability to trace the chain of influences in analyzing results of a simulation is critically important. System Dynamics models tend to be better in this respect.  A modeler with little experience in insurgency modeling will find System Dynamics models easier to construct, to understand, and to debug.  Many insurgencies, and many strategies for managing insurgencies, cannot be properly modeled without representing multiple players-individuals and groups-in the insurgency. In such cases, an agent-based model is most appropriate.  Always model international actors' influences on insurgents and counterinsurgents. Few insurgencies ever unfold without a major impact by one or more international participant.  Before constructing a model, ask your analysis's customer about the actions her organization considers taking with respect to the insurgency, or within an insurgency-plagued region. Make sure your model is capable of representing those actions and their effects.  For every action your customer plans to take, include modeling of undesirable side-effects. Thus, if the analysis' customer plans to provide food aid to refugees, include in your model a possible diversion of food by insurgents.  In selecting variables and attributes for your model, give preference to most tangible and measurable ones. Include the metrics that your customer expects to use or to affect. Also, consider including historical or current data available to you for calibration and validation.  Allocate adequate time and resources to review and improve your model with Subject Matter Experts (SMEs). Produce model's visualizations specifically designed for easy comprehension by SMEs.  Insurgencies are emotional topics, and SMEs tend to hold strong, passionate opinions. Collaborate with several SMEs and welcome widely divergent views. Try to find SMEs who had experiences on both sides of an insurgency.  Early in the modeling project, work with insurgency SMEs to create a set of test cases. Ask several SMEs, independently, to produce their estimates of probable evolution of the insurgency in each case. Expect to receive widely divergent estimates. Resources Conflict Analysis Resource Center http://www.cerac.org.co/datasets.htm Pointers to multiple depositories of datasets related to armed conflicts, inlcuding insurgencies Ethnicity, Insurgency and Civil War Project http://www.stanford.edu/group/ethnic/publicdata/publicdata.html data relating ethnic fractionalization and insurgency The American Political Science Association, Task Force on Political Violence and Terrorism http://www.apsanet.org/content_29436.cfm Pointers to multiple depositories of datasets related to political violence, inlcuding insurgencies David C. Gompert and John Gordon IV, "War by Other means," RAND Publication 2008 http://www.rand.org/pubs/monographs/MG595.2/ Appendices A and B provide data and analysis of outcomes and correlations for 89 insurgencies, and data on counterinsurgency capabilities of world states and organizations Political Instability Task Force http://globalpolicy.gmu.edu/pitf/pitfdata.htm http://globalpolicy.gmu.edu/pitf/pitfpset.htm Multiple datasets related to internal wars, including insurgencies Minorities at Risk Data http://www.cidcm.umd.edu/mar/data.asp Datasets characterizing multiple minorities, their conditions and potential risks, including potential or ongoing insurgency Genocide and Politicide project http://globalpolicy.gmu.edu/genocide/ Data describing cases of genocide, many related to insurgencies War and Health website http://warandhealth.com/civilian-victims-in-an-asymmetrical-conflict-data/ Civilian victims in an asymmetrical conflict http://www.systemicpeace.org/inscr/inscr.htm Armed Conflict and Intervention (ACI) Datasets http://www.systemicpeace.org/ multiple datasets on state instability and conflict Correlates of War Project (COW) http://www.umich.edu/~cowproj/dataset.html Wolrd Bank Datasets http://econ.worldbank.org/WBSITE/EXTERNAL/EXTDEC/EXTRESEARCH/EXTPROGRAMS/EXTCONFLICT/0, ,contentMDK:20336174~menuPK:637270~pagePK:64168182~piPK:64168060~theSitePK:477960,00.html Data on civil wars post-WWII Center for Computational Analysis of Social and Organizational Systems (CASOS) at Carnegie Mellon University. URL: http://www.casos.cs.cmu.edu/computational_tools/tools.html Listings and/ or repositories of software tools and libraries International Network for Social Network Analysis (INSNA). URL: http://www.insna.org/software/index.html Listings and/ or repositories of software tools and libraries Network Workbench (NWB) Tool https://nwb.slis.indiana.edu/community/?n=Main.NWBTool network analysis, modeling, and visualization toolkit References Andrew Silke, Rebel's dilemma: The changing relationship between the IRA, Sinn Féin and paramilitary vigilantism in Northern Ireland, Terrorism and Political Violence, Volume 11, Issue 1, 1999, Pages 55 – 93 Jonathan Tonge, Northern Ireland, Polity 2006 Rainer Hegselmann, Ulrich Krause, OPINION DYNAMICS AND BOUNDED CONFIDENCE MODELS, ANALYSIS, AND SIMULATION, Journal of Artifical Societies and Social Simulation (JASSS) vol.5, no. 3, 2002 http://jasss.soc.surrey.ac.uk/5/3/2.html Friedkin, N. E. (1999). Choice Shift and Group Polarization. American Sociological Review, 64(6), 856-875. Elbadawi, I. and Sambanis, N., "External Interventions and the Duration of Civil Wars," Working Paper 2433, The World Bank, September 2000 D. DUONG, R. MARLING, L. MURPHY, J. JOHNSON, M. OTTENBERG, B SHELDON, S STEPHENS, "NEXUS: AN INTELLIGENT AGENT MODEL OF SUPPORT BETWEEN SOCIAL GROUPS," in Proceedings of the Agent 2007 Conference on Complex Interaction and Social Emergence, Evanston, Illinois, November 15-17, 2007, pp.241-246 isee systems, Stella Software Website, 2009, http://www.iseesystems.com/softwares/Education/StellaSoftware.aspx MICHAEL DZIEDZIC, BARBARA SOTIRIN, JOHN AGOGLIA, MEASURING PROGRESS IN CONFLICT ENVIRONMENTS (MPICE), Report ADA488249, US Army of Corps of Engineers, 2008 Gabriel Marcella, THE UNITED STATES AND COLOMBIA:THE JOURNEY FROM AMBIGUITY TO STRATEGIC CLARITY, Strategic Studies Institute of the U.S. Army War College, May 2003, http://www.strategicstudiesinstitute.army.mil/pdffiles/PUB10.pdf P. W. Singer, Children at War, University of California Press, 2006, p. 46 Parkman, J., (2005). Peace Support Operations Study. McLean, Virginia, MORS Workshop on Agent- Based Models and Othyer Analytical Tools in Support of Stability Operations. Hartley 2008 http://www.mors.org/meetings/es_2008/pres/hartley_dime.pdf John Benedict, L. Dean Simmons, Enterprise-Wide Opportunities for, Advancing Irregular Warfare Analyses, 13 December 2007 Brief at MORS Workshop Collier, P., Hoeffler, A. and Söderbom, M. 2001. "On the Duration of Civil War." Policy Working Paper 2861, World Bank, Washington DC. Mark Abdollahian, Michael Baranick, Brian Efird, and Jacek Kugler, A Predictive Political Simulation Model, Center for Technology and National Security Policy, National Defense University, 2005 http://www.ndu.edu/ctnsp/Def_Tech/DTP%2032%20Senturion.pdf Sentia Group, Implications of a U.S. Drawdown in Iraq, July 2008, Report, Sentia Group, Inc., http://www.sentiagroup.com/pdf/SentiaInsightMonthly-USDrawdownInIraq-July2008.pdf Sterman, John D. (2001). "System dynamics modeling: Tools for learning in a complex world". California management review 43 (1): 8–25. Jason Campbell, Michael O'Hanlon, Jeremy Shapiro, Assessing Counterinsurgency and Stabilization Missions, POLICY PAPER Number 14, May 2009, The Brookings Institution Washington, D.C. JAMES CLANCY and CHUCK CROSSETT, Measuring Effectiveness in IrregularWarfare, Parameters, Summer 2007, pp.88-100 Jason H. Campbell and Jeremy Shapiro, Afghanistan Index: Tracking Variables of Reconstruction and Security in Post-9/11 Afghanistan, a website, http://www.brookings.edu/about/programs/foreign- policy/afghanistan-index Michael E. O'Hanlon, Jason H. Campbell, Iraq Index: Tracking Variables of Reconstruction & Security in Post-Saddam Iraq, a website, http://www.brookings.edu/iraqindex F. C. Cuny and R. B. Hill, "Famine, Conflict and Response: a Basic Guide," Kumarian Press, West Hartford, Connecticut, 1999 Nagl, John A. Learning to Eat Soup with a Knife: Counterinsurgency Lessons from Malaya and Vietnam. Chicago: University of Chicago Press, 2002. Lawrence, T. E. "Evolution of a Revolt." Army Quarterly and Defense Journal. Tavistock, United Kingdom (October 1920). Accessed via Command and General Staff College, US Army, Ft Leavenworth KS, at http://cgsc.leavenworth.army.mil/carl/resources/csi/ Lawrence/lawrence.asp on 23 April 2008. Lawrence, T. E. Seven Pillars of Wisdom: A Triumph. Oxford: Alden Press, 1935. Leites, Nathan and Wolf, Charles. Rebellion and Authority: An Analytical Essay on Insurgent Conflict. Arlington: RAND Corporation, 1970 (R-462-ARPA). Long, Austin. On "Other War" Lessons from Five Decades of RAND Counterinsurgency Research. Arlington: RAND Corporation, 2006. Galula, David. Counterinsurgency Warfare: Theory and Practice. Westport, CT: Praeger, 1964. Peters, Ralph. "The hearts-and-minds myth - Sorry, but winning means killing." Armed Forces Journal (September 2006). Accessed via News Bank, Inc. - Armed Services and Government News at http://infoweb.newsbank.com/iw-search/we/InfoWeb on 7 March 2008. Pierson, Brett M. "OSD/Joint Special Session: Irregular Warfare Activities in OSD and the Joint Staff," Proceedings of the 76th Annual Military Operations Research Society Symposium. Alexandria, VA: MORS, 2008. Edward G. Anderson Jr, A Proof-of-Concept Model for Evaluating Insurgency Management Policies Using the System Dynamics Methodology Strategic Insights, Volume VI, Issue 5 (August 2007) Baker, A., "How the Taliban Thrives," Time, vol.174, no.9, September 7, 2009, pp.46-53 Choucri, N., C. Electris, D. Goldsmith, D. Mistree, S. Madnick, J. Morrison, M. Siegel, & M. Sweitzer- Hamilton. 2006. "Understanding & Modeling State Stability: Exploiting System Dynamics." Proc. of 2006 Institute of Electrical and Electronics Engineers Aerospace Conference. Big Sky, MT: IEEE, 2006. US Army, Counterinsurgency Field Manual FM-3-24, Paladin Press, 2006 Hegre, H., T. Ellingsen, S. Gates, N. Gleditsch, "Toward a Democratic Civil Peace? Democracy, Political Change, and Civil War, 1816-1992," American Political Science Review, Vol. 95, Issue 1, March, 2001: 33- 48. Available online: http://www.worldbank.org/research/conflict/papers/CivilPeace2.pdf Iyengar, R., & Monten, J. (2008). Is there an" emboldenment" effect? Evidence from the insurgency in Iraq (No. w13839). National Bureau of Economic Research. JAMES T. QUINLIVAN, Force Requirements in Stability Operations, From Parameters, Winter 1995, pp. 59-69. Alex Grynkewich, USAF and Chris Reifel, USAF, Modeling Jihad: A System Dynamics Model of the Salafist Group for Preaching and Combat Financial Subsystem, Strategic Insights, Volume V, Issue 8 (November 2006) Tara A. Leweling, USAF and Otto Sieber, USN, Calibrating a Field-level, Systems Dynamics Model of Terrorism's Human Capital Subsystem: GSPC as Case Study, Strategic Insights, Volume V, Issue 8 (November 2006) Michael Kowalchuck, Siddhartha Singh and Kathleen M. Carley, CORES – Complex Organizational Reasoning System, Report CMU-ISRI-04-131, Carnegie Mellon University, School of Computer Science, September 2004 Robbins, M. (2005). Investigating the Complexities of Nationbuilding: A Sub-National Regional Perspective. Master Thesis, AIR FORCE INSTITUTE OF TECHNOLOGY, Wright-Patterson Air Force Base, Ohio US Department of Defense (12 July 2007) (PDF), Joint Publication 1-02 Department of Defense Dictionary of Military and Associated Terms, JP 1-02, http://www.dtic.mil/doctrine/jel/new_pubs/jp1_02.pdf retrieved 2007-11-21 Hegselmann, R., & Krause, U. (2002). Opinion Dynamics and Bounded Confidence Models, Analysis, and Simulation. Journal of Artificial Societies and Social Simulation, 5(3). Friedkin, N. E. (1999). Choice Shift and Group Polarization. American Sociological Review, 64(6), 856- 875. CAIN Web Service (May, 2006). Political party support in Northern Ireland, 1969 to the present. http://cain.ulst.ac.uk/issues/politics/election/electsum.htm Grier, R. A., Skarin, B., Lubyansky, A., & Wolpert, L. (2008). SCIPR: A computational model to simulate cultural identities for predicting reactions to events. Proceedings of the Second International Conference on Computational Cultural Dynamics Kott, A., & Corpac, P. S. (2007). COMPOEX technology to assist leaders in planning and executing campaigns in complex operational environments. 12th International Command and Control Research and Technology Symposium, 2007 Kott, A., Hansberger, J., Waltz, E., & Corpac, P. (2010). Whole-of-Government Planning and Wargaming of Complex International Operations: Experimental Evaluation of Methods and Tools. International Journal of Command and Control, The International C2 Journal, Vol 4, No 3, March 2010 Kott, A., Hawley, L., Brown, G., Citrenbaum, G., & Corpac, P. S. (2007). Next State Planning: A" Whole of Government" Approach for Planning and Executing Operational Campaigns. 12th International Command and Control Research and Technology Symposium, 2007 Kott, A., & Citrenbaum, G. (Eds.). (2010). Estimating Impact: A Handbook of Computational Methods and Models for Anticipating Economic, Social, Political and Security Effects in International Interventions. Springer Science & Business Media.
1903.05561	1	1903	2019-02-26T01:41:40	Learning Multi-agent Communication under Limited-bandwidth Restriction for Internet Packet Routing	[ "cs.MA", "cs.LG" ]	Communication is an important factor for the big multi-agent world to stay organized and productive. Recently, the AI community has applied the Deep Reinforcement Learning (DRL) to learn the communication strategy and the control policy for multiple agents. However, when implementing the communication for real-world multi-agent applications, there is a more practical limited-bandwidth restriction, which has been largely ignored by the existing DRL-based methods. Specifically, agents trained by most previous methods keep sending messages incessantly in every control cycle; due to emitting too many messages, these methods are unsuitable to be applied to the real-world systems that have a limited bandwidth to transmit the messages. To handle this problem, we propose a gating mechanism to adaptively prune unprofitable messages. Results show that the gating mechanism can prune more than 80% messages with little damage to the performance. Moreover, our method outperforms several state-of-the-art DRL-based and rule-based methods by a large margin in both the real-world packet routing tasks and four benchmark tasks.	cs.MA	cs	Learning Multi-agent Communication under Limited-bandwidth Restriction for Internet Packet Routing Hangyu Mao 1 , Zhibo Gong 2 , Zhengchao Zhang 1 , Zhen Xiao 1 and Yan Ni 1,3 1 Peking University 2 Huawei Technologies Co., Ltd. {hy.mao, zhengchaozhang, xiaozhen, yanni}@pku.edu.cn, [email protected] 3 Microsoft 9 1 0 2 b e F 6 2 ] A M . s c [ 1 v 1 6 5 5 0 . 3 0 9 1 : v i X r a Abstract Communication is an important factor for the big multi-agent world to stay organized and produc- tive. Recently, the AI community has applied the Deep Reinforcement Learning (DRL) to learn the communication strategy and the control policy for multiple agents. However, when implementing the communication for real-world multi-agent applica- tions, there is a more practical limited-bandwidth restriction, which has been largely ignored by the existing DRL-based methods. Specifically, agents trained by most previous methods keep sending messages incessantly in every control cycle; due to emitting too many messages, these methods are un- suitable to be applied to the real-world systems that have a limited bandwidth to transmit the messages. To handle this problem, we propose a gating mech- anism to adaptively prune unprofitable messages. Results show that the gating mechanism can prune more than 80% messages with little damage to the performance. Moreover, our method outperforms several state-of-the-art DRL-based and rule-based methods by a large margin in both the real-world packet routing tasks and four benchmark tasks. 1 Introduction Communication is an essential human intelligence, while Re- inforcement Learning (RL) and especially the Deep Neural Network (DNN) based Reinforcement Learning (DRL) are emergent techniques that can achieve artificial intelligence. Recently, inspired by the communication among humans, DRL-based methods have been successfully applied to learn the communication among multiple intelligent agents. However, it remains an open question to apply these meth- ods to the real-world multi-agent systems, because real-world applications usually impose many constraints on communica- tion, such as the bandwidth limitation for transmitting mes- sages, the random delay of messages, and even the protection of private messages. Only by resolving these constraints can we develop practical communication strategy. In this paper, we focus on resolving the limited-bandwidth restriction in multi-agent communication. We then try to ap- ply our method to the real-world packet routing systems, be- cause packet routing is not only a representative task with the above property, but also one of the most essential and critical tasks on the Internet. Formally, we define the limited-bandwidth restriction as follows: the bandwidth (or more generally, the resource) for transmitting the communication messages is limited, there- fore the agents should generate as few messages as possible on the premise of maintaining the performance. We are interested in the limited-bandwidth restriction due to two reasons. On the one hand, it is ubiquitous in real sys- tems. For example, in the packet routing systems, the link has a limited transmission capacity; in the Internet of Things, the sensor has a limited battery capacity. Once we figure out a principled method to address this problem, many fields can benefit from this work. On the other hand, this problem has been largely ignored by the existing DRL-based methods. There is a great need to devote attention to this problem. We take two steps to address this problem. Firstly, we ag- gregate the merits of the existing methods to form a basic model named Actor-Critic with Message Learning (ACML). However, the proposed ACML is still not practical because it does not change the communication pattern of the existing methods. That is to say, ACML keeps sending messages in- cessantly in every control cycle, regardless whether the mes- sage is beneficial enough for the whole agent team. Secondly, we extend ACML with a Gating mechanism to design a more flexible and practical GACML model. The gating mechanism is trained based on a novel auxiliary task, which tries to open the gate to encourage communication when the message is beneficial enough to the whole agent team, and close the gate to discourage communication otherwise. As a result, after the gating mechanism is trained well, it can prune unprofitable messages adaptively to control the message quantity around a desired threshold. Consequently, GACML is applicable to real-world systems with limited-bandwidth restriction. We evaluate our method in the real-world packet routing and benchmark tasks. It outperforms several state-of-the-art DRL-based and rule-based methods by a large margin. Fur- thermore, the proposed gating mechanism can prune more than 80% messages with little damage to the performance. 2 Background DEC-POMDP. We consider a partially observable cooper- ative multi-agent setting that can be formulated as DEC- POMDP [Bernstein et al., 2002]. It is formally defined as a tuple (cid:104)N, S, (cid:126)A, T, R, (cid:126)O, Z, γ(cid:105), where N is the number of agents; S is the set of state s; (cid:126)A = [A1, ..., AN ] represents the set of joint action (cid:126)a, and Ai is the set of local action ai that agent i can take; T (s(cid:48)s, (cid:126)a) : S × (cid:126)A × S → [0, 1] rep- resents the state transition function; R : S × (cid:126)A × S → R is the reward function shared by all agents; (cid:126)O = [O1, ..., ON ] is the set of joint observation (cid:126)o controlled by the observation function Z : S × (cid:126)A → (cid:126)O; γ ∈ [0, 1] is the discount factor. In a given state s, agent i can only observe an observation oi ∈ s, and each agent takes an action ai based on its own ob- servation oi, resulting in a new state s(cid:48) and a shared reward r. The agents try to learn a policy πi(aioi) : Oi × Ai → [0, 1] that can maximize E[G] where G is the discount return de- t=0 γtrt, and H is the time horizon. In prac- tice, we map the observation history rather than the current observation to an action (namely, oi represents the observa- tion history of agent i in the rest of the paper). fined as G =(cid:80)H Reinforcement Learning (RL). RL [Sutton and Barto, 1998] is generally used to solve special DEC-POMDP prob- lems where N = 1. In practice, we usually define the Q-value function as Qπ(s, a) = Eπ[GS = s, A = a], then the opti- mal policy can be derived by π∗ = arg maxπ Qπ(s, a). Deterministic Policy Gradient (DPG) [Silver et al., 2014] is a special actor-critic algorithm where the actor adopts a de- terministic policy µθ : S → A and the action space A is con- tinuous. Deep DPG (DDPG) [Lillicrap et al., 2015] applies DNN µθ(s) and Q(s, a; w) to approximate the actor and the critic, respectively. DDPG is an off-policy method. It adopts target network and experience replay to stabilize training and to improve data efficiency. Specifically, the critic's parame- ters w and the actor's parameters θ are updated based on: L(w) = E(s,a,r,s(cid:48))∼D[δ2] δ = r + γQ(s(cid:48), a(cid:48); w−)a(cid:48)=µθ− (s(cid:48)) − Q(s, a; w) (1) (2) (3) ∇θJ(θ) = Es∼D[∇θµθ(s) ∗ ∇aQ(s, a; w)a=µθ(s)] where D is the replay buffer containing recent experience tu- ples (s, a, r, s(cid:48)); Q(s, a; w−) and µθ−(s) are the target net- works whose parameters w− and θ− are periodically updated by copying w and θ. A merit of actor-critic algorithms is that the critic is only used during training, while only the actor is needed during execution. Due to this merit, we only need to prune messages among the actors in our GACML. 3 Related Work Traditional Communication Model. The communication models have been widely studied in the RL community, e.g., MTDP-COM [Pynadath and Tambe, 2002] and DEC- POMDP-COM [Goldman and Zilberstein, 2004]. However, traditional methods usually either predefine the communica- tion message [Wu et al., 2011] or optimize the communi- cation message for a predefined control policy [Roth et al., 2005], which are inapplicable to the real-world multi-agent systems. Previous studies also try to address the limited- bandwidth restriction by pruning the messages [Roth et al., 2005; Becker et al., 2009; Wu et al., 2011]. A general method is the Value of Communication (VoC) [Becker et al., 2009]. VoC measures the difference between the expected values of communicating and remaining silent. Communication is a better choice when VoC is larger than zero. Nevertheless, calculating VoC requires the knowledge of the environment, which is not easy to get since multi-agent systems are usually too complex. In contrast, we focus on model-free learning. Deep Communication Model. Recently, combining DNN with RL, DRL-based communication models have been ex- plored in model-free setting, such as CommNet [Sukhbaatar et al., 2016], DIAL [Foerster et al., 2016], BiCNet [Peng et al., 2017], AMP [Peng et al., 2018] and [Mao et al., 2017; Mordatch and Abbeel, 2018; Mao et al., 2018]. They adopt a DNN with hard-coded structure to represent the policy, and the policy takes as input the messages from all agents to gen- erate a single control action. Thus, the agents have to keep sending messages incessantly in every control cycle, without alternatives to reduce the message quantity. Due to emitting too many messages, they are inflexible to be applied to multi- agent systems with the limited-bandwidth restriction. Methods for addressing the limited-bandwidth restriction can be roughly divided into two types. The simple type does not generate messages at all. For example, MADDPG [Lowe et al., 2017] and COMA [Foerster et al., 2017] adopt an inde- pendent DNN to generate the policy for each individual agent. Because the policy is independent of other agents' informa- tion, these methods can generate control action without any communication. Therefore, they are applicable in tasks with strict limit-bandwidth restriction. However, as the policies do not exchange messages, these methods suffer from the par- tially observable problem. Obviously, it is impossible to fig- ure out the best control policies based on partial information. The principled type applies special mechanisms to adap- tively decide whether to send the messages (equivalently, whether to prune the messages), so the message quantity can be controlled to some extent. Our GACML is an instance of such method, and the most relevant studies include ATOC [Jiang and Lu, 2018], MADDPG-M [Kilinc and Montana, 2019], SchedNet [Kim et al., 2019] and IC3Net [Singh et al., 2019]. Both ATOC and IC3Net learn a binary action to specify whether the agent wants to communicate with others, which is similar to our gating mechanism. However, ATOC and IC3Net are only suitable for homogeneous agents due to their DNN structures, while our GACML is a general model for both homogeneous and heterogeneous agents. More im- portantly, both ATOC and IC3Net cannot control how many messages will be pruned, while GACML has this ability be- cause we introduce a special threshold for that purpose in our method. MADDPG-M adopts a two-level policy to learn whether the agent's private observation is sufficiently infor- mative to be shared with others. However, MADDPG-M is only applicable to 2D tasks where the distance between the agents and their target agents is measurable and available. SchedNet generates a weight w for each agent, and the top M agents in terms of their weights w will communicate with each other. However, it is not easy to apply a hard-coded pa- rameter M to real-world multi-agent systems. Summary. Most of the existing RL-based communication models are inflexible to be applied in the real-world multi- agent systems with limited-bandwidth restriction. A few pos- sible DRL-based methods seem to be still preliminary, while our GACML is the first formal method that can control the message quantity to a desired threshold, as far as we know. 4 The Proposed Method 4.1 ACML The key variables used in this section are as follows. ai is the local action of agent i. (cid:126)a−i is the joint action of other agents except for agent i. (cid:126)a is the joint action of all agents, i.e., (cid:126)a = (cid:104)ai, (cid:126)a−i(cid:105). The observation history (cid:126)o, oi, (cid:126)o−i, and the policy µθi are denoted similarly. s(cid:48) is the next state after s, and (cid:126)o(cid:48), o(cid:48) The Design. ACML is motivated by combining the mer- its of the existing DRL-based communication methods men- tioned in Section 3. As can be seen from Figure 1, ACML adopts the following designs. −i are denoted similarly. −i, (cid:126)a(cid:48), a(cid:48) i, (cid:126)a(cid:48) i, (cid:126)o(cid:48) (1) Each agent is made up of an ActorNet and a Message- GeneratorNet. This design represents the policy and the mes- sage with two separated network branches. Another choice is to take the hidden layer of policy network as the message. We notice that most previous methods adopt the former, which usually outperforms the latter. (2) All agents share the same CriticNet and MessageCo- ordinatorNet, which are placed in a coordinator (i.e., a spe- cially designed agent). The shared MessageCoordinatorNet is similar to many previous methods such as the CommNet, AMP and ATOC, while the shared CriticNet is similar to the well-known MADDPG and COMA. Although each separate design is common, aggregating them together properly is novel. For example, MADDPG and COMA do not adopt the MessageGeneratorNet and Mes- sageCoordinatorNet, making them suffer from the partially observable problem during execution; while AMP and ATOC do not adopt the shared CriticNet, making them suffer from the non-stationary problem [Hernandez et al., 2017] during training. ACML is fully observable and training stationary. ACML works as follows during execution 1. (1) mi = MessageGeneratorNet(oi), i.e., agent i generates the local message mi based on its own observation oi. (2) All agents send their mi to the coordinator. (3) M1, .., MN = MessageCoordinatorNet(m1, .., mN ), i.e., the coordinator extracts the global message Mi for each agent i based on all local messages mi. (4) The coordinator sends Mi back to agent i. (5) ai = ActorNet(oi, Mi), i.e., agent i generates action ai based on its local observation oi and the global message Mi, which encodes all (cid:104)o1, .., oN(cid:105) for full observability. The Training. As described above, the agents generate ai based on oi and Mi to interact with the environment, and the environment will feed a shared reward r back to the agents. −i(cid:105) are Then, the experience tuples (cid:104)oi, (cid:126)o−i, ai, (cid:126)a−i, r, o(cid:48) used to train ACML. Specifically, as the agents exchange messages with each other, the actor and the shared critic can be represented as µθi(oi, Mi) and Q((cid:126)o, (cid:126)a; w), respectively. i, (cid:126)o(cid:48) 1Please note that the CriticNet is only used during training, while other components are needed during execution. Figure 1: The proposed ACML. For clarity, we show this model using a two-agent example. All components are made up of DNN. h is the hidden layer of the DNN; mi is the local message; Mi is the global message. The red arrows imply the message sending process. θ − i We can extend Equation (1 -- 3) to multi-agent formulations shown in Equation (4 -- 6), where the parameters w, w−, θi and θ− i have similar meaning to these of single-agent setting. i) − Q((cid:126)o, (cid:126)a; w) (4) (o(cid:48) (5) δ = r + γQ((cid:126)o(cid:48), (cid:126)a(cid:48); w−)a(cid:48) i,(cid:126)o(cid:48) L(w) = E(oi,(cid:126)o−i,ai,(cid:126)a−i,ri,o(cid:48) i=µ −i)∼D[δ2] ∇θiJ(θi) = E(oi,(cid:126)o−i)∼D[ ∇θiµθi(oi, Mi) ∗ ∇aiQ((cid:126)o, (cid:126)a; w)ai=µθi (oi) ] (6) Since all components in ACML are implemented by DNN, ACML is end-to-end differentiable, and the communication message and the control policy can be optimized jointly using back propagation (BP) based on the above equations. 4.2 GACML Motivation. As can be seen from the execution process, ACML takes as input the communication messages from all agents to generate a single control action. Thus, the agents have to keep sending messages incessantly in every control cycle, regardless whether the messages are beneficial enough to the performance of the agent team. This is also a common problem of most deep communication models as mentioned in Section 3. As a result of too many messages, these meth- ods are inflexible to be applied to the real-world multi-agent systems with limited-bandwidth restriction. GACML is motivated by handling this problem through pruning unprofitable messages in a principled manner. The Design. We propose a gating mechanism to adaptively prune unprofitable messages among ActorNets, such that the agents can maintain the performance while pruning as many messages as possible to resolve the bandwidth constraints. As shown in Figure 2, except for the original components, each agent is equipped with an additional GatingNet. Specif- ically, GACML works as follows. (1) The agent generates a local message mi as well as a probability score p ∈ (0, 1) based on its own observation oi. hhhhhhh𝑜2𝑜1𝑎2𝑎1𝑀1𝑀2𝑚1𝑚1𝑚2𝑚2𝑀1𝑀2𝑎1𝑜1𝑎2𝑜2hQActorNetCriticNetMessageGeneratorNetMessageCoordinatorNetActorNetcontrolling how many messages should be pruned. setting, the label of this auxiliary task can be formulated as In this Y (o) = I(Q(o, aC) − Q(o, aI ) > T ) (7) where I is the indicator function. Then we can train p by minimizing the following loss function: Lθop (o) = −Eo[Y (o) log p(o; θop) +(1 − Y (o)) log(1 − p(o; θop))] (8) where θop are the parameters between the observation o and the probability p as shown in Figure 2. The insight of the above loss function is that if ∆Q(o) = Q(o, aC)−Q(o, aI ) is really larger than T (i.e., aC can obtain at least T Q-values more than aI, and the corresponding label is Y (o) = 1), the network should try to generate a probabil- ity p(o; θop) that is larger than Tp = 0.5 to encourage com- munication. In other word, GACML only prunes messages that contribute less Q-values than the threshold T . Therefore, after the gating mechanism is trained well, it can prune un- profitable messages adaptively to control the message quan- tity around a desired threshold specified by T . Consequently, GACML needs much fewer messages to achieve a desirable performance. It is our key contribution that makes GACML a novel and principled method. The Key Implementation. The above training method re- lies on correct labels of the auxiliary task. It means that we should provide suitable Q(o, aC), Q(o, aI ) and T as indi- cated by Equation (7). For the Q-values, we firstly set g = 1 (i.e., without message pruning) to train other components except for the GatingNet based on Equation (4 -- 6). After the model is trained well, we can get approximately correct ActorNet and CriticNet. After- wards, for a specific observation o, the ActorNet can generate aC and aI when we set g = 1 and g = 0, respectively; then the CriticNet can estimate an approximately correct Q-value Q(o, aC) and Q(o, aI ). For the threshold T , we propose two methods to set a fixed T and a dynamic T , respectively. To calculate a fixed T , we firstly sort the ∆Q(o) of the latest K observations o encoun- tered during training, resulting in a list of ∆Q(o), which is called L∆Q(o). Then, we set T by splitting L∆Q(o) in terms of the index. For example, if we want to prune Tm% messages, we set T = L∆Q(o)[K × Tm%]. We do not split L∆Q(o) in terms of the value, since ∆Q(o) usually has a non-uniform distribution. The advantage of a fixed T is that the actual number of the pruned messages is ensured to be close to the desired Tm%. Besides, this method is friendly to a large K. For the dynamic T , we adopt the exponential moving aver- age technique 3 to set T : t ) − Q(ot, aI t )) Tt = (1 − β)Tt−1 + β(Q(ot, aC (9) where β is a coefficient for discounting older T . We test some β in [0.6, 0.9], and they all work well. The advantage of a dynamic T is that Y (o) becomes an adaptive training label even for the same observation o. This is very important for the dynamically changing environments, because T and Y (o) can quickly adapt to these environments. Figure 2: The actor part of the proposed GACML. For clarity, we only show one agent's structure, and we do not show the critic part because it is the same as that of ACML. following process is the same as that of ACML. (2) A gate value g ∈ {0, 1} is generated by a threshold function Tp = 0.5. That is to say, if p ≤ Tp, we set g = 0, otherwise, we set g = 1. (3) The agent sends mi (cid:12) g to the coordinator, and the In the above process, if g = 0, mi(cid:12) g will be a zero vector, and the agent has no need to send mi to the coordinator. Ac- cordingly, the coordinator does not send the global message Mi back to the agent, because we replace Mi with a zero vector to "tell" our GACML model that the message Mi is pruned indeed. We call this design zero padding. Please note that zero padding is a crucial design for prun- ing a lot of messages. Suppose that we have pruned many messages, the time span of two valid messages would be rel- atively large. If we simply adopt the message caching method [Dowling and Haridi, 2008] to replace M t i gen- erated a long time ago (i.e., t − t is relatively large), M t i can hardly approximate M t i anymore. In contrast, the zero padding will always "tell" GACML a correct signal: the mes- sages have been really pruned. Using another GatingNet to decide whether to prune Mi will be evaluated in the future. i with M t The Training Method with a Novel Auxiliary Task. In order to make the above design work, a suitable probability p must be trained for each observation o 2, otherwise GACML may degenerate to ACML (in the extreme case where p is always larger than Tp). However, as the threshold function Tp is non-differentiable, it makes the end-to-end BP method inapplicable. We also tried the approximate gradient [Hubara et al., 2016], the sparse regularization [Makhzani and Frey, 2015] and several other methods without success. Finally, we decided to embrace the auxiliary task technique [Jaderberg et al., 2016] to provide training signal for p di- rectly. Recall that we want to prune the messages on the premise of maintaining the performance. For RL, the per- formance could be measured by the Q-value, so we design the following auxiliary task. Let p indicate the probability of that ∆Q(o) = Q(o, aC)− Q(o, aI ) is larger than T , where aC is the action generated based on communication, aI is the action generated indepen- dently (i.e., without communication), and T is a threshold 2We remove the subscript i to represent that the following de- scription is suitable for all agents. 3https://en.wikipedia.org/wiki/Moving average#Exponential moving average hhhh𝑜1𝑎1𝑀1𝑀2𝑚1𝑚1𝑚2𝑀1ActorNet𝒑𝒈{0, 1}(0, 1)𝑻𝒑= 0.5𝜽𝒐𝒑GatingNetMessageCoordinatorNet5 Experiments 5.1 The Experimental Settings Due to limited space, we only show the evaluation on real- world packet routing systems. Four benchmark tasks such as traffic control and predator prey are shown in the appendix. Baseline. We adopt Independent Actor-Critic (IND-AC) [Peng et al., 2017], MADDPG [Lowe et al., 2017] and AMP [Peng et al., 2018] for comparison. These methods can be regarded as the ablation models of ACML. In IND-AC, the agent learns its own actor-critic network independently with- out communication. We can know the effect of communica- tion by comparing with IND-AC. MADDPG adopts central- ized critics to share information among multiple agents, while the actors are independent. In AMP, only the actors exchange messages, while the critics are independent. In contrast, both actors and critics in ACML can exchange messages. Methods like ATOC are not compared, since they are unsuitable for the routing tasks due to the reasons analyzed in Section 3. Parameter. The probability p in GACML is a hidden layer with 1 neuron, and the activation function is Sigmoid. The first hidden layer of our model has 64 neurons, while other hidden layers have 32 neurons, and the activation function is Relu. Learning rates of actor, critic and target networks are 0.001, 0.01 and 0.001, respectively. Replay buffer size, batch size and discount factor are 106, 128 and 0.95, respectively. 5.2 The Packet Routing System Environment Description. In the information era, packet routing is a very fundamental and critical task on the Internet. We evaluate our methods on three routing tasks. As shown in Figure 3, the small topology and the moderate topology are the most classical topologies in the Internet Traffic Engineer- ing community [Kandula et al., 2005]; the large topology is based on the real needs of our industrial collaborator, and it is more complex than the real-world Abilene Network 4 in terms of the numbers of routers, links and paths. In each topology, there are several edge routers. Each edge router has an aggregated flow that should be transmitted to other edge routers through available paths. For example, in Figure 3(a), B is set to transmit flow to D, and the available paths are BEF D and BD. Each path is made up of several links, and each link has a link utilization, which equals to the ratio of the current flow on this link to the maximum flow transmission capacity of this link. The necessity of cooperation among routers is as follows: one link can be used to transmit the flow from more than one router, so the routers must not split too much or too little flow on the same link at the same time; otherwise this link will be either overloaded or underloaded. Note that there is a strict limited-bandwidth restriction in these tasks, because the capacity of the network is limited. Problem Definition. The routers are controlled by our al- gorithm, and they try to learn a good flow splitting policy to minimize the Maximum Link Utilization in the whole network (MLU). The intuition behind this goal is that high link utiliza- tion is bad for dealing with bursty flow. The observation in- cludes the latest ten steps' flow demands, the latest ten steps' 4A backbone net https://en.wikipedia.org/wiki/Abilene Network (a) The small topology. (b) The moderate topology. (c) The large topology. Figure 3: The packet routing environments. link utilizations and their average, and the latest action taken by the router. The action is a flow splitting ratio on each avail- able path (e.g., in Figure 3(b), B will generate an action like [a%, b%, 1 − a% − b%] for paths [B12G, BG, B34G]). The reward is set as 1 − M LU to minimize MLU. Besides MLU, we also care about the convergence ratio of all experiments. 5.3 The Experimental Results Results without Message Pruning. In this experiment, we use synthetic flow to evaluate different methods on the small and moderate topologies. The flow has a shape of A sin(wx+ ϕ) + b with different settings of A, w, ϕ, b. (a) The MLU. (b) The convergence ratio. Figure 4: The average results of 30 independent experiments. A smaller MLU is better. A larger convergence ratio is better. Figure 4(a) shows the MLU of 30 independent experi- ments. In the small topology, all DRL-based methods (i.e., excluding the rule-based WCMP and TeXCP) have a simi- lar performance. The reason is that this topology is rather simple, and all DRL-based methods can find the near-optimal control policies easily after they have been trained, regard- less whether the methods adopt communication. Thus, the more advanced methods, such as MADDPG and ACML, do not have enough space to further improve IND-AC. In the moderate topology, the performances of IND-AC, AMP and MADDPG drop severely, while ACML keeps the performance and achieves much smaller MLU. This is be- CDBEFAPaths:ACAEFCBEFDBDABCDEFGHIJ12345678Paths:AFA12FB12GBGB34GC34HCHC56HD56IDID78IE78JEJPaths:A16CA146CA147CA246CA247CA257CA147DA247DA257DA258DB246CB247CB257CB357CB247DB257DB258DB357DB358DB38DABCD167854320.650.70.750.80.850.90.951moderate topologysmall topology00.20.40.60.81moderate topologysmall topologycause other models suffer from either the partially observable problem or the non-stationary problem, making them ineffec- tive in complex environment. In contrast, ACML can address these problems as analyzed before, resulting in stronger abil- ity to deal with complex tasks. Furthermore, compared with the rule-based TeXCP 5 [Kan- dula et al., 2005] and WCMP 6, the DRL-based models achieve much smaller MLU in both topologies. The reason lies in that the DRL-based methods can take the actions' fu- ture effect into consideration, which is in favor of accom- plishing the routing task, whereas the rule-based methods can only consider the current effect of the actions. Even worse, WCMP gets a MLU larger than 1.0 in the moderate topology, and the simulation system crashed. Figure 4(b) shows the convergence ratio. As can be seen, ACML achieves the largest convergence ratio in both topolo- gies. Furthermore, when the evaluation turns to the moderate topology, the decrease of convergence ratio of ACML is much smaller than that of other methods. It means that our ACML is more stable than other DRL-based methods during training. Overall, the above results demonstrate ACML with general applicability, scalability and stability. GACML using real flow trajectory and the large topology. Message Pruning Evaluation. In this experiment, we test The results for adopting a fixed threshold T = L∆Q(o)[K× Tm%] are shown in Table 1. We draw the following conclu- sions. (1) For a predefined threshold Tm%, the actual num- ber of the pruned messages is close to Tm%. It means that GACML can control the message to our desired quantity. (2) GACML can prune a large number of messages (e.g., more than 82%) with little damage to the performance (e.g., less than 7%). It implies that the pruned messages are usually not beneficial enough to the performance. (3) When we pruned all messages (i.e., the last row of Table 1), the reward has a large decrease. Compared with the second to the last row of Table 1, it indicates that the remaining 1.5% messages are very important for keeping the performance, and GACML has learnt to share these important messages with all agents. The results for adopting a dynamic threshold T are shown in Figure 5. As can be seen, GACML performs much better than the state-of-the-art MADDPG. In addition, GACML can keep the performance close to ACML while pruning quite a lot of messages (in this case, 74.3%). To conclude, the above results demonstrate that the pro- posed gating mechanism can prune unnecessary messages on the premise of maintaining the performance indeed, and that our GACML is very suitable for the real-world routing sys- tems with limited-bandwidth restriction. Message Pruning Analysis. In this experiment, we ana- lyze GACML using synthetic flow and the small topology. We adopt this setting because the results are easy to under- stand. For example, as shown in Figure 6(a), when the total flow in the network is decreasing during timestep 100 to 300, GACML adapts to this situation more quickly than MAD- DPG, and it obtains similar rewards as ACML after dozens of steps. Besides, since we use A sin(wx + ϕ) + b to generate 5We slightly modify it for the testing environment. 6https://en.wikipedia.org/wiki/Equal-cost multi-path routing Tm% # of Pruned Message Reward Decrease 80% 6.84% 90% 8.05% 95% 11.30% 98% 15.98% 99% 14.86% 100% 61.52% 82.13% 87.68% 93.39% 99.14% 98.53% 100.00% Table 1: The pruned message and the reward decrease in terms of ACML for a fixed (i.e., a predefined) pruning threshold Tm%. Figure 5: The testing rewards. For GACML, we set β = 0.8. (a) The average rewards. (b) The gate values of 2 tests. Figure 6: The testing results of 30 message pruning experiments based on the small topology and synthetic flow. GACML gets 8.2% fewer rewards than ACML, while the message decrease is 56.3%. It also outperforms MADDPG. Please zoom in for better view. synthetic flow, the reward curves shown in Figure 6(a) and the gate value curves shown in Figure 6(b) are similar to the shape of sin(x). We also notice that the bursty flow usually generates a gate value g = 1. The results imply that GACML has learnt a subtle communication strategy. 6 Conclusion This paper presents a method to jointly learn the communica- tion strategy and the control policy for multiple cooperative agents. In contrast to most previous methods, we focus on ad- dressing the limited-bandwidth restriction that exists in many real-world applications. Specifically, we firstly aggregate the merits of the existing methods to form a new method that out- performs several state-of-the-art DRL-based and rule-based methods. Then, we propose a gating mechanism with several crucial designs that can prune quite a lot of unprofitable mes- sages with little damage to the performance. Consequently, our method is applicable to the real-world packet routing sys- tems with limited bandwidth. As far as we know, it is the first formal method to achieve this in a novel and principled way, and it is the key contribution of this work. 0.150.20.250.30.350.40.450.505001000150020002500300035004000450050005500mean average rewardtimestepACMLMADDPGGACML00.10.20.30.40.50.60.7020406080100120140160180200220240260280300320340360380400420440460480500520540560580600mean average rewardtimestepACMLMADDPGGACML01020406080100120140160180200220240260280300320340360380400420440460480500520540560580600620gate valuerouter Arouter Btimestep01020406080100120140160180200220240260280300320340360380400420440460480500520540560580600620gate valuerouter Arouter BtimestepReferences [Becker et al., 2009] Raphen Becker, Alan Carlin, Victor Lesser, and Shlomo Zilberstein. Analyzing myopic ap- proaches for multi-agent communication. Computational Intelligence, 25(1):31 -- 50, 2009. [Bernstein et al., 2002] Daniel S Bernstein, Robert Givan, Neil Immerman, and Shlomo Zilberstein. The complexity of decentralized control of mdp. Mathematics of opera- tions research, 27(4):819 -- 840, 2002. [Dowling and Haridi, 2008] Jim Dowling and Seif Haridi. Decentralized reinforcement learning for the online opti- mization of distributed systems. In Reinforcement Learn- ing. InTech, 2008. [Foerster et al., 2016] Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learn- ing to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pages 2137 -- 2145, 2016. [Foerster et al., 2017] Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. arXiv preprint arXiv:1705.08926, 2017. [Goldman and Zilberstein, 2004] Claudia V Goldman and Shlomo Zilberstein. Decentralized control of cooperative systems: Categorization and complexity analysis. Journal of Artificial Intelligence Research, 22:143 -- 174, 2004. [Hernandez et al., 2017] Pablo Hernandez, Michael Kaisers, Tim Baarslag, and Enrique Munoz de Cote. A survey of learning in multiagent environments: Dealing with non- stationarity. arXiv preprint arXiv:1707.09183, 2017. [Hubara et al., 2016] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Bina- rized neural networks. In Advances in neural information processing systems, pages 4107 -- 4115, 2016. [Jaderberg et al., 2016] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. [Jiang and Lu, 2018] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent coop- eration. arXiv preprint arXiv:1805.07733, 2018. [Kandula et al., 2005] Srikanth Kandula, Dina Katabi, Bruce Davie, and Anna Charny. Walking the tightrope: Respon- In ACM SIGCOMM sive yet stable traffic engineering. Computer Communication Review, volume 35, pages 253 -- 264. ACM, 2005. [Kilinc and Montana, 2019] Ozsel Kilinc and Giovanni Montana. Multi-agent deep reinforcement learning with extremely noisy observations. In International Conference on Learning Representations, 2019. [Kim et al., 2019] Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan Son, and Yung Yi. Learning to schedule communication in multi-agent reinforcement learning. In International Con- ference on Learning Representations, 2019. [Lillicrap et al., 2015] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, Continuous con- David Silver, and Daan Wierstra. arXiv preprint trol with deep reinforcement learning. arXiv:1509.02971, 2015. [Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi- agent actor-critic for mixed cooperative-competitive envi- ronments. In Advances in Neural Information Processing Systems, pages 6379 -- 6390, 2017. [Makhzani and Frey, 2015] Alireza Makhzani and Brendan J Frey. Winner-take-all autoencoders. In Advances in Neural Information Processing Systems, pages 2791 -- 2799, 2015. [Mao et al., 2017] Hangyu Mao, Zhibo Gong, Yan Ni, and Zhen Xiao. Accnet: Actor-coordinator-critic net for "learning-to-communicate" with deep multi-agent rein- arXiv preprint arXiv:1706.03235, forcement learning. 2017. [Mao et al., 2018] Hangyu Mao, Zhengchao Zhang, Zhen Xiao, and Zhibo Gong. Modelling the dynamic joint pol- icy of teammates with attention multi-agent ddpg. arXiv preprint arXiv:1811.07029, 2018. [Mordatch and Abbeel, 2018] Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent populations. In Thirty-Second AAAI Con- ference on Artificial Intelligence, 2018. [Peng et al., 2017] Peng Peng, Quan Yuan, Ying Wen, Yaodong Yang, Zhenkun Tang, Haitao Long, and Jun Wang. Multiagent bidirectionally-coordinated nets for learning to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017. [Peng et al., 2018] Zhaoqing Peng, Libo Zhang, and Tiejian Luo. Learning to communicate via supervised attentional In Proceedings of the 31st Inter- message processing. national Conference on Computer Animation and Social Agents, pages 11 -- 16. ACM, 2018. [Pynadath and Tambe, 2002] David V Pynadath and Milind Tambe. The communicative multiagent team decision problem: Analyzing teamwork theories and models. Jour- nal of artificial intelligence research, 16:389 -- 423, 2002. [Roth et al., 2005] Maayan Roth, Reid Simmons, and joint beliefs for Manuela Veloso. In Proceed- execution-time communication decisions. ings of the fourth international joint conference on Au- tonomous agents and multiagent systems, pages 786 -- 793. ACM, 2005. Reasoning about [Silver et al., 2014] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In ICML, 2014. [Singh et al., 2019] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Individualized controlled continu- ous communication model for multiagent cooperative and competitive tasks. In International Conference on Learn- ing Representations, 2019. [Sukhbaatar et al., 2016] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backprop- In Advances in Neural Information Processing agation. Systems, pages 2244 -- 2252, 2016. [Sutton and Barto, 1998] Richard S Sutton and Andrew G Introduction to reinforcement learning, volume Barto. 135. MIT press Cambridge, 1998. [Wu et al., 2011] Feng Wu, Shlomo Zilberstein, and Xi- aoping Chen. Online planning for multi-agent systems Artificial Intelligence, with bounded communication. 175(2):487 -- 511, 2011.
1801.02693	1	1801	2018-01-08T21:21:05	Stable Marriage with Multi-Modal Preferences	[ "cs.MA", "cs.DS" ]	We introduce a generalized version of the famous Stable Marriage problem, now based on multi-modal preference lists. The central twist herein is to allow each agent to rank its potentially matching counterparts based on more than one "evaluation mode" (e.g., more than one criterion); thus, each agent is equipped with multiple preference lists, each ranking the counterparts in a possibly different way. We introduce and study three natural concepts of stability, investigate their mutual relations and focus on computational complexity aspects with respect to computing stable matchings in these new scenarios. Mostly encountering computational hardness (NP-hardness), we can also spot few islands of tractability and make a surprising connection to the \textsc{Graph Isomorphism} problem.	cs.MA	cs	Stable Marriage with Multi-Modal Preferences∗ Jiehua Chen† Ben-Gurion University of the Negev Beer-Sheva, Israel Rolf Niedermeier TU Berlin Berlin, Germany Piotr Skowron‡ TU Berlin Berlin, Germany Abstract We introduce a generalized version of the famous STABLE MARRIAGE problem, now based on multi- modal preference lists. The central twist herein is to allow each agent to rank its potentially matching counterparts based on more than one "evaluation mode" (e.g., more than one criterion); thus, each agent is equipped with multiple preference lists, each ranking the counterparts in a possibly different way. We introduce and study three natural concepts of stability, investigate their mutual relations and focus on computational complexity aspects with respect to computing stable matchings in these new scenarios. Mostly encountering computational hardness (NP-hardness), we can also spot few islands of tractability and make a surprising connection to the GRAPH ISOMORPHISM problem. Keywords: Stable matching, concepts of stability, multi-layer (graph) models, NP-hardness, parameter- ized complexity analysis, exact algorithms. 1 Introduction Information about the same "phenomenon" can come from different, possibly "contradicting", sources. For instance, when evaluating candidates for an open position, data concerning experience and so far achieved successes of the candidates may give different candidate rankings than data concerning their formal qualifi- cations and degrees. In other words, one has to deal with a multi-modal data scenario. Clearly, in maximally objective and rationality-driven decision making, it makes sense to take into account several information resources in order to achieve best possible results. In this work we systematically apply this point of view to the STABLE MARRIAGE problem [25]; a key observation here is that several natural and well-motivated "multi-modal variants" of STABLE MARRIAGE need to be studied. We investigate the complexity of com- puting matchings that are stable according to the considered definitions. In the classic (conservative) STABLE MARRIAGE problem [25], we are given two disjoint sets U and W of n agents each, where each of the agents has a strict preference list that ranks every member of the other set. The goal is to find a bijection (which we call a matching) between U and W without any blocking pair which can endanger the stability of the matching. A pair of agents is blocking a matching if they are not matched to each other but rank each other higher than their respective partners in the matching. ∗Work started when all authors were with TU Berlin. †Supported by the People Programme (Marie Curie Actions) of the European Union's Seventh Framework Programme (FP7/2007-2013) under REA grant agreement number 631163.11, and by the Israel Science Foundation (grant number 551145/14). ‡Supported by a postdoctoral fellowship of the Alexander von Humboldt Foundation, Bonn, Germany. 1 Gale and Shapley [25] introduced the STABLE MARRIAGE problem in the fields of Economics and Computer Science in the 1960s. One of their central results was that every STABLE MARRIAGE instance with 2n agents admits a stable matching, which can be found by their algorithm in O(n2) time. Since then STABLE MARRIAGE has been intensively studied in Economics, Computer Science, and Social and Political Science [1, 27, 29, 30, 35, 37, 39]. Practical applications of STABLE MARRIAGE (and its variants) include partnership issues in various real-world scenarios, matching graduating medical students (so-called residents) with hospitals, students with schools, and organ donors with patients [27, 35, 39], and the design of content delivery systems [36] and other distributed markets [45]. The original model of STABLE MARRIAGE assumes, roughly speaking, that there is a (subjective) cri- terion and that each agent has a single preference list depending on this criterion. In typically complex real-world scenarios, however, there are usually multiple aspects one takes into account when making a de- cision. For instance, if we consider the classical partnership scenario, then there could be different criteria such as working hours, family background, physical appearance, health, hobbies, etc. In other words, we face a much more complex multi-modal scenario. Accordingly, the agents may have multiple preference lists, each defined by a different criterion; we call each of these criteria a layer. For an illustration, let us consider the following stable marriage example with two sets of two agents each, denoted as u1, u2, w1, and w2, and three layers of preferences, denoted as P1, P2, and P3. w1 w2 u1 w2 w1 u2 w1 w2 u1 w1 w2 u2 w2 w1 u1 w2 w1 u2 P1: w1 u1 u2 w2 u1 u2 P2: w1 u1 u2 w2 u1 u2 P3: w1 u2 u1 w2 u1 u2 In the above diagram the preferences are depicted right above (respectively, right below) the corresponding agents; preferences are represented through vertical lists where more preferred agents are put above the less preferred ones. For example, in the first layer, all agents from the same set have the same preference list, i.e. both u1 and u2 rank w1 higher than w2 while both w1 and w2 rank u1 higher than u2. Similarly, in the second layer, both u1 and u2 rank w2 higher than w1 while both w1 and w2 rank u1 higher than u2. In the last layer, the preference lists of two agents from the same set are reverse to each other. For instance, u1 ranks w1 higher than w2, which is opposite to u2. In terms of the classic stable marriage problem, we will have three independent instances, one for each layer. The corresponding stable matching(s) for each instance are depicted through the edges between the agents. For instance, the first layer admits exactly one stable matching, which matches u1 with w1, and u2 with w2. Yet, if we want to take all these layers jointly into account, then we need to extend the traditional concept of stability. With multiple preference lists for each agent, there are many natural ways to extend the original stability concept. We propose three naturally emerging concepts of stability. Assume each agent has ℓ (possibly different) preferences lists. All three concepts are defined for a certain threshold α with 1 ≤ α ≤ ℓ, which quantifies "the strength" of stability. In the following, we briefly describe our three concepts and defer the formal definitions to Section 2. - The first one, called α-layer global stability, extends the original stability concept in a straightforward way. It assumes that the matched pairs agree on a set S of α layers where no unmatched pair is blocking the matching in any layer from S. In our introductory example, the matching M1 = {{u1, w1}, {u2, w2}} is stable in the first and the last layer, and thus it is a 2-layer globally stable matching. 2 - The second one, called α-layer pair stability, changes the "blocking ability" of the unmatched pairs. It forbids an unmatched pair to block more than ℓ − α layers. In other words, each pair of matched agents needs to be stable in some α layers, but the choice of these layers can be different for different pairs. Considering again our running example, we can verify that the 2-layer globally stable matching M1 is also 2-layer pair stable as each unmatched pair is blocking at most one layer. Indeed, we will see that α-layer pair stability is strictly weaker than α-layer global stability (Proposition 3.1 and Example 3.1). - The last one, called α-layer individual stability, focuses on the "willingness" of an agent to stay with its partner. It requires that for each unmatched pair, at least one of the agents prefers to stay with its partner in at least α layers. In our introductory example, the matching M1 is also 2-layer individually stable. Thus, it is tempting to assume that α-layer individual stability also generalizes α-layer global stability. This is, however, not true as the following matching M2 = {{u1, w2}, {u2, w1}} is 2-layer globally stable but not 2- layer individually stable. Neither does the latter implies the former. We refer to Example 3.2 for more explanations. 1.1 Related work While we are not aware of research on an arbitrary number ℓ of layers, there is some work on ℓ = 2 layers. Weems [48] considered the case where each agent has two preference lists that are the reverse of each other. He provided a polynomial-time algorithm to find a bistable matching, i.e. a matching that is stable in both layers. Thus, while his concept falls into our α-layer global stability concept for α = ℓ = 2, it is a special case since the preference lists in the two layers are the reverse of each other. In fact, for α = ℓ = 2, we show that the complexity of determining α-layer global stability is NP-hard. Aggregating the preference lists of multiple layers into one (by comparing each pair of agents) and then searching for a "stable" matching for the agents with aggregated preferences is a plausible approach to multi-modal stable marriages. As already noted by Farczadi et al. [23], the aggregated preferences may be intransitive or even cyclic. Addressing this situation, they consider a generalized variant of STABLE MARRIAGE, where each agent u of one side, say U , has a strict preference list ≻u (as in the original STABLE MARRIAGE) while each agent w of the other side, say W , may order each possible pair of partners separately, expressed by a subset Bw ⊆ U × U of ordered pairs. They defined a matching M to be stable if no unmatched pair {u, w} satisfies "w ≻u M (u) and (u, M (u)) ∈ Bw". It turns out that our concept of individual stability and their concept for a more generalized case where both sides of the agents may have intransitive preferences are related, and we can use one of their results as a subroutine. In a way, our analysis provides a more fine-grained view, since we consider a richer model and thus are able to discuss how certain assumptions on elements of this model (e.g., the number of layers, the threshold value α, etc.) affect the computational complexity of the problem. Aziz et al. [2] considered a variant of STABLE MARRIAGE, where each agent has a probability for each ordered pair of potential partners. Assigning a probability of 1 to either (x, y) or (y, x) for each x and y, their variant is closely related to the one of Farczadi et al. [23] and is shown to be NP-hard. We refer to several expositions [35, 27, 31, 39, 33, 7] for a broader overview on STABLE MARRIAGE and related problems. 1.2 Our contributions We introduce three main concepts of stability for STABLE MARRIAGE with multi-modal preferences. In Section 2 we formally define these concepts, global stability, pair stability, and individual stability, and provide motivating and illustrating examples. In Section 3, we study the relations between the three concepts 3 Table 1: The computational complexity of finding matchings stable according to the three consideerd definitions-α-layer global stability, α-layer pair stability, and α-layer individual stability-for instances with 2n agents and ℓ layers. All results hold for each value of α specified in the first column. Results marked with ∗ hold even if we assume that each agent of one side has the same preference list in all layers. The NP-hardness results hold even for a fixed number of layers. Parameters global stability pair stability individual stability Arbitrary O(n2) [25] NP-h [T. 4.2+P. 4.4] 1 = α 2 ≤ α = ℓ ⌊ℓ/2⌋ < α < ℓ NP-h [P. 5.1] 2 ≤ α ≤ ⌊ℓ/2⌋ NP-h [P. 5.1] Single-layered O(n2) [25] NP-h [C. 4.3+P. 4.4] NP-h [P. 5.4+P. 5.5] NP-h∗ [C. 5.3] O(n2) [25] O(ℓ · n2) [T. 4.1] ? NP-h∗ [T. 5.2] NP-h for unbounded α [T. 6.2] NP-h∗ when 2 ≤ α ≤ ⌊ℓ/2⌋ [C. 5.3] NP-h∗ when 2 ≤ α ≤ ⌊ℓ/2⌋ [T. 5.2] W[1]-h & in XP for α [T. 6.2] O(ℓ · n2) when α > ⌊ℓ/2⌋ [P. 6.3] O(ℓ · n2) when α > ⌊ℓ/2⌋ [P. 6.3] Uniform α ≥ ℓ/2 + 1 α = ℓ/2 O(ℓ · n) [P. 6.7] O(ℓ · n) [P. 6.7] ? ? nO(log (n)) + O(ℓ · n2) [C. 6.6] GRAPH ISOM.-hard [T. 6.5] and show that pair stability is the least restrictive form while global and individual stability are in general incomparable (also see Figure 1 for a much refined picture). In Section 4, we consider the special case of all-layers stability (α = ℓ) for the three concepts. On the one hand, we provide a polynomial-time algorithm for checking individual stability for arbitrary large number of preference lists. On the other hand, through an involved construction, we show NP-hardness for the other two stability concepts, even if there are only two layers. The hardness results demonstrate a complexity dichotomy for both global and pair stability since for single-layer preference lists, all three concepts of stability are the same and polynomial- time computable. In Section 5 we investigate the case of finding stable matchings with respect to less than all layers and only find NP-hardness results. In Section 6, we identify two special scenarios with strong but natural restrictions on the preference lists. For the fist scenario we assume that one side of the agents has single-layered preferences, i.e. on one side the preference list of each agent remains the same in all layers. We find that under such restrictions two out of three studied concepts are equivalent, and can be computed in polynomial time; for global stability we obtain W[1]-hardness (and also NP-hardness) and XP membership for the threshold parameter α. In the second scenario, we assume that the preferences of all agents on each side are uniform in each layer, i.e. when for each fixed layer and side all agents have the same preference list, and when considering individual stability we find surprising tight connections to the complexity of the GRAPH ISOMORPHISM problem. Table 1 gives a broad overview on our complexity results. 2 Definitions and Notations For each natural number t by [t] we denote the set {1, 2, . . . , t}. Let U = {u1, . . . , un} and W = {w1, . . . , wn} be two n-element disjoint sets of agents. There are ℓ lay- ers of preferences, where ℓ is a non-negative integer. For each i ∈ [ℓ] and each u ∈ U , let ≻(i) u be a linear order on W that represents the ranking of agent u over all agents from W in layer i. Analogously, for each i ∈ [ℓ] and each w ∈ W , the symbol ≻(i) w represents a linear order on U that encodes preferences of w in layer i. We refer to such linear orders as preference lists. A preference profile Pi of layer i ∈ [ℓ] is a collection of preference lists of all the agents in layer i, {≻(i) a a ∈ U ∪ W }. Let U ⋆ W = {{u, w} u ∈ U ∧ w ∈ W }. A matching M ⊆ U ⋆ W is a set of pairwisely disjoint pairs, i.e. for each two pairs p, p′ ∈ M it holds that p ∩ p′ = ∅. If {u, w} ∈ M , then we also use M (u) to 4 refer to w and M (w) to refer to u, and we say that u and w are their respective partners under M ; otherwise we say that {u, w} is an unmatched pair. Example 2.1 below shows an example matching and introduces a graphical notation that we will use throughout the paper. Example 2.1. Consider two sets of agents, U = {u1, u2, u3} and W = {w1, w2, w3}, and two layers of preference profiles, P1 and P2. Let us recall that in the following diagram the preferences are represented through vertical lists where more preferred agents are put above the less preferred ones. For instance, in the diagram the preference list of agent u3 in the first layer (profile P1) is w2 ≻(1) u3 w3 ≻(1) u3 w1. P1: w3 w2 w1 u1 w1 u2 u3 u1 w1 w2 w3 u2 w2 u3 u1 u2 w2 w3 w1 u3 w3 u3 u1 u2 P2: w2 w3 w1 u1 w1 u1 u2 u3 w3 w1 w2 u2 w2 u2 u3 u1 w1 w2 w3 u3 w3 u3 u1 u2 In our diagrams we will depict stable matchings in each layer through edges between matched nodes. If a layer has more than one stable matching, then we will use different types of lines (solid, dashed, dotted) and different colors to distinguish between them. For instance, in the above example profile P1 has one stable matching M1 = {{u1, w3}, {u2, w1}, {u3, w2}}, and P2 has three stable matchings: (1) M2 = ⋄ {{u1, w1}, {u2, w2}, {u3, w3}}, (2) M3 = {{u1, w2}, {u2, w3}, {u3, w1}}, and (3) M1. Let us now introduce two notions that we will use when defining various concepts of stability. Definition 2.1 (Dominating pairs and blocking pairs). Let M be a matching over U ∪ W . Consider an unmatched pair {u, w} ∈ (U ⋆ W ) \ M and a layer i ∈ [ℓ]. We say that {u, w} dominates {u, v} in layer i if w ≻(i) (1) u is unmatched in M or {u, w} dominates {u, M (u)} in layer i, and (2) w is unmatched in M or {u, w} dominates {w, M (w)} in layer i. u v. We say that {u, w} is blocking matching M in layer i if it holds that For a single layer i, a matching M is stable in layer i if no unmatched pair is blocking M in layer i. Let us illustrate the concept of dominating and blocking pairs through Example 2.1. Consider the matching M3 = {{u1, w2}, {u2, w3}, {u3, w1}} and profile P1 of layer 1. Here, pair {u1, w3} dominates both {u1, w2} (since u1 prefers w3 to w2) and {u2, w3} (since w3 prefers u1 to u2). Thus, {u1, w3} is a blocking pair and so it witnesses that M is not stable in profile P1. We are interested in matchings which are stable in multiple layers, i.e. we aim at generalizing the classic STABLE MARRIAGE problem [25, 35, 27, 39] which is defined for a single layer to the case of multiple layers. The idea behind each of the concepts defined below is similar: in order to call a matching stable for multiple layers we require that it must be stable in at least a certain, given number of layers α (α is a number indicating the "strength" of the stability). However, for different concepts we require a different level of agreement with respect to which layers are required for stability. Informally speaking, on the one end of the spectrum we have a variant of stability where we require a global agreement of the agents regarding the set of α layers for which the matching must be stable. On the other end of the spectrum we have a variant where we assume that the agents act independently: an agent a would deviate if it would find another agent, say b, such that a prefers b to its matched partner in some α layers, and b prefers a to its matched partner in another, possibly different, set of α layers. In the intermediate case, we require that a deviating pair must agree on the subset of layers which form the reason for deviation. We formally define the three concepts below. 5 2.1 α-layer global stability Informally speaking, a matching M is α-layer globally stable if there exist α layers in each of which M is stable. Definition 2.2 (global stability). A matching M is α-layer globally stable if there exists a set S ⊆ [ℓ] of α layers, such that for each layer i ∈ S and for each unmatched pair {u, w} ∈ U ⋆ W \ M at least one of the two following conditions holds: (1) pair {u, M (u)} dominates {u, w} in layer i, or (2) pair {w, M (w)} dominates {w, u} in layer i. The following example describes a scenario where the above concept of multi-layer stability appears to be useful. Example 2.2. Assume that the preferences of the agents depend on external circumstances which are not known a priori. Assume that each layer represents a different possible state of the universe. If we want to find a matching that is stable in as many states of the universe as possible, then we need to find an α-layer ⋄ globally stable matching for the highest possible value of α. Already for α-layer global stability we see substantial differences compared to the original concept of stability for a single layer. It is guaranteed that such a matching always exists for α = 1; indeed this would be a matching that is stable in an arbitrary layer. However, one can observe that as soon as α > 1 an α-layer globally stable matching might not exist (see Example 3.1). 2.2 α-layer pair stability While α-layer global stability requires that the agents globally agree on a certain subset of α layers for which the matching should be stable, pair stability forbids each unmatched pair to block more than a certain number of layers. The formal definition, using the domination concept, is as follows: Definition 2.3 (pair stability). A matching M is α-layer pair stable if for each unmatched pair {u, w} ∈ (U ⋆ W ) \ M , there is a set S ⊆ [ℓ] of α layers such that for each layer i ∈ S at least one of the following conditions holds: (1) pair {u, M (u)} dominates {u, w} in layer i, or (2) pair {w, M (w)} dominates {w, u} in layer i. Definition 2.3 can be equivalently formulated using a generalization of the concept of blocking pairs. Let β ∈ [ℓ] be an integer bound. We say that a pair {u, w} ∈ (U ⋆ W ) \ M is β-blocking M if there exists a subset S ⊆ {1, 2, . . . , ℓ} of β layers such that for each i ∈ S, pair {u, w} is blocking M in layer i. Proposition 2.1. A matching M is α-layer pair stable if and only if no unmatched pair p is (ℓ − α + 1)- blocking M . Proof. To prove the statement, we show that a matching M is not α-layer pair stable if and only if there is an unmatched pair p that is (ℓ − α + 1)-blocking M . For the "if" direction, assume that {u, w} is an unmatched pair and R ⊆ [ℓ] is a subset of ℓ−α+1 layers such that {u, w} is blocking every layer in R. Now consider an arbitrary subset S ⊆ [ℓ] of size α. By the cardinalities of R and S, it is clear that S ∩ R 6= ∅. Let i ∈ S ∩ R be such a layer. Then, by assumption, we have that {u, w} is blocking M in layer i. This means that none of the conditions stated in Definition 2.3 holds. Thus, {u, w} is an unmatched pair witnessing that M is not α-layer pair stable. For the "only if" direction, assume that M is not α-layer pair stable and let {u, w} be an unmatched pair that witnesses the non-α-layer pair stability of M . We claim that {u, w} is (ℓ−α+1)-blocking M . Towards a contradiction, suppose that {u, w} is not (ℓ − α + 1)-blocking M . Then, there must be a subset S ⊆ [ℓ] of at least α layers where the pair {u, w} is not blocking M in each layer in S. Equivalently, we can say that 6 for each layer i ∈ S, {u, M (u)} dominates {u, w} in layer i or {w, M (w)} dominates {u, w} in layer i-a contradiction to {u, w} being a witness. The following example motivates α-layer pair stability. Example 2.3. Consider the case when the preferences of the agents depend on a context, yet a context is pair-specific. For instance, in matchmaking a woman may have different preferences over men depending on which country they will decide to live in. Thus, a pair of a man and a woman is blocking if they agree on certain conditions, and if they will find each other more attractive than their current partners according to ⋄ the agreed conditions. In Section 3, we show that α-layer global stability implies α-layer pair stability (Proposition 3.1). This, among other things, implies that for α = 1 an α-layer pair stable matching always exists. However, as soon as α ≥ 2 the existence is no longer guaranteed (see Example 3.1). 2.3 α-layer individual stability We move to the third and last concept of stability. Definition 2.4 (individual stability). A matching M is α-layer individually stable if for each unmatched pair {u, w} ∈ (U ⋆ W ) \ M there is a set S ⊆ [ℓ] of α layers such that at least one of the following conditions holds: (1) pair {u, M (u)} dominates {u, w} in each layer of S, or (2) pair {w, M (w)} dominates {w, u} in each layer of S. The following example illustrates a potential application in the domain of partnership agencies. Example 2.4. Assume that each layer describes a single criterion for preferences. The preferences of each agent may differ depending on the criterion. For instance, the two sets of agents can represent, respectively, men and women, as in the traditional stable marriage problem. Different criteria may correspond, for in- stance, to the intelligence, sense of humor, physical appearance etc. Assume that an agent a will have no incentive to break his or her relationship with b, and to have an affair with c if he or she prefers b to c according to at least α criteria. In order to match men with women so that they form stable relationships, ⋄ one needs to find an α-layer individually stable matching. Definition 2.4 can be equivalently formulated using a generalization of the concept of dominating pairs. Let β ∈ [ℓ] be an integer bound. We say that a pair {u, w} is β-dominating {u, w′} if there is a subset R ⊆ [ℓ] of β layers such that for each i ∈ R the pair {u, w} dominates {u, w′} in layer i. Proposition 2.2. A matching M is α-layer individually stable if and only if no unmatched pair {u, w} exists that is both (ℓ − α + 1)-dominating {u, M (u)} and (ℓ − α + 1)-dominating {w, M (w)}. Proof. To prove the statement, we show that a matching M is not α-layer individually stable if and only if there is an unmatched pair p that is (ℓ − α + 1)-dominating {u, M (u)} and (ℓ − α + 1)-dominating {w, M (w)}. For the "if" direction, assume that {u, w} is an unmatched pair and R1, R2 ⊆ [ℓ] are two (possibly different) subsets of ℓ − α + 1 layers each, such that {u, w} is dominating {u, M (u)} in each layer i ∈ R1 and is dominating {w, M (w)} in each layer j ∈ R2. Now consider an arbitrary subset S ⊆ [ℓ] of size α. By the cardinalities of R1, R2, and S , it is clear that S ∩ R1 6= ∅ and S ∩ R2 6= ∅. Let i ∈ S ∩ R1 and j ∈ S ∩ R2 be two layers in the intersections. Then, by assumption, we have that {u, w} is dominating {u, M (u)} in layer i and {u, w} is dominating {w, M (w)} in layer j. This means that none of the conditions stated in Definition 2.4 holds. Thus, {u, w} is an unmatched pair witnessing that M is not α-layer individually stable. For the "only if" direction, assume that M is not α-layer individually stable and let {u, w} be an un- matched pair that witnesses the non-α-layer individual stability of M . We claim that {u, w} is (ℓ − α + 1)- dominating {u, M (u)} and is (ℓ − α + 1)-dominating {w, M (w)}. Towards a contradiction, first suppose 7 that {u, w} is not (ℓ − α + 1)-dominating {u, M (u)}, meaning that there are at most ℓ − α layers where {u, w} dominates {u, M (u)}. This implies that there is a subset S ⊆ [ℓ] of α layers such that for each i ∈ S, the pair {u, M (u)} is dominating {u, w}, a contradiction to {u, w} being a witness of the non-α- layer individual stability of M . Analogously, if {u, w} was not (ℓ − α + 1)-dominating {w, M (w)}, then we could obtain the same contradiction. For α = 1 an α-layer individually stable matching always exists (it will follow from Propositions 3.1 and 3.3); however, this is no longer the case when α ≥ 2 (see Proposition 3.2 and Example 3.1). Observe that according to α-layer individual stability the preferences of the agents can be represented as sets of linear orders: it does not matter which preference order comes from which layer. This is not the case for the other two concepts. 2.4 Central computational problems In this paper, we study the algorithmic complexity of finding matchings that are stable according to the above definitions. To this end, we first investigate how the three concepts relate to each other. Next, we formally define the search problem of finding an α-layer globally stable matching. GLOBALLY STABLE MARRIAGE Input: Two disjoint sets of n agents each, U and W , ℓ preference profiles, and an integer bound α ∈ [ℓ]. Output: Return an α-layer globally stable matching if one exists, or claim there is no such. The other two problems, PAIR STABLE MARRIAGE and INDIVIDUALLY STABLE MARRIAGE, are defined analogously. 3 Relations Between the Multi-Layer Concepts of Stability Below we establish relations among the three concepts. We start by showing that α-layer pair stability is a weaker notion than α-layer global stability and α-layer individual stability. Proposition 3.1. An α-layer globally stable matching is α-layer pair stable. Proof. Let M be an α-layer globally stable matching and let S ⊆ {1, . . . , ℓ} be such that S = α and that for each i ∈ S, matching M is stable in layer i. For the sake of contradiction let us assume that M is not α-layer pair stable. By Proposition 2.1, let {u, w} be an (ℓ − α + 1)-blocking pair for M . Let S′ ⊆ {1, . . . , ℓ} be such that S′ = ℓ − α + 1 and that for each i ∈ S′, pair {u, w} blocks M in layer i. Since S′ + S = ℓ + 1, we get that S ∩ S′ 6= ∅. Let i ∈ S ∩ S′. This gives a contradiction since on the one hand M is stable in layer i, and on the other hand {u, w} blocks M in layer i. Proposition 3.2. An α-layer individually stable matching is α-layer pair stable. Proof. Consider an α-layer individually stable matching M . Towards a contradiction, suppose that M is not α-layer pair stable. By Proposition 2.1, this means that there exists an unmatched pair {u, w} and a subset S′ ⊆ [ℓ] of ℓ − α + 1 layers such that {u, w} is blocking M in each layer from S′. This means that {u, w} is both (ℓ − α + 1)-dominating {u, M (u)} and (ℓ − α + 1)-dominating {w, M (w)}. Then, by Proposition 2.2, M is not α-layer individually stable, a contradiction. Example 3.1, below, shows a matching which is α-layer pair stable, but which is not α-layer globally sta- ble. This example, together with Proposition 3.1, also shows that α-layer global stability is strictly stronger than α-layer pair stability. 8 Example 3.1. Consider an instance with six agents and two layers of preference profiles. P1: w1 w2 w3 u1 w1 u2 u1 u3 w2 w1 w3 u2 w2 u2 u1 u3 w3 w1 w2 u3 w3 u3 u1 u2 P2: w2 w1 w3 u1 w1 u3 u1 u2 w3 w1 w2 u2 w2 u1 u2 u3 w1 w3 w2 u3 w3 u2 u1 u3 Observe that matching M = {{u1, w1}, {u2, w3}, {u3, w2}} is 1-layer individually stable and, thus 1-layer pair stable. However, M is blocked by pair {u2, w1} in the first layer and by {u1, w2} in the second. Thus, M is not 1-layer globally stable. Indeed the only 1-layer globally stable matchings are indicated by the solid lines, which are also 1-layer individually stable (and thus 1-layer pair stable). As soon as α ≥ 2, α-layer pair stability is not guaranteed to exist, even if ℓ > α. To see this we augment the instance with one more layer whose preference lists are identical to the first layer given in Example 2.1. One can verify that for each of all six possible matchings, there is always an unmatched pair that is blocking ⋄ at least two layers. For α = 1, we observe that 1-layer pair stability is equivalent to 1-layer individual stability. Proposition 3.3. A matching is 1-layer pair stable if and only if it is 1-layer individually stable. Proof. By Proposition 3.2, we know that 1-layer individual stability implies 1-layer pair stability. It remains to show the other direction. Let M be a 1-layer pair stable matching. Suppose, for the sake of contradic- tion, that M is not 1-layer individually stable. By Proposition 2.2, this means that there is an unmatched pair {u, w} that is both ℓ − 1 + 1 = ℓ-dominating {u, M (u)} and ℓ − 1 + 1 = ℓ-dominating {w, M (w)}. This implies that the pair {u, w} is indeed ℓ-blocking M , which by Proposition 2.1, is a contradiction to M being 1-layer pair stable. Example 3.2 shows that for α > 1, individual stability and pair stability are not equivalent, neither is individual stability equivalent to global stability. Example 3.2. Consider the example given in Section 1. Recall that the first layer admits exactly one stable matching, namely M1 = {{u1, w1}, {u2, w2}} (depicted by solid lines). The second layer also admits exactly one (different) stable matching, namely M2 = {{u1, w2}, {u2, w1}} (also depicted by solid lines). The third layer has two stable matchings, M1 and M2 (depicted by solid lines and dashed lines, resp.). Thus, both M1 and M2 are 2-layer globally stable (and 2-layer pair stable). However, M1 is 2-layer individually stable while M2 is not. To see why M2 is not 2-layer individually stable, we can verify that the unmatched pair p = {u1, w1} dominates {u1, w2} in the first and the third layer and it dominates {u2, w1} in the first two layers. By Proposition 2.2, M2 is not 2-layer individually stable since ℓ − α + 1 = 2. If we restrict the example to the last two layers only, then matching M2 is also evidence that an ℓ-layer globally stable (which, by Proposition 3.4, implies ℓ-layer pair stability) is not ℓ-layer individually stable. ⋄ For α = ℓ global stability and pair stability are equivalent. Proposition 3.4. For α = ℓ, a matching is α-layer globally stable if and only if it is α-layer pair stable. Proof. By Proposition 3.1, ℓ-layer global stability implies ℓ-layer pair stability. Now, assume that a match- ing M is ℓ-layer pair stable. For the sake of contradiction, suppose that M is not ℓ-layer globally stable. This means that there exists a pair, say {u, w}, and a layer, say i, such that {u, w} is blocking in layer i. Thus, {u, w} is 1-blocking M , and so M cannot satisfy ℓ-layer pair stability. This gives a contradiction. 9 It is somehow counter-intuitive that even an ℓ-layer globally stable matching (i.e. a matching that is stable in each layer) may not be ℓ-layer individually stable (see Example 3.2). By Proposition 3.5 we can thus infer that ℓ-layer global stability is strictly weaker than ℓ-layer individual stability. Proposition 3.5. For α = ℓ, an α-layer individually stable matching is α-layer globally stable. Proof. Proposition 3.2 and Proposition 3.4 imply the statement since α = ℓ. By Example 3.2, ℓ-layer global stability does not imply ℓ-layer individual stability. However, it implies ⌈ℓ/2⌉-layer individual stability. Proposition 3.6. Every ℓ-layer globally stable matching is ⌈ℓ/2⌉-layer individually stable. There are in- stances where ℓ-layer globally stable matchings are not (⌈ℓ/2⌉ + 1)-layer individually stable. Proof. For the first statement, let M be an ℓ-layer globally stable matching. Suppose, for the sake of contradiction, that M is not ⌈ℓ/2⌉-layer individually stable. Let β = ℓ − ⌈ℓ/2⌉ + 1, which is ⌊ℓ/2⌋ + 1. By Proposition 2.2, let {u, w} be an unmatched pair that is both β-dominating {u, M (u)} and β-dominating {w, M (w)}. Since 2 · β > ℓ, there is at least one layer i where {u, w} is dominating both {u, M (u)} and {w, M (w)}, meaning that {u, w} is blocking layer i, a contradiction to M being ℓ-layer globally stable. To see why ℓ-layer global stability may not imply (⌈ℓ/2⌉ + 1)-layer individual stability, consider the following instance with four agents and ℓ = 4 layers. P1: w1 w2 u1 w1 u1 u2 w1 w2 u2 w2 u1 u2 P2: w1 w2 u1 w1 u1 u2 w2 w1 u2 w2 u1 u2 P3: w2 w1 u1 w1 u2 u1 w2 w1 u2 w2 u2 u1 P4: w2 w1 u1 w1 u1 u2 w2 w1 u2 w2 u2 u1 M = {{u1, w1}, {u2, w2}} is the only 4-layer globally stable matching. However, it is not 3-layer individ- ually stable as the unmatched pair {u1, w2} dominates {u1, w1} in layers 3 and 4 and dominates {u2, w2} in layers 1 and 2. By Proposition 2.2, M is not 3-layer individually stable since ℓ − α + 1 = 2. The relations among the different concepts of multi-layer stability are depicted in Figure 1. A 1-layer globally stable matching always exists. Together with Propositions 3.1 and 3.2, we obtain the following. Proposition 3.7. A preference profile with ℓ layers always admits a matching, which is 1-layer globally stable, 1-layer pair stable, and 1-layer individually stable. 4 All-Layers Stability (α = ℓ) In this section, we discuss the special case when α = ℓ. It turns out that deciding whether a given instance admits an ℓ-layer individually stable matching can be solved in polynomial time. For the other two concepts of stability, however, the problem becomes NP-hard even when ℓ = 2. 4.1 Algorithm for ℓ-layer individual stability The algorithm for deciding ℓ-layer individual stability is based on the following simple lemma. Lemma 4.1. Let u ∈ U and w ∈ W be two agents such that w is the first ranked agent of u in some layer i ∈ [ℓ], and let u′ ∈ U \ {u} be another agent such that w prefers u over u′ in some layer j ∈ [ℓ]. Then, no ℓ-layer individually stable matching contains {u′, w}. 10 ℓ-global Prop 3.5, Ex 3.2 Prop 3.6 for α ≤ ⌈ℓ/2⌉ ℓ-individual α-global α-individual 1-global Props 3.1 and 3.3, Ex 3.1 1-individual 3 . 3 p P r o Prop Pro p P r o P r o p 3 . 4 p 3 . 1 , E x 3 . 1 3.1, E x 3.1 1-pair α-pair ℓ-pair 3.2 3.5,Ex 3.2 3.2, E x and 3.4 Props Figure 1: Relations among the different multi-layer concepts of stability for different values of α. Herein, an arc is to be read like an implication: one property implies the other. Proof. Let u, w, u′ be the three agents and let i, j be the two (possibly equal) layers as described by the assumption. Suppose towards a contradiction that there is an ℓ-layer individually stable matching with {u′, w} ∈ M . This implies that {u, w} is an unmatched pair under M . However, w prefers u over u′ = M (w) in layer j and u prefers w over M (u) in layer i-a contradiction to M being ℓ-layer individually stable. Lemma 4.1 leads to Algorithm 1 which looks quite similar to the so-called extended Gale-Shapley al- gorithm by Irving [28]. The crucial difference is that we loop into different layers and we cannot delete a pair p of agents that does not belong to any stable matching, as it may still serve to block certain matchings. Instead of deleting such pair, we will mark it. Herein, marking a pair {u, w} means marking the agent u (resp. w) in the preference list of w (resp. u) in every layer. The correctness of Algorithm 1 follows from Lemmas 4.2 to 4.4. Lemma 4.2. If a pair {u′, w} is marked during the execution of Algorithm 1, then no ℓ-layer individually stable matching contains this pair. Proof. Each pair is marked within two "foreach" loops in Line 2 and Line 3, respectively (we will refer to them as the "outer" loop and the "inner" loop). Let us fix an arbitrary u ∈ U and i ∈ [ℓ] and consider the pairs which was marked when the outer and the inner loops were run for u and i, respectively. We show the statement via induction on the sequence of pairs which were marked when u and i was considered for the two loops. For the induction to begin, let {u′, w} with u′ ∈ U and w ∈ W be the first pair that is marked during the execution. This implies that agent u ranks w in the first position in layer i and that w prefers u to u′ in some (possibly different) layer. By Lemma 4.1, no ℓ-layer individually stable matching contains {u′, w}. For the induction assumption, let {u′, w} be the mth pair (for given u and i) that is marked during the execution and no ℓ-layer individually stable matching contains a pair that is marked prior to {u′, w}. Suppose for the sake of contradiction that there is an ℓ-layer individually stable matching M which contains the marked pair {u′, w}. The fact that {u′, w} has been marked implies that 11 Algorithm 1: Algorithm for finding an ℓ-layer individually stable matching. Input: A set of agents U ∪ W and ℓ layers of preferences. 1 repeat 2 3 4 5 6 7 8 9 10 11 foreach agent u ∈ U do foreach layer i = 1, 2, . . . , ℓ do w ← the first ranked agent in u's preference list in layer i r ← 1 repeat foreach u′ with w : u ≻(j) w u′ for some layer j do mark {u′, w} r ← r + 1 w ← the rth ranked agent in u's preference list in layer i until {u, w} is not marked 12 until (some agent's preference list consists of only marked agents) or (no new pair was marked in the last iteration) 13 if some agent's preference list consists of only marked agents then no ℓ-layer individually stable matching exists 14 else return M = {{u, w} w ← the first unmarked agent in u's preference list in any layer} as an ℓ-layer individually stable matching 1. u ranks w in the pth position in layer i for some p, and 2. w prefers u over u′ in some layer j. However, by the description of the algorithm in layer i (Lines 4–11) for each agent w′ that u prefers to w in layer i, i.e. w′ ≻(i) u w, we have that {u, w′} is marked (see the "until" condition in Line 11). The induction assumption implies that M does not contain any {u, w′} with w′ ≻(i) u w. Thus, it follows that u prefers w to M (u) in layer i, i.e. w ≻(i) u M (u). This is a contradiction to M being ℓ-layer individually stable on the unmatched pair {u, w} since there is a layer j ∈ [ℓ] such that u ≻(j) w u′ = M (w). The following lemma ensures that in Line 14 if w is matched to an agent u, then it is the most preferred unmarked agent of u in all layers. Lemma 4.3. If no agent's preference list consists of only marked agents and there is an agent u and two different layers i, j ∈ [ℓ], i 6= j such that the first unmarked agent in the preference list of u in layer i differs from the one in layer j, then Algorithm 1 will mark at least one more pair. Proof. Suppose towards a contradiction that no new pair is marked, but there is an agent u ∈ U such that the first agent unmarked by u is different in different layers, say w and w′ in layers i and j, with w 6= w′ and i 6= j. Since no new pair will be marked, u is the last unmarked agent in the preference lists of w and w′ in all layers (see Line 8 of Algorithm 1). Since U = W there is a different agent u′ ∈ U \ {u} such that for each agent w ∈ W we have that u′ is not the last unmarked agent in the preference list of w in any layer. Since no agent's preference list consists of only marked agents, the preference list of u′ (in some layer) contains an agent which is unmarked. Denote this agent as w′. Again, since no new pair will be marked, in the preference list of w′′, agent u′ is the last unmarked agent-a contradiction. When Algorithm 1 terminates and no agent contains a preference list that consists of only marked agents, then we can construct an ℓ-layer individually stable matching by assigning to each agent u its first unmarked agent in any preference list (note that Lemma 4.3 ensures on termination that for each agent it holds that in all layers is first unmarked agent is the same). Lemma 4.4. If upon termination no agent's preference list consists of only marked agents, then the match- ing M computed by Algorithm 1 is ℓ-layer individually stable. 12 u M (u) and w : u ≻(j) Proof. Towards a contradiction suppose that M returned by Algorithm 1 is not ℓ-layer individually stable. That is, there is an unmatched pair {u, w} ∈ (U ⋆ W ) \ M with u ∈ U and w ∈ W and two layers i, j ∈ [ℓ] such that u : w ≻(i) w M (w). Observe that agent u is matched with the first agent, denoted as x, in the preference list of u such that the pair {u, x} is not marked, and so we infer that {u, w} is marked. Thus, the innermost loop of the algorithm has been run for the pair {u, w} (see Line 11). By Line 8 of Algorithm 1 for all agents u′ where w : u ≻(j) w u′ for some layer j′ ∈ [ℓ], the pair {u′, w} is marked. This includes the pair {M (w), w} since w : u ≻(j) w M (w)-a contradiction to Lemma 4.2. Finally, we obtain that Algorithm 1 computes an ℓ-layer individually stable matching if one exists. Theorem 4.1. For α = ℓ, Algorithm 1 solves INDIVIDUALLY STABLE MARRIAGE in O(ℓ · n2) time. Proof. Let I = (U, W, P1, P2, . . . , Pℓ) with 2 · n agents be the input of Algorithm 1. By Lemma 4.2, no ℓ- layer individually stable matching contains a marked pair. If there is an agent whose preference list consists of only marked agents, then we can immediately conclude that the given instance is a no-instance. Otherwise, Lemma 4.4 proves that the algorithm returns an ℓ-layer individually stable matching. It remains to show that the algorithm terminates and has running time O(ℓ · n2). Since there are in total O(n2) pairs, the algorithm will eventually terminate, either because some agent's preference list consists of only marked agents or because no new new pair will be marked. By using a list that points to the first unmarked agent of each agent u in each layer, and by using a table that stores pairs which are already marked and reconsidered by Line 8, the algorithm needs to "touch" each pair at most twice, once when it is not yet marked and a second time when it is already marked. 4.2 NP-hardness for ℓ-layer global stability and ℓ-layer pair stability In contrast to ℓ-layer individual stability, in this section we show that deciding GLOBALLY STABLE MAR- RIAGE is NP-hard as soon as ℓ = 2. We establish this by reducing the NP-complete 3-SAT problem [26] to the decision variant of GLOBALLY STABLE MARRIAGE. The idea behind this reduction is to introduce for each variable four agents that admit exactly two possible globally stable matchings, one for each truth value. Then, we construct a satisfaction gadget for each clause by introducing six agents. These agents will have three possible globally stable matchings. We use a layer for each literal contained in the clause to enforce that setting the literal to false will exclude exactly one of the three globally stable matchings. Therefore, unless one of the literals in the clause is set to true, no globally stable matching remains. Using the above idea, we can already show hardness of deciding ℓ-layer global stability for ℓ = 3. With some tweaks and using a restricted variant of 3-SAT [26] (see Lemma 4.5), we can strengthen our hardness result to hold even for ℓ = 2. From here on, we call a clause monotone if the contained literals are either all positive or all negative. Lemma 4.5. 3-SAT is NP-hard even if each clause has either two or three literals, and no size-three clause is monotone while all size-two clauses are monotone Proof. We start with a 3-SAT instance and do the following. For each variable xi introduce a helper vari- able zi, and make sure that the helper variable zi is set to false if and only if the original variable xi is set to true. To achieve this, we add to the instance two new clauses (xi ∨ zi) and (xi ∨ zi). Finally, for each original clause (note that it has size three) that contains only positive literals (resp. only negative literals), say xi∨xj ∨xk (resp. xi∨xj ∨xk), we replace an arbitrary literal, say xi (resp. xi), with zi (resp. zi). Observe that in the new instance, each original clause has size three and contains at least one negative and at least one positive literal, and that the newly introduced clauses have size two and are monotone. It is straightforward to see that the original instance is a yes-instance if and only if the new instance is a yes-instance. Now, we are ready to present one of our main results. 13 Theorem 4.2. GLOBALLY STABLE MARRIAGE is NP-hard even if α = ℓ = 2. Proof. We provide a polynomial-time reduction from an NP-complete restricted variant of 3-SAT as given by Lemma 4.5 to the decision version of GLOBALLY STABLE MARRIAGE. Further, without loss of gener- ality we assume that no clause contains two literals of the form x and x as it will be satisfied anyway and can be ignored from the input instance. Let (X, C) be an instance of the aforementioned 3-SAT variant with X = {x1, x2, . . . , xn} being the set of variables and C = {C1, C2, . . . , Cm} being the set of clauses of size at most three each. To unify the expression, for each size-three clause Cj = ℓ1 j we order the literals so that the first literal is positive and the second one is negative. For each size-two clause Cj = ℓ1 j (note that it is monotone), we order the literals arbitrarily, and we call it a positive clause if it contains only positive literals, otherwise we call it a negative clause. j ∨ ℓ2 j ∨ ℓ2 j ∨ ℓ3 For each variable xi ∈ X, we create four variable agents xi, xi, yi, yi (we will make it clear when using xi and xi whether we are referring to the literals or the variable agents). We will construct the preference lists of the variable agents so that each globally stable matching contains either M true := {{xi, yi}, {xi, yi}} or M false will correspond to setting the variable xi to true or false, respectively. := {{xi, yi}, {xi, yi}}. Briefly put, using M true and M false i i i i For each clause Cj ∈ C, we create six clause agents aj, bj, cj, dj, ej, fj . We will construct preference lists for these clause agents so that for each size-three-clause, there are exactly three different ways in which these agents are matched in a globally stable matching, and for each size-two-clause, there are exactly two such ways. We use two layers to enforce that setting a different literal contained in the clause Cj to false excludes exactly one of these ways. In total, we have 4n + 6m agents and we divide them into two groups U and W with U = {xi, xi i ∈ [n]}} ∪ {ai, bi, cj j ∈ [m]} and W = {yi, yi i ∈ [n]} ∪ {dj, ej, fj j ∈ [m]}. Preference lists of the variable agents. The preference lists of the variable agents restricted to the variable agents have the same pattern. We use the symbol "· · · " to denote some arbitrary order of the remaining agents (that is, agents which were not yet explicitly mentioned in the preference list). Layer (1) : ∀i ∈ [n]: Layer (2) : ∀i ∈ [n]: xi : yi ≻ Di ≻ yi ≻ · · · , xi : yi ≻ yi ≻ · · · , yi : xi ≻ Ai ≻ xi ≻ · · · , yi : xi ≻ xi ≻ · · · . xi : yi ≻ yi ≻ · · · , xi : yi ≻ D′ i ≻ yi ≻ · · · , yi : xi ≻ xi ≻ · · · , yi : xi ≻ A′ i ≻ xi ≻ · · · . It remains to specify the meaning of symbols Ai, Di, A′ i, and D′ i. Ai denotes a list (in an arbitrary order) of all clause agents aj that satisfy either of the following conditions: (a) aj corresponds to a size-three-clause Cj such that the second literal of clause Cj (which is a negative literal) is xi, or (b) it corresponds to a negative size-two-clause Cj such that the first literal of clause Cj is xi. Di denotes a list (in an arbitrary order) of all clause agents dj such that the first literal of clause Cj is xi (note that in this case Cj has either three literals or exactly two positive literals). A′ i denotes a list (in an arbitrary order) of all clause agents aj such that the last literal of Cj is xi (note that in this case Cj has either three literals or exactly two positive literals). D′ i denotes a list (in an arbitrary order) of all clause agents dj such that the last literal of Cj is xi (note that in this case Cj has either three literals or exactly two negative literals). 14 To illustrate the above notation, suppose that variable xi appears in four size-three-clauses, call them C1, C2, C3, and C5, and in two size-two-clauses: C4 and C6. The positive literal xi is the first literal in clauses C1, C3, and C4. The negative literal xi is the second literal in C2, and the last literal in C5 and C6. In this case, Ai = a2, Di could be Di = d1 ≻ d3 ≻ d4, A′ i is empty, and D′ i could be D′ i = d5 ≻ d6. Preference lists of the clause agents. The preference lists for the clause agents in the first layer are "fixed" when restricted to the clause agents; they only differ in the positions of variable agents. For a clause Cj and an integer t ∈ {1, 2, 3} let C (t) if Cj has two literals). For a literal ℓi which is xi or xi, by X(ℓi), Y (ℓi), X(ℓi), and Y (ℓi), we denote the variable agents xi, yi, xi, and yi, respectively, all corresponding to variable xi. For instance, for a clause Cj = x2 ∨ x4 ∨ x5, we have that Y (C (1) j denote the t-th literal in Cj (there will be no C (3) ) = y2, and Y (C (2) ) = y4. j j j Layer (1), ∀j ∈ [m]: Cj = 3: ) ≻ fj ≻ · · · , dj : bj ≻ cj ≻ X(C (1) ) ≻ aj ≻ · · · , j aj : dj ≻ ej ≻ Y (C (2) bj : ej ≻ fj ≻ dj ≻ · · · , cj : fj ≻ dj ≻ ej ≻ · · · , j Cj = 2 and Cj is positive: aj : dj ≻ ej ≻ · · · , bj : ej ≻ dj ≻ · · · , cj : fj ≻ · · · , Cj = 2 and Cj is negative: aj : dj ≻ Y (C (1) j ) ≻ ej ≻ · · · , bj : ej ≻ dj ≻ · · · , cj : fj ≻ · · · , ej : cj ≻ aj ≻ bj ≻ · · · , fj : aj ≻ bj ≻ cj ≻ · · · , dj : bj ≻ X(C (1) ej : aj ≻ bj ≻ · · · , fj : cj ≻ · · · , j ) ≻ aj ≻ · · · , dj : bj ≻ aj ≻ · · · , ej : aj ≻ bj ≻ · · · , fj : cj ≻ · · · . The preference lists for the second layer depends on the "positiveness" of the last literal. There are two variants: Layer (2), ∀j ∈ [m] with Cj = 3: Variant 1 (C (3) j is positive) : aj : fj ≻ dj ≻ Y (C (3) j ) ≻ ej ≻ · · · , dj : cj ≻ aj ≻ bj ≻ · · · , bj : dj ≻ ej ≻ fj ≻ · · · , cj : ej ≻ fj ≻ dj ≻ · · · , Variant 2 (C (3) j is negative) : aj : ej ≻ fj ≻ dj ≻ · · · , bj : fj ≻ dj ≻ ej ≻ · · · , cj : dj ≻ ej ≻ fj ≻ · · · , Layer (2), ∀j ∈ [m] with Cj = 2: ej : aj ≻ bj ≻ cj ≻ · · · , fj : bj ≻ cj ≻ aj ≻ · · · , dj : aj ≻ bj ≻ X(C (3) ej : bj ≻ cj ≻ aj ≻ · · · , fj : cj ≻ aj ≻ bj ≻ · · · , j ) ≻ cj ≻ · · · , Variant 1 (Cj is positive) : j aj : dj ≻ Y (C (2) bj : ej ≻ dj · · · , cj : fj ≻ · · · , Variant 2 (Cj is negative) aj : dj ≻ ej ≻ · · · , 15 ) ≻ ej ≻ · · · , dj : bj ≻ aj ≻ · · · , ej : aj ≻ bj · · · , fj : cj ≻ · · · , dj : bj ≻ X(C (2) j ) ≻ aj ≻ · · · , bj : ej ≻ dj · · · , cj : fj ≻ · · · , ej : aj ≻ bj · · · , fj : cj ≻ · · · . This completes the construction which can be done in polynomial time. Before we show the correctness of our construction, we first discuss some properties that each 2-layer globally stable matching M must satisfy. Claim 1. Let M be a 2-layer globally stable matching for our two-layer preference profiles. For each variable xi ∈ X, it holds that either M true i ⊆ M or M false i ⊆ M . Proof. To see this, we distinguish between two cases, depending on whether the partner of xi, M (xi), is yi or not. If M (xi) 6= yi, then by the stability of M for the first layer, it follows that yi prefers its partner M (yi) to xi in the first layer as otherwise xi and yi are forming a blocking pair for the first layer. Since xi is the only agent that yi prefers to xi in this layer, we have that M (yi) = xi. Then, it must hold that M (xi) = yi as otherwise xi and yi would block M in the first layer. Thus, M false i ⊆ M . Similarly, if M (xi) = yi, then by the stability of M and by construction of the preference lists of xi and (of Claim 1) ⋄ yi in the second layer we must have that M (xi) = yi. This leads to M true i ⊆ M . We obtain a similar result for the clause agents. For each clause Cj ∈ C with Cj = 3, let N 1 j = {{aj, ej }, {bj, fj}, {cj , dj}}. j = {{aj , fj}, {bj , dj}, {cj , ej}}, N 3 {{aj, dj }, {bj , ej}, {cj , fj}}, N 2 j = Claim 2. Let Cj ∈ C be a size-three-clause, and let xi be a variable that appears (as either a positive or a negative literal) in Cj . For a 2-layer globally stable matching M the following conditions hold: (i) If xi is the first literal in Cj and if M false (ii) If xi is the second literal in Cj and if M true (iii) If xi is the third literal in Cj and if M false (iv) If xi is the third literal in Cj and if M true i ⊆ M , then either N 2 j ⊆ M or N 3 j ⊆ M . i ⊆ M , then either N 1 i ⊆ M , then either N 1 i ⊆ M , then either N 1 j ⊆ M or N 3 j ⊆ M or N 2 j ⊆ M or N 2 j ⊆ M . j ⊆ M . j ⊆ M . Proof. We consider the four cases separately: i (i) Assume that xi is the first literal in Cj and M false ⊆ M . This implies that {xi, yi} ∈ M . Consider the preference list of xi in the first layer, and observe that dj appears in Di. Since xi prefers dj to its partner yi in the first layer, it follows that dj must obtain a partner that it prefers to xj in the first layer. By the preference list of dj in the first layer, we have that M (dj) ∈ {bj, cj }. If M (dj) = bj, then by the preference list of bj in the first layer it follows that both ej and fj must obtain partners that they find better than bj in the first layer. This means that M (fj) = aj and M (ej) ∈ {aj, cj }, implying that M (ej ) = cj . Analogously, if M (dj) = cj , then {aj, ej} ∈ M as otherwise they will block the first layer since the most preferred agents of both aj and ej are already assigned to someone else, and aj and ej are each other's second most preferred agents. Then, bj must obtain a partner that it prefers to dj . Since fj is the only agent left that bj prefers to dj, we get that M (bj) = fj, and so N 3 j ⊆ M . (ii) Assume that xi is the second literal in Cj and M true i ⊆ M (thus, in particular, {xi, yi} ∈ M ). Since aj appears in Ai in the preference list of yi in the first layer, we infer that yi prefers aj to its partner xi in the first layer. Thus, it follows that aj must obtain a partner that it prefers to yi in the first layer, i.e. that M (aj) ∈ {dj, ej }. If M (aj) = dj , then by considering the preference list of dj in the first layer we infer that both bj and cj obtain partners that they find better than dj in the first layer. This means that M (cj) = fj , and M (bj) ∈ {ej, fj}. Since fj is taken by cj , we get that M (bj) = ej . Together, we have that N 1 Analogously, if M (aj) = ej , then {bj, fj} ∈ M as otherwise they would block the first layer since the most preferred agents of both bj and fj are already assigned to someone else, and bj and fj are each other's second most preferred agent. Moreover, cj must obtain a partner that it prefers to ej . Since dj is the only agent left that cj prefers to ej , we obtain that M (cj ) = dj . This leads to N 3 j ⊆ M . j ⊆ M . 16 (iii) Assume that xi is the third literal in Cj and M false i ⊆ M . Observe that in this case aj appears in A′ i in the preference list of yi in the second layer, and since {xi, yi} ∈ M , that yi prefers aj to its partner xi in the second layer. It follows that aj must obtain a partner that it prefers to yj in the second layer. By investigating the preference list of aj in the second layer (note that we are in Variant 1), we have that M (aj) ∈ {fj, dj }. If M (aj) = fj, then by looking at the preference list of fj in the second layer we infer that both bj and cj must obtain partners that they find better than fj in the second layer. Thus M (cj) = ej , and consequently, M (bj) = dj . Summarizing, in this case we have that N 2 Analogously, if M (aj) = dj , then {bj, ej} ∈ M as otherwise they would block the second layer. Moreover, fj must obtain a partner that it prefers to aj . Since cj is the only agent left that fj prefers to aj , we obtain that M (cj ) = fj. This leads to N 1 j ⊆ M . j ⊆ M . (iv) Finally, assume that xi is the third literal in Cj and M true i ⊆ M . In this case, we have that dj appears in D′ i in the preference list of xi in the second layer, and that {xi, yi} ∈ M . This means that xi prefers dj to its partner yi in the second layer, and so it must be the case that dj obtains a partner that it prefers to xj in the second layer. As a result, we have that M (dj) ∈ {aj, bj}. If M (dj) = aj , then the preference list of aj indicates that both ej and fj must obtain partners that they find better than aj in the second layer. Thus, M (fj) = cj , and M (ej) ∈ {bj, cj}. Consequently, M (ej ) = bj, and we get that N 1 j ⊆ M . Finally, if M (dj) = bj , then {cj, ej} ∈ M as otherwise they would block the second layer. Further, aj must obtain a partner that it prefers to dj , thus M (aj) = fj. Consequently, N 2 j ⊆ M . (of Claim 2) ⋄ For each clause Cj ∈ C with Cj = 2, let N 1 j = {{aj , dj}, {bj , ej}}, N 2 j = {{aj, ej }, {bj , dj}}. Claim 3. Let Cj ∈ C be a size-two-clause, and let xi be a variable that appears (as either a positive or a negative literal) in Cj . Assume that M is a 2-layer globally stable matching. The following holds: (i) If xi is the first literal in Cj and if M false (ii) If xi is the last literal in Cj and if M true (iii) If xi is the last literal in Cj and if M false (iv) If xi is the first literal in Cj and if M true i ⊆ M , then N 2 i ⊆ M , then N 2 i ⊆ M , then N 1 i ⊆ M , then N 1 j ⊆ M . j ⊆ M . j ⊆ M . j ⊆ M . Proof. We show the first two statements together and the last two statements together. Assume that one of the conditions in the first two statements holds, that is, 1. xi is the first literal in Cj and M false 2. xi is the last literal in Cj and M true i ⊆ M , or i ⊆ M . This implies that 1. either {xi, yi} ∈ M , and in the first layer the list Di contains dj, or 2. {xi, yi} ∈ M , and in the second layer the list D′ i contains dj and we are in Variant 2. Since xi prefers all agents from Di to yi in the first layer and xi prefers all agents from D′ i to yi in the second layer, we must have that dj obtains a partner that it prefers to xi in the first layer or to xi in Variant 2 of the second layer. In either case, bj is the only agent that fulfills the requirement, implying that {bj, dj } ∈ M . By looking at the preference lists of bj and ej in the first layer, we derive that {aj, ej} ∈ M . Thus, N 2 j ⊆ M . Analogously, assume that one of the conditions in the last two statements holds, that is, 1. xi is the last literal in Cj and M false 2. xi is the first literal in Cj and M true i ⊆ M , or i ⊆ M . This implies that 1. {xi, yi} ∈ M , and in the second layer we have Variant 1 such that the list A′ 2. {xi, yi} ∈ M , and in the first layer the list Ai contains aj. i contains aj, or Since yi prefers all agents from A′ i to xi in the second layer and yi prefers all agents from Ai to xi in the first layer, we must have that aj obtains a partner that it prefers to yi in the second layer (Variant 1) 17 or to yi in the first layer. In either case, dj is the only agent that fulfills the requirement, implying that {aj, dj} ∈ M . By the preference lists of dj and bj in the first layer, we further derive that {bj, ej } ∈ M . Thus, N 1 (of Claim 3) ⋄ j ⊆ M . Now, we are ready to show that (X, C) admits a satisfying truth assignment if and only if there exists a 2-layer globally stable matching for the so-constructed instance. (⇒) For the "only if" direction, assume that σ : X → {T, F } is a satisfying truth assignment for (X, C). We claim that the matching M constructed as follows is all-layer globally stable. (1) For each variable xi ∈ X with σ(xi) = T , let M true i ⊆ M . (2) For each size-three-clause Cj , identify a literal ℓj such that σ(ℓj) makes Cj satisfied. If ℓj is Cj's first i ⊆ M ; otherwise let M false literal, then let N 1 j ⊆ M . If ℓj is the second literal, then let N 2 j ⊆ M . Otherwise let N 3 j ⊆ M . (3) For each size-two-clause Cj , let {cj, fj} ∈ M . Identify a literal ℓj such that σ(ℓj) makes Cj satisfied. j ⊆ M ; otherwise let N 2 If ℓj is the first literal in Cj , then let N 1 Towards a contradiction suppose that M is not 2-layer globally stable, and let p = {u, w} be a possible blocking pair with u ∈ U and w ∈ W . First, we observe that p involves neither two variable agents that correspond to different variables nor two clause agents that correspond to different clauses. Second, p does not involve two variable agents that belong to the same variable since for each layer and for each two variable agents that correspond to the same variable, exactly one of both is already matched to its most preferred agent. Third, p also does not involve two clause agents that belong to the same size-two-clause as either of such clause agents is already matched with its most preferred agent. j ⊆ M . Next, consider the case that u and w are two clause agents that belong to the same size-three-clause, say Cj . If {u, w} is blocking M in the first layer, then we know that N 3 j ⊆ M as otherwise either u or w already obtains its most preferred agent. But then {u, w} cannot be blocking M in the first layer as for each agent w′ that is preferred to M (u) by u in the first layer, we have that w′ prefers M (w′) to u in the first layer. Similarly, if {u, w} is blocking M in the second layer with Variant 1 (resp. Variant 2), then we know that N 1 j ⊆ M ) as otherwise either u or w already obtains its most preferred agent. Since for each agent w′ such that u prefers w′ to M (u) in the second layer we have that w′ prefers M (w′) to u, it follows that {u, w} cannot be blocking the second layer. j ⊆ M (resp. N 2 Now, suppose that p involves a variable agent and a clause agent. By the construction of the preference lists, we can assume that the clause agent involved in the blocking pair p is either an aj or a dj for some j ∈ [m]. We distinguish between two cases, depending on the size of Cj. Case 1: Cj = 3. If aj ∈ p and p is blocking the first layer, then by the preference list of aj in the first layer, we have that N 2 j ⊆ M . By the construction of the matching M , we have that the truth assignment of the second literal, say xi, in Cj makes Cj satisfied; note that by our convention, the second literal is always negative. This implies that M false i ⊆ M . Since p involves aj and is blocking the first layer, it follows that the agent yi that corresponds to variable xi prefers ai to M (yi) in the first layer. This means that M true i ⊆ M , a contradiction. Analogously, if aj ∈ p and p is blocking the second layer, then by the preference list of aj in the second layer, we have that the preference list of aj comes from Variant 1 and N 3 j ⊆ M . By the construction of the matching M , we have that the truth assignment of the third literal which is positive in Cj (recall that Variant 1 was used) makes Cj satisfied. Let this literal be xi. Then, it must hold that M true i ⊆ M . Since p involves aj and is blocking the second layer (due to Variant 1), it follows that the agent yi that corresponds to variable xi prefers aj to M (yi) in the second layer. This means that M false ⊆ M , which results in a contradiction. i If dj ∈ p and p is blocking the first layer, then by the preference list of dj in the first layer, we have that N 1 j ⊆ M . By the construction of the matching M , we have that the truth assignment of the first literal, say xi, in Cj makes Cj satisfied; note that by convention, the first literal is always positive. This implies that M true i ⊆ M . Since p involves dj and is blocking the first layer, it follows that the agent xi that corresponds 18 to variable xi prefers dj to M (xi) in the first layer. By the preference list of xi in the first layer, it follows that M false i ⊆ M , which also yields a contradiction. Finally, if dj ∈ p and p is blocking the second layer, then by the preference list of dj in the second layer, we have that the preference list of dj comes from Variant 2 and N 3 j ⊆ M . By the construction of the matching M , we have that the truth assignment of the third literal which is negative in Cj makes Cj satisfied. Denote this literal by xi. Then, it follows that M false i ⊆ M . Since p involves dj and is blocking the second layer (due to Variant 2), it follows that the agent xi that corresponds to variable xi in Cj prefers dj to M (xi) in the second layer. By the preference list of xi in the first layer, this means that M true i ⊆ M , which is a contradiction. Case 2: Cj = 2. If aj ∈ p, then by the preference lists of aj in any of the two layers, we have that N 2 j ⊆ M . Hence, the second literal in Cj makes it satisfied. We distinguish between two cases. If the second literal in Cj is positive, say xi, then M true i ⊆ M and the other involved agent in p must be either yi or yi. Since yi prefers its partner xi to aj in both layers, it follows that yi is the other involved agent. However, by the preference list of yi in the first layer, Ai does not contain aj , meaning that yi also prefers its partner xi to aj in both layers, which is a contradiction to p being a blocking pair. If the second literal in Cj is negative, say xi, then M false ⊆ M and our reasoning is very similar. First we infer that the other involved agent in p must be either yi or yi. Since yi prefers its partner xi to aj in both layers, it follows that the other agent in p is yi. However, by the preference list of yi in the second layer, A′ i does not contain aj , which means that yi prefers its partner to aj in both layers. Thus, in this case we also get a contradiction. i If dj ∈ p, then our reasoning is very similar. First, by looking at the preference lists of dj in any of the two layers, we infer that N 1 j ⊆ M . By the construction of the matching M , we get that the first literal in Cj makes it satisfied. We consider two cases. If this literal is positive, say xi, then M true i ⊆ M and the other involved agent in p must be either xi or xi. Agent xi prefers its partner yi to dj in both layers and so it cannot be involved in the blocking pair. Thus, xi is the other involved agent. However, by the preference list of xi in the first layer, D′ i does not contain dj, meaning that xi also prefers its partner yi to dj in both layers, a contradiction to p being a blocking pair. If the first literal in Cj is negative, say xi, then M false ⊆ M and the other involved agent in p must be either xi or xi. Since xi prefers its partner yi to dj in both layers, it follows that xi is the other involved agent. However, by the preference list of xi in the last layer, Di does not contain dj, meaning that xi prefers its partner yi to dj in both layers, which is again a contradiction. i (⇐) For the "if" direction, let M be a 2-layer globally stable matching. We construct a truth assign- ment σ as follows. For each variable agent xi, if M true i ⊆ M , then let σ(xi) = T ; otherwise by Claim 1 we have that M false ⊆ M , and let σ(xi) = F . Suppose, towards a contradiction, that σ is not a satisfying assignment and let Cj be a clause where none of the literals is evaluated to true. We distinguish between two cases. i Case 1: Cj = 3. Let xr, xs, and ℓt be the first, second, and the third literal in Cj. Since none of these literals is evaluated to true, it follows that M false s ⊆ M . By statements (i) and (ii) in Claim 2, we must have that N 3 j ⊆ M . However, by the statements (iii) and (iv) in Claim 2, applied for ℓt, we must have that either N 1 Case 2: Cj = 2. Let the first and the second literals in Cj correspond to variables xi and xk, respectively. If Cj is positive, then since Cj is not satisfied, we have that M false k ⊆ M . By Claim 3, we have that both N 1 j must belong to M , which leads to a contradiction. j ⊆ M , a contradiction. j ⊆ M or N 2 j and N 2 , M false , M true r i Analogously, if Cj is negative, then since Cj is not satisfied, we have that M true , M true k ⊆ M . By i Claim 3, we have that both N 1 j and N 2 j must belong to M , a contradiction. Altogether, we showed that the constructed matching is globally stable for two layers. This concludes the proof of Theorem 4.2. 19 Since ℓ-layer global stability equals ℓ-layer pair stability (Proposition 3.4), by Theorem 4.2 we obtain the following corollary for the pair stability. Corollary 4.3. PAIR STABLE MARRIAGE is NP-hard even if α = ℓ = 2. By adding to the profile constructed in the proof of Theorem 4.2 an arbitrary number of layers with preferences that are stable anyway, we can deduce hardness for arbitrary α = ℓ ≥ 2. Proposition 4.4. For each α = ℓ ≥ 2, both GLOBALLY STABLE MARRIAGE and PAIR STABLE MAR- RIAGE are NP-hard. Proof. We add to the profile constructed in the proof of Theorem 4.2 ℓ − 2 layers with preferences of the following form: Layers (3)–(ℓ),∀i ∈ [n] : ∀j ∈ [m] : xi : yi ≻ yi ≻ · · · , xi : yi ≻ yi ≻ · · · , aj : dj ≻ ej ≻ fj ≻ · · · , bj : ej ≻ fj ≻ dj ≻ · · · , cj : fj ≻ dj ≻ ej ≻ · · · , yi : xi ≻ xi ≻ · · · , yi : xi ≻ xi ≻ · · · . dj : bj ≻ cj ≻ aj ≻ · · · , ej : cj ≻ aj ≻ bj ≻ · · · , fj : aj ≻ bj ≻ cj ≻ · · · . It is straightforward that a matching is ℓ-layer globally stable if and only if it is 2-layer globally stable for the first two layers. 5 Multi-Layer Stable Marriage with α < ℓ In this section we show that for each of the three concepts that we introduced in Section 2 the problem of computing a multi-layer stable matching is computationally hard as soon as 2 ≤ α < ℓ. 5.1 NP-hardness for α-layer global stability To find a matching M that is α-layer globally stable, even if α < ℓ, the main difficulty is not just to determine α layers where M should be stable. In fact, we sometimes need to find a matching that is stable in some specific layers. This requirement allows us to adapt the construction in the proof of Theorem 4.2 to show hardness for deciding α-layer global stability for the case when 2 ≤ α < ℓ. Proposition 5.1. For each fixed number α with α ≥ 2, GLOBALLY STABLE MARRIAGE is NP-hard. Proof. To prove the NP-hardness, we adapt the reduction in the proof of Theorem 4.2 which shows that deciding α-layer global stability for α = ℓ = 2 is NP-hard. Let P be the constructed two-layer instance in the proof of Theorem 4.2. Besides the original agents from P, we introduce two sets U and W of dummy agents with U = W = 2·(ℓ−α+1), where U = {u1, u2, . . . , uℓ−α+1} and W = {w1, w2, . . . , wℓ−α+1}. The idea of introducing such dummy agents is to make sure that each α-layer globally stable matching must include all pairs {uj, wj}, j ∈ [ℓ − α + 1]. However, this is the case only when such matching is stable in the two layers constructed in the NP-hardness proof of Theorem 4.2; we denote these two layers as layers (1) and (2). In the following, we use "· · · " to denote an arbitrary order of the unmentioned agents. Preferences of the original agents. The preferences of the original agents in the first two layers are the same as in P in the proof of Theorem 4.2. For each other layer, the preferences of the original agents are as follows. Layers (3)–(ℓ),∀i ∈ [n] : xi : yi ≻ yi ≻ · · · , yi : xi ≻ xi ≻ · · · , 20 ∀j ∈ [m] : xi : yi ≻ yi ≻ · · · , aj : dj ≻ ej ≻ fj ≻ · · · , bj : ej ≻ fj ≻ dj ≻ · · · , cj : fj ≻ dj ≻ ej ≻ · · · , yi : xi ≻ xi ≻ · · · . dj : bj ≻ cj ≻ aj ≻ · · · , ej : cj ≻ aj ≻ bj ≻ · · · , fj : aj ≻ bj ≻ cj ≻ · · · . Preferences of the dummy agents. The preferences of the dummy agents are as follows; let ℓ = ℓ−α+1: Layers (1)–(α), ∀j ∈ [ℓ] : Layer (i + α), 1 ≤ i ≤ ℓ − 1, ∀j ∈ [ℓ] : uj : wj ≻ · · · , uj : w(j mod ℓ)+i ≻ · · · , wj : uj ≻ · · · , wj : u(j−1 mod ℓ)+i ≻ · · · . Observe that each dummy agent obtains a different partner in different layers with indices higher than α. More precisely, for each layer (i + α) with 1 ≤ α ≤ ℓ − 1, the only stable matching in this layer must include Mi = {{uj, w(j mod ℓ)+1} 1 ≤ j ≤ ℓ}} since uj and w(j mod ℓ)+1 are each other's most preferred agent. Moreover, by the same reasoning, each layer with index at most α admits exactly the same stable matching regarding the dummy agents which is different from any layer with index higher than α, namely M0 = {{uj , wj} 1 ≤ j ≤ ℓ}. For each two distinct values i, j ∈ {0, 1, . . . , ℓ − 1}, however, we have that Mi ∩ Mj = ∅. This means that each α-layer globally stable matching must include M0 and must be stable in the first α layers, including the first two layers. Now, it is straightforward to see that a matching M is 2-layer globally stable for P if and only if M ∪M0 is α-layer globally stable for our new instance. We remark that our proof for Proposition 5.1 also implies hardness for α = ℓ for arbitrary ℓ ≥ 2. 5.2 NP-hardness for α-layer individual stability For α-layer individual stability, we also obtain a hardness result by reducing from the NP-hard PERFECT SMTI problem, the problem of finding a perfect SMTI-stable matching with (possibly) incomplete prefer- ence lists and ties [32, 38] which is defined as follows. A preference list is incomplete if not all agents from one side are considered acceptable to an agent from the other side. A preference list has a tie if there are two agents in the list which are considered to be equally good. As a result, a preference list of an agent u from one side can be considered as a weak (i.e. transitive and complete) order (cid:23)u on a subset of the agents on the other side. We use ≻u and ∼u to denote the asymmetric and symmetric part of the preference list, respectively. Equivalently, two agents x and y are said to be tied by u if x (cid:23)u y and y (cid:23)u x, denoted as x ∼u y. Formally, we say that a matching M for a PERFECT SMTI with two disjoints sets U and W of agents is SMTI-stable if there are no SMTI-blocking pairs for M . A pair {u, w} is SMTI-blocking M if all of the following three conditions are satisfied: (i) u and w appear in the preference lists of each other, (ii) w ≻u M (u) or u is not matched to any agent from W , and (iii) u ≻w M (w) or w is not matched to any agent from U . The reduction is based on the following ideas: In a PERFECT SMTI instance I the preference list of an agent z may have ties. To encode ties, we first "linearize" the preference list of z in I to obtain two linear preference lists such that the resulting lists restricted to the tied agents are reverse to each other. Then, we let half of the ℓ layers have one of the preference lists and let the remaining half to have the other list. Since α < ℓ/2, it is always possible to find half layers which fulfill our α-layer individual stability constraint. In I, two agents, say x and y, may not be acceptable to each other. To encode this, we introduce to the source instance ℓ pairs of dummy agents with ℓ layers of preferences that preclude x and y from being matched together. However, to make sure that an agent x is not matched to any dummy agent, we have to require that ℓ ≥ 4. 21 Theorem 5.2. For each fixed number ℓ of layers with ℓ ≥ 4 and for each fixed value α with 2 ≤ α ≤ ⌊ℓ/2⌋, INDIVIDUALLY STABLE MARRIAGE is NP-hard, even if on one side, the preference list of each agent is the same in all layers. Proof. Assume that ℓ ≥ 4 and that 2 ≤ α ≤ ⌊ℓ/2⌋. We give a polynomial-time reduction from the NP-hard PERFECT SMTI problem [32, 38]. Let I be an instance of PERFECT SMTI. In I we are given two disjoint sets of agents, U and W , with U = W = n; each agent u ∈ U is endowed with a weak order (cid:23)u on a subset of W . By the SMTI- stability, an agent u prefers not to be assigned to any agent rather than to be assigned to an agent outside of its preference list. We assume that ties occur in the preferences of the agents from the side U only, that there is at most one tie per list, and each tie is of length two as this variant remains NP-hard [38, Theorem 2.2]. From I we construct an instance I ′ of the problem of finding an α-layer individually stable matching in the following way. We copy the sets of agents U and W ; further, we introduce a set of 2ℓ · n dummy agents P ∪ R with P = R = ℓ · n. We denote the elements in these sets as: P = {pi,j 1 ≤ i ≤ n ∧ 1 ≤ j ≤ ℓ} and R = {ri,j 1 ≤ i ≤ n ∧ 1 ≤ j ≤ ℓ}. One side of the bipartite "acceptability graph" will consist of the agents from U ∪ P , and the other side of the agents from W ∪ R. We construct the preference orders of the agents as follows: Agents from U . Consider an agent u ∈ U , and let Lu denote its preference list in I. In I ′ we construct the preference list of u from Lu as follows. We iterate over Lu starting from the top position. If agent u prefers x over y, then we assume that u also prefers x over y in all layers in I ′. If x and y are tied in Lu, then we assume that u prefers x over y in the first ⌈ℓ/2⌉ layers and y over x in the remaining ⌊ℓ/2⌋ layers (or the other way around). The remaining parts of the preference orders of u are constructed as follows: first, let us assume that u = ui ∈ U . Right after all agents from Lu, agent u puts in all layers ri,1, ri,2, . . . , ri,ℓ, respectively in the following orders: (1) ri,1 ≻ ri,2 ≻ . . . ≻ ri,ℓ, (2) ri,2 ≻ ri,3 ≻ . . . ≻ ri,ℓ ≻ ri,1, and so on, until (ℓ) ri,ℓ ≻ ri,1 ≻ ri,2 ≻ . . . ≻ ri,ℓ−1. Next, u puts all the remaining agents in an arbitrary order. For example, if ℓ = 5 and Lu is equal to w1 ≻ w4 ≻ w3 ∼ w5, then the preference lists of u in the five layers will be as follows, where "· · · " denotes some arbitrary but fixed order of the remaining agents. Layer (1) : agent ui : w1 ≻ w4 ≻ w3 ≻ w5 ≻ ri,1 ≻ ri,2 ≻ ri,3 ≻ ri,4 ≻ ri,5 ≻ · · · , Layer (2) : agent ui : w1 ≻ w4 ≻ w3 ≻ w5 ≻ ri,2 ≻ ri,3 ≻ ri,4 ≻ ri,5 ≻ ri,1 ≻ · · · , Layer (3) : agent ui : w1 ≻ w4 ≻ w3 ≻ w5 ≻ ri,3 ≻ ri,4 ≻ ri,5 ≻ ri,1 ≻ ri,2 ≻ · · · , Layer (4) : agent ui : w1 ≻ w4 ≻ w5 ≻ w3 ≻ ri,4 ≻ rr,5 ≻ ri,1 ≻ ri,2 ≻ ri,3 ≻ · · · , Layer (5) : agent ui : w1 ≻ w4 ≻ w5 ≻ w3 ≻ rr,5 ≻ ri,1 ≻ ri,2 ≻ ri,3 ≻ ri,4 ≻ · · · . Observe that in the first three layers w3 is preferred to w5 and in the remaining layers w5 is preferred to w3. Agents from W . For each agent wi from W , we recall that the preferences of wi in I do not have ties (that was one of the assumptions in the problem we reduce from). Let Lwi denote the preference list of wi. The preferences of wi in I ′ are the same in all layers, where the second "· · · " denotes some arbitrary but fixed order of the remaining agents. Layers (1)-(ℓ) : agent wi : Lwi ≻ pi,1 ≻ pi,2 ≻ · · · ≻ pi,ℓ ≻ · · · . Agents from P ∪ R. Consider pi,j ∈ P : this agent ranks wi first, next ri,j, and next all the remaining agents in some arbitrary order: Layers (1)-(ℓ) : agent pi,j : wi ≻ ri,j ≻ · · · . 22 The preferences of an agent from ri,j ∈ R are constructed analogously: Layers (1)-(ℓ) : agent ri,j : ui ≻ pi,j ≻ · · · . This completes the description of the construction of instance I ′. Obviously, the preferences of each agent from W ∪ R are the same in all layers. Now, we will show that there exists a perfect SMTI-stable matching in I if and only if there exists an α-layer individually stable matching in I ′. For the "only if" direction, assume that the instance I has a perfect SMTI-stable matching M . We claim that M ′ = M ∪ (cid:8){ri,j, pi,j} 1 ≤ i ≤ n ∧ 1 ≤ j ≤ ℓ(cid:9) is α-layer individually stable for I ′. First, M ′ is a perfect matching for I ′ as M is a perfect matching for I. Second, observe that no two agents from R ∪ P are blocking M ′ as for each agent a in R ∪ P it prefers its partner M ′(a) to every other agent in R ∪ P in all layers. Third, no agent from U ∪ W can form a blocking pair with an agent from R ∪ P for the matching M ′ as each agent a from U ∪ W prefers its partner M ′(a) = M (a) ∈ La to every other agent in R ∪ P in all layers. Now, consider two arbitrary agents u and w with u ∈ U and w ∈ W such that {u, w} is unmatched. We show that we can always find a set S of ⌊ℓ⌋ ≥ α layers such that M ′(u) ≻u w in each layer from S or M ′(w) ≻w u in each layer from S. We can assume that u and w are acceptable to each other in I as otherwise both u and w prefer to be with their respective partner M ′(u) and M ′(w) rather than with each other in all layers. Let us consider the following cases; note that the preference list of w does not have ties and remains the same in all layers. Case 1: u ≻w M ′(w) = M (w). By the stability of M we have that M ′(u) = M (u) (cid:23)u w in in- stance I. For the case that M ′(u) and w tied by u we can identify ⌊ℓ/2⌋ layers, either the first ⌊ℓ/2⌋ layers or the last ⌊ℓ/2⌋ layers where u prefers M ′(u) to w. For the case that M ′(u) ≻u w we have that u prefers M ′(u) to w in all layers. Case 2: M ′(w) ≻w u. In this case we know that w prefers M ′(w) to u in all layers. For the "if" direction, assume that M ′ is an α-layer individually stable matching for I ′. First, observe that no agent ui ∈ U can be matched with an agent that it ranks lower than any agent from {ri,j 1 ≤ j ≤ ℓ} in any layer. Indeed, for such matching each pair {ui, ri,j}, with j ∈ [ℓ], would be blocking M ′ in all layers. Similarly, ui cannot be matched with ri,j for j ∈ [ℓ] as the pair {ui, ri,(j+ℓ−2 mod ℓ)+1} would be blocking M ′ in exactly ℓ−1 layers, namely those layers other than the jth layer, which are more than ℓ−α layers. For instance, if ui was matched with ri,1, then {ui, ri,ℓ} would be blocking the 2nd, 3rd, . . . , ℓth layers; if ui was matched with ri,2, then (ui, ri,1) would be blocking the 1st, 3rd, 4th, . . . , ℓth layers, etc. Thus, each agent u from U must be matched with someone from Lu, where Lu is the preference list of u in I. Consequently, no agent in W is matched with an agent in P as otherwise, by the pigeonhole principle, some agent in U must be matched with an agent in R which is not possible by our reasoning above. Also, by a similar reasoning we deduce that each agent w from W must be matched with someone from Lw, where Lw is the preference list of w in I. Now, we show that M = {{u, w} ∈ M ′ u, w ∈ U ∪ W } is a perfect SMTI-stable matching for I. Since M ′ is a perfect matching in I ′, by the reasoning above, it follows that M is a perfect matching in I. Suppose, for the sake of contradiction, that there is an unmatched pair {x, y} /∈ M that is SMTI-blocking M . This means that x and y are acceptable to each other, and that y ≻x M (x) = M ′(x) and x ≻y M (y) = M ′(y) in I. By the preference lists of x and y in I ′ and by the definition of M , it follows that in each layer x prefers y to M ′(x) and y prefers x to M ′(y)-a contradiction to M ′ being α-layer individually stable since α > 2. 5.3 NP-hardness of α-layer pair stability We note that in the instance constructed in the proof of Theorem 5.2 every agent from the side W ∪R has the same preference list in all layers. Later on, in Proposition 6.1 we will show that in such a case the concepts 23 of α-layer individual stability and α-layer pair stability are equivalent, Thus, we obtain the same hardness result for pair stability when α ≤ ⌊ℓ/2⌋. Corollary 5.3. For each fixed number ℓ of layers with ℓ ≥ 4 and for each fixed value α with 2 ≤ α ≤ ⌊ℓ/2⌋, PAIR STABLE MARRIAGE is NP-hard even if on one side the preference list of each agent is the same in all layers. For α > ⌈ℓ/2⌉, we use an idea similar to the one used for showing Proposition 5.1. Proposition 5.4. For each fixed number α with ⌈ℓ/2⌉ + 1 ≤ α ≤ ℓ, PAIR STABLE MARRIAGE is NP-hard. Proof. Assume that α ≥ ⌈ℓ/2⌉ + 1. To prove the NP-hardness, we slightly adapt the reduction in the proof of Theorem 4.2 which shows that deciding α-layer global stability is NP-hard for α = ℓ = 2. Let P be the two-layer instance constructed in the proof of Theorem 4.2. The idea is to copy ℓ/2 times profile P and make sure that an α-layer pair stable matching must be stable in the two layers of the original profile. For the preference lists in the ℓ layers, we do the following; let k = ⌊ℓ/2⌋. 1. We make k copies of the profile P. 2. If ℓ is odd, then we add an ℓth layer with the following preference lists: Layer (ℓ),∀i ∈ [n] : ∀j ∈ [m] : xi : yi ≻ yi ≻ · · · , xi : yi ≻ yi ≻ · · · , aj : dj ≻ ej ≻ fj ≻ · · · , bj : ej ≻ fj ≻ dj ≻ · · · , cj : fj ≻ dj ≻ ej ≻ · · · , yi : xi ≻ xi ≻ · · · , yi : xi ≻ xi ≻ · · · . dj : bj ≻ cj ≻ aj ≻ · · · , ej : cj ≻ aj ≻ bj ≻ · · · , fj : aj ≻ bj ≻ cj ≻ · · · . This completes the construction, which can clearly be done in polynomial time. We claim that profile P with two layers has a 2-layer globally stable matching if and only if the new profile with ℓ layers has an α-layer pair stable matching. For the "only if" direction, it is straightforward to see that each 2-layer globally stable matching for P is ℓ-layer globally stable for the new profile. By Proposition 3.1, M is also ℓ-layer pair stable for the new instance. For the "if" direction, assume that M is an α-layer pair stable matching for the new instance with ℓ layers. We claim that M is 2-layer globally stable for instance P. First, for each unmatched pair {u, w} /∈ M , let Sunblock({u, v}) be the set that consists of all layers that are not blocked by {u, w}: Sunblock({u, v}) := {i ∈ [ℓ] M (u) ≻(i) w u}. Since M is α-layer pair stable it follows that Sunblock({u, v}) ≥ α ≥ ⌈ℓ/2⌉ + 1; the last inequality holds by our assumption on the value of α. We claim that Sunblock({u, v}) contains at least k + 1 layers from the first 2 · k layers. If ℓ is odd, then ⌈ℓ/2⌉ + 1 = k + 2, and the claim follows; otherwise ℓ = 2k and ⌈ℓ/2⌉ + 1 = k + 1, implying our claim. Since the first 2k layers are k copies of the two layers from P, we can assume without loss of generality that Sunblock({u, v}) contains at least the first k + 1 layers. Now, it is obvious that T{u,v} /∈M Sunblock({u, v}) contains at least the first two layers. This implies that no unmatched pair is blocking the first two layers, and hence that M is 2-layer globally stable for the instance P. u w or M (w) ≻(i) Corollary 5.3 and Proposition 5.4 cover the whole range of the potential values of α except for α = ⌊ℓ/2⌋ + 1 when ℓ is odd. As we will see in the next section Theorem 5.2 cannot be strengthened to cover the value α = ⌊ℓ/2⌋ + 1 (see Proposition 6.3), and so, also Corollary 5.3 cannot be directly strength- ened. However, we can tweak the construction from Theorem 5.2, breaking the restriction that on one side the preference list of each agent is the same in all layers, and obtain hardness for α-layer pair stability for α = ⌊ℓ/2⌋ + 1. Proposition 5.5. For each fixed odd number ℓ of layers with ℓ ≥ 5 and for the case when α = ⌊ℓ/2⌋ + 1, PAIR STABLE MARRIAGE is NP-hard. 24 Proof. We provide a reduction from an instance I of the NP-hard PERFECT SMTI problem which is very similar to the reduction given in the proof of Theorem 5.2. Consequently, instead of describing the new reduction from scratch, we will only explain how it differs from the one from the proof of Theorem 5.2. For each agent wi ∈ W in the proof of Theorem 5.2, the preference list of wi was the same in all layers. Now, in the last ℓ − 1 layers this list is constructed in exactly the same way as in the proof of Theorem 5.2. However, in the first layer we set the first part of the list to be reversed in comparison to the remaining (ℓ − 1) layers; ←−− Lwi be the reverse of the strict preference list of w in the input instance of PERFECT SMTI and let the let second "· · · " denote some arbitrary but fixed order of the remaining agents: Layer (1) : agent wi : ←−− Lwi ≻ pi,1 ≻ pi,2 ≻ · · · ≻ pi,ℓ ≻ · · · . The preference lists of all other agents are constructed in exactly the same way as in the proof of Theorem 5.2. We prove that I admits a perfect SMTI-stable matching if and only if the constructed in- stance I ′ with ℓ layers admits an α-layer pair stable matching with α = ⌊ℓ/2⌋ + 1. For the "only if" direction, assume that M is a perfect SMTI-stable matching for I and consider match- ing M ′ = M ∪ (cid:8){ri,j, pi,j} 1 ≤ i ≤ n ∧ 1 ≤ j ≤ ℓ(cid:9). We will show that M ′ is α-layer pair stable for I ′. First of all, just as in the proof of Theorem 5.2 we deduce that no agent from P ∪ R and no agent from U ∪ W will form a blocking pair. Neither will any two agents that are not acceptable to each other in I form a blocking pair. Now, consider two arbitrary agents u and w with u ∈ U and w ∈ W that are acceptable to each other in I. We need to find a subset of ⌊ℓ/2⌋ + 1 layers, where in each layer in the subset u prefers M ′(u) to w or w prefers M ′(w) to u . We distinguish between two cases concerning the preference list of w in I; note that it does not have ties: Case 1: u ≻w M ′(w) = M (w) in I. This implies that w prefers M ′(w) to u in the first layer in I ′. Thus, it suffices if we can find ⌊ℓ/2⌋ layers in the last ℓ − 1 layers, where u prefers M ′(u) to w. By the stability of M we have that M ′(u) = M (u) (cid:23)u w in instance I. For the case that M ′(u) ≻u w in I we have that u prefers M ′(u) to w in all layers. For the case that M ′(u) and w are tied by u we can identify ⌊ℓ/2⌋ layers, either the layers from 2 to ⌊ℓ/2⌋ + 1 or the layers from ⌊ℓ/2⌋ + 2 to ℓ, where u prefers M ′(u) to w. Additionally, in the first layer w prefers M (w) to u, which gives in total the subset of ⌊ℓ/2⌋ + 1 layers. Case 2: M ′(w) ≻w u. In this case we know that w prefers M ′(w) to u in the last ℓ − 1 layers. For the "if" direction, assume that M ′ is an α-layer pair stable matching for I ′. By the same reasoning as in the proof of Theorem 5.2 we deduce that each agent u ∈ U must be matched with someone that it prefers to each agent from R in each layer, and that w ∈ W must be matched with an agent that it prefers to each agent from P in all layers. Thus, u and M ′(u) (resp. w and M ′(w)) must be acceptable to each other in I, meaning that M = {{u, w} ∈ M ′ u ∈ U, w ∈ W } is a perfect matching for I. We show that M is SMTI-stable. For the sake of contradiction, suppose that an unmatched pair {u, w} is SMTI-blocking M . This means that w prefers u to M (w) = M ′(w) in the last ℓ − 1 layers in I ′, and that u prefers w to M (u) = M ′(u) in all ℓ layers, a contradiction to M ′ being α-layer pair stable, since α ≥ 2. 6 Two Special Cases of Preferences In this section we consider two well-motivated special cases of our general multi-layer framework for sta- ble matchings. We will discuss how the corresponding simplifying assumptions affect the computational complexity of finding multi-layer stable matchings. Interestingly, even under seemingly strong assumptions some variants of our problem remain computationally hard. 25 6.1 Single-layered preferences on one side Consider the special case where the preferences of the agents from U can be expressed through a single layer. Formally, we model this by assuming that for each agent from U , its preference list is the same in all layers. In this case we say that the agents from U have single-layered preferences. Single-layered preferences on one side can arise in many real-life scenarios. For instance, consider the standard example of matching residents with hospitals and, similarly as in Example 2.4, assume that each layer corresponds to a certain criterion. It is reasonable to assume that the hospitals evaluate their potential employees with respect to the level of their qualifications only (thus, having a single layer of preferences), while the residents take into account a number of factors such as how far is a given hospital from the place they live, the level of compensation, the reputation of the hospital, etc. First, we observe that in such a case two out of our three solution concepts from Section 2 are equivalent. Proposition 6.1. If each agent from U has single-layered preferences, then α-layer pair stability and α- layer individual stability are equivalent for each α. Proof. The fact that α-layer individual stability implies α-layer pair stability follows from Proposition 3.2. Let us prove the other direction. Let M be an α-layer pair stable matching. Towards a contradiction, suppose that M is not α-layer individually stable. By Proposition 2.2, let p = {u, w} be an unmatched pair of agents such that there is a subset S1 of ℓ − α + 1 layers where p is dominating {u, M (u)}, and that there is a subset S2 of (ℓ − α + 1) layers where p is dominating {w, M (w)}. By our assumption of single-layered preferences for each agent in U , we have that p is dominating {u, M (u)} in every layer. Thus, the pair p is blocking M in each layer in S2-a contradiction to Proposition 2.1. For profiles with single-layered preferences of the agents on one side, α-layer global stability is strictly stronger than the other two concepts: Example 6.1. Consider four agents with the following three layers of preference profiles: P1: w1 w2 u1 w1 u2 u1 w2 w1 u2 w2 u1 u2 P2: w1 w2 u1 w1 u2 u1 w2 w1 u2 w2 u2 u1 P3: w1 w2 u1 w1 u1 u2 w2 w1 u2 w2 u1 u2 Let M = {{u1, w2}, {u2, w1}}. We show that M is 2-layer pair stable but not 2-layer globally stable. Indeed, M is stable only in the first layer, so it cannot be 2-layer globally stable. To see why M is 2-layer pair stable, consider the unmatched pairs {u1, w1} and {u2, w2}. The first one only blocks the third layer, ⋄ and the latter one only blocks the second layer. 6.1.1 Global stability By Propositions 3.4 and 6.1, we can find out whether an instance with single-layered preferences on one side admits an α-layer globally stable matching in time O(ℓα · α · n2) by guessing a subset of α layers and using Algorithm 1. However, the following result shows that fixed-parameter tractability (FPT) for the parameter α, i.e., the existence of an algorithm running in f (α) · (ℓ · n)O(1) time for some computable function f , is unlikely (for details on parameterized complexity we refer to the books of Cygan et al. [17], Downey and Fellows [19], Flum and Grohe [24], and Niedermeier [42]). Theorem 6.2. Even if the preferences of the agents from U are single-layered GLOBALLY STABLE MAR- RIAGE is NP-hard and is W[1]-hard for the threshold parameter α. It can be solved in O(ℓα · α · n2) time. 26 Proof. For the running time statement, let I be an instance of GLOBALLY STABLE MARRIAGE with ℓ layers. We guess a subset S ⊆ [ℓ] of α layers and check whether a matching M exists that is stable in all layers of S. Now consider the instance I ′ restricted to the α layers in S. By Proposition 3.4, M is α-layer globally stable in I ′ if and only if M is α-layer pair stable in I ′. Since each agent on one side has the same preference list in all layers, by Proposition 6.1, M is α-layer pair stable in I ′ if and only if M is α-layer individually stable in I ′. By Theorem 4.1, we can use Algorithm 1 to decide whether M is α-layer individually stable in I ′ in O(α · n2) time. The total running time is thus O(ℓα · α · n2). For the last statement, we provide a parameterized reduction1 from the W[1]-complete INDEPENDENT SET problem parameterized by the size k of the solution. We will see that the reduction is also a polynomial- time reduction, showing the first statement since INDEPENDENT SET is also NP-hard. In the INDEPENDENT SET problem we are given an undirected graph G = (V, E) with vertex set V and edge set E, and a non-negative integer k, and we ask whether G admits a k-vertex independent set, i.e. a vertex subset V ′ ⊆ V with k pairwisely non-adjacent vertices. Given an INDEPENDENT SET instance (G = (V, E), k), we construct an instance with V layers with the set of agents U ⊎ W as follows. Assume that V = {v1, v2, . . . , vn}. For each vertex vi ∈ V we construct six agents, three for each side, denoted by ui, ui, wi, wi, ai and bi. Let U = {ui, ui, ai 1 ≤ i ≤ n} and W = {wi, wi, bi 1 ≤ i ≤ n}. We create a layer for each vertex so that if a matching is stable for a layer, then an independent set solution has to include the corresponding vertex and exclude any of its adjacent vertices. Again, the notation "· · · " denotes an arbitrary order of the unmentioned agents. Agents from U . The preference list of each agent in U is the same for all layers. ∀i ∈ [n] : agent ui : wi ≻ wi ≻ · · · , agent ui : wi ≻ wi ≻ · · · , agent ai : wi ≻ bi ≻ · · · . Agents from W . For each layer i, 1 ≤ i ≤ n, the preference list of agents wi and wi and of agents wj and wj for vj 6= vi are constructed in such a way that they encode the adjacency structure of graph G. agent wi : ui ≻ ai ≻ · · · , agent wi : ui ≻ · · · , ∀j with {vi, vj} ∈ E : agent wj : uj ≻ · · · , agent wj : uj ≻ · · · , ∀k with {vi, vk} /∈ E : agent wk : uk ≻ uk ≻ · · · , agent wk : uk ≻ uk ≻ · · · . The preference list of each bj with j ∈ [n] remains the same in all layers and is as follows. agent bj : aj ≻ · · · . Observe that for each layer i ∈ [n], wi is the most preferred agent of ai. Thus, a matching M is stable in layer i only if M (wi) ∈ {ui, ai}. From this we infer that for M to be stable in layer i it must also hold that M (ui) = wi. This implies that M (wi) = ui as otherwise {ui, wi} would be blocking layer i. Further, if M is stable in layer i, then for each vertex vj that is adjacent to vi, it must hold that M (wj) = uj and M (wj) = 1A parameterized reduction from a parameterized problem Π1 to a parameterized problem Π2 is an algorithm mapping an instance (x, k) to an instance (x′, k′) in f (k) · xO(1) time such that k′ ≤ g(k), and that (x, k) ∈ Π1 if and only if (x′, k′) ∈ Π2, where f and g are some computable functions. 27 uj . In the following, for each i ∈ [n], let M vc Then, we have the following observation for the case that a matching M is stable in layer i: i = {{ui, wi}, {ui, wi}} and M ind i = {{ui, wi}, {ui, wi}}. M vc i ⊆ M and for each vertex vj incident with {vi, vj} ∈ E it holds that M ind j ⊆ M . (1) Intuitively, the matching M ind ing M vc i means that vertex vi belongs to an independent set. i means that vertex vj does not belong to an independent set while the match- To complete the construction we let α = k. Clearly, the construction can be done in polynomial time. Now, we claim that the given graph G has an independent set of size k if and only if the constructed instance has an α-layer globally stable matching. For the "only if" part, assume that V ′ ⊆ V is a size-k independent set for G. We show that the j ∪ {{ai, bi} i ∈ [n]} is stable in each layer i with vi ∈ V ′. matching M = Svi∈V ′ M vc Suppose, for the sake of contradiction, that there is a layer i with vi ∈ V ′ and there is an unmatched pair p = {x, y} with x ∈ U and y ∈ W that is blocking layer i, i.e. the following holds. i ∪ Svj∈V \V ′ M ind In layer i, agent x prefers y to M (x) and agent y prefers x to M (y). First, y /∈ {bj j ∈ [n]} since M (bj) = aj is the most preferred agent of bj. Second, y /∈ {wj, wj {vi, vj} ∈ E} as both wj and wj already obtain their most preferred agents, namely uj and uj in layer i. Third, y /∈ {wj, wj vj ∈ V ′}, as in this way wj and wj obtain uj and uj as their partners, and since V ′ is an independent set, we have that {vi, vj} /∈ E, and so uj and uj are the most preferred partners of wj and wj. Now, the only possible choice for y would be agent wj or wj such that {vi, vj} /∈ E and vj /∈ V ′. By our construction of M we have that M (wj) = uj and M (wj) = uj. First, consider the case when y = wj. By the preference list of wj in layer i, if {x, y} is a blocking pair, then x = uj . However, uj already obtains its most preferred agent, namely, wj; a contradiction. Analogously, the case when y = wj also results in a contradiction. Summarizing, matching M is stable in each layer i with vi ∈ V ′. For the "if" part, assume that our constructed instance admits an α-layer globally stable matching M . We show that the vertex subset V ′ := {vi ∈ V M is stable in layer i} is an independent set of size k. Clearly, V ′ = α = k. Suppose, for the sake of contradiction, that V ′ contains two adjacent vertices vi and vj with {vi, vj} ∈ E. Since M is stable in layer i and {vi, vj} ∈ E, by our observation (1) we deduce that M vc j ⊆ M . However, this is a contradiction to M being stable in layer j. i ⊆ M and M ind 6.1.2 Individual stability and pair stability Observe that, by Theorem 5.2, we know that finding an α-layer individually stable matching with prefer- ences single-layered on one side is NP-hard if ℓ ≥ 4 and α ≤ ⌊ℓ/2⌋. Next, we establish a relation between the individual (and thus pair) stability for preferences single- layered on one side, and the stability concept in the traditional single-layer setting, but with general (in- complete and possibly intransitive) preferences as studied by Farczadi et al. [23]. This relation allows us to construct a polynomial-time algorithm for finding an α-layer individually stable (and hence α-layer pair stable) matching with single-layer preferences on one side when α ≥ ⌊ℓ/2⌋ + 1. Proposition 6.3. If the preferences of the agents on one side are single-layered and α ≥ ⌊ℓ/2⌋ + 1, then PAIR STABLE MARRIAGE and INDIVIDUALLY STABLE MARRIAGE can be solved in O(ℓ · n2) time. Proof. We will reduce our problem to the STABLE MARRIAGE WITH GENERAL PREFERENCES (SMG) problem [23]. Let I be an instance of our problem, let U ∪ W denote the set of agents in I, and let ℓ be the number of layers. We construct an instance I ′ of SMG as follows. In I ′ we have the same set of agents as in I. For each agent u ∈ U we copy the preferences from I to I ′ (such an agent has the same preference 28 list in all layers). The preferences of the agents from W can be arbitrary binary relations on the set V . Thus, we use an ordered pair (u, u′) of two agents to denote that u is regarded as good as u′. Formally, for each agent w ∈ W we define the general preferences of w, denoted as Rw, as follows: For each two agents u, u′ ∈ U we let (u, u′) ∈ Rw if and only if w prefers u to u′ in at least α layers. First, by our assumption on the value of α, we observe that the preferences of the agents from W are asymmetric as, i.e., for each pair of agents u, u′ ∈ U we have that {(u, u′), (u′, u)} ∩ Rw ≤ 1. Now, we will prove that a matching M is α-layer pair stable in I if and only if it is stable in I ′ according to the definition of stability by Farczadi et al. [23]. Indeed, according to Farczadi et al. [23] a pair {u, w} is SMG-blocking if and only if the following two conditions hold: (1) w ≻u M (u) and (2) (M (w), u) /∈ Rw. The first condition is equivalent to saying that u prefers w to M (u) in each layer in I. The second condition holds if and only if w prefers M (w) to u in less than α layers. This is equivalent to saying that w prefers u to M (w) in at least ℓ − α + 1 layers, and so the two conditions are equivalent to saying that {u, w} is (ℓ − α + 1)-blocking M in I. For asymmetric preferences SMG can be solved in O(n2) time [23, Theorem 2], where the number of agents is 2n. The reduction to an SMG instance takes O(ℓ · n2) time. The number of agents in I equals the number of agents in I ′. Thus, the problem (deciding α-layer pair stability) can be solved in O(ℓ · n2) time. By Proposition 6.1, we obtain the same result for α-layer individual stability. 6.2 Uniform preferences in each layer In this section we consider the case when for each layer the preferences of all agents from U (resp. all agents from W ) are the same-we call such preferences uniform in each layer. This special case is motivated with the following observation pertaining to Example 2.4: preferences uniform in a layer can arise if the criterion corresponding to the layer is not subjective. For instance, if a layer corresponds to the preferences regarding the wealth of potential partners, it is natural to assume that everyone prefers to be matched with a wealthier partner (and so the preferences of all agents for this criterion are the same); similarly it is natural to assume that the preferences of all hospitals are the same: candidates with higher grades will be preferred by each hospital. First, we observe that for uniform preferences no two among the three concepts are equivalent. Example 6.2. Consider four agents with the following four layers of uniform preferences: P1: w1 w2 u1 w1 u1 u2 w1 w2 u2 w2 u1 u2 P2: w1 w2 u1 w1 u2 u1 w1 w2 u2 w2 u2 u1 P3: w2 w1 u1 w1 u1 u2 w2 w1 u2 w2 u1 u2 P4: w2 w1 u1 w1 u2 u1 w2 w1 u2 w2 u2 u1 Let M1 = {{u1, w1}, {u2, w2}} and M2 = {{u1, w2}, {u2, w1}}. We observe that M1 is stable only in layers (1) and (4) whereas M2 is stable only in layers 2 and 3. Thus, neither is 3-layer globally stable. One can check that neither is 3-layer individually stable. However, both M1 and M2 are 3-layer pair stable. To see why this is the case we observe that for M1, both unmatched pairs {u1, w2} and {u2, w1} are each blocking only one layer, namely layers 2 and 3, respectively. Moreover, if we restrict the instance to only the first three layers, then we have the following results: 1. M1 is not even 2-layer globally stable but it is 2-layer individually stable as for each unmatched pair p with respect to M1, at least one agent in p obtains a partner which is its most preferred agent in at least two layers. 2. M2 is 2-layer globally stable (it is stable in layers 2 and 3) but it is not 2-layer individually stable. To see why it is not 2-layer individually stable, we consider the unmatched pair {u1, w1}. There are 29 3 − 2 + 1 = 2 layers where u1 prefers w1 to its partner M2(u1) = w2 and there are two layers where w1 prefers u1 to its partner M2(w1). ⋄ 6.2.1 Individual stability For preferences that are uniform in each layer, we find that there is a close relation between INDIVIDUALLY STABLE MARRIAGE and GRAPH ISOMORPHISM, the problem of finding an isomorphism between graphs or deciding that there exists none. Herein, an instance of GRAPH ISOMORPHISM consists of two undirected graphs G and H with the same number of vertices and the same number of edges. We want to decide whether there is an edge-preserving bijection f : V (G) → V (H) between the vertices, i.e. for each two vertices u, v ∈ V (G) it holds that {u, v} ∈ E(G) if and only if {f (u), f (v)} ∈ E(H). We call such bijection an isomorphism between G and H. We explore this relation through the following construction. For an instance I of INDIVIDUALLY STA- BLE MARRIAGE we construct two directed graphs GI and HI as follows. The agents from U and W will correspond to the vertices in GI and HI , respectively. For each two vertices u, u′ ∈ U we add to GI an arc (u, u′) if the agents from W prefer u to u′ in at least (ℓ − α + 1) layers. Analogously, for each two agents w and w′, we add to HI an arc (w, w′) if the agents from U prefer w to w′ in at least (ℓ − α + 1) layers. Let E(GI ) and E(HI ) denote the arc sets of GI and HI , respectively. We first explain how the so constructed graphs can be used to find an α-layer individually stable match- ing in the initial instance I, or to claim there is no such a matching. Proposition 6.4. A matching M for instance I is α-layer individually stable if and only if the following two properties hold. 1. For each two vertices u, u′ ∈ U , it holds that: (u, u′) ∈ E(GI ) implies (M (u′), M (u)) /∈ E(HI ). 2. For each two vertices w, w′ ∈ W , it holds that: (w, w′) ∈ E(HI ) implies (M (w′), M (w)) /∈ E(GI ). Proof. For the "only if" direction, assume that M is α-layer individually stable. Towards a contradiction, suppose that one of both properties stated above does not hold. That is, there exist two vertices u, u′ ∈ U such that (u, u′) ∈ E(GI ) and (M (u′), M (u)) ∈ E(HI ) or there exist two vertices w, w′ ∈ W such that (w, w′) ∈ E(HI ) and (M (w′), M (w)) ∈ E(GI ). For the first case, it follows that there is a subset S of at least ℓ − α + 1 layers such that all agents (including M (u′)) in W prefer u to u′ in each layer of S, and there is a (possibly different) subset R of at least ℓ − α + 1 layers such that all agents (including u) in U prefer M (u′) to M (u) in each layer of R. This means that the pair {u, M (u′)} is (ℓ − α + 1)-dominating {u′, M (u′)} and is (ℓ − α + 1)-dominating {u, M (u)}-a contradiction to Proposition 2.2. Analogously, the second case also leads to a contradiction to Proposition 2.2 regarding the unmatched pair {M (w′), w}. For the "if" direction, assume that both properties hold. Towards a contradiction and by Proposition 2.2, suppose that there is an unmatched pair, call it {u, w} with u ∈ U and w ∈ W , which is (ℓ − α + 1)- dominating {u, M (u)} and is (ℓ − α + 1)-dominating {w, M (w)}. By our construction of GI and HI , it follows that (w, M (u)) ∈ E(HI ) and (u, M (w)) ∈ E(GI )-a contradiction to the second property. Using a construction by McGarvey [40], given an arbitrary directed graph, we can indeed construct multi-layer preferences that induce this graph. Lemma 6.1. For each two directed graphs GI and HI with m arcs each there exists an instance with ℓ = 2m layers of preferences that induces GI and HI via McGarvey's construction, where m denotes the number of arcs in GI (and thus in HI ). Proof. Our profile uses the vertex sets of GI and HI as the two disjoint subsets of agents, denoted as U and W , respectively. For each arc in the graphs GI (resp. HI ) we construct two layers of preference lists 30 Arcs to all other vertices except for the deputy. primeG deputyG connectorG {v,v′} . . . connectorG ∗ v sinkG {v,v′} v′ Figure 2: Construction of the graphs GI and HI in the proof of Theorem 6.5. The diagram shows the case {v, v′} ∈ G. for the agents from W (resp. U ) to "encode" it. We describe how to construct the preference lists for an arc in GI , say (u, u′). The construction for the arcs in HI works analogously. To encode the arc (u, u′), −→ we construct two layers where the preference lists for the agents in W are as follows; denote by X some arbitrary but fixed order of the agents other than u and u′ and let ←− X be the reverse of −→ X . One layer : u ≻ u′ ≻ ←− X ≻ u ≻ u′. The other layer : −→ X . Finally, we set α = ℓ/2 = m. Note that for each two agents u and u′ in GI all agents from W prefer u to u′ in either exactly m − 2 layers, or exactly m layers, or exactly m + 2 layers. It is straightforward to see that our constructed instance induces the two input graphs GI and HI . Using Proposition 6.4 and Lemma 6.1 as tools, we can show that for the preferences of the agents being uniform in all layers, the problem of finding an α-layer individually stable matching is at least as hard as the graph isomorphism problem. Theorem 6.5. GRAPH ISOMORPHISM is polynomial-time reducible to INDIVIDUALLY STABLE MAR- RIAGE, where the preferences of the agents are uniform in all ℓ layers, ℓ is even, and α = ℓ/2. Proof. Let G, H be two undirected graphs that form an instance of GRAPH ISOMORPHISM. Without loss of generality, assume that n ≥ 4 (number of vertices) and m ≥ 3 (number of edges). From G and H we construct an instance I of the problem of deciding whether there exists an α-layer individually stable matching for preferences being uniform in all ℓ layers, where α = ℓ/2. By Lemma 6.1 we can describe this instance by providing the corresponding directed graphs GI and HI . We show how to construct GI from G. The graph HI is constructed from H analogously. Let V and E denote the sets of vertices and edges in G. We copy all vertices from G to GI (we will refer to these vertices as non-special) and we additionally introduce five types of "special" vertices, which we call 31 2(cid:1) + 1 connectors, m sinks, and (cid:0)n p . One remaining connector is distinguished and denoted as connectorG primeG, deputyG, (cid:0)n and the sources formally: With each pair of vertices p = {v, v′} ∈ (cid:0)V as connectorG edge e ∈ E we have one sink, denoted as sinkG denoted as sourceG SINKG = {sinke e ∈ E}, and CONNECTORG = {connectorG the set of all sinks, and the set of non-special connectors, respectively: • For each pair e = {v, v′} ⊆ V of vertices, we do the following: 2(cid:1) − m sources. We specify the connectors, the sinks, 2(cid:1) we associate one connector, denoted ∗ . Further, for each e , and for each non-edge e = {v, v′} /∈ E we have a source, 2(cid:1) \ E}, 2(cid:1)} denote the set of all sources, e . The arcs in GI are constructed as follows. Let SOURCEG = {sourcee e ∈ (cid:0)V e e ∈ (cid:0)V 1. If {v, v′} ∈ E, then we add to GI the following five arcs: (v, sinkG (v′, connectorG (sourceG e ), and (sinkG e , connectorG e , v′), (v, connectorG e , v), (sourceG 2. For each other non-special vertex u ∈ V \ {v, v′}, we add to GI the arc (connectorG 3. For each other source or sink x ∈ SOURCEG ∪ SINKG \ {sinkG e , sourceG e }, we add to GI the e ), (v, connectorG e ), e ). Otherwise, we add to GI the following five arcs: e ), (v′, connectorG e ), and (sourceG e ), (v′, sinkG e ). e , connectorG e , u). arc (connectorG e , x). • For each source and each sink x ∈ SOURCEG ∪ SINKG we add to GI an arc (x, connectorG ∗ ). For each non-special vertex v ∈ V , we additionally add to G the arc (connectorG ∗ , v). • For each vertex x ∈ V ∪ SOURCEG ∪ SINKG ∪ CONNECTORG ∪ {connectorG ∗ } except for the deputy, we add to GI an arc (primeG, x). • For each vertex x ∈ SOURCEG ∪ SINKG ∪ {primeG} we add to GI the arc (deputyG, x). For each connector y ∈ CONNECTORG ∪ {connectorG ∗ } we add to GI the arc (y, deputyG). This completes the construction of GI (crucial elements of this construction are illustrated in Figure 2). Observe a useful property: Each vertex in GI except primeG has at least two incoming arcs; primeG has exactly one incoming arc, namely from deputyG. Indeed, since GI has at least three edges (and so at least three connectors), deputyG has at least three incoming arcs from the connectors. Each connector connectorG ∗ has arcs from all sources and sinks (there are at least six of them). For each sink, source, and each non-special vertex there are at least two incoming arcs, one from the prime and the other from the deputy. {v,v′} has incoming arcs from v and v′; The special connector connectorG The arcs in HI are constructed analogously. We claim that there is an α-layer individually stable match- ing in the instance (GI , HI) with α = ℓ/2 if and only if there is an isomorphism between G and H. {v,v′} ∈ SOURCEG ∪ SINKG we let M (cid:16)sG (⇐) For the "if" direction, assume that there is an isomorphism f : V (G) → V (H) between G to H. We show that the following matching M is m-layer individually stable. For each vertex v ∈ V (G) let M (v) = {v,v′}(cid:17) = f (v) ∈ V (H). Further, for each source or sink sG {v,v′} is a sink, then {v, v′} ∈ E(G); since f is sH {f (v),f (v′ )}. Why should such a matching exist? If sG an isomorphism, we have that {f (v), f (v′)} ∈ E(H) and so the sink sH {f (v),f (v′ )} exists. Analogously, if sG {v,v′} is a source, then sH {v,v′} is also a source and exists in H. Similarly, we match the corresponding connectors: for each pair of vertices {v, v′} we let M (cid:16)connectorG {f (v),f (v′ ))}, and we ∗ . Finally, we let M (deputyG) = deputyH and M (primeG) = primeH . let M (connectorG Since f is an isomorphism between G and H, by the construction of GI and HI , an arc (x, y) in GI (resp. HI ) implies no arc (M (x), M (y)) in HI (resp. GI ). By Proposition 6.4, M is m-layer individually stable. {v,v′}(cid:17) = connectorH ∗ ) = connectorH (⇒) For the "only if" direction, assume that M is an m-layer individually stable matching for (GI , HI ). We start by showing M (primeG) = primeH . For the sake of contradiction, suppose that this is not the case, and let y = M (primeG) ∈ V (HI ), with y 6= primeH . Since y 6= primeH , there are at least two vertices w1, w2 ∈ V (HI) with (w1, y), (w2, y) ∈ E(HI ). Thus, one of them is not matched to deputyG, say M (w1) 6= deputyG. Since (w1, y) ∈ E(HI ), by Proposition 6.4, it must hold that (primeG, M (w1)) = 32 (M (y), M (w1)) /∈ E(GI ). This is a contradiction with M (w1) 6= deputyG since primeG has outgoing arcs to all vertices but deputyG. Second, we show that M (deputyG) = deputyH . The reason for this is that using Proposition 6.4 for the arc (deputyG, primeG) implies that (M (primeG), M (deputyG)) = (primeH, M (deputyG)) cannot exists in HI . By construction, primeH has an arc to all vertices except deputyH . Thus, M (deputyG) = deputyH . By an analogous reasoning focusing on the deputy vertex, we deduce that each connector in GI must be matched with some connector in HI . Further, for each source or sink sG ∈ SOURCEG ∪ SINKG since ∗ ), M (sG)) does not exist (sG, connectorG in HI ; thus there are at least 6 forbidden arcs from M (connectorG ∗ ) to some vertices which are sinks, sources, or non-special vertices. This means that connectorG ∗ because each other connector in HI has all but three outgoing arcs to such vertices. Thus, we deduce that each source/sink must be matched with a source/sink only, and consequently, that non-special vertices in GI are matched with non-special vertices in HI . We show that the bijection f : V (GI ) → V (HI ) derived from M by setting f (v) = M (v) for each v ∈ V is an isomorphism between G and H. ∗ ) ∈ E(GI ), by Proposition 6.4, it follows that (M (connectorG ∗ can only be matched to connectorH e . We know that M (connectorG e ) ∈ E(GI ) and (connectorG e ) ∈ CONNECTORH , say M (connectorG e ) = connectorH Consider an arbitrary pair e = {v, v′} ⊆ V (G) of non-special vertices and its corresponding connector connectorG {w,w′}. We claim that {w, w′} = {M (v), M (v′)}. Towards a contradiction, suppose that M (v) /∈ {w, w′}. Then, (v, connectorG {w,w′}, M (v)) ∈ E(HI ), a contradiction to Proposition 6.4. Fi- nally, we show that e ∈ E(G) if and only if {f (v), f (v′)} = {M (v), M (v′)} = {w, w′} ∈ E(H). If e ∈ E(G), then sinkG e ) ∈ E(GI ). By Proposition 6.4, it follows that e ) ∈ SOURCEH ∪ {w,w′}, M (sinkG (connectorH SINKH , we have that M (sinkG e ) ∈ E(GI ) {w,w′}, w) /∈ E(GI ), and so s is a sink, and w and w′ are connected in H. Similarly, if we infer that (sH e /∈ E(G), then we deduce that sourceG {w,w′}, where s is a source or a sink. Further, since (v, sinkG e exists and (sinkG e )) = (M (connectorG e ), M (sinkG e )) /∈ E(HI ). Since M (sinkG e and sourceH {w,w′} exist, and thus {w, w′} /∈ E(H). e , connectorG e ) = sH Theorem 6.5 implies that developing a polynomial-time algorithm for our problem is currently out of scope, since the question of whether GRAPH ISOMORPHISM is solvable in polynomial time is still open. Besides Theorem 6.5 there are other interesting implications of Proposition 6.4 and Lemma 6.1. For α ≥ ℓ/2 + 1 our problem can be reduced to the TOURNAMENT ISOMORPHISM problem, which, given two tournament graphs, asks whether there is an arc-preserving bijection between the vertices of the two tour- naments [3, 46, 47]. TOURNAMENT ISOMORPHISM has been studied extensively in the literature (for a more detailed discussion see, e.g., [3, 47, 46]), but to the best of our knowledge it is still open whether it is solvable in polynomial time [3]. The best known algorithm solving TOURNAMENT ISOMORPHISM runs in nO(log n) time, where n denotes the number of vertices. Corollary 6.6. If the preferences of the agents are uniform in all ℓ layers and α ≥ ℓ/2 + 1, then INDIVID- UALLY STABLE MARRIAGE can be solved in nO(log n) + O(ℓ · n2) time, where n denotes the number of agents. Proof. Assume that α ≥ ℓ/2 + 1 and construct the two directed graphs as we did for Proposition 6.4. Since ℓ − α + 1 ≤ ℓ/2, for each pair of vertices x and y in GI (resp. HI ) at least one arc from (x, y) and (y, x) exists. Note that such graphs are not necessarily tournaments, where for each two vertices x and y exactly one of (x, y) and (y, x) exists. But we can deal with the case when one of the graphs is not a tournament. If both (x, y) and (y, x) exist in GI or in HI , then by Proposition 6.4 no α-layer individually stable matchings exists. Thus, the only non-trivial case is when the graphs GI and HI are tournaments. Moreover, in such a case, the condition from Proposition 6.4 can be reformulated as (x, y) ∈ GI if and only if (M (x), M (y)) ∈ HI and this is the condition that M is a tournament isomorphism between GI and HI . Consequently, for α ≥ ℓ/2 + 1 the problem of finding an α-layer individually stable matching can be reduced to TOURNAMENT ISOMORPHISM. Note that the number of vertices in the constructed graphs 33 equals the number of agents in our problem. By the result of Babai and Luks [3], we obtain an algorithm for our problem with the desired running time. 6.2.2 Global stability There exists a fairly straightforward polynomial-time algorithm for finding α-layer globally stable match- ings. Proposition 6.7. If the preferences of the agents are uniform in all ℓ layers, then GLOBALLY STABLE MARRIAGE can be solved in O(ℓ · n) time, where n denotes the number of agents. Proof. It is apparent that for uniform preferences of the agents, each layer admits a unique stable matching: for each i ∈ [n] the i-th most preferred agent from U is matched with the i-th most preferred agent from W . Further, such a matching can be computed in O(n) time. Thus, our algorithm proceeds as follows. For each layer we compute a unique stable matching, and we pick the matching which is stable in the largest number of layers. Our algorithm returns this matching if it is stable in at least α layers; otherwise, the algorithm outputs that there exists no α-layer globally stable matchings for the given instance. 7 Open Problems and Conclusions We have considered a new multi-layer model of preferences in the context of the STABLE MARRIAGE problem. We identified three natural concepts of stability and discussed their relations with each other. Our results show that the algorithmic problem of finding stable matchings according to each of the three concepts is, in general, computationally hard. On the positive side, we also managed to identify a number of natural special cases which are tractable (Table 1 summarizes our results). Interestingly, while in the world of multi-layer stable matchings the case of two layers already leads to most computational hardness results, in the world of maximum-cardinality matching in two-layer graphs one obtains polynomial-time solvability, while in the case of three layers one encounters NP-hardness [14]. Our work provides a rich structure for analyzing computational properties of the problems we consid- ered, and we view our work as only initiating this line of research. Indeed, it directly leads to the following open questions: (1) How hard is it to find an α-layer individually stable matching for ⌈ℓ/2⌉ < α < ℓ? (2) When the preferences of the agents are uniform in each layer and α ≥ ℓ/2 + 1, we have shown that the decision variant of INDIVIDUALLY STABLE MARRIAGE is solvable in quasi-polynomial time nO(log n) + O(ℓ · n2)) which implies that the problem is in the complexity class LOGSNP [43]. It would be interesting to know whether it is also complete for LOGSNP.2 However, LOGSNP-hardness for our problem would also imply LOGSNP-hardness for GRAPH ISOMORPHISM (see Theorem 6.5). (3) When the preferences of the agents are uniform in each layer, how hard is it to search for an α-layer pair stable matching for arbitrary α > 1 or an α-layer individually stable matching when α < ℓ/2, or in general when the number of layers is constant. We also believe that a number of other parameters and special cases can be motivated naturally in the context of our model, in particular parameters quantifying the degree to which the preferences of the agents differ. Analogous parameterizations have been studied in computational social choice, for instance for the NP-hard KEMENY SCORE problem [6, 5]. 2LOGSNP-hardness has been encountered and discussed for natural problems in Computational Social Choice [10, 11]. 34 Continuing our research on special cases of input preferences (Section 6), it might be interesting to study stable matching with multi-layer structured preferences, such as single-peaked [8], single-crossing [44, 41], and 1-Euclidean [16, 34, 15] preferences. We note that it can be detected in polynomial time whether a pref- erence profile has any of these structure [4, 18, 22, 18, 20, 9, 34] and we refer the reader to Bredereck et al. [12] and Elkind et al. [21] for an overview of the literature on single-peakedness and single-crossingness. We also note that Bartholdi III and Trick [4] worked on stable roommates for narcissistic and single-peaked preferences, while Bredereck et al. [13] extended this line by also studying other structured preferences and including preferences with ties and incompleteness. Finally, our multi-modal view on the bipartite variant (STABLE MARRIAGE) can be generalized to the non-bipartite variant (STABLE ROOMMATES) and the case with incomplete preferences with ties. It would be interesting to see whether our computational tractability results can transfer to these cases. References [1] D. J. Abraham, P. Bir´o, and D. Manlove. "Almost stable" matchings in the roommates problem. In Pro- ceedings of the Third International Workshop on Approximation and Online Algorithms (WAOA '05), pages 1–14, 2005. 2 [2] H. Aziz, P. Bir´o, T. Fleiner, S. Gaspers, R. de Haan, N. Mattei, and B. Rastegari. Stable matching with uncertain pairwise preferences. In Proceedings of the 16th International Conference on Autonomous Agents and Multiagent Systems (AAMAS '17), pages 344–352. AAAI Press, 2017. 3 [3] L. Babai and E. Luks. Canonical labeling of graphs. In Proceedings of the 15th Annual ACM Sympo- sium on Theory of Computing (STOC '83), pages 171–183. ACM Press, 1983. 33, 34 [4] J. J. Bartholdi III and M. Trick. Stable matching with preferences derived from a psychological model. Operations Research Letters, 5(4):165–169, 1986. 35 [5] N. Betzler, J. Guo, C. Komusiewicz, and R. Niedermeier. Average parameterization and partial ker- nelization for computing medians. Journal of Computer and System Sciences, 77(4):774–789, 2011. 34 [6] N. Betzler, R. Bredereck, and R. Niedermeier. Theoretical and empirical evaluation of data reduction for exact kemeny rank aggregation. Autonomous Agents and Multi-Agent Systems, 28(5):721–748, 2014. 34 [7] P. Bir´o. Applications of matching under preferences. In Trends in Computational Social Choice, pages 345–373. AI Access, 2017. 3 [8] D. Black. The Theory of Committees and Elections. Cambridge University Press, 1958. 35 [9] R. Bredereck, J. Chen, and G. J. Woeginger. A characterization of the single-crossing domain. Social Choice and Welfare, 41(4):989–998, 2013. 35 [10] R. Bredereck, J. Chen, P. Faliszewski, J. Guo, R. Niedermeier, and G. J. Woeginger. Parameterized algorithmics for computational social choice: Nine research challenges. Tsinghua Science and Tech- nology, 19(4):358–373, 2014. 34 [11] R. Bredereck, J. Chen, S. Hartung, S. Kratsch, R. Niedermeier, O. Such´y, and G. J. Woeginger. A multivariate complexity analysis of lobbying in multiple referenda. Journal of Artificial Intellgence Research, 50:409–446, 2014. 34 35 [12] R. Bredereck, J. Chen, and G. J. Woeginger. Are there any nicely structured preference profiles nearby? Math. Soc. Sci., 79:61–73, 2016. 35 [13] R. Bredereck, J. Chen, U. P. Finnendahl, and R. Niedermeier. Stable roommate with narcissistic, single-peaked, and single-crossing preferences. In Proceedings of the 5th International Conference on Algorithmic Decision Theory (ADT '17), pages 315–330. Springer, 2017. 35 [14] R. Bredereck, C. Komusiewicz, S. Kratsch, H. Molter, R. Niedermeier, and M. Sorge. Assessing the computational complexity of multi-layer subgraph detection. In Proceedings of the 10th International Conference on Algorithms and Complexity (CIAC '17), pages 128–139. Springer, 2017. 34 [15] J. Chen, K. Pruhs, and G. J. Woeginger. The one-dimensional Euclidean domain: Finitely many obstructions are not enough. Social Choice and Welfare, 48(2):409–432, 2017. 35 [16] C. H. Coombs. A Theory of Data. John Wiley and Sons, 1964. 35 [17] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer, 2015. 26 [18] J. Doignon and J. Falmagne. A polynomial time algorithm for unidimensional unfolding representa- tions. Journal of Algorithms, 16(2):218–233, 1994. 35 [19] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Springer, 2013. 26 [20] E. Elkind, P. Faliszewski, and A. Slinko. Clone structures in voters' preferences. In Proceedings of the 13th ACM Conference on Electronic Commerce (EC '12), pages 496–513. ACM Press, 2012. 35 [21] E. Elkind, M. Lackner, and D. Peters. Structured preferences. In Trends in Computational Social Choice, pages 187–207. AI Access, 2017. 35 [22] B. Escoffier, J. Lang, and M. Ozturk. Single-peaked consistency and its complexity. In Proceedings of the 18th European Conference on Artificial Intelligence (ECAI '08), pages 366–370. IOS Press, 2008. 35 [23] L. Farczadi, K. Georgiou, and J. K onemann. Stable marriage with general preferences. Theory of Computing Systems, 59(4):683–699, 2016. 3, 28, 29 [24] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, 2006. 26 [25] D. Gale and L. S. Shapley. College admissions and the stability of marriage. The American Mathe- matical Monthly, 120(5):386–391, 1962. 1, 2, 4, 5 [26] M. R. Garey and D. S. Johnson. Computers and Intractability-A Guide to the Theory of NP- Completeness. W. H. Freeman and Company, 1979. 13 [27] D. Gusfield and R. W. Irving. The Stable Marriage Problem–Structure and Algorithms. Foundations of Computing Series. MIT Press, 1989. 2, 3, 5 [28] R. W. Irving. Stable marriage and indifference. Discrete Applied Mathematics, 48(3):261–272, 1994. 11 [29] R. W. Irving. Optimal stable marriage. In M. Kao, editor, Encyclopedia of Algorithms, pages 1470– 1473. Springer, 2016. 2 36 [30] R. W. Irving. Stable marriage. In M. Kao, editor, Encyclopedia of Algorithms, pages 2060–2064. Springer, 2016. 2 [31] K. Iwama and S. Miyazaki. A survey of the stable marriage problem and its variants. In International Conference on Infomatics Education and Research for Knowledge-Circulating Society, pages 131–136, 2008. 3 [32] K. Iwama, S. Miyazaki, Y. Morita, and D. Manlove. Stable marriage with incomplete lists and ties. In Proceedings of the 26th International Colloquium on Automata, Languages, and Programming (ICALP '99), pages 443–452, 1999. 21, 22 [33] B. Klaus, D. F. Manlove, and F. Rossi. Matching under preferences. In Handbook of Computational Social Choice. Cambridge University Press, 2016. 3 [34] V. Knoblauch. Recognizing one-dimensional Euclidean preference profiles. Journal of Mathematical Economics, 46(1):1–5, 2010. 35 [35] D. Knuth. Mariages Stables. Les Presses de L'Universit´e de Montr´eal, 1976. 2, 3, 5 [36] B. Maggs and R. Sitaraman. Algorithmic nuggets in content delivery. SIGCOMM Computer Commu- nication Review, 45(3):52–66, 2015. 2 [37] D. Manlove. Hospitals/residents problem. In M. Kao, editor, Encyclopedia of Algorithms. Springer, 2016. 2 [38] D. Manlove, R. W. Irving, K. Iwama, S. Miyazaki, and Y. Morita. Hard variants of stable marriage. Theoretical Computer Science, 276(1-2):261–279, 2002. 21, 22 [39] D. F. Manlove. Algorithmics of Matching Under Preferences, volume 2 of Series on Theoretical Computer Science. WorldScientific, 2013. 2, 3, 5 [40] D. C. McGarvey. A theorem on the construction of voting paradoxes. Econometrica, 21(4):608–610, 1953. 30 [41] J. A. Mirrlees. An exploration in the theory of optimal income taxation. Review of Economic Studies, 38:175–208, 1971. 35 [42] R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford University Press, 2006. 26 [43] C. H. Papadimitriou and M. Yannakakis. On limited nondeterminism and the complexity of the V-C dimension. Journal of Computer and System Sciences, 53(2):161–170, 1996. 34 [44] K. W. Roberts. Voting over income tax schedules. Journal of Public Economics, 8(3):329–340, 1977. 35 [45] A. E. Roth and M. A. O. Sotomayor. Two-Sided Matching: A Study in Game-Theoretic Modeling and Analysis. Cambridge University Press, 1992. Part of Econometric Society Monographs. 2 [46] P. Schweitzer. A polynomial-time randomized reduction from tournament isomorphism to tournament asymmetry. Technical report, arXiv:1704.08529, 2017. 33 [47] F. Wagner. Hardness results for tournament isomorphism and automorphism. In Proceedings of the 32nd International Symposium on Mathematical Foundations of Computer Science (MFCS '07), pages 572–583. Springer, 2007. 33 [48] B. P. Weems. Bistable versions of the marriages and roommates problems. Journal of Computer and System Sciences, 59(3):504–520, 1999. 3 37
1111.7033	1	1111	2011-11-30T02:07:22	Stability of Evolving Multi-Agent Systems	[ "cs.MA", "cs.NE" ]	A Multi-Agent System is a distributed system where the agents or nodes perform complex functions that cannot be written down in analytic form. Multi-Agent Systems are highly connected, and the information they contain is mostly stored in the connections. When agents update their state, they take into account the state of the other agents, and they have access to those states via the connections. There is also external, user-generated input into the Multi-Agent System. As so much information is stored in the connections, agents are often memory-less. This memory-less property, together with the randomness of the external input, has allowed us to model Multi-Agent Systems using Markov chains. In this paper, we look at Multi-Agent Systems that evolve, i.e. the number of agents varies according to the fitness of the individual agents. We extend our Markov chain model, and define stability. This is the start of a methodology to control Multi-Agent Systems. We then build upon this to construct an entropy-based definition for the degree of instability (entropy of the limit probabilities), which we used to perform a stability analysis. We then investigated the stability of evolving agent populations through simulation, and show that the results are consistent with the original definition of stability in non-evolving Multi-Agent Systems, proposed by Chli and De Wilde. This paper forms the theoretical basis for the construction of Digital Business Ecosystems, and applications have been reported elsewhere.	cs.MA	cs	Stability of Evolving Multi-Agent Systems Philippe De Wilde, Senior Member, IEEE, and Gerard Briscoe 1 1 1 0 2 v o N 0 3 ] A M . s c [ 1 v 3 3 0 7 . 1 1 1 1 : v i X r a Abstract-A Multi-Agent System is a distributed system where the agents or nodes perform complex functions that cannot be written down in analytic form. Multi-Agent Systems are highly connected, and the information they contain is mostly stored in the connections. When agents update their state, they take into account the state of the other agents, and they have access to those states via the connections. There is also external, user-generated input into the Multi-Agent System. As so much information is stored in the connections, agents are often memory-less. This memory-less property, together with the randomness of the external input, has allowed us to model Multi-Agent Systems using Markov chains. In this paper, we look at Multi-Agent Systems that evolve, i.e. the number of agents varies according to the fitness of the individual agents. We extend our Markov chain model, and define stability. This is the start of a methodology to control Multi-Agent Systems. We then build upon this to construct an entropy-based definition for the degree of instability (entropy of the limit probabilities), which we used to perform a stability analysis. We then investigated the stability of evolving agent populations through simulation, and show that the results are consistent with the original definition of stability in non- evolving Multi-Agent Systems, proposed by Chli and De Wilde. This paper forms the theoretical basis for the construction of Digital Business Ecosystems, and applications have been reported elsewhere. I. INTRODUCTION Multi-Agent Systems is a growing field primarily because of recent developments of the Internet as a means of circulat- ing information, goods and services. Many researchers have contributed valuable work to the area in recent years [39]. However, despite both Evolutionary Computing and Multi- Agent Systems being mature research areas [38], [30] their integration, creating evolving agent populations, is a recent development [47]. This integration is non-trivial, because agents can be considered as state machines and Evolutionary Computing algorithms have been developed to work on nu- merical data and strings without memory effects. So, our aim here is to determine, for Multi-Agent Systems which make use of Evolutionary Computing [29], [7], [46], macroscopic variables that characterise their stability. While there are several definitions of stability defined for Multi-Agent Systems [40], [1], [34], [54], [33], they are not applicable because of the Evolutionary Computing dynamics inherent in the context of evolving agent populations. Chli and De Wilde [15] model Multi-Agent Systems as Markov chains, which are an established modelling approach in Evo- lutionary Computing [44]. They model agent evolution in time as Markov processes, and so view a Multi-Agent System as a discrete time Markov chain with potentially unknown Philippe De Wilde is with the Intelligent Systems Lab, Department of Com- puter Science, Heriot Watt University, United Kingdom, [email protected]. Gerard Briscoe is with the Systems Research Group, Computer Laboratory, University of Cambridge, United Kingdom, [email protected]. transition probabilities, considered stable when its state has converged to an equilibrium distribution [15]. Also, while there is past work on modelling Evolutionary Computing algorithms as Markov chains [43], [36], [24], [20], we have found none including Multi-Agent Systems. Therefore, we decided to extend the existing Chli-DeWilde definition of agent stability to include the dynamics of Evolu- tionary Computing, including population dynamics and macro- states of the population state-space. We have applied our efforts reported here to the construction of Digital Business Ecosystems [35], [6], Digital Ecosys- tems (distributed adaptive open socio-technical systems, with properties of self-organisation, scalability and sustainability, inspired by natural ecosystems) for conducting business that enable network-based economies, levelling the playing field for Small and Medium sized Enterprises (SMEs) [5], [10]. SMEs provide substantial employment and conduct much innovative activity, but struggle in global markets on a far from level playing field, where large companies have distinct advantages [49]. So, we created Digital Ecosystems to be the digital counterparts of natural ecosystems [12], [8], which are considered to be robust, self-organising and scalable architectures that can automatically solve complex, dynamic problems [11], [5], [7]. This lead to a novel optimisation tech- nique inspired by natural ecosystems, where the optimisation works at two levels: a first optimisation, migration of agents which are distributed in a decentralised peer-to-peer network, operating continuously in time; this process feeds a second optimisation based on evolutionary computing that operates locally on single peers and is aimed at finding solutions to satisfy locally relevant constraints. So, the local search is improved through this twofold process to yield better local optima faster, as the distributed optimisation provides prior sampling of the search space through computations already performed in other peers with similar constraints. Therefore, our extended Chli-DeWilde stability was invaluable in un- derstanding and developing the evolving agent populations in these Ecosystem-Oriented Architectures (EOAs) [9], because it was important for us to be able to understand, model, and define stability, including the determination of macroscopic variables to characterise the stability, of the order constructing processes within, the evolving agent populations. II. MULTI-AGENT SYSTEMS A software agent is a piece of software that acts, for a user in a relationship of agency, autonomously in an environment to meet its designed objectives [57]. So, a Multi- Agent System is a system composed of several software agents, collectively capable of reaching goals that are difficult to achieve by an individual agent or monolithic system [57]. Examples of problems which are appropriate to Multi- Agent Systems research include online trading [42], disaster response [45], and modelling social structures [50]. Multi- agent systems are applied in the real world to graphical applications such as computer games and films. They are also used for coordinated defence systems, with other applications including transportation, logistics, as well as in many other fields. It is widely being advocated for use in networking and mobile technologies, to achieve automatic and dynamic load balancing, high scalability, and self-healing networks. The systems studied by Wiener [56] model information flow between an actor and the environment. Input, state and output and defined, and consequently an evolution equation can be established. In such a system it is possible to distinguish an observational sequence (of the inputs), followed by a decisional sequence of outputs [53]. III. CHLI-DEWILDE STABILITY We will now briefly introduce Chli-DeWilde stability for Multi-Agent System and Evolutionary Computing, sufficiently to allow for the derivation of our extensions to Chli-DeWilde stability to include Multi-Agent Systems with Evolutionary Computing. Chli-DeWilde stability was created to provide a clear notion of stability in Multi-Agent Systems [15], because while computer scientists often talk about stable or unstable systems [52], [4], they did so without having a concrete or uniform definition of stability. So, the Chli-DeWilde definition of stability for Multi-Agent Systems was created [15], derived [14] from the notion of stability defined by De Wilde [19], [28], based on the stationary distribution of a stochastic system, making use of discrete-time Markov chains, which we will now introduce1. If we let I be a countable set, in which each i ∈ I is called a state and I is called the state-space. We can then say that λ = (λi : i ∈ I) is a measure on I if 0 ≤ λi < ∞ for all i∈I λi = 1 [14]. So, if X is a random variable taking values in I and we have λi = Pr(X = i), then λ is the distribution of X, and we can say that a matrix P = (pij : i, j ∈ I) is stochastic if every row (pij : j ∈ I) is a distribution [14], [37]. We can then extend familiar notions of matrix and vector multiplication to cover a general index set I of potentially infinite size, by defining the multiplication of a matrix by a measure as λP , which is given by i ∈ I, and additionally a distribution if(cid:80) (cid:88) j∈I (λP )i = λjpij. (1) We can now describe the rules for a Markov chain by a definition in terms of the corresponding matrix P [14], [37]. Definition 1. We say that (X t)t≥0 is a Markov chain with initial distribution λ = (λi : i ∈ I) and transition matrix P = (pij : i, j ∈ I) if: 1A more comprehensive introduction to Markov chain theory and stochastic processes is available in [37] and [16]. 2 1) Pr(X 0 = i0) = λi0 and 2) Pr(X t+1 = it+1 X 0 = i0, . . . , X t = it) = pitit+1. We abbreviate these two conditions by saying that (X t)t≥0 is Markov(λ, P ). definition the Markov From this first process resulting in only the current state of is memoryless2, the system being required to describe its subsequent behaviour. We say that a Markov process X 0, X 1, . . . , X t has a stationary distribution if the probability distribution of X t becomes independent of the time t [15]. So, the following theorem is an easy consequence of the second condition from the first definition. Theorem 1. A discrete-time random process (X t)t≥0 is Markov(λ, P ), if and only if for all t and i0, . . . , it we have (2) Pr(X 0 = i0, . . . , X t = it) = λi0pi0i1 ··· pit−1it. This first theorem depicts the structure of a Markov chain [14], [37], [16], illustrating the relation with the stochastic matrix P . The next Theorem shows how the Markov chain evolves in time, again showing the role of the matrix P . Theorem 2. Let (X t)t≥0 be M arkov(λ, P ), then for all t, s ≥ 0: 1) Pr(X t = j) = (λP t)j and 2) Pr(X t = j X 0 = i) = Pr(X t+s = j X s = i) = (P t)ij. For convenience (P t)ij can be denoted as p(t) ij . Given this second theorem we can define p(t) ij as the t-step transition probability from the state i to j [14], and we can now introduce the concept of an invariant distribution [14], in which we say that λ is invariant if λP = λ. (3) theorem will link the existence of an invariant The next distribution, which is an algebraic property of the matrix P , with the probabilistic concept of an equilibrium distribution. This only applies to a restricted class of Markov chains, those with irreducible and aperiodic stochastic namely, matrices. However, there is a multitude of analogous results for other types of Markov chains which we can refer [37], [16], and the following theorem is provided as an indication, of the family of theorems that apply. An irreducible matrix P is one for which, for all i, j ∈ I there exist a sufficiently large t such that p(t) ij > 0. I matrix P and is aperiodic if for 2Markov systems with probabilities is a very powerful modelling technique, applicable in large variety of scenarios, and it is common to start memoryless, in which the output probability distribution only depends on the current input. However, there are scenarios in which alternative modelling techniques, like queueing systems, are more suitable, such as when there is asynchronous communications, and to fully characterise the system state at time (t), the history of states at (t-1), (t-2), ... might also need to be considered. all states i ∈ I we have p(t) ii > 0 for all sufficiently large t [14], [37], [16]. The meaning of these properties can broadly be explained as follows. An irreducible Markov chain is a chain where all states intercommunicate. For this to happen, there needs to be a non-zero probability to go from any state to any other state. This communication can happen in any number t of time steps. This leads to the condition p(t) ij > 0 for all i and j. An aperiodic Markov chain is a chain where all states are aperiodic. A state is aperiodic if it is not periodic. Finally, a state is periodic if subsequent occupations of this state occur at regular multiples of a time interval. For this to happen, p(t) ii has to be zero for t an integral multiple of a number. This leads to the condition p(t) ii > 0 for a-periodicity. For further explanations, please refer to [16]. Theorem 3. Let P be irreducible, aperiodic and have an invariant distribution, λ an arbitrary distribution, and suppose that (X t)t≥0 is Markov(λ, P ) [14], then Pr(X t = j) → p∞ j as t → ∞ for all j ∈ I and ij → p∞ p(t) j as t → ∞ for all i, j ∈ I. (4) (5) We can now view a system S as a countable set of states I with implicitly defined transitions between them, and at time t the state of the system is the random variable X t, with the key assumption that (X t)t≥0 is Markov(λ, P ) [14], [37], [16]. Definition 2. The system S is said to be stable when the distribution of its states converge to an equilibrium distribution, Pr(X t = j) → p∞ j as t → ∞ f or allj ∈ I. (6) More intuitively, the system S, a stochastic process X 0, ... is stable if the probability distribution of X t X 1, X 2, becomes independent of the time index t for large t [15]. Most Markov chains with a finite state-space and positive transition probabilities are examples of controllable stable systems, because after an initialisation period they reach a stationary distribution [14]. A Multi-Agent System can be viewed as a system S, with the system state represented by a finite vector X, having dimensions large enough to represent the states of the agents in the system. The state vector will consist of one or more elements for each agent, and a number of elements to define general properties3 of the system state. Hence there are many more states of the system (different state vectors) than there are agents. We can then model agent death, i.e. not being present in the system, by setting the vector elements for that agent to a fixed value that is shared by no other agent, called d [14]. 3These general properties are intended to represent properties that are external to the agents, and as such could include the coupling between the agents. However, we would expect such properties, as the coupling between the agents, to be stored within the agents themselves, and so be part of the elements defining the agents. 3 IV. INCLUDING EVOLUTION Having now introduced Chli-DeWilde stability, we will now briefly introduce Evolutionary Computing, sufficiently to allow for the derivation of our extensions to Chli-DeWilde stability to include Multi-Agent Systems with Evolutionary Computing. Evolution is the source of many diverse and creative solutions to problems in nature [17], [22]. However, it can also be useful as a problem-solving tool in artificial systems. Computer scientists and other theoreticians realised that the selection and mutation mechanisms that appear so effective in biological evolution could be abstracted to a computational algorithm [30]. This Evolutionary Computing is now recognised as a sub-field of artificial intelligence (more particularly computational intelligence) that involves combinatorial optimisation problems [3]. A. Evolutionary Algorithms Evolutionary algorithms are based upon several fundamental principles from biological evolution, including reproduction, mutation, recombination (crossover), natural selection, and survival of the fittest. As in biological populations, evolution occurs by the repeated application of the above operators [2]. An evolutionary algorithm operates on the collection of individuals making up a population. An individual, in the natural world, is an organism with an associated fitness [27]. So, candidate solutions to an optimisation problem play the role of individuals in a population, and a cost (fitness) function determines the environment within which the solutions live, analogous to the way the environment selects for the fittest individuals. The number of individuals varies between different implementations and may also vary during the use of an evolutionary algorithm. Each individual possesses some characteristics that are defined through its its genetic composition, which will be passed genotype, onto the descendants of that individual [2]. Processes of mutation (small random changes) and crossover (generation of a new genotype by the combination of components from two individuals) may occur, resulting in new individuals with genotypes differing from the ancestors they will come to replace. These processes iterate, modifying the characteristics of the population [2]. Which members of the population are kept, and used as parents for offspring, depends on the fitness (cost) function of the population. This enables improvement to occur [2], and corresponds to the fitness of an organism in the natural world [27]. Recombination and mutation create the necessary diversity and thereby facilitate novelty, while selection acts as a force increasing quality. Changed pieces of information resulting from recombination and mutation are randomly chosen. Selection operators can be either deterministic, or stochastic. In the latter case, individuals with a higher fitness have a higher chance to be selected than individuals with a lower fitness [2]. B. Evolving Agent Populations While the construction of Multi-Agent Systems that make use of Evolutionary Computing differs [47], [7], they uni- versally include populations of agents evolving to provide i.e. evolving agent populations. So, desired functionality, extending Chli-DeWilde stability to the class of Multi-Agent Systems that make use of Evolutionary Computing requires understanding population dynamics and macro-states. 1) Population Dynamics: Change of notation. Section III has defined the main properties of Markov chains that we use to model Multi-Agent Systems. We have used the standard notation, as used in [37] and many other textbooks on stochastic processes. This notion uses integers to denote states. These integers are easily used to label the rows and columns of the transition matrix. At the end of Section II, we made the step from Markov chain to multi-agent system. There are many more states than agents, and both ought to be labeled by integers. This turns out to be confusing, and therefore we make a change of notation here. We will now label the agents with integers i, and the states by vectors. Each agent i is in a scalar state ξi. All n agents together are in an n-dimensional vector state ξ. (An agent can be described by multiple numbers, but we group these here into a single scalar ξi, just as multiple binary digits can be read as a single integer.) The states of the agents are random variables. The actual values of the random variable ξi will be denoted by numbers Xi or Yi. The actual values of the vector of random variables ξ will be denoted by vectors X or Y. All the states can vary in time, and time will be denoted by the superscript t. The symbol for t is the same as in Section II, but the symbols i and X have a new meaning. The symbol i now denotes the agent, and Xi is a number describing the state of the agent i. The change of notation is essential because each agent is described by a random variable that has the Markov property. The multi-agent system is a network of interacting Markov chains. Without our change of notation, one would need multiple subscripts counting different things. A Multi-Agent System that makes use of Evolutionary Computing, an evolving agent population, is composed of n agents, with each agent i in a state ξt i at time t, where i = 1, 2, . . . , n. The states of the agents are random variables, and so the state vector for the Multi-Agent System is a vector of random variables ξt, with the time being discrete, t = 0, 1, . . . . The interactions among the agents are noisy, and are given by the probability distributions i = Xiξt = Y), Pr(XiY) = Pr(ξt+1 i = 1, . . . , n, (7) where Xi is a value for the state of agent i, and Y is a value for the state vector of the Multi-Agent System. The probabilities implement a Markov process [51], with the noise caused by mutations. Furthermore, the agents are individually subjected to a selection pressure from the environment of the system, which is applied equally to all the agents of the population. So, the probability distributions are statistically independent, and Pr(XY) = Πn i=1 Pr(ξt+1 i = Xiξt = Y). (8) (cid:88) Y 4 This is a discrete time equation used to calculate the evolution of the state occupation probabilities from t = 0, while equation (8) is the probability of moving from one state to another. The Multi-Agent System is self-stabilising if the limit distribution of the occupation probabilities exists and is non-uniform, i.e. p∞ X = limt→∞pt X (10) exists for all states X, and there are states X and Y such that X (cid:54)= p∞ p∞ Y . (11) These equations imply that some configurations of the system after an extended time will be more likely than others, because the likelihood of their occurrence no longer changes. Such a system is stable, because the likelihood of states occurring no longer changes with time, and this is the definition of stability developed in [15]. This is stability in the stochastic sense, because there is some indeterminacy in the future evolution described by the probability distributions. So, even with knowing the initial conditions, there are many directions in which the process might evolve, but still some paths will be more probable than others. Equation (10) is the probabilistic equivalent of an attractor4 in a system with deterministic interactions, which we had to extend to a stochastic process because mutation is inherent in Evolutionary Computing. While the number of agents in the Chli-DeWilde formalism can vary, we require it to vary according to the selection pressure acting upon the evolving agent population. We must therefore formally define and extend the definition of dead agents, by introducing a new state d for each agent. If an i = d, then it is dead and does not agent is in this state, ξt affect the state of other agents in the population. If an agent i has low fitness then that agent will likely die, because Pr(dY) = Pr(ξt+1 i = dξt = Y) (12) will be high for all Y. Conversely, if an agent has high fitness, then it will likely replicate, becoming a similarly successful agent (mutant), or crossover might occur changing the state of the successful agent and another agent. 2) Population Macro-States: The state of the system, an evolving agent population, S is determined by the collection of agents of which it consists at a specific time t, which potentially changes as the time increases, t+1. This collection of agents will have varying fitness values, and so the one with the highest fitness at the current time t is the current maximum fitness individual. For example, an evolving agent population with individuals ranging in fitness between 36.2% and 45.8%, the current maximum fitness individual (agent) is the one with a fitness of 45.8%. So, we can define a macro-state M as a set of states (evolving agent populations) with a common property, here possessing at least one copy of the current maximum fitness individual. Therefore, by its definition, each macro-state M must also have a maximal state composed entirely of copies of the current maximum fitness individual. If the population size is not fixed (not in nature, can be in evolutionary computing), the state space of the evolving agent If the occupation probability of state X at time t is denoted by pt X, then pt X = Pr(XY)pt−1 Y . (9) 4An attractor is a set of states, invariant under the dynamics, towards which neighbouring states asymptotically approach during evolution [55]. 5 where N is the number of possible states, and taking log to the base N normalises the degree of instability. The degree of instability will range between zero (inclusive) and one (exclusive), because a maximum instability of one would only occur in the theoretical extreme scenario of a non- discriminating selection pressure [25]. V. SIMULATION AND RESULTS A. Evolutionary Dynamics A simulated agent population was evolved relative to an arti- ficial selection pressure created by a fitness function generated from a user request R. An individual (agent) of the population consisted of a set of attributes, a1, a2, ..., and a user request consisted of a set of required attributes, r1, r2, .... The fitness function for evaluating an individual agent A, relative to a user request R, was f itness(A, R) = (14) 1 +(cid:80) 1 r∈R r − a , where a is an attribute of the agent A such that the differ- ence to the required attribute r of R was minimised. The abstract agent descriptions was based on existing and emerg- ing technologies for semantically capable Service-Oriented Architectures [41], such as the OWL-S semantic markup for web services [31]. We simulated an agent's semantic description with an abstract representation consisting of a to simulate the properties of a semantic set of attributes, description. Each attribute representing a property of the se- mantic description, ranging between one and a hundred. Each simulated agent was initialised with a semantic description of between three and six attributes, which would then evolve in number and content. Equation 14 was used to assign fitness values between 0.0 and 1.0 to each individual of the current generation of the population, directly affecting their ability to replicate into the next generation. The Evolutionary Computing process was encoded with a low mutation rate, a fixed selection pressure and a non-trapping fitness function (i.e. did not get trapped at local optima5). The type of selection used was fitness-proportional and non- elitist, with fitness-proportional meaning that the fitter the individual the higher its probability of surviving to the next generation. Non-elitist means that the best individual from one generation was not guaranteed to survive to the next generation; it had a high probability of surviving into the next generation, but it was not guaranteed as it might have been mutated [21]. These initial parameters where chosen to focus 5These constraints can be considered in abstract using the metaphor of the fitness landscape, in which individuals are represented as solutions to the problem of survival and reproduction [58]. All possible solutions are distributed in a space whose dimensions are the possible properties of individuals. An additional dimension, height, indicates the relative fitness (in terms of survival and reproduction) of each solution. The fitness landscape is envisaged as a rugged, multidimensional landscape of hills, mountains, and valleys, because individuals with certain sets of properties are fitter than others [58]. Here the ruggedness of the fitness landscape is not so severe, relative to the population diversity (population size and mutation rate), to prevent the evolving population from progressing to the global optima of the fitness landscape. Fig. 1. State-Space of an Evolving Agent Population: A possible evolutionary path through the state-space is shown, with the selection pressure of the evolutionary process driving it towards the maximal state of the maximum macro-state Mmax. population is infinite, but in practise would be bounded by resource availability. So, there is also an infinite number of configurations for an evolving agent population that has the same current maximum fitness individual. So, the state-space I of the system (evolving agent pop- ulation) S can be grouped to a set macro-states {M}. For one macro-state, which we will call the maximum macro- state Mmax, the current maximum fitness individual will be the global maximum fitness individual, which is the optimal solution (fittest individual) that the evolutionary computing process can reach through the evolving agent population (system) S. For example, an evolving agent population at its maximum macro-state Mmax, with individuals ranging in fitness between 88.8% and 96.8%, the global maximum fitness individual (agent) is the one with a fitness of 96.8% and there will be no fitter agent. Also, we can therefore refer to all other macro-states of the system S as sub-optimal macro-states, as there can be only one maximum macro-state Mmax. We can consider the macro-states of an evolving agent population visually through the representation of the state- space I of the system S shown in Figure 1, which includes a possible evolutionary path through the state-space. Traversal through the state-space I is directed by the selection pressure of the evolutionary process acting upon the population S, driving it towards the maximal state of the maximum macro- state Mmax, consisting entirely of copies of the optimal solution. It is the equilibrium state that the system S is forever falling towards without ever quite reaching, because of the noise (mutation) within the system. Yet, the maximum macro- state Mmax, in which this maximal state is located, will be reached, provided the system does not get trapped at local optima, i.e. the probability of being in the maximum macro- state Mmax at infinite time is one, p∞ = 1. Furthermore, we can define quantitatively the probability distribution of the macro-states that the system occupies at infinite time. For a stable system, as defined by equation (11), the degree of instability, δ, is the entropy of its probability distribution at infinite time, Mmax δ = H(p∞) = −(cid:88) X p∞ X logN (p∞ X ), (13) macrostatemaximal stateevolutionary pathstatemaximummacro-stateMmaxglobal maximum fitness, p∞ Mmax = limt→∞p(t) Mmax = 1, 6 while the states of the macro-state Mhalf each possessed at least one individual with fitness equal to half of the global maximum fitness, p∞ = limt→∞p(t) = 0, Mhalf Mhalf Fig. 2. Graph of Evolutionary Dynamics: This shows both the maximum and average fitness increasing over the generations of a typical evolving agent population, and as expected the average fitness remains below the maximum fitness because of variation in the evolving agent population [23]. on studying our extended Chli-DeWilde stability, rather than the evolutionary computing process itself. Crossover (recombination) was then applied to a randomly chosen 10% of the surviving population. Mutations were then applied to a randomly chosen 10% of the surviving population; point mutations were randomly located, consisting of inser- tions (an attribute was inserted into an agent), replacements (an attribute was replaced in an agent), and deletions (an attribute was deleted from an agent). Rates of 10% where chosen, because they would provide the necessary behaviour to a sufficient degree. A dynamic population size was used, with an initial population size of 300, to ensure exploration of the available agent attribute combination space, which increased with the average size of the population's agents, because fixed population sizes can sometimes fail to sufficient search available combination spaces. The issue of bloat [26] was controlled by augmenting the fitness function with a parsimony pressure [48], which biased the search to smaller agents, evaluating larger than average agents with a reduced fitness, and thereby providing a dynamic control limit which adapted to the average size of the individuals of the evolving agent population. We first plotted the fitness of the evolutionary process for a typical evolving agent population to elucidate its inherent evolutionary dynamics. The graph in Figure 2 shows both the maximum and average fitness increasing over the generations of a typical evolving agent population, and as expected the average fitness remains below the maximum fitness because of variation in the evolving agent population [23], showing that the inherent dynamism of evolutionary processes applies to evolving agent populations. B. Stability 1) Initial Parameters: An evolving agent population was called stable if the distribution of the limit probabilities existed and was non-uniform, as defined by equations (10) and (11). The simplest case was a typical evolving agent population with a global optimum, which was stable if there were at least two macro-states with different limit occupation probabilities. So, we considered the maximum macro-state Mmax and one of the sub-optimal macro-states, Mhalf . Where the states of the macro-state Mmax each possessed at least one individual with thereby fulfilling the requirements of equations (10) and (11). A value of t = 1000 was chosen to represent t = ∞ experimentally, because the simulation has often been observed to reach the maximum macro-state Mmax within 500 generations. Therefore, the probability of the system S being in the maximum macro-state Mmax at the thousandth generation is expected to be one, p1000 = 1. Furthermore, the probability of the system being in the sub-optimal macro- state Mhalf at the thousandth generation is expected to be zero, p1000 Mhalf 2) Predictions: The sub-optimal macro-state Mhalf , having a lower fitness, is predicted to be seen earlier in the evolu- tionary process before disappearing as higher fitness macro- states are reached. The system S will take longer to reach the maximum macro-state Mmax, but once it does will likely remain, leaving only briefly depending on the strength of the mutation rate, as the selection pressure was non-elitist. = 0. Mmax 3) Results: Figure 3 shows, for a typical evolving agent population, a graph of the probability as defined by equation (9) of the maximum macro-state Mmax and the sub-optimal macro-state Mhalf at each generation, averaged from ten thou- sand simulation runs for statistical significance. The behaviour of the simulated system S was as expected, being in the maximum macro-state Mmax only after generation 178 and always after generation 482. It was also observed being in the sub-optimal macro-state Mhalf only between generations 37 and 113, with a maximum probability of 0.053 at generation 61. This is because the evolutionary path (state transitions) could avoid visiting the macro-state. 4) Conclusions: As expected the probability of being in the maximum macro-state Mmax at the thousandth generation was one, p1000 = 1, and so the probability of being in Mmax Fig. 3. Graph of the Probabilities of the Macro-States Mmax and Mhalf at each Generation: The system S, a typical evolving agent population, was in the maximum macro-state Mmax only after generation 178 and always after generation 482. It was also observed being in the sub-optimal macro-state Mhalf only between generations 37 and 113. 020406080100100200300FitnessGenerationMaximumFitnessAverageFitness00.20.40.60.812004006008001000ProbabilityGenerationpMmaxpMhalf7 where t = 1000 was an effective estimate for t = ∞. This result was expected because the maximum macro-state Mmax at the thousandth generation was one, p1000 = 1, and so the probability of being in any other macro-state at the thousandth generation was zero. Mmax 2) Conclusions: The system therefore showed no instabil- ity, as there is no entropy in the occupied macro-states at infinite time. We can therefore conclude that the degree of instability of our extended Chli-DeWilde stability can provide a macroscopic value to characterise the level of stability of evolving agent populations (Multi-Agent Systems with Evolutionary Computing). E. Stability Analysis 1) Initial Parameters: We then performed a stability analy- sis (similar to a sensitivity analysis [13]) of a typical evolving agent population, by varying its key parameters while mea- suring its stability. We varied the mutation and crossover rates from 0% to 100% in 10% increments to provide a sufficient density of measurements to identify any trends that might be present, calculating the degree of instability, δ from (13), at the thousandth generation. These degree of instability values were averaged over 10 000 simulation runs to ensure statistical significance, and graphed against the mutation and crossover rates in Figure 5. to 60%, 2) Results: It shows that the crossover rate had little effect on the stability of the evolving agent population, whereas the mutation rate did significantly affect stability. With the mutation rate under or equal the evolving agent population showed no instability, with δ values equal to zero as the system S was always in the same macro-state M at infinite time, independent of the crossover rate. With the mutation rate above 60% the instability increased significantly, with the system being in one of several different macro-states at infinite time; with a mutation rate of 70% the system was still very stable, having low δ values ranging between 0.08 and 0.16, but any other macro-state, including the sub-optimal macro-state Mhalf , at the thousandth generation was zero, p1000 = 0. Mhalf We can therefore conclude that our extended Chli-DeWilde stability accurately models the stability over time of evolving agent populations (Multi-Agent Systems with Evolutionary Computing). C. Visualisation 1) Results: A visualisation for the state of a typical evolv- ing agent population, from the experiment of previous subsec- tion, at the thousandth generation is shown in Figure 4, with each line representing an agent and each shade representing an agent attribute, with the identical agents grouped for clarity. It shows that the evolving agent population reached the maximum macro-state Mmax and remained there, but as expected never reached its maximal state, where all the agents are identical and have maximum fitness, which is indicated by the lack of total uniformity in the visualisation. Fig. 4. Visualisation of an Evolving Agent Population at the 1000th Generation: The population consists of 323 agents, with each line representing an agent, and each shade representing an agent attribute. So, we see a population consisting of multiple agents, many of which are identical, having the same maximum global fitness. The identical agents were grouped for clarity, and as expected the system S reached the maximum macro-state Mmax. 2) Conclusions: The result was as expected, a lack of total uniformity in the visualisation, because of the mutation (noise) within the evolutionary process, which is necessary to create the opportunity to find fitter (better) sequences and potentially avoid getting trapped at any local optima that may be present. We can therefore conclude that the macro-state interpretation of our extended Chli-DeWilde stability accurately models the state-space of evolving agent populations (Multi-Agent Systems with Evolutionary Computing). D. Degree of Instability 1) Results: Given that our simulated evolving agent pop- ulation is stable as defined by equations (10) and (11), we can determine the degree of instability as defined by equation (13). So, calculated from its limit probabilities, the degree of instability was δ = H(p1000) = −(cid:88) X p1000 X logN (p1000 X ) = −1logN (1) = 0, Fig. 5. Graph of Stability with Different Mutation and Crossover Rates: With the mutation rate under or equal to 60%, the evolving agent population showed no instability, with δ values equal to zero as the system S was always in the same macro-state M at infinite time, independent of the crossover rate. With the mutation rate above 60% the instability increased significantly. 1451278910111213141516179341415131545678910111213141516179341415135678910111213141516179341415134578910111213141516179341415134568910111213141516179341415134567910111213141516179341415134567810111213141516179341415134567891091112131415161793414151345678910121314151617934141513456789101112141314151617934141513456789101112813141516179341415134567891011121213141516179341415134567891011121415161793414151345678910111213314151617934141513456789101112131614151617934141513456789101112131516179341415134567891011121314161793414151345678910111213141516121793414151345678910111213141516174934141513456789101112131415161734141513456789101112131415161794141513456789101112131415161793141513456789101112131415161793141513456789101112614151617934141513456789103121314151617934141513456789101112131415166934141513456789101112131415161593414151345678101011121314151617934141513456789101112131415161791141415134567891011121314151617974141513456789101112131415161793314151345678910111213141516179312141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415EvolvingAgentPopulationatMaximumMacro-state:MmaxpMmax(1000)=1AgentPopulation020406080100204060801000.10.20.30.40.5DegreeofInstabilityCrossoverRateMutationRateDegreeofInstabilityonce the mutation rate was 80% or greater the system became quite unstable, shown by high δ values nearing 0.5. 3) Conclusions: As one would have expected, an extremely high mutation rate had a destabilising affect on the sta- bility of evolving agent populations. Also, as expected the crossover rate had only a minimal effect, because variation from crossover was limited once the population matured, consisting of agents identical or very similar to one another. It should also be noted that the stability of the system is different to its performance at optimising, because while showing no instability with mutation rates below 60% (inclusive), it reached the maximum macro-state Mmax only with a mutation rate of 10% or above, while at 0% it was stable at sub-optimal macro-states between Mtenth (inclusive) and Mtwentieth (inclusive), i.e. with at least one individual with a fitness between twentieth (inclusive) to a tenth (inclusive) of the global maximum fitness, as indicated by all of the 0% Mutation Rate experiments have degree of instability, δ, values of zero in Figure 5. This is because there was no mutation (mutate rate = 0%), and so the evolving agent populations always remained near the sub-optimal macro-state to which they where initialised (seeded), with any crossover rate having little effect. We can therefore conclude that the degree of instability of our extended Chli-DeWilde stability can used of perform stability analyses (similar to a sensitivity analyses [13]) of evolving agent populations (Multi-Agent Systems with Evolu- tionary Computing). VI. CONCLUSIONS Our extension of Chli-DeWilde stability was developed to provide a greater understanding of stability in Multi- Agent Systems that make use of Evolutionary Computing, i.e. evolving agent populations. We then built upon this to construct an entropy-based definition for the degree of instability, which provides information about the level of stability, applicable to Multi-Agent Systems with or without Evolutionary Computing. Furthermore, it can be used to perform a stability analysis, similar to a sensitivity analysis, of Multi-Agent Systems. Collectively, the experimental results confirm that Chli- DeWilde stability has been successfully extended to evolving agent populations, while our definition for the degree of insta- bility provides a macroscopic value to characterise the level of stability. These findings also support the proposition that Chli- DeWilde stability can be widely applied to different classes of Multi-Agent Systems. So, our extended Chli-DeWilde stability is a useful tool for analysing Multi-Agent Systems, with or without Evolutionary Computing, providing an effective understanding and quantification to help better understand the stability of such systems. Overall, an insight has been achieved into the stability of Multi-Agent Systems that make use of Evolutionary Comput- ing, which is a first step in being able to control such systems. For example, say one wanted to avoid a number of bad states. If the probability of being in those states kept changing with time, it would be difficult to devise a strategy to avoid these 8 states. However, if the probability converges in time, while one could not guarantee to avoid those states, one could at least calculate the expected damage, i.e. the probability of being in a state times by the penalty for being in it, summed over all the states one wishes to avoid. Stochastic control of multi-agent systems will be the subject of further work. Our future work will also consider include more experi- mental scenarios to further consolidate the conclusions, with a range of different fitness landscapes [58], including flat ones (from the neutral theory of molecular evolution [25]) and ones with multiple global optima. VII. ACKNOWLEDGMENTS The authors would like to thank the following for encour- agement and suggestions; Dr Paolo Dini of the London School of Economics and Political Science, Dr Maria Chli of Aston University, and Dr Jon Rowe of the University of Birmingham. This work was supported by the EU-funded OPAALS Network of Excellence (NoE), Contract No. FP6/IST-034824. REFERENCES [1] D. Angeli and P.A. Bliman. Stability of leaderless discrete-time multi- agent systems. Mathematics of Control, Signals, and Systems (MCSS), 18(4):293–322, 2006. [2] T. Back. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary Programming, Genetic Algorithms. Oxford University Press, 1996. [3] T Baeck, D Fogel, and Z Michalewicz, editors. Handbook of Evolutionary Computation. CRC Press, 1997. [4] H. Balakrishnan, M. Stemm, S. Seshan, and R. Katz. Analyzing stability in wide-area network performance. In Scott Leutenegger, editor, International Conference on Measurement and Modeling of Computer Systems, pages 2–12. ACM Press, 1997. [5] G Briscoe. Digital Ecosystems. PhD thesis, Imperial College London, 2009. [6] G Briscoe. Complex adaptive digital ecosystems. In ACM Management of Emergent Digital Ecosystems Conference, 2010. [7] G Briscoe and P De Wilde. Digital Ecosystems: Evolving service- In IEEE Bio Inspired Models of Network, oriented architectures. Information and Computing Systems Conference, 2006. [8] G Briscoe and P De Wilde. Computing of applied digital ecosystems. In ACM Management of Emergent Digital Ecosystems Conference, 2009. [9] G Briscoe and P De Wilde. Digital Ecosystems: Stability of evolving agent populations. In ACM Management of Emergent Digital Ecosystems Conference, 2009. [10] G Briscoe and P De Wilde. The computing of digital ecosystems. International Journal of Organizational and Collective Intelligence, 1(4), 2010. [11] G Briscoe and S Sadedin. Natural science paradigms. In Digital Business Ecosystems, pages 48–55. European Commission, 2007. [12] G Briscoe, S Sadedin, and G Paperin. Biology of applied digital ecosystems. In IEEE Digital Ecosystems and Technologies Conference, pages 458–463, 2007. [13] Dan Cacuci, Mihaela Ionescu-Bujor, and Ionel Navon. Sensitivity and Uncertainty Analysis. CRC Press, 2003. [14] M Chli. Convergence and Interactivity of Multi-Agent Systems. PhD thesis, Imperial College London, 2006. [15] M Chli, P De Wilde, J Goossenaerts, V Abramov, N Szirbik, L Correia, P Mariano, and R Ribeiro. In E Santos Jr and P Willett, editors, International Conference on Systems, Man, and Cybernetics, pages 551–556. IEEE Press, 2003. Stability of multi-agent systems. [16] D. Cox and H. Miller. The Theory of Stochastic Processes. CRC Press, 1977. John Murray, 1859. [17] C Darwin. On the Origin of Species by Means of Natural Selection. [18] K De Jong, D Fogel, and H Schwefel. A history of evolutionary computation. In Baeck et al. [3], pages 1–12. 9 [19] P. De Wilde, H. Nwana, and L. Lee. Stability, fairness and scalability International Journal of Knowledge-Based of multi-agent systems. Intelligent Engineering Systems, 3:84–91, 1999. [20] A Eiben, E Aarts, and K Van Hee. Global convergence of genetic algorithms: A Markov chain analysis. In H Schwefel and R Manner, editors, Parallel Problem Solving from Nature, pages 4–12. Springer, 1991. [21] A. Eiben and J. Smith. Introduction to Evolutionary Computing. Springer, 2003. [22] D Futuyma. Evolutionary Biology. Sinauer Associates, 1998. [23] D Goldberg. Genetic algorithms in search, optimization, and machine learning. Addison-Wesley, 1989. [24] D Goldberg and P Segrest. Finite Markov chain analysis of genetic In John Grefenstette, editor, International Conference on algorithms. Genetic Algorithms and their application, pages 1–8. Lawrence Erlbaum Associates, 1987. [25] M. Kimura. The neutral theory of molecular evolution. Cambridge University Press, 1983. [47] R. Smith and N. Taylor. A framework for evolutionary computation In Janice Glasgow, editor, International in agent-based systems. Conference on Intelligent Systems, pages 221–224. AAAI Press, 1998. [48] T. Soule and J. Foster. Effects of code growth and parsimony pressure on populations in genetic programming. Evolutionary Computation, 6:293– 309, 1998. [49] J Stanley and G Briscoe. The abc of digital business ecosystems. Com- munications Law - Journal of Computer, Media and Telecommunications Law, 15(1), 2010. [50] R. Sun and I. Naveh. Simulating organizational decision-making using a cognitively realistic agent model. Journal of Artificial Societies and Social Simulation, 7(3), 2004. [51] J Suzuki. A Markov chain analysis on simple genetic algorithms. IEEE Transactions on Systems, Man, and Cybernetics, 25:655–659, 1995. [52] James Thomas and Katia Sycara. Heterogeneity, stability, and efficiency International in distributed systems. Conference on Multi Agent Systems, pages 293 – 300. IEEE Press, 1998. [53] R. Vallee. Cognition et systeme (cognition and systems). Paris: In Yves Demazeau, editor, [26] J Koza. Genetic Programming: On the Programming of Computers by l'Interdisciplinaire Systeme (s), 1995. Means of Natural Selection. MIT Press, 1992. [54] G. Weiss. Multiagent Systems: A Modern Approach to Distributed [27] E. Lawrence. Henderson's dictionary of biological terms. Pearson Artificial Intelligence. MIT Press, 1999. [55] E Weisstein. CRC Concise Encyclopedia of Mathematics. CRC Press, Education, 2005. [28] LC Lee, HS Nwana, DT Ndumu, and P. De Wilde. The stability, scalability and performance of multi-agent systems. BT Technology Journal, 16:94–103, 1998. [29] S. Mabu, K. Hirasawa, and J. Hu. A graph-based evolutionary algorithm: Genetic network programming (GNP) and its extension using reinforcement learning. Evolutionary Computation, 15:369–398, 2007. [30] P Marrow. Nature-inspired computing technology and applications. BT Technology Journal, 18:13–23, 2000. [31] D. Martin, M. Paolucci, S. McIlraith, M. Burstein, D. McDermott, D. McGuinness, B. Parsia, T. Payne, M. Sabou, M. Solanki, Naveen Srinivasan, and Katia Sycara. Bringing semantics to web services: The OWL-S approach. In Jorge Cardoso and Amit Sheth, editors, Semantic Web Services and Web Process Composition, pages 6–9. Springer, 2004. [32] T. Marwala, P. De Wilde, L. Correia, P. Mariano, R. Ribeiro, Scalability and V. Abramov, N. Szirbik, and J. Goossenaerts. In optimisation of a committee of agents using genetic algorithm. ICSC Symposium D Campbell and C Fyfe, editors, Soft Computing and Intelligent Systems For Industry. ICSC-NAISO Academic Press, 2001. International [33] G. Mohanarajah and T. Hayakawa. Formation stability of multi-agent In American Control Conference, systems with limited information. 2008, pages 704–709, 2008. [34] L. Moreau. Stability of multiagent systems with time-dependent communication links. IEEE Transactions on Automatic Control, 50:169– 182, 2005. [35] F Nachira, A Nicolai, P Dini, M Le Louarn, and L Rivera Le´on, editors. Digital Business Ecosystems. European Commission, 2007. [36] A. Nix and M. Vose. Modeling genetic algorithms with Markov chains. Annals of Mathematics and Artificial Intelligence, 5:79–88, 1992. [37] J. Norris. Markov Chains. Cambridge University Press, 1997. [38] H Nwana. Software agents: An overview. Knowledge Engineering [39] H.S. Nwana. Software agents: An overview. Knowledge Engineering Review, 11:205–244, 1996. Review, 11(3):205–244, 1996. [40] R. Olfati-Saber, J. Fax, and R. Murray. Consensus and cooperation in networked multi-agent systems. Proceedings of the IEEE, 95:215–233, 2007. [41] P Rajasekaran, J Miller, K Verma, and A Sheth. Enhancing web services description and discovery to facilitate composition. In Jorge Cardoso and Amit Sheth, editors, Semantic Web Services and Web Process Composition, pages 55–68. Springer, 2004. [42] A. Rogers, E. David, N.R. Jennings, and J. Schiff. The effects of proxy bidding and minimum bid increments within eBay auctions. ACM Transactions on the Web (TWEB), 1(2):9, 2007. [43] G Rudolph. Convergence analysis of canonical genetic algorithms. IEEE Transactions on Neural Networks, 5:96–101, 1994. [44] G. Rudolph. Finite Markov chain results in evolutionary computation: A tour d'horizon. Fundamenta Informaticae, 35:67–89, 1998. [45] N. Schurr, J. Marecki, M. Tambe, P. Scerri, N. Kasinadhuni, and J. Lewis. The future of disaster response: Humans working with In AAAI Spring Symposium on multiagent teams using DEFACTO. Homeland Security, 2005. [46] R. Smith, C. Bonacina, P. Kearney, and W. Merlat. Embodiment of evolutionary computation in general agents. Evolutionary Computation, 8:475–493, 2000. 2003. [56] N. Wiener. Cybernetics or Control and Communication in the Animal and the Machine. MIT Press, 1948. [57] M Wooldridge. Introduction to MultiAgent Systems. Wiley, 2002. [58] S Wright. The roles of mutation, inbreeding, crossbreeding and selection in evolution. In D. Jones, editor, International Congress on Genetics, pages 356–366. Brooklyn botanic garden, 1932. BIOGRAPHIES Philippe De Wilde is a Professor at the Intelligent Systems Lab, Department of Computer Science, and Head of the School of Mathematical and Computer Sciences, Heriot-Watt University, Edinburgh, United Kingdom. Research interests: stability, scalability and evolution of multi-agent systems; networked populations; coordination mechanisms for popula- tions; group decision making under uncertainty; neural networks, neuro-economics. He tries to dis- cover biological and sociological principles that can improve the design of decision making and of networks. Research Fellow, British Telecom, 1994. Laureate, Royal Academy of Sciences, Letters and Fine Arts of Belgium, 1988. Senior Member of IEEE, Member of IEEE Computational Intelligence Society and Systems, Man and Cybernetics Society, ACM, and British Computer Society. Associate Editor, IEEE Transactions on Systems, Man, and Cybernetics, Part B, Cybernetics. Gerard Briscoe is a Research Associate at the Systems Research Group of the Computer Labora- tory, University of Cambridge, UK, and a Visiting Scholar at Intelligent Systems Lab of the School of Mathematical and Computer Sciences, Heriot-Watt University, UK. Before this he was a postdoctoral researcher at the Department of Media and Com- munications of the London School of Economics and Political Science, UK. He received his PhD in Electrical and Electronic Engineering from Imperial College London, UK. Before which he worked as a Research Fellow at the MIT Media Lab Europe, after completing his B/MEng in Computing also from Imperial College London. His research interests include Sustainable Computing, Cloud Computing, Social Media and Natural Computing.
1902.00380	3	1902	2019-11-24T12:17:38	ACSEE: Antagonistic Crowd Simulation Model with Emotional Contagion and Evolutionary Game Theory	[ "cs.MA", "cs.MM" ]	Antagonistic crowd behaviors are often observed in cases of serious conflict. Antagonistic emotions, which is the typical psychological state of agents in different roles (i.e. cops, activists, and civilians) in crowd violent scenes, and the way they spread through contagion in a crowd are important causes of crowd antagonistic behaviors. Moreover, games, which refers to the interaction between opposing groups adopting different strategies to obtain higher benefits and less casualties, determine the level of crowd violence. We present an antagonistic crowd simulation model, ACSEE, which is integrated with antagonistic emotional contagion and evolutionary game theories. Our approach models the antagonistic emotions between agents in different roles using two components: mental emotion and external emotion. We combine enhanced susceptible-infectious-susceptible (SIS) and game approaches to evaluate the role of antagonistic emotional contagion in crowd violence. Our evolutionary game theoretic approach incorporates antagonistic emotional contagion through deterrent force, which is modelled by a mixture of emotional forces and physical forces defeating the opponents. Antagonistic emotional contagion and evolutionary game theories influence each other to determine antagonistic crowd behaviors. We evaluate our approach on real-world scenarios consisting of different kinds of agents. We also compare the simulated crowd behaviors with real-world crowd videos and use our approach to predict the trends of crowd movements in violence incidents. We investigate the impact of various factors (number of agents, emotion, strategy, etc.) on the outcome of crowd violence. We present results from user studies suggesting that our model can simulate antagonistic crowd behaviors similar to those seen in real-world scenarios.	cs.MA	cs	JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 1 ACSEE: Antagonistic Crowd Simulation Model with Emotional Contagion and Evolutionary Game Theory 9 1 0 2 v o N 4 2 ] A M . s c [ 3 v 0 8 3 0 0 . 2 0 9 1 : v i X r a Chaochao Li, Pei Lv, Dinesh Manocha, Hua Wang, Yafei Li, Bing Zhou, and Mingliang Xu* Abstract -- Antagonistic crowd behaviors are often observed in cases of serious conflict. Antagonistic emotions, which is the typical psychological state of agents in different roles (i.e. cops, activists, and civilians) in crowd violent scenes, and the way they spread through contagion in a crowd are important causes of crowd antagonistic behaviors. Moreover, games, which refers to the interaction between opposing groups adopting different strategies to obtain higher benefits and less casualties, determine the level of crowd violence. We present an antagonistic crowd simulation model, ACSEE, which is integrated with antagonistic emotional contagion and evolutionary game theories. Our approach models the antagonistic emotions between agents in different roles using two components: mental emotion and external emotion. We combine enhanced susceptible-infectious-susceptible (SIS) and game approaches to evaluate the role of antagonistic emotional contagion in crowd violence. Our evolutionary game theoretic approach incorporates antagonistic emotional contagion through deterrent force, which is modelled by a mixture of emotional forces and physical forces defeating the opponents. Antagonistic emotional contagion and evolutionary game theories influence each other to determine antagonistic crowd behaviors. We evaluate our approach on real-world scenarios consisting of different kinds of agents. We also compare the simulated crowd behaviors with real-world crowd videos and use our approach to predict the trends of crowd movements in violence incidents. We investigate the impact of various factors (number of agents, emotion, strategy, etc.) on the outcome of crowd violence. We present results from user studies suggesting that our model can simulate antagonistic crowd behaviors similar to those seen in real-world scenarios. Index Terms -- Group violence, emotional contagion, evolutionary game theory ! 1 INTRODUCTION Crowd simulation has received increased attention in virtual reality, games, urban modeling, and pedestrian dy- namics. One of the most important tasks in crowd sim- ulation is to generate realistic crowd behaviors. Physical methods [1], [2], [3], psychology principles [4], [5], [6], or approaches from other relatively matured disciplines [7], [8], [9], [10] are leveraged into the crowd simulation to improve the similarity between simulation results and real- world crowd movements. As pointed out in [4], emotion has a great influence on crowd behavior and it often invokes an agent to implement either a positive or negative behavioral response. Thus, the emotion modeling in crowd simulation is always the main focus in latest research work. How- ever, the emotional aspect of antagonistic crowd behaviors among people in different roles is left unexplored [11]. Analyzing the emotions of antagonistic crowd behaviors is indeed extremely important, as it can help us understand evolution process of antagonistic crowd behaviors and pre- dict trends of crowd movements. In this paper, we mainly deal with the problem of sim- *Corresponding author. • • Chaochao Li, Pei Lv, Hua Wang, Yafei Li, Bing Zhou, and Mingliang Xu are with Center for Interdisciplinary Information Science Research, ZhengZhou University, 450000. • Dinesh Manocha is with Department of Computer Science and Electrical & Computer Engineering, University of Maryland, College Park, MD, USA. E-mail: {ielvpei, ieyfli, iebzhou, iexumingliang} @zzu.edu.cn; [email protected]; [email protected]; [email protected] ulating antagonistic crowd behaviors. Such behaviors are associated with acts of violation and destruction and are typically carried out as a sign of defiance against a central authority or an indication of conflict between opposing groups [12]. Our goal is to develop a new crowd simulation model that can predict trends of crowd movement in these situations while ignoring the trajectory of a particular indi- vidual, discuss the conditions of winning and losing sides, and help to develop measures to quell incidents of crowd violence. Not only will such a method be useful for training police officers, but it could also predict the trends of crowd movements and provide the decision for controlling crowd violence incidents. It is difficult to simulate realistic antagonistic crowd be- haviors because of complex influencing factors. In practice, such behaviors are closely related to antagonistic emotions [13], [14], i.e. the emotions between opposed groups, and evolutionary game theory [5], [12]. In the pursuit of more realistic antagonistic behaviors in virtual agents, antagonis- tic emotion simulation should be incorporated into crowd simulation models [6]. Most prior crowd simulation models ignore antagonistic emotions and individuals' antagonistic behaviors. Fu et al. [15] focus on agents' emotions for only one role without involving the antagonistic emotions be- tween different types of agents. Some empirical methods of modeling antagonistic behaviors are presented in the form of riot games [16], game theoretic models [17], and social networks. These methods are based on statistical spatial- temporal analysis and role-playing dynamics in crowds and can generate emergent social phenomena. Other models use JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 2 evolutionary game theory to simulate the behaviors and interactions between different kinds of agents [12], [18]. Evolutionary game theory has successfully helped explain many complex and challenging aspects of biological and social phenomena in recent decades [18]. Inspired by the idea that more offspring will be produced by more fit biolog- ical organisms in a given environment, evolutionary game theory provides us with the methodology to study strategic interactions among agents in incidents of crowd violence [19]. There is considerable work on evaluating individuals' emotions [20], [21]. Panic, for example, destroys an individ- ual's normal mental function, transforms the individual into an irrational state, and can lead to unpredictable abnormal behaviors. Furthermore, Durupinar et al. [5] point out that agents' emotions dominate their decision-making process in games and other behaviors. However, the relationship between antagonistic emotion and antagonistic crowd be- havior has not been fully explored [15]. It is challenging to accurately model antagonistic emotions among agents in different roles because the antagonism is complex and changes constantly and dynamically [22]. Moreover, prior methods do not consider the effect of antagonistic emotion on evolutionary game theory despite the fact that emotion influences an agent's behavior significantly. For example, Lv et al. [23] calculate emotions based on political viewpoints of individuals at political rallies. Their method doesn't involve evolutionary game theory or explore the relationship be- tween evolutionary game theory and antagonistic emotion. Recent advances in crowd simulation models attempt to simulate plausible human behavior by introducing psy- chological phenomena to virtual agents [6]. Inspired by the psychological theory in [24], which shows the char- acterization and simulation of emotional contagion, these psychological phenomena effectively improve the reliability of simulations. Fu et al. [15] integrate emotional contagion with individual movement to obtain realistic emotions and behaviors in a crowd during an emergency. Based on this method, we propose an enhanced SIS model considering the benefits of games and test our antagonistic emotional contagion method among agents in different roles. More- over, we integrate antagonistic emotional contagion with evolutionary game theory through deterrent force, which describes the differences between agents more accurately. Finally, we determine antagonistic crowd behaviors com- bining antagonistic emotional contagion and evolutionary theories. Our main contributions include: • We propose a method to simulate emotion contagion between antagonistic agents in different roles, which is determined by combing the enhanced SIS and game approaches. • We propose a kind of deterrent force, which is mod- elled by a mixture of emotional forces and physical forces defeating the opponents, and it helps individ- uals make reasonable decisions in the games. • We integrate antagonistic emotions into an evolu- tionary game theoretic method through deterrent force to model antagonistic crowd behaviors more realistically. Our model can generate simulation results that are clos- est to the real-world scenarios in the overall trends of crowd movements. We have implemented our model and tested it on several outdoor scenarios with varying ratios of different types of roles. Our simulation results are compared with the real-world videos and we evaluate the benefits of our model by performing user studies. The results indicate that the behaviors of agents generated by our model are closer to real-world scenes in the overall trend of crowd movements than those seen in other methods. The rest of this paper is organized as follows. We review related work in Section 2. We give an overview of our method in Section 3. We introduce our model in Section 4 and show it can be used to generate antagonistic crowd behaviors. We describe the implementation and highlight its performance on complex scenarios in Section 5. We also present results from our preliminary user studies in Section 5. 2 RELATED WORK In this section, we provide a brief overview of prior works in emotional contagion, evolutionary game theory, and agent- based crowd simulation. 2.1 Emotional contagion Emotion is a psychological parameter and has a significant influence on individuals in a crowd [5], [25], [26], [27]. Emotional contagion is closely related to human movement [28], [29], [30]. In this subsection, we introduce some repre- sentative works about emotional contagion [4]. The epidemiological susceptible-infectious-recovered (SIR) model [31] divides the individuals in a crowd into three categories: infected, susceptible, and recovered. The analysis of the spread of epidemic among these three groups has also been extended to other fields. In [32], the extended model is used to simulate the spread of rumors. Some re- searchers use the epidemiological SIR model in conjunction with other models to describe emotion propagation under specific situations. In [5], the epidemiological SIR model is improved by combining it with the OCEAN (openness, conscientiousness, extroversion, agreeableness, neuroticism) personality model [33]. In [34], a qualitatively simulated approach to model emotional contagion is proposed for large-scale emergency evacuation. The method shows that the effectiveness of rescue guidance is influenced by the leading emotions in a crowd. Moreover, in [15], the cellular automata model based on the SIR model (CA-SIRS) is used to describe emotional contagion in a crowd during an emer- gency, capturing the dynamic process "susceptible-infected- recovered-susceptible." However, in some case, people only need to consider two emotional states: infected and sus- ceptible. Hill et al. [35] evaluate the spread of long-term emotional states across a social network based on the clas- sical SIS (Susceptible Infected Susceptible) model. Cai et al. [36] combine the OCEAN and SIS models to simulate emotional contagion on crowd evacuation. Song et al. [37] discuss the factors influencing individual evacuation de- cision making in the view of social contagion based on Susceptible-Infective (SI) model. Emotional contagion can also be used in traffic simulation [38], [39], [40] and crowd queuing simulation [41]. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 3 Fig. 1: Overview of our ACSEE model. (a) Crowd violence occurs. Civilians, activists, and cops are specified in our model and they can perceive environmental information. The basic attributes of these agents are introduced in Section 4.2. (b) Our antagonistic emotional contagion method consists of two components: mental emotion and external emotion. After the outbreak of the crowd violence, the antagonistic emotions of each agent are updated at each time step (Section 4.3). (c) We calculate the antagonistic emotion and deterrent force of each agent. The situation of the cops and activists is determined based on the deterrent force, which is defined according to antagonistic emotion. The game between antagonistic cops and activists is carried out in different situations. Agents use different strategies to carry out game interactions and their benefits are analyzed (Section 4.4). Their strategies are evolved over time using the evolutionary learning method. (d) We present behavior control method based on deterrent force, strategy, and benefit (Section 4.5). Agents choose rational movement directions and actions and their positions are updated. We use step (b) to calculate the positions of all the agents at the next time step. (e) If the crowd violence subsides, the incident ends. A thermodynamic-based emotional contagion model is introduced by Bosse et al. [42] in the ASCRIBE system. The authors use a multi-agent based approach to define emo- tional contagion within groups. Their study focuses on the emotions of a collective entity rather than the emotions of single individuals. Neto et al. [6] adapt the model proposed by Bosse et al. [43] into BioCrowds and cope with different groups of agents. Tsai et al. [44] present an emotional con- tagion model that spreads the highest level of emotion to surrounding agents in their ESCAPES framework. In [45], dynamic emotion propagation is described from the per- spective of social psychology, combining thermodynamic- based models and epidemiological-based models. According to the behaviors of different agents in antag- onistic scenarios, we define three kinds of agents. Each kind of agents is assigned a role separately: civilians, activists, and cops. Because of the antagonism between agents with different roles, the above-mentioned emotional contagion models can't be directly applied to crowd violent scenes. In this paper, we propose antagonistic emotions. Antagonistic emotions are the opposite emotions of agents in different roles. Different roles of agents come from opposing groups in crowd violent scenes. In our model the emotions of cops and activists are the antagonistic emotions. Antagonistic emotional contagion is the emotional contagion process of antagonistic agents in different roles under certain situa- tions. Our antagonistic emotional contagion method can quantitatively characterize dynamic changes in the antag- onistic emotions between different roles. 2.2 Evolutionary game theory In this subsection, we summarize some representative works about evolutionary game theory. Evolutionary game theory has solved many biological problems. For example, in [46], genome-driven evolutionary game theory helps to explain the rise of metabolic interde- pendencies in microbial communities. Some researchers have applied evolutionary game the- ory to the recommendation system. Saab et al. [47] use classical and spatial evolutionary game theory as a possible solution to the Sybil attack in recommender systems. Li et al. [48] propose a new stability analysis of repeated games and evolutionary games based on a subset of Nash equilibrium. Evolutionary game theory can help us understand the behaviors of individuals better. Huang et al. [49] present a model in which every mutation leads to a new game between the mutants and the residents based on a evolu- tionary game theoretic approach. Evolutionary game theory is one of the most effective approaches to understand and analyze widespread cooperative behaviors among individ- uals [50]. In [51], the combination of evolutionary game theory and graph theory provides an extended framework to investigate cooperative behavior in social systems. Quek et al. [12] focus on the development of a spatial evolution- ary multiagent social network to study the macroscopic- JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 behavioral dynamics of civil violence due to microscopic game-theoretic interactions between goal-oriented agents. Some previous works [12], [52] study antagonistic crowd behaviors in crowd violence incidents based on evolution- ary game theory but do not fully consider the differences of antagonistic emotions and deterrent forces of agents with different roles. In this paper, we further improve the evolutionary game theoretic method. At first, game in our model is established between opposing groups (cops and activists). When these agents confront different scenarios and situations, they adopt different strategies such as defec- tion or cooperation to get higher benefits and less casualties. Evolutionary game theory is a modeling approach of strate- gic interactions among agents in crowd violent incidents based on natural selection mechanism. Natural selection mechanism means that more offspring will be produced by more fit biological organisms in a given environment. We use evolutionary game theory to analyze the strategies and benefits of agents. Our model incorporates antagonistic emotions into evolutionary game theory to describe the differences between agents, which estimates situations of antagonistic scenes more accurately. 2.3 Agent-based crowd simulation Agent-based model is a kind of versatile method can sim- ulate complex scenarios. In an agent-based model, all the agents are endowed with greater autonomy and have their own inherent attributes and properties. Agents can receive information from the surrounding environment, which will influence their actions and decisions [53]. They have sep- arate velocities and moving directions. Hence, an agent- based model can produce complex crowd behaviors. In con- trast to agent-based models, flow-based and particle-based models cannot accurately describe the differences between agents [54]. A flow-based model is mainly used to simulate high-density crowds and has no individuals or groups. A particle-based model cannot model high-level decision- making behaviors. Therefore, we integrate the proposed model with an agent-based method. In this subsection, we summarize this kind of methods. An agent-based approach is the most common way to simulate crowd movements [55]. Kountouriotis et al. [54] use the agent-based model to simulate thousands of agents in real time, which integrates a high level of indi- vidual parametrization, such as group behaviors between friends and between a leader and a follower. Luo et al. [56] introduce a novel framework for proactive steering in agent-based crowd simulation. The Social Forces Model is combined with an agent-based method to simulate crowds [57]. In this model, repulsive and tangential forces of each agent are introduced to avoid collisions with surrounding agents and obstacles. However, all the agents share the same attributes and move with the same speed, which doesn't conform to real-world scenarios. Some agent-based methods are used to simulate crowd movements in emergency scenarios. In these scenarios, the emotional state of an agent is a very important influencing factor in simulating realistic behaviors. Shendarkar et al. [58] present a novel crowd simulation model for emergency response using BDI (belief, desire, intention) agents. Luo 4 et al. [59] describe the human-like decision-making process based on various physiological, emotional, and social group attributes for agents under normal and emergency situa- tions. Aydt et al. [60] propose an emotion model integrated with an agent-based method in serious games based on modern appraisal theory. In contrast to the above meth- ods, which don't involve antagonistic scenes or emotions between agents in different roles, our method focuses on antagonistic emotions and the relationship between antago- nistic emotions and evolutionary game theory. TABLE 1: The parameters used in our ACSEE model. Notation Description PR eex i eme i ∆eex i,j (t) ∆eex c (t) ∆benei (t) ∆eme i (t) ∆ei (t) ei (t) Ta2c Tc2a fi (t) Fi (t) Fi (t) ∆Fi Pdie Twarn warn time Twarn time The radius of perceived range External emotion of agent i Mental emotion of agent i The increase in the strength of agent i's external emotion received from agent j at time t The increment of external emotion of agent c at time t The difference of the benefits of the games at time t and t − 1 for agent i The increment of the mental emotion of agent i at time t The increment of the total emotion of agent i at time t The total emotion of agent i at time t If the emotion value of an activist exceeds the threshold Ta2c, role transition from activist to civilian occurs. If the emotion value of a civilian less than the threshold Tc2a , role transition from civilian to activist occurs. The deterrent force of agent i at time t The total deterrent force of agents of the same type for agent i at time t The total deterrent force of his or her opponents in the cells neighboring agent i at time t The difference of the total deterrent forces between cops and activists that agent i can perceive at time t The death probability The early warning threshold The time of early warning The time of early warning threshold 3 OVERVIEW OF OUR APPROACH In this section, we introduce some basic and important concepts about crowd behavior simulation and crowd emo- tional contagion. We also give an overview of our method. 3.1 Crowd behavior simulation Crowd behavior simulation can be defined as a process of emulating or simulating the movement of large amount of entities, characters or agents [61]. At a broad level, crowd movement is governed by psychological status of individu- als and their surrounding environment [22]. When humans form a crowd, interaction becomes an essential part of the overall crowd movement [5]. For agent-based methods of crowd simulation used in this paper, each agent is as- sumed as an independent decision-making entity, which has knowledge of the environment and a desired goal position JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 at each step of the simulation. The interactions between an agent with others or with the environment are often performed at a local level [62]. A typical crowd simulation model can be defined as in Equation 1. P t represents the positions of all the agents in the scene at time t and P t+1 is the positions of all the agents at time t + 1, which can be induced by crowd simulation model f. P t+1 = f (P t) (1) 3.2 Crowd emotional contagion Crowd simulation research has recently taken a new di- rection for modeling emotion of individuals to generate believable, heterogeneous crowd behaviors. Emotion of an agent can greatly affect its ability to perceive, learn, behave, and communicate within the surrounding environment [5]. The emotion owned by one agent provides information about other agents' behavioral intentions and modulates his or her behavioral decision-making process. Based on their appraisal of the environment, emotions of the agents in a crowd are updated dynamically at different time. In antagonistic scenes, such emotional changes become more obvious and play a vital role in crowd interaction behaviors. As shown in Equation 2, the emotion values of all the agents Et+1 at time t + 1 can be computed according to their relative positions P t and emotions of the agents Et at time t. Et+1 = g(Et, P t) (2) 3.3 Overview of our ACSEE model This paper mainly discusses the influence of antagonistic emotions on agents' behaviors, whose purpose is to com- pute and update the status of all the agents at different time steps according to their emotions and roles. The crowd emotion is fully integrated into crowd behavior simulation, also with considering the confrontation between agents with different roles, such as civilians, activists, and cops. Given the positions P t, the emotions Et, and the roles Rt of all the agents at time t, our crowd simulation model ACSEE(P t, Et) estimates the positions P t+1, the emotions Et+1, and the roles Rt+1 of all the agents at next time step as Equation 3. {P t+1, Et+1, Rt+1} = ACSEE(f (P t−1), g(Et−1, P t−1), Rt) = ACSEE(P t, Et, Rt) (3) (ACSEE) based on Emotional 4 ACSEE MODEL We present a novel Antagonistic Crowd behavior Simulation model contagion and Evolutionary game theories. Our model consists of three important modules: antagonistic emotional contagion, an- tagonistic evolutionary gaming, and behavior control. The antagonistic emotional contagion method is designed by combining the enhanced SIS and game approaches. Using the antagonistic emotions of agents, we define their deter- rent forces in Section 4.4. The enhanced evolutionary game theoretic approach is determined based on the deterrent forces of agents. Our ACSEE model computes the behavior of each agent by modeling the influence from antagonistic emotional contagion and evolutionary game theories. The flowchart of our ACSEE model is presented in Figure 1. 5 4.1 Symbols and Notations For convenience, the important parameters used in the ACSEE model and their descriptions are listed in Table 1. 4.2 Agent modeling in antagonistic scenes In this section, we mainly describe the role of different kinds of agents and the assumptions we formulated. 4.2.1 Civilians, activists, and cops Crowd violence incidents are often caused by some serious social contradictions, where a certain amount of activists challenge or break the normal and peaceful social order or stability in different ways of violence such as large- scale gathering, group activities, and physical conflicts. We classify the agents in the crowd under such situations as civilians, activists, or cops based on their roles according to [12], [63]. Civilians are neutral agents in the environment and pose no danger to the central authority. In general, civilians are vulnerable groups and do not participate in confrontation. The cops do their best to protect civilians while the activists persecute them. However, civilians may change their roles if conditions are favorable to express their anger and frustra- tion publicly. For example, because of the instigation of the surrounding activists, the civilians may turn into activists to participate in the riot. Activists aim to create havoc and fuel the ongoing unrest while avoiding being defeated by cops. Cops maintain public order by suppressing activists and play a key role in preventing terrorist attack. In real- world scenarios, cops and activists can represent any two an- tagonistic groups [13]. Civilians can also represent onlookers and neutral parties. 4.2.2 Assumptions Antagonistic crowds arise for many complex influencing factors. The simulation of antagonistic crowd behaviors considering all the influencing factors is an insoluble prob- lem. From the observations of the real antagonistic crowd behaviors, we formulate the following assumptions to make this problem solvable. • Emotions of cops are positive while those of activists are negative. Civilians are neutral agents. The pro- cess of emotional contagion can change their emo- tions. Cops with high positive emotions will make the agents around them more positive. Activists with high negative emotions will make the agents around them more negative. Civilians don't actively affect the emotions of surrounding agents [5]. • An agent can maintain a perceived range that is centered around it. We regard the perceived range of an agent as a circular area with a fixed radius PR [5], [64]. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 6 • Agents use different strategies to interact with their opposing agents. Agents' strategies refer to the ac- tions taken during the games. According to the en- countered situation, activists and cops can adopt one of two different strategies: cooperation or defection [12], [65]. The cooperation strategy of activists means accepting peaceful settlements and running away from the cops. The activists with defection strategy will revolt aggressively and instigate civilians to revolt. The cops with cooperation strategy keep away from large gatherings of activists to protect civilians. The defection strategy of cops means pursuing ac- tivists. • The agent may be classified as death. Dead agents are subdued by their opponents and pose no threat any more. It doesn't mean biological death. 4.3 Antagonistic emotional contagion module Since emotion has an important influence on people's be- havior decision-making, accurate emotion modeling is es- sential and fundamental for a crowd simulation model [5]. In this section, we present our antagonistic emotional contagion module. Emotions in our model incorporate the antagonism between agents in different roles. ei denotes the emotion of agent i. ei ∈ (−1, 1) and ei (cid:54)= 0. There are two different types of emotion: positive emotion and negative emotion. When the emotion value is greater than 0, the emotion is positive. The higher the emotion value of an agent, the more positive his/her emotion. When the emotion value is less than 0, the emotion of the agent becomes negative. The lower the emotion value of an agent, the more negative his/her emotion. When an emotion value is closer to 0, the agent is regarded as being in a peaceful state and he or she tends to be conservative. The descriptions of different emotion values of different roles are listed in Table 2. TABLE 2: The descriptions of different emotion values of different roles Descriptions of emotion Roles Cops Activists Civilians Closer to 1 high morale; brave in subduing the activists escape from cops not challenging cops brave; not fear of activists Emotion Closer to 0 peaceful peaceful peaceful Closer to −1 low morale; afraid of activists; dare not to fight with activists arrogant and attack cops uneasy, suffering, fear of activists, and obeyed In such an antagonistic game scenario, individuals will be influenced by external and internal stimuli. External stimuli mainly come from the external environment and are often accepted by individuals passively. Internal stimuli come from the subjective perception and judgment of indi- viduals by themselves. Both the internal and external stimuli in the antagonistic scenarios are able to produce emotions, so the emotion of an agent ei consists of two parts. The first part is the external emotion eex , which is influenced by i surrounding agents. The second part is the mental emotion eme [66], which is determined by an agent's own subjective i consciousness. Therefore, the final emotion value is defined as follows [67], [68]: ei = eex i + eme i (4) 4.3.1 External emotion Our method for calculating external emotion is inspired by the emotional contagion model in [15]. An agent can be affected by others in his or her perceived range. The increase in the strength of external emotion of agent i (∆eex i,j (t)) received from agent j at time t is defined as: ∆eex i,j(t) = [1 − 1 (1 + exp(−L)) ] × ei(t) × Aj,i × Bi,j (5) where L represents the distance between agent i and j, ei denotes the emotion of agent i, Aj,i is the intensity of emotion received by i from sender j, and Bi,j is the intensity of emotion which is sent from i to receiver j. Civilians can only passively receive the emotional conta- gion from surrounding agents and cannot actively influence others. The increment of external emotion of agent o at time t is denoted as ∆eex o (t) includes emotional influences received from all the cops and activists in the perceived range of agent o . ∆eex o (t) is defined as follows: o (t). ∆eex k(cid:88) ∆eex o (t) = n(cid:88) ∆eex o,ci (t) + ∆eex o,aj (t) (6) i=1 j=1 (t) and ∆eex o,aj where ∆eex (t) denote the increase in the o,ci strength of the external emotion transmitted from cop ci and activist aj to agent o. 4.3.2 Mental emotion Each agent establishes game play with agents in the oppos- ing group who are in his or her perceived range. Because civilians are neutral members and remain peaceful, we assume that no game interaction will occur between civilian agents. The benefits of each game are determined according to the method outlined in Section 4.4. Mental emotion is defined as the difference between the benefits of two games and the mental emotion of civilians is a constant [69]. The difference between the benefits of the games at time t and t− 1 for agent i is denoted as ∆benei (t) = benei (t)− benei (t − 1). The threshold that leads to emotional fluctua- tions is δ. The relationship between the increment of mental emotion and the difference between the benefits at time t is defined by:  rand(−0.01, 0.01), ∆benei(t) < δ δ+exp(δ/∆benei(t)) , ∆benei(t) ≥ δ δ+exp(∆benei(t)/δ) , ∆benei(t) ≤ −δ − 0.1 0.1 (7) ∆eme i (t) = In Equation 7, ∆benei (t) < δ means that the difference between benefits fails to reach the emotional fluctuation threshold δ. Therefore, there is little change in the emotion of agent i. In this case, the mental emotion value is a random number on the interval (-0.01, 0.01). ∆benei (t) ≥ δ means that the benefit at time t is higher than that at time t− 1. The benefit increase makes the cops more positive and activists more negative. ∆benei (t) ≤ −δ means that the difference between the benefits of time t and t − 1 is higher than the emotional fluctuation threshold δ. The benefit at time t is lower than that at time t − 1. The benefit decrease makes cops more negative and activists more positive. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 4.3.3 Emotion updating Our emotion updating method for agents is presented in Figure 2. The mental and external emotions of each agent are updated according to the evolution of the games and changes in agents' locations. The external emotion of an agent is determined by the emotional contagion (external stimulus) of surrounding cops and activists. The differences between the benefits of games (internal stimulus), which is defined in Section 4.3.2, lead to the changes in the mental emotions of cops and activists. Fig. 2: Emotion updating method. The benefit of the game between cops and activists is the internal stimulus, which leads to the changes in mental emotions. Emotional contagion among different kinds of agents is the external stimulus, which leads to the changes in external emotions. The increment of the total emotion (∆eo (t)) of agent o at time t is defined as follows: ∆eo(t) = ∆eex o (t) + ∆eme o (t) (8) For each time step, the total emotion value is updated. At time t, the total emotion value of agent o is defined as follows: eo(t) = eo(t − 1) + ∆eo(t) (9) 4.4 Antagonistic evolutionary gaming module In our model an agent establishes game play with all the agents from the opposing group within his or her perceived range, according to game theory. When agents confront dif- ferent scenarios and situations, they adopt different strate- gies and get varying benefits. They aim to maximize their benefits and minimize casualties according to the current situation. In this subsection, we present the evolutionary game theoretic module of our model, which is used to analyze the strategies and benefits of agents. At first, agents estimate surrounding situations based on the deterrent forces of the agents in their perceived range. Deterrent force is a kind of power by which an agent can beat his or her opponents and is closely related to the agent's behavior. Emotion plays a crucial role in agents' deterrent forces [70]. The deterrent force of agent i is defined in the following equation: (cid:16)ei (t) · π (cid:17) fi (t) = sin 2 (10) where ei (t) is the emotion of agent i at time t. The more positive or negative the emotion of an agent is, the greater the deterrent force the agent possesses [70]. The total deter- rent forces are defined as follows: 7 (cid:88) (cid:88) k∈A k∈ A Fi (t) = Fi (t) = fk (t) fk (t) (11) (12) where the set A denotes agents of the same type in the perceived range of agent i and the set of the opposing agents in the perceived range of agent i is A. The situation is defined according to the difference be- tween the total deterrent forces of cops and activists per- ceived by agent i, which is expressed in Equation 13, as follows. ∆Fi = Fi (t) − Fi (t) (13) Under different situations, the benefits gained during the games are different. The benefit matrix is defined according to varying situations. In contrast, the benefit matrix defined in [12] is based on the number of cops and activists and assumes that the deterrent forces of all the agents are the same. Instead, we define the benefit matrix based on the deterrent forces of cops and activists. We fully account for the differences between the deterrent forces of all the agents, which conforms to real-world scenes. The benefit matrix corresponding to different situations is shown in Table 3. TABLE 3: Benefit matrix of cops and activists. ∆F denotes the situations of cops and activists. ∆F > 0 means that the total deterrent force of the cops is higher than that of the activists. ∆F < 0 and ∆F = 0 are defined in a similar way. Activists and cops can adopt two different strategies: cooperation or defection. We list the benefits of cops and activists corresponding to different situations. For example, the ratio of the benefits of the cops using a cooperation strategy to the activists using a cooperation strategy is 1:4 when ∆F > 0. The ratio of the benefits of the cops using a cooperation strategy to the activists using a defection strategy is 2:2 when ∆F > 0 [12] Situations Benefit ∆F > 0 ∆F < 0 ∆F = 0 Strategy of cops Cooperation Defection Cooperation Defection Cooperation Defection Strategy of activists Cooperation Defection 1,4 3,3 4,1 2,2 3,3 5,0 2,2 4,1 3,3 1,4 0,5 1,1 When the total deterrent force of the cops is higher than that of the activists (∆F > 0), if both groups adopt a strategy of cooperation, cops miss an opportunity to make arrests. Compared with activists, cops reap fewer benefits. If both groups defect, cops gain the upper hand because they have a higher total deterrent force and therefore reap more benefits. If cops cooperate and activists defect, both groups' JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 benefits remain relatively neutral. Cops should defect to confront activists while activists should cooperate to avoid challenging cops and inviting casualties. If cops defect and activists cooperate, cops exert dominance over activists and activists avoid direct conflict. Therefore, both groups obtain benefits. When the total deterrent force of cops is less than or equal to that of activists (∆F < 0 or ∆F = 0), their benefits are defined similarly. After a game, the strategies of cops and activists are updated according to the results of that game. Each agent is defined by a binary string. This string encodes the strategy bits in different situations. This string suggests the strategies an agent should adopt when ∆F = 0, when ∆F > 0, and when ∆F < 0. Then the effectiveness or benefit of each strategy is calculated. The more beneficial strategy is chosen and it will be passed on to the offspring in an attempt to create a better strategy. 4.5 Behavior control module The behavior control method of agents is determined by antagonistic emotional contagion and evolutionary game theoretic approaches. In this section, we present some rules about how agents determine their positions at the next time step and their living states. Agents determine their positions at the next time step based on the cellular automaton model [15]. A cellular space of M × N cells is defined and each agent occupies one cell. At each time step, agents choose to move to their neighboring cells or stay still. Whether an agent moves or not depends on the deterrent forces exhibited by his or her neighboring agents. For a cop or an activist, this is divided into the following possible cases according to real-world videos: • If the total deterrent force of agents in the opposing group is higher than that of agents of the same type in agent i' s neighboring cells, he or she has to move. The moving direction of agent i is determined by the expected benefits of his or her neighboring cells. At first, the neighboring cells around agent i will be checked to find the empty ones (an empty cell means that there is not an agent in it). Then the expected benefits of all the empty cells are calculated. The cell with the highest benefit is the position to which agent i will move. • Next we consider the situation where the total de- terrent force of the opposing agents is less than that of agents of the same type in agent i's neighboring cells. If agent i's strategy is defection, he or she will move to the nearest opposing agent (i.e. attack the opposing agent). If agent i's strategy is cooperation, he or she will stay away from opposing agents and move to the nearest civilian. If agent i is a cop, he or she will protect the civilian. If agent i is an activist, he or she will attack the civilian. In this case, agent i may also choose to stay still. • Agent i with no neighbors chooses to move. The moving direction is the same as the situation where the total deterrent force of the opposing agents is less than that of agents of the same type. (cid:32) (cid:33) (cid:80) Fi(cid:80) (cid:95) F i 8 • Civilians move to safer positions where there are more cops around them. The agent with a defection strategy attacks his or her opponents. The agent may be dead. A dead agent in this case means being subdued by his or her opponents and therefore posing no threat to these opponents. It doesn't mean real death. At each time step, the death probability of each agent is calculated and denoted as Pdie. In contrast to the definition in [12], which is based on the number of cops and activists, we define Pdie based on the total deterrent forces of cops and activists. ln0.1 · Pdie = 1 − exp (14) where(cid:80) Fi represents the total deterrent force of agents of the same type, and(cid:80) (cid:98)Fi represents the total deterrent force (cid:80) Fi(cid:80) (cid:95) ensure a plausible value (Pdie = 0.9) when(cid:80) Fi = (cid:80) (cid:95) of his or her opponents in the cells neighboring agent i. denotes the total deterrent force of the cop-to-activist ratio within the perceived range of agent i. ln0.1 is set to F i F i [12], [71]. Each agent has an early warning threshold Twarn. When the value of Pdie exceeds the threshold Twarn, the value of warn time increases by 1. When the value of warn time exceeds the threshold Twarn time, the agent will die. Because the endurance of each agent is different, the values of the thresholds Twarn and Twarn time are also different for each agent. 5 IMPLEMENTATION AND PERFORMANCE We have implemented our model using Visual C++ to sim- ulate antagonistic crowd behaviors based on Unity3D . The computing environment is a common PC with a quadcore 2.50 GHz CPU,16 GB memory, and an Nvidia GeForce GTX 1080 Ti graphics card. TABLE 4: List of parameter values used in our simulation Scenario Number of agents Civilians Activists Cops Size of 2-D Grid Ta2c Tc2a Emotion Civilians Activists Cops No.1:Activists attack civilians No.2:Role transitions No.3:Cops encircling activists No.4:Real-world 1 No.5:Real-world 2 No.6:Real-world 3 No.7:Real-world 4 80 80 80 10 80 3 0 No.8:Real-world 5 100 No.9:ACSEE vs. CVM 80 50 50 50 50 50 14 30 30 50 40 70 70 30 30 40 20×20 squares 20×20 squares 20×20 squares 20×20 squares 20×20 squares 20×20 squares 100 40×40 squares 100 40×40 squares 0.1 0.1 0.1 0.1 0.1 1 1 1 -0.5 -0.5 -0.5 -0.5 -0.5 -1 -1 -1 30 20×20 squares 0.1 -0.5 0.1 0.1 0.1 0.1 0.1 0.1 0 0.1 0.1 -0.5 -0.5 -0.5 -0.5 -0.5 -0.1 and -0.3 -0.2 -0.9 and -0.2 -0.5 0.5 0.5 0.5 0.5 0.5 0.9 0.8 0.9 0.5 We run a series of experiments involving varying role number ratios in outdoor scenarios. The parameter val- ues in different scenarios used in the simulation runs are listed in Table 4. The perceived range PR of agents is 10, Twarn ∈ [0.7, 0.9], and Twarn time ∈ [8, 20]. Figures 3, 6, JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 9 (a) (e) (b) (f) (c) (g) (d) (h) Fig. 3: The numbers of three types of agents at different time steps according to different initial Rca (the ratio of cops to activists): (a) Rca is 0 (0 cops), (b) Rca is 0.6 (30 cops), (c) Rca is 0.8 (40 cops), (d) Rca is 1 (50 cops), (e) Rca is 1.2 (60 cops), (f) Rca is 1.4 (70 cops), and (g) Rca is 1.6 (80 cops). (h) Active ratio curves for different initial Rcas. (a) shows a steep drop in the number of civilians. The number of activists remains the same because there are no cops to fight with the activists. We can see from (a) to (g) that the number of dead activists increases as Rca increases. When Rca is 1.2, all the activists have been subdued by cops. The time required for all the activists to be subdued by cops decreases as Rca increases. The civilian survival time lengthens and there are more surviving civilians as Rca increases. We can see that there is an initial rise in the number of civilians in (f) and (g) because the total deterrent force of the cops is much higher than that of the activists. Some activists change their roles (from activist to civilian). (h) shows the overview of active ratios (ratios of the activists to the sum of the activists and civilians) corresponding to different Rcas. An inverse relationship between the active ratio and the Rca is presented. Therefore, increasing the ratio of cops can help subdue activists. 7, 9, and 12 report the results of 200 runs. The average of the 200 runs is the final result. We investigate the effect of varying role number ratios on the crowd violence results (in Section 5.1). By analyzing the simulation results, we find that our model can account for many emergent phenomena (discussed in Section 5.3). Next, we study how different important factors influence crowd violence. In Section 5.4 we show the impact of emotional contagion on the results of crowd violence. We compare the different simulation results with or without considering emotional contagion. In Section 5.5, the relationship between the deterrent force and the strategy is revealed. The strategies adopted by each agent in different situations are analyzed. In Section 5.6, our simulation results are compared with the real-world videos and the simulation results generated by different models. The real-world videos are chosen from the public dataset and real antagonistic events. More details about comparisons can be seen in the supplementary video. User studies are performed to evaluate our method in Section 5.7. Fig. 4: The positions of all the agents at the 168th frame (Rca is 1). The green, purple, and blue circles represent civilians, activists, and cops, respectively. The stronger the color intensity of a circle, the higher the deterrent force of the agent. 5.1 The impact of ratios of agents in different roles on the result of crowd violence We investigate the effects of varying ratios of agents in different roles on the results of crowd violence. By analyzing a large number of real-world antagonistic videos, we select several representative values of Rca (the ratio of cops to activists). In this section the initial Rcas are 0, 0.6, 0.8, 1, 1.2, 1.4, and 1.6. The initial numbers of civilians and activists are 80 and 50, respectively. The initial emotion values of all the activists and cops are -0.5 and 0.5, respectively. We can learn from Figure 3 that the number of different kinds of agents determines the outcome of crowd violence to some extent. When we increase the ratio of cops, they can JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 10 (a) (d) (g) (b) (e) (h) (c) (f) (i) Fig. 5: Some fascinating emergent phenomena our model uncovers. (a), (b), and (c) are the simulations of activists attacking civilians. (d), (e), and (f) are the simulations of role transitions. (g), (h), and (i) are the simulations of cops encircling activists. The green, purple, blue, and grey circles are civilians, activists, cops, and dead agents, respectively. The stronger the color intensity of a circle, the higher the deterrent force of the agent. subdue activists quickly, stabilize the situation, and reduce civilian casualties. When Rca is 1, it means that the initial total deterrent force of the cops is equal to that of the activists. Both cops and activists have the same probability of winning. However, in the end, the cops fail which means that all the cops are subdued by the activists. We offer a detailed explanation in the following analysis. Figure 4 shows the positions of all the agents at the 168th frame when Rca is 1. At this time step, the numbers of cops and activists are the same. Activists with strong deterrent forces are mainly located in the upper right corner of the scene. Some cops with weak deterrent forces are among the crowd of activists (upper right corner of the scene). Some surrounding civilians turn into activists because of the incitement of the activists with high deterrent forces. The cops with strong deterrent forces are in the upper left corner of the scene, which makes it difficult for them to rescue the cops in the upper right corner. When the cops with weak deterrent forces are killed by the activists with strong deterrent forces, there are fewer cops left. Finally, all the activists get together to attack the remaining cops and the activists win. We learn from Figure 4 that activists participate in collective behavior to create regions of low cop-to-activist ratios, which reduces the chances of activist death. The conglomeration of scattered activists into small groups and the amalgamation of small groups into large ones make it difficult to wipe them out [71]. Although the number of agents determines the outcome of crowd violence to a certain extent, some take advantage of agents' spatial distributions to affect the outcome of crowd violence. 5.2 The impact of agent parameters on the result of crowd violence In this section we analyze the impact of agent parameters on the result of crowd violence. We choose two parameters which have great influences on the simulation results: PR (the radius of perceived range) and A or B in Equation 5. In Equation 5, Aj,i is the strength attribute by which an emotion is received by i from sender j and Bi,j is the strength attribute by which an emotion is sent from i to receiver j. We discuss the relationship between active ratios (ratios of activists to the sum of the activists and civilians) and the parameters. The initial numbers of the civilians, activists, and cops are 80, 50, and 70, respectively. The initial emotions of the civilians, activists, and cops are 0.1, -0.5, and 0.5, respectively. The initial total deterrent force of the cops is higher than that of the activists. We show active ratios according to various values of PR in Figure 6. We assume that all individuals have the same PR. When PR of agents are different, the results of active ratios are also different. With the increase of PR, agents can more accurately estimate the situations. Therefore, the high total deterrent force of cops plays a role in defeating the activists. The active ratio decreases with the increase of PR. In addition, we also discuss the relationship between the values of A or B and active ratios in Figure 7. A and JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 11 Fig. 6: Active ratios corresponding to different PRs (the radius of perceived range). The active ratio decreases with the increase of PR. When PR is greater than 10, the active ratio tends to be stable. We choose PR = 10 in this paper. Fig. 7: Active ratios corresponding to different values of A. The active rate decreases with the increases of A. When A is greater than 0.8, the active ratio tends to be stable. We choose A = 0.8 in this paper. B are very important parameters for emotional contagion in Equation 5. The values of them are positively correlated with the values of emotional contagion. We suppose all the strength attributes of A between any two individuals are the same and those of B between any two individuals are the same. The relationship between the active ratio and A is the same as that between the active ratio and B. We take A as an example to discuss this relationship. As the value of A increases, agents' emotions also increase and the difference between the total deterrent forces of cops and activists is greater. Due to the high deterrent force of cops, more and more activists are subdued by cops. Therefore when the value of A increases, the active rate decreases. 5.3 Emergent phenomena uncovered by our model Our model can simulate many emergent phenomena that conform to real-world scenes. Figures 5a, 5b, and 5c show Fig. 8: Simulation of antagonistic crowd behavior using 3D character models. that activists with strong deterrent forces attack civilians. At the top left corner of this scene, there are a lot of civilians. Many activists are at the lower right corner of Figure 5a. Agents of the same type gather together according to [72]. Therefore, it is reasonable that agents of the same type gather together. The number of activists is larger than that of the cops. We can see from Figure 5a that the total deterrent force of the activists is stronger than that of the cops. Activists attack civilians (Figure 5b) and more and more civilians die (Figures 5b and 5c). At the lower right corner of Figure 5d, there are plenty of cops with high deterrent forces near the highlighted activist and his deterrent force is weak. If he continues to resist, he will die. He therefore transitions roles (from activist to civilian). When all the other activists die, he survives (Figure 5f). When the number of surrounding activists is large enough, it may impel civilians to become activists [73]. In Figure 5g, there are many more cops than activists. The total deterrent force of the cops is much stronger than that of the activists. At first, the cops divide the activists into two groups (Figure 5g) and prevent the activists from gathering together to form a larger group. Next, the cops eliminate these two groups of activists individually. There are some activists in the left side of the scene (the first group). Cops encircle these activists and more activists will die. When all the activists on the left side are eliminated, the cops return to the activists on the right side of the scene (the second group). These activists are encircled by the cops. Finally, all the activists are killed by the cops and the cops win. Figure 8 shows the simulation of antagonistic crowd behavior using 3D character models. The activists with high deterrent forces on the left of the scene are not afraid of the cops with low deterrent forces. In the upper left corner of the scene, the activists with high deterrent forces attack a civilian. The civilian runs away from the activists. In the middle and lower part of the scene, the cop with a high deterrent force attacks the activist and he runs away from the cop. 5.4 The impact of emotional contagion on the result of crowd violence The impact of emotional contagion on the results of crowd violence is presented in Figure 9. The initial numbers of civilians, activists, and cops are 80, 50, and 40, respectively. In Figure 9a, the initial emotion values of the cops are higher than those of the activists. At the 53rd time step, the number JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 12 (a) (b) Fig. 9: The changes in the number of civilians, activists, and cops: (a) considering emotional contagion and (b) not considering emotional contagion. (a) (b) Fig. 10: Simulation results (a) considering emotional con- tagion and (b) not considering emotional contagion. In (a) the deterrent forces of the agents are different and the same types of agents are more likely to gather together. In (b) the deterrent forces of all the agents are the same and the same type of agents are dispersed. of activists is zero. All the activists have been wiped out by the cops. At first the number of civilians increases because the total deterrent force of the cops is much higher than that of the activists. Some activists change their role (from activist to civilian). In Figure 9b, all the agents without emotion have the same deterrent force. At the 41st time step, the number of cops is zero and the activists win. We can learn from Figure 9 that the emotion module of our model can describe the differences observed between the agents. In our model, the deterrent forces of all the agents are different according to their emotions. The greater the absolute value of an agent's emotion, the higher the deterrent force of that agent. Although the number of agents is small, it is still possible to overcome a larger group of opponents by improving the emotions and deterrent forces of them. Therefore, our model can simulate situations in which agents overcome their more numerous opponents. Figure 10 shows simulation results with and without considering emotion. In Figure 10a the deterrent forces of the agents are different. The stronger the color intensity of a circle, the higher the deterrent force of the agent. Because of emotional contagion, the same types of agents are more likely to gather together, which is similar to what happens in real-world scenarios [72]. In Figure 10b the deterrent forces of all the agents are the same and the same type of agents are more dispersed. Figure 11 shows the heat maps of antagonistic emotion. (a) (c) (b) (d) Fig. 11: The heat maps of antagonistic emotion: (a) heat map at the 5th time step, (b) heat map at the 17th time step, (c) heat map at the 27th time step, and (d) heat map at the 42th time step. The red area represents the emotions of the cops. The blue area represents the emotions of the activists. The stronger the color intensity, the larger the value of the emotion. Fig. 12: Cooperation ratios of cops and activists at different time steps: (a) when the total deterrent force of the activists is higher than that of the cops; (b) when the total deterrent force of the activists is equal to that of the cops; (c) when the total deterrent force of the activists is less than that of the cops. Different colors represent different types of agents' emo- tions. The stronger the color intensity, the larger the value of the emotion. At first, the emotions of the cops and the activists are very weak. Later, as a result of the confrontation between cops and activists, both types of emotions increase. Since the initial emotions of the cops are higher than those of activists, the overall emotional scope of the cops becomes wider and wider and that of the activists becomes smaller and smaller. Finally, all the activists are wiped out by the cops and there is no blue area on the map. 5.5 The influence of deterrent force on strategy selec- tion We present the relationship between deterrent force and strategy selection in this subsection. We analyze the strat- egy (cooperation or defection) adopted by each agent with respect to their different deterrent forces. Figure 12 shows the overview of cooperation ratios (the ratio of the number of agents adopting a cooperation strat- egy to the total number of this type of agents) in relation to JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 different deterrent forces. When the total deterrent force of the activists is higher than that of the cops (Figure 12a), most of the cops adopt a cooperation strategy to avoid causalities and conserve their fighting forces. When the total deterrent force of the activists is equal to that of the cops, the cooperation ratios of cops and activists are almost the same from the 5th time step to the 45th time step. After the 45th time step, the cops defeat the activists. More and more cops adopt defection strategy, the cooperation ratio of the cops decreases, and the cooperation ratio of the activists increases. When the total deterrent force of the activists is less than that of the cops, the cooperation ratio of the activists is higher than that of the cops. In summary, the agent with a higher deterrent force is more likely to adopt a defection strategy and the agent with a lower deterrent force is more likely to adopt a cooperation strategy. 5.6 Comparisons To validate our approach, we compare our simulation re- sults with those generated by the CVM [12] and ABEC [74] models and real-world videos. Main goal of our model is to predict trends of crowd movements in these situations without explicitly modeling the trajectory of a particular individual. Simulation results obtained by our model are closest to the real-world scenes in the overall trends of crowd movements. The real scenes shown in Figures 13a and 13c are chosen from the public web dataset [75]. The real scenes of Figures 13e, 13g, and 13i are chosen from real antagonistic incidents on YouTube. For the real-world scenarios in Figure 13, we use different colors to distinguish different roles of indi- viduals. The green, purple, and blue circles and cylinders represent civilians, activists, and cops, respectively. The red arrows are the dominant paths of crowd movements. Cops and activists in this case represent two opposing groups. Civilians in this case represent onlookers and neutrals. There are no deaths of activists or cops in these scenes. The scenes can be simulated by increasing the values of the thresholds of Twarn and Twarn time for our model. The parameter values used in the simulation runs are listed in Table 4. More details can be seen in the supplementary video. We use the dominant path and entropy metric to quan- titively evaluate our simulation results. Dominant path is defined based on collectiveness of crowd movements. Collectiveness, which indicates the degree to which indi- viduals acting as a unit, is a fundamental and universal measurement for various crowd systems, including crowds in antagonistic scenes [76]. Individuals locally coordinate their movements and behaviors with their neighbors, and then the crowd is self-organized into collective movements without external control. The method proposed in [77] is used to calculate the collectiveness of our simulation results. When the collectiveness of individual movements in a cer- tain area is significantly higher than that of surrounding area, a small group of agents with similar movements is formed. The center of the group is determined according to the average of the positions of all the agents in this group. The trajectory of the group center forms its dominant path 13 (b) (d) (f) (h) (j) (a) (c) (e) (g) (i) Fig. 13: Comparisons between real scenes and our simula- tion results. (a), (c), (e), (g), and (i) are five real antagonistic scenes. (b), (d), (f), (h), and (j) are our corresponding simula- tion results generated by our ACSEE model. The red arrows are the dominant paths of crowd movements. and can be treated as the movement trend of the whole group. Using this method, we can get the dominant paths of the real-world videos. We use entropy metric [78], angular error (AE) [79], and inter-group distance metric (IDM) [80] to evaluate our crowd simulation results. They are used to evaluate the errors of trajectories, movement directions, and distances between each group, respectively. These three evaluation methods complement each other, which are used to evaluate our results more comprehensively. An entropy metric [78] is adopted to evaluate the error between the dominant paths of simulation results and that of the real-world videos. A lower entropy value implies a higher similarity between the simulation results and the real-world scenarios. Table 5 shows the entropy metric of JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 the simulation results achieved by the CVM, ABEC, and ACSEE (ours) models on different scenarios in Figure 13. The simulation results obtained by our model conform to the real-world videos best. The angular error (AE) [79] between the movement direction of the simulation result (Vx, Vy) and that of the ground-truth (Vxgt, Vygt) is also used to evaluate our crowd simulation method. This AE is defined in Equation 15. The inter-group distance metric (IDM) [80] compares the difference in the average distances between each pair of clustered agents. Tables 6 and 7 show the AE and IDM of our simulation results on different scenarios in Figure 13. Tables 5, 6, and 7 show that our method consistently outperforms the CVM and ABEC models. Compared with the CVM model, our ACSEE model considers the agents' an- tagonistic emotions and accurately describe the differences between agents' deterrent forces. Compared with the ABEC model only considering emotional contagion in antagonistic scenarios, our method considers not only antagonistic emo- tion, but also the relationship between antagonistic emotion and evolutionary game theory. The agent is able to choose a more reasonable strategy according to the situation. AE = cos xgt + V 2 ygt) (15) −1((Vx · Vxgt + Vy · Vygt)/(cid:112)V 2 x + V 2 y (cid:112)V 2 TABLE 5: Entropy metric for our simulation algorithms on different scenarios from Figure 13. A lower entropy value implies higher similarity between the simulation results and the real-world scenarios. Simulations with an entropy score less than 1.000 are considered very visually similar to the source data and those with a score greater than 6.000 are visually very different [78]. Scene No. 2 (Figure 13c) is too large and chaotic and the collectiveness of crowd movement is not so obvious. Therefore, the entropy value of Scene No. 2 is larger than 1.000, but this value is far less than 6.000. Scene No. 1 ACSEE ABEC CVM 0.193 0.232 0.259 2 1.310 1.320 1.369 3 0.255 0.268 0.270 4 0.104 0.139 0.151 5 0.117 0.180 0.187 TABLE 6: AE for the simulation algorithms on different scenes from Figure 13. A lower value for AE implies higher similarity with respect to the real-world crowd videos. We report mean and variance of AE at different time steps. Scene No. 1 2 3 4 5 ACSEE ABEC CVM 0.132/0.150 0.285/0.329 0.210/0.099 0.090/0.003 0.036/0.003 0.214/0.155 0.463/3.027 0.999/12.822 0.500/0.002 0.558/0.006 1.329/20.709 1.579/1.500 1.464/14.452 0.948/0.007 0.856/0.062 The simulation result generated by our model is com- pared with those of the CVM [12] and ABEC [74] models in Figure 14. More details can be seen in the supplementary video. In contrast to the CVM and ABEC models, our method can better quantify the differences of the deterrent forces of all the agents. Besides, we also integrate antag- onistic emotional contagion into evolutionary game theory TABLE 7: IDM (pixel) for our simulation algorithms on different scenes from Figure 13. Lower numbers are better. 14 Scene No. ACSEE ABEC CVM 1 1 31 40 2 30 32 42 3 2 21 32 4 6 18 49 5 12 35 63 to estimate situations of antagonistic scenes more accurately, which helps agents make reasonable strategies in the games. Therefore our simulation result is the most similar to the real-world scenario. 5.7 User Studies In this section we describe user studies conducted to demon- strate the perceptual benefits of our ACSEE model com- pared to other models in simulating antagonistic crowd behaviors. Experiment Goals & Expectations: Our main goal is to measure how close the crowd movement tendencies gener- ated using different models are to those observed in real- world videos. We hypothesize that in both studies, agents simulated with our ACSEE model will exhibit overall more plausible antagonistic crowd movements than other models. Therefore, participants will strongly prefer our model to the other models. Experimental Design: Two user studies were conducted based on a paired-comparison design. In each study, partic- ipants were shown pre-recorded videos in a side-by-side comparison of simulation results generated by different models and a real-world video. In particular, we asked the users to compare the crowd movement tendencies gen- erated by different crowd simulation models with those observed in real-world videos. The studies had no time constraints and the participants were free to take breaks between the benchmarks. We encouraged the users to watch these videos as many times as they wanted and finally give a stable score. Comparison Methods: The first study compares our ACSEE model considering emotion with our model that does not consider emotion. The second study compares our model with the CVM [12] and ABEC [74] models. Environments: We use outdoor scenarios without ob- stacles. In these scenarios, the green, purple, blue, and grey circles are civilians, activists, cops, and dead agents, respectively. Metrics: Participants were shown two pre-recorded videos in a side-by-side comparison of the simulation result and a real-world video. We asked the users to first watch the real-world video and then rate each simulation result on a scale of 1 - 5 in terms of the similarity of movement tendencies between the real-world video and the simulation result video. A score of 1 indicates most dissimilar and a score of 5 indicates most similar movement tendencies. Results: There are 39 participants (20 female) with a mean age of 25.12± 3.26 years in these studies. We measure the mean and the variance of their scores and then compute the p-values using a two-tailed t-test. The means of their JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 15 (a) Real scenario (b) CVM result (c) ABEC result (d) Our result Fig. 14: Comparisons between the real scenario and sim- ulation results of different crowd simulation models. The simulation result obtained by our model conform to the real- world video best. Fig. 15: User evaluation of simulation scenes. (a) Compari- son of simulation results considering emotion and without considering emotion. (b) Comparison of simulation results generated by the CVM, ABEC, and ACSEE models. Partic- ipants were asked to rate each simulation result on a scale of 1 - 5 in terms of the similarity of movement tendencies between the real-world videos and the simulation result videos. We can see from the results that our simulation results are most similar to real-world scenarios. scores for our model without emotion and with emotion are 2.87 ± 0.67 and 3.73 ± 0.93, respectively. The means of their scores for the CVM, ABEC, and ACSEE models are 2.56 ± 1.44, 3.30 ± 0.50, and 3.70 ± 0.90, respectively. The p-value for Figure 15a comparison is 2.97e−26. The p- values for the comparison of the CVM and ACSEE models and that of the ABEC and ACSEE models in Figure 15b are 2.97e−26 and 1.05e−17, respectively. We observe that the antagonistic crowd behavior simulations generated by our ACSEE model score much higher than the other models at a statistically significant rate (p-value<0.05). The result indicates that the addition of antagonistic emotion and the integration of antagonistic emotion and evolutionary game theory improve the perceptual similarity of our simulations to the crowd movement tendencies in real-world scenes. Our model gets higher scores than the CVM and ABEC models, as detailed in Figure 15b. Participants indicate their preference for our ACSEE model. 6 CONCLUSION AND LIMITATIONS We present a new model for antagonistic crowd behavior simulation integrated with emotional contagion and evolu- tionary game theory. Our approach builds on well-known psychological theories to present a comprehensive and an- tagonistic emotional contagion model. Based on the emo- tional calculation method, we propose the deterrent force to determine the situation of cops and activists. According to the situation, an enhanced evolutionary game theoretic ap- proach incorporated with antagonistic emotional contagion is determined. Finally, we present a behavior control deci- sion method based on the antagonistic emotional contagion and evolutionary game theoretic approaches. Our proposed model is verified by simulations. We in- vestigate the impact of different factors (number of agents, emotion, strategy, etc.) on the outcome of crowd violence. Our model is compared with real-world videos and previ- ous approaches. Results show that our proposed model can reliably generate realistic antagonistic crowd behaviors. However, our model still has several limitations. Al- though our simulation results are closer to the real-world scenes in the overall trend of crowd movement, the antago- nistic emotions in a crowd violent scene cannot be obtained directly or inferred accurately. One of the main reasons is that the quality of most of the real videos is poor, since they are often captured by moving phones. At present, there is no effective methods to identify and quantify the emotion values of all the individual in such videos with poor quality. Thus, the initial state of our model is set empirically according to real-world videos, which is time- consuming and not very accurate. In the future, we plan to use the latest wearable equipment to collect these data and provide a new method that can quickly and accurately obtain the initial state. Moreover, the strategies and benefits calculated by our antagonistic evolutionary game theoretic approach are the ideal situations. Game theory assumes that all individuals are rational. However, some people in real scenes are irrational and extreme, which doesn't fully satisfy the precondition of game theory. In practice, people do not necessarily adopt the optimal strategy because of the limitations of perception and other complex factors. Our current calculation result is optimal, which is only one of the possible results. In fact, it is impossible for all simulation results to be consistent with the real results. We will continue to improve our prediction results considering more actual situations of antagonistic crowds. At present, the behavior control in our model is proposed based on the cellular automata [15]. Other more complex behavioral control methods will be further considered. 7 ACKNOWLEDGMENTS This work was supported by National Natural Science Foundation of China under Grant Number 61672469, 61772474, 61822701, 61872324, Program for Science & Tech- nology Innovation Talents in Universities of Henan Province (20HASTIT021, 18HASTIT020) and Youth Talent Promotion Project in Henan Province (2019HYTP022). [3] JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 REFERENCES [1] D. Helbing, I. Farkas, and T. Vicsek, "Simulating dynamical fea- tures of escape panic," Nature, vol. 407, no. 6803, pp. 487 -- 490, 2000. [2] M. Zucker, J. Kuffner, and M. Branicky, "Multipartite RRTs for rapid replanning in dynamic environments," in Proc. IEEE Conf. Robotics and Automation, 2007, pp. 1603 -- 1609. I. Karamouzas and M. Overmars, "Simulating and evaluating the local behavior of small pedestrian groups," IEEE Transactions on Visualization and Computer Graphics, vol. 18, no. 3, pp. 394 -- 406, 2012. [4] G. Zhang, D. Lu, and H. Liu, "Strategies to utilize the positive emotional contagion optimally in crowd evacuation," IEEE Trans- actions on Affective Computing, vol. 14, no. 8, pp. 1 -- 14, 2017. [5] F. Durupinar, U. Gudukbay, A. Aman, and N. I. Badler, "Psycho- logical parameters for crowd simulation: from audiences to mobs," IEEE Transactions on Visualization and Computer Graphics, vol. 22, no. 9, pp. 2145 -- 2159, 2016. [6] A. B. F. Neto, C. Pelachaud, and S. R. Musse, "Giving emotional contagion ability to virtual agents in crowds," in Proc. Intelligent Virtual Agents, vol. 10498, 2017, pp. 63 -- 72. J. Bruneau, A.-H. Olivier, and J. Pettre, "Going through, going around: A study on individual avoidance of groups," IEEE Trans- actions on Visualization and Computer Graphics, vol. 21, no. 4, pp. 520 -- 528, 2015. S. J. Guy, J. Chhugani, S. Curtis, P. Dubey, M. Lin, and D. Manocha, "PLEdestrians: a least-effort approach to crowd simulation," in Proc. ACM SIGGRAPH/Eurographics Symp. Comput. Animation, 2010, pp. 119 -- 128. J. Ondrej, J. Pettre, A.-H. Olivier, and S. Donikian, "A synthetic- vision based steering approach for crowd simulation," vol. 29, no. 4, pp. 123:1 -- 123:9, 2010. [10] S. A. Stuvel, N. Magnenat-Thalmann, D. Thalmann, A. F. van der Stappen, and A. Egges, "Torso crowds," IEEE Transactions on Visualization and Computer Graphics, vol. 23, no. 7, pp. 1823 -- 1837, 2017. [11] Y. Zhang, L. Qin, R. Ji, S. Zhao, Q. Huang, and J. Luo, "Exploring coherent motion patterns via structured trajectory learning for crowd mood modeling," IEEE Transactions on Circuits and Systems for Video Technology, vol. 27, no. 3, pp. 635 -- 648, MAR 2017. [12] H. Y. Quek, K. C. Tan, and H. A. Abbass, "Evolutionary game theoretic approach for modeling civil violence," IEEE Transactions on Evolutionary Computation, vol. 13, no. 4, pp. 780 -- 800, 2009. [13] H. Situngkir, "On massive conflict: Macro-micro link," Journal of [7] Social Complexity, vol. 1, no. 4, FEB 2004. [8] [9] [14] M. Xu, H. Jiang, X. Jin, and Z. Deng, "Crowd Simulation and Its Applications: Recent Advances," Journal of Computer Science and Technology, vol. 29, no. 5, pp. 799 -- 811, SEP 2014. [15] L. Fu, W. Song, W. Lv, and S. Lo, "Simulation of emotional conta- gion using modified SIR model: A cellular automaton approach," Physica A Statistical Mechanics & Its Applications, vol. 405, pp. 380 -- 391, 2014. [16] S. Parikh and C. Cameron, "Riot games: A theory of riots and mass political violence," in Proc. 7th Wallis Inst. Conf. Political Economy, New York: Univ. Rochester, OCT 2000. [17] R. B. Myerson, Game theory : analysis of conflict. Harvard Univer- sity Press, 1997. [20] [18] S. Y. Chong, J. Humble, G. Kendall, J. Li, and X. Yao, "Iterated prisoner's dilemma and evolutionary game theory," The Iterated Prisoners' Dilemma: 20 Years On, pp. 23 -- 62, 2007. [19] J. Li, G. Kendall, and R. John, "Computing nash equilibria and evolutionarily stable states of evolutionary games," IEEE Transac- tions on Evolutionary Computation, vol. 20, no. 3, pp. 460 -- 469, 2016. ´A. G´omez, L. L´opez-Rodr´ıguez, H. Sheikh, J. Ginges, L. Wilson, H. Waziri, A. V´azquez, R. Davis, and S. Atran, "The devoted ac- tor's will to fight and the spiritual dimension of human conflict," Nature Human Behaviour, vol. 1, no. 9, pp. 673 -- 679, SEP 2017. [21] H. Jiang, Z. Deng, M. Xu, X. He, T. Mao, and Z. Wang, "An emotion evolution based model for collective behavior simulation," in Proc. Symp. Interactive 3D Graphics and Games, 2018, pp. 10:1 -- 10:6. [22] S. Kim, S. J. Guy, D. Manocha, and M. C. Lin, "Interactive simulation of dynamic crowd behaviors using general adaptation syndrome theory," in Proc. Symp. Interactive 3D Graphics and Games, 2012, pp. 55 -- 62. [23] P. Lv, Z. Zhang, C. Li, Y. Guo, B. Zhou, and M. Xu, "Crowd behavior evolution with emotional contagion in political rallies," IEEE Transactions on Computational Social Systems, vol. 6, no. 2, pp. 377 -- 386, APR 2019. 16 [24] S. Huerre, J. Lee, M. Lin, and C. O'Sullivan, "Simulating believable crowd and group behaviors," in Proc. ACM SIGGRAPH Asia, 2010, pp. 13:1 -- 13:92. [25] C. M. D. Melo, J. Gratch, and P. J. Carnevale, "Humans vs. computers: Impact of emotion expressions on peoples decision making," IEEE Transactions on Affective Computing, vol. 6, no. 2, pp. 127 -- 136, 2015. [26] J. Tang, Y. Zhang, J. Sun, J. Rao, W. Yu, Y. Chen, and A. C. M. Fong, "Quantitative study of individual emotional states in social networks," IEEE Transactions on Affective Computing, vol. 3, no. 2, pp. 132 -- 144, APR 2012. [27] M. Xu, X. Xie, P. Lv, J. Niu, H. Wang, C. Li, R. Zhu, Z. Deng, and B. Zhou, "Crowd behavior simulation with emotional contagion in unexpected multi-hazard situations," IEEE Transactions on Systems, Man, and Cybernetics: Systems, pp. 1 -- 15, 2019. [28] M. Xu, Y. Wu, Y. Ye, I. Farkas, H. Jiang, and Z. Deng, "Collec- tive Crowd Formation Transform with Mutual Information-Based Runtime Feedback," Computer Graphics Forum, vol. 34, no. 1, pp. 60 -- 73, FEB 2015. [29] M. Xu, Y. Wu, P. Lv, H. Jiang, M. Luo, and Y. Ye, "miSFM: On combination of Mutual Information and Social Force Model towards simulating crowd evacuation," Neurocomputing, vol. 168, pp. 529 -- 537, 2015. [30] I. J. Farkas, J. Kun, Y. Jin, G. He, and M. Xu, "Keeping speed and distance for aligned motion," Phys. Rev. E, vol. 91, p. 012807, Jan 2015. [31] W. O. Kermack and A. G. Mckendrick, "Contributions to the math- ematical theory of epidemics," Bulletin of Mathematical Biology, vol. 53, no. 1, pp. 57 -- 87, MAR 1991. [32] L. Zhao, H. Cui, X. Qiu, X. Wang, and J. Wang, "SIR rumor spreading model in the new media age," Physica A Statistical Mechanics & Its Applications, vol. 392, no. 4, pp. 995 -- 1003, 2013. [33] F. Durupinar, N. Pelechano, J. Allbeck, U. Gudukbay, and N. I. Badler, "How the ocean personality model affects the perception of crowds," IEEE Computer Graphics and Applications, vol. 31, no. 3, pp. 22 -- 31, 2011. [34] J. H. Wang, S. M. Lo, J. H. Sun, Q. S. Wang, and H. L. Mu, "Qual- itative simulation of the panic spread in large-scale evacuation," Simulation, vol. 88, no. 12, pp. 1465 -- 1474, 2012. [35] A. L. Hill, D. G. Rand, M. A. Nowak, and N. A. Christakis, "Emotions as infectious diseases in a large social network: the SISa model," in Proceeding of the Royal Society B-Biological Sciences, vol. 277, no. 1701, 2010, pp. 3827 -- 3835. [36] M. Cao, G. Zhang, M. Wang, D. Lu, and H. Liu, "A method of emotion contagion for crowd evacuation," Physica A Statistical Mechanics & Its Applications, vol. 483, pp. 250 -- 258, OCT 2017. [37] Y. Song and X. Yan, "A method for formulizing disaster evacu- ation demand curves based on SI model," International Journal of Environmental Research and Public Health, vol. 13, no. 10, pp. 986:1 -- 986:21, 2016. [38] M. Xu, H. Wang, S. Chu, Y. Gan, X. Jiang, Y. Li, and B. Zhou, "Traf- fic simulation and visual verification in smog," ACM Transactions on Intelligent Systems and Technology, vol. 10, no. 1, pp. 3:1 -- 3:17, 2019. [39] X. Lu, Z. Wang, M. Xu, W. Chen, and Z. Deng, "A personality model for animating heterogeneous traffic behaviors," Comput. Animat. Virtual Worlds, vol. 25, no. 3-4, pp. 363 -- 373, 2014. [40] H. Wang, M. Xu, F. Zhu, Z. Deng, Y. Li, and B. Zhou, "Shadow traffic: A unified model for abnormal traffic behavior simulation," Computers & Graphics, vol. 70, pp. 235 -- 241, 2018. [41] J. Xue, H. Yin, P. Lv, M. Xu, and Y. Li, "Crowd queuing simulation with an improved emotional contagion model," Science China Information Sciences, vol. 62, no. 4, pp. 044 101:1 -- 044 101:3, 2019. [42] T. Bosse, R. Duell, Z. A. Memon, J. Treur, and C. N. V. D. Wal, "A multi-agent model for mutual absorption of emotions," in Proc. 23rd EuropeanConf. Mod. and Sim., 2009, pp. 212 -- 218. [43] T. Bosse, R. Duell, Z. A. Memon, J. Treur, and N. van der Wal, "A multi-agent model for emotion contagion spirals integrated within a supporting ambient agent model," in Proc. Principles of Practice in Multi-Agent Systems, vol. 5925, 2009, pp. 48 -- 67. [44] J. Tsai, N. Fridman, E. Bowring, M. Brown, S. Epstein, G. Kaminka, S. Marsella, A. Ogden, I. Rika, A. Sheel, M. E. Taylor, X. Wang, A. Zilka, and M. Tambe, "ESCAPES: evacuation simulation with children, authorities, parents, emotions, and social comparison," in Proc. Autonomous Agents and Multiagent Systems, 2011, pp. 457 -- 464. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 [45] J. Tsai, E. Bowring, S. Marsella, and M. Tambe, "Empirical eval- uation of computational fear contagion models in crowd disper- sions," Autonomous Agents and Multi-Agent Systems, vol. 27, no. 2, pp. 200 -- 217, 2013. [46] A. R. Zomorrodi and D. Segre, "Genome-driven evolutionary game theory helps understand the rise of metabolic interdepen- dencies in microbial communities." Nature Communications, vol. 8, no. 1, pp. 1563 -- 1574, NOV 16 2017. [47] F. Saab, A. Kayssi, I. Elhajj, and A. Chehab, "Playing with sybil," ACM Sigapp Applied Computing Review, vol. 16, no. 2, pp. 16 -- 25, 2016. [48] J. Li, G. Kendall, and R. John, "Computing nash equilibria and evolutionarily stable states of evolutionary games," IEEE Transac- tions on Evolutionary Computation, vol. 20, no. 3, pp. 460 -- 469, 2016. [49] W. Huang, B. Haubold, C. Hauert, and A. Traulsen, "Emergence of stable polymorphisms driven by evolutionary games between mutants," Nature Communications, vol. 3, JUN 2012. [50] M. H. Vainstein, C. Brito, and J. J. Arenzon, "Percolation and cooperation with mobile agents: Geometric and strategy clusters," Physical Review E, vol. 90, no. 2, pp. 860 -- 877, 2014. [51] J. Li, C. Zhang, Q. Sun, Z. Chen, and J. Zhang, "Changing the intensity of interaction based on individual behavior in the iter- ated prisoners dilemma game," IEEE Transactions on Evolutionary Computation, vol. 21, no. 4, pp. 506 -- 517, 2017. [52] C. K. Goh, H. Y. Quek, K. C. Tan, and H. A. Abbass, "Modeling civil violence: An evolutionary multi-agent, game theoretic ap- proach," in Proc. IEEE Congress on Evolutionary Computation, 2006, pp. 1624 -- 1631. [53] F. Tian, H. Li, W. Chen, T. Qin, E. Chen, and T. Y. Liu, "Agent behavior prediction and its generalization analysis," in Twenty- eighth AAAI Conference on Artificial Intelligence, 2014, pp. 1300 -- 1306. [54] V. Kountouriotis, S. C. A. Thomopoulos, and Y. Papelis, "An agent- based crowd behaviour model for real time crowd behaviour simulation," Pattern Recognition Letters, vol. 44, pp. 30 -- 38, JUL 15 2014. [55] D. Helbing and S. Balietti, "How to do agent-based simulations in the future: From modeling social mechanisms to emergent phenomena and interactive systems design," Chapter "Agent-Based Modeling" of the book "Social Self-Organization" by Dirk Helbing (Springer, Berlin, 2012), pp. 25 -- 70, 2013. [56] L. Luo, C. Chai, J. Ma, S. Zhou, and W. Cai, "Proactivecrowd: Modelling proactive steering behaviours for agent-based crowd simulation," Computer Graphics Forum, vol. 37, no. 1, pp. 375 -- 388, FEB 2018. [57] D. Helbing and P. Molnar, "Social force model for pedestrian dynamics," Physical Review E, vol. 51, no. 5, pp. 4282 -- 4286, 1995. [58] A. Shendarkar, K. Vasudevan, S. Lee, and Y. J. Son, "Crowd simulation for emergency response using BDI agent based on virtual reality," in Simulation Conference, 2006. [59] L. Luo, S. Zhou, W. Cai, M. Y. H. Low, F. Tian, Y. Wang, X. Xiao, and D. Chen, "Agent-based human behavior modeling for crowd simulation," Computer Animation & Virtual Worlds, vol. 19, pp. 271 -- 281, AUG 2008. [60] H. Aydt, M. Lees, L. Luo, W. Cai, M. Y. H. Low, and S. K. Kadirve- len, "A computational model of emotions for agent-based crowds in serious games," in IEEE/WIC/ACM International Conference on Web Intelligence & Intelligent Agent Technology, 2011. [61] F. M. Nasir and M. S. Sunar, "A survey on simulating real-time crowd simulation," in International Conference on Interactive Digital Media, 2015. [62] S. Patil, J. van den Berg, S. Curtis, M. C. Lin, and D. Manocha, "Directing crowd simulations using navigation fields," IEEE Trans- actions on Visualization and Computer Graphics, vol. 17, no. 2, pp. 244 -- 254, FEB 2011. [63] F. Bu and J. Sun, "The psychological behaviour research of individ- uals in mass violence events," in Proceedings of the 8th International Conference on Intelligent Computing Theories and Applications, ser. ICIC, 2012, pp. 625 -- 632. [64] X. Wang, N. Liu, S. Liu, Z. Wu, M. Zhou, J. He, P. Cheng, C. Miao, and N. M. Thalmann, "Crowd formation via hierarchical plan- ning," in ACM SIGGRAPH Conference on Virtual-Reality Continuum and ITS Applications in Industry, 2016, pp. 251 -- 260. [65] S. Schutte, "Violence and Civilian Loyalties: Evidence from Afghanistan," Journal of Conflict Resolution, vol. 61, no. 8, pp. 1595 -- 1625, SEP 2017. 17 [66] I. Funahashi, "Kayika vedana (physical emotion) and cetasika vedana (mental emotion)," Journal of Indian and Buddhist Studies, vol. 10, pp. 510 -- 511, 1962. [67] E. M. Reiman, R. D. Lane, G. L. Ahern, G. E. Schwartz, R. J. Davidson, K. J. Friston, L. S. Yun, and K. Chen, "Neuroanatomical correlates of externally and internally generated human emotion," The American journal of psychiatry, vol. 154, no. 7, pp. 918 -- 925, 1997. [68] C. E. Salas, D. Radovic, and O. H. Turnbull, "Inside-out: compar- ing internally generated and externally generated basic emotions." Emotion, vol. 12, no. 3, pp. 568 -- 578, 2012. [69] R. W. Picard, Affective computing. Cambridge, MA, USA: MIT Press, 1997. [70] L. J. Carlson and R. Dacey, "The use of fear and anger to alter crisis initiation," Conflict Management and Peace Science, vol. 31, no. 2, pp. 168 -- 192, APR 2014. [71] J. M. Epstein, "Modeling civil violence: An agent-based computa- tional approach," in Proceedings of the National Academy of Sciences of the United States of America, vol. 99, 2002, pp. 7243 -- 7250. [72] G. L. Bon, The Crowd: A Study of the Popular Mind. Flint Hill, VA: Fraser Publishing Company, 1982. [73] P. Sunita and C. Charles, "Riot games: A theory of riots and mass political violence," in Proc. 7th Wallis ins. Conf. Political Economy, 2000. [74] L. Huang, G. Cai, H. Yuan, and J. Chen, "From public gatherings to the burst of collective violence: An agent-based emotion contagion model," in IEEE International Conference on Intelligence and Security Informatics, 2018, pp. 193 -- 198. [75] R. Mehran, A. Oyama, and M. Shah, "Abnormal crowd behavior detection using social force model," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit, 2009, pp. 935 -- 942. [76] B. Zhou, X. Tang, H. Zhang, and X. Wang, "Measuring crowd collectiveness," in Proc. Computer Vision and Pattern Recognition, 2013, pp. 3049 -- 3056. [77] M. Xu, C. Li, P. Lv, N. Lin, R. Hou, and B. Zhou, "An efficient method of crowd aggregation computation in public areas," IEEE Transactions on Circuits & Systems for Video Technology, vol. 28, no. 10, pp. 2814 -- 2825, 2018. [78] S. J. Guy, J. V. D. Berg, W. Liu, R. Lau, M. C. Lin, and D. Manocha, "A statistical similarity measure for aggregate crowd dynamics," ACM Transactions on Graphics, vol. 31, no. 6, pp. 190:1 -- 190:11, 2012. [79] I. Kajo, A. S. Malik, and N. Kamel, "An evaluation of optical flow algorithms for crowd analytics in surveillance system," in International Conference on Intelligent and Advanced Systems, 2016. [80] D. Wolinski, S. Guy, A.-H. Olivier, M. Lin, D. Manocha, and J. Pettre, "Parameter estimation and comparative evaluation of crowd simulations," Computer Graphics Forum, vol. 33, no. 2, pp. 303 -- 312, MAY 2014. Chaochao Li is a Ph.D candidate in the School of Information Engineering of Zhengzhou Uni- versity, China, and his research interest is com- puter graphics and computer vision. He got his B.S. degree in computer science and technology from the School of Information Engineering of Zhengzhou University. Pei Lv is an associate professor in School of Information Engineering, Zhengzhou University, China. His research interests include video anal- ysis and crowd simulation. He received his Ph.D in 2013 from the State Key Lab of CAD&CG, Zhejiang University, China. He has authored more than 20 journal and conference papers in these areas, including IEEE TIP, IEEE TCSVT, IEEE TACACM MM, etc. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, NOVEMBER 2019 Dinesh Manocha is the Paul Chrisman Iribe Chair in Computer Science & Electrical and Computer Engineering at the University of Mary- land College Park. He is also the Phi Delta Theta/Matthew Mason Distinguished Professor Emeritus of Computer Science at the Univer- sity of North Carolina - Chapel Hill. He has won many awards, including Alfred P. Sloan Re- search Fellow, the NSF Career Award, the ONR Young Investigator Award, and the Hettleman Prize for scholarly achievement. His research in- terests include multi-agent simulation, virtual environments, physically- based modeling, and robotics. His group has developed a number of packages for multi-agent simulation, crowd simulation, and physics- based simulation that have been used by hundreds of thousands of users and licensed to more than 60 commercial vendors. He has published more than 510 papers and supervised more than 35 PhD dissertations. He is an inventor of 9 patents, several of which have been licensed to industry. His work has been covered by the New York Times, NPR, Boston Globe, Washington Post, ZDNet, as well as DARPA Legacy Press Release. He is a Fellow of AAAI, AAAS, ACM, and IEEE and also received the Distinguished Alumni Award from IIT Delhi. See http://www.cs.umd.edu/∼dm 18 Mingliang Xu is a full professor in the School of Information Engineering of Zhengzhou Univer- sity, China, and currently is the director of CIISR (Center for Interdisciplinary Information Science Research) and the vice General Secretary of ACM SIGAI China. He received his Ph.D. degree in computer science and technology from the State Key Lab of CAD&CG at Zhejiang Univer- sity, Hangzhou, China. He previously worked at the department of information science of NSFC (National Natural Science Foundation of China), Mar.2015-Feb.2016. His current research interests include computer graphics, multimedia and artificial intelligence. He has authored more than 60 journal and conference papers in these areas, including ACM TOG, ACM TIST, IEEE TPAMI, IEEE TIP, IEEE TCYB, IEEE TCSVT, ACM SIGGRAPH (Asia), ACM MM, ICCV, etc. Hua Wang was born in 1982. She received her PhD degree in Computer Science from the Institute of Computing Technology, Chinese Academy of Sciences, in 2015. She is currently an associate professor of Zhengzhou University of Light Industry, Zhengzhou, China. Her main research interests include traffic animation and environment modeling. Yafei Li is an assistant professor in the School of Information Engineering, Zhengzhou University, Zhengzhou, China. He received the PhD degree in computer science from Hong Kong Baptist University, in 2015. He held a visiting position in the Database Research Group with Hong Kong Baptist University. His research interests span mobile and spatial data management, location- based services, and urban computing. He has authored more than 20 journal and conference papers in these areas, including IEEE TKDE, IEEE TSC, ACM TWEB, ACM TIST, PVLDB, IEEE ICDE, WWW, etc. Bing Zhou is currently a professor at the School of Information Engineering, Zhengzhou Univer- sity, Henan, China. He received the B.S. and M.S. degrees from Xi'an Jiao Tong University in 1986 and 1989, respectively,and the Ph.D. degree in Beihang University in 2003, all in computer science. His research interests cover video processing and understanding, surveil- lance, computer vision, multimedia applications.
0912.4553	1	0912	2009-12-23T02:42:33	Consensus Dynamics in a non-deterministic Naming Game with Shared Memory	[ "cs.MA", "cs.AI" ]	In the naming game, individuals or agents exchange pairwise local information in order to communicate about objects in their common environment. The goal of the game is to reach a consensus about naming these objects. Originally used to investigate language formation and self-organizing vocabularies, we extend the classical naming game with a globally shared memory accessible by all agents. This shared memory can be interpreted as an external source of knowledge like a book or an Internet site. The extended naming game models an environment similar to one that can be found in the context of social bookmarking and collaborative tagging sites where users tag sites using appropriate labels, but also mimics an important aspect in the field of human-based image labeling. Although the extended naming game is non-deterministic in its word selection, we show that consensus towards a common vocabulary is reached. More importantly, we show the qualitative and quantitative influence of the external source of information, i.e. the shared memory, on the consensus dynamics between the agents.	cs.MA	cs	Consensus Dynamics in a non-deterministic Naming Game with Shared Memory Regina ldo J. da Silva Filho, Matthias R. Brust and Carlos H.C. Ribeiro Computer Science Division Technological Institute of Aeronautics (ITA) São José dos Campos, Brazil Abstract—In the naming game, individuals or agents exchange pairwise local information in order to communicate about objects in their common environment. The goal of the game is to reach a consensus about naming these objects. Originally used to investigate language formation and self-organizing vocabularies, we extend the classical naming game with a globally shared memory accessible by all agents. This shared memory can be interpreted as an external source of knowledge like a book or an Internet site. The extended naming game models an environment similar to one that can be found in the context of social bookmarking and collaborative tagging sites where users tag sites using appropriate labels, but also mimics an important aspect in the field of human-based image labeling. Although the extended naming game is non-deterministic in its word selection, we show that consensus towards a common vocabulary is reached. More importantly, we show the qualitative and quantitative influence of the external source of information, i.e. the shared memory, on the consensus dynamics between the agents. Keywords-distributed collaboration; consensus; naming game; collaborative tagging; shared memory; multi-agent system I. INTRODUCTION The natural emergence of a common language between individuals still remains an unexplained phenomenon. However, a deeper understanding of the evolutionary processes of indispensable for developing is language formation autonomous multi-agent systems where each agent can potentially have different origins and where no knowledge about the language used in an open-ended environment is provided. To put in a question: How can these agents build a common language through local negations and reach a consensus about the meaning of their vocabulary? A promising model for a deeper understanding of the common language phenomenon is the naming game [1]. It describes a model in which individuals can reach a consensus on how to name different objects. All individuals (or agents) exist in the same environment and sense the same set of objects. The agents are able to invent or create words for the objects. An interaction between two agents is a word transmission from one agent (the speaker) to the second agent (the hearer). The goal of the game is to reach an agreement between speaker and hearer about the object-word association used for a single object [2]. In this way, a self-organized vocabulary or even a common language with syntactic and semantic levels can be built [3]. The naming game thus is a microscopic model for the interaction dynamics among autonomous agents that communicate without any centralized control [4]. The interesting fact is that a consensus can be reached by means of local interactions. Distributed models can be used to understand, for instance, how large populations reach an agreement with respect to the usage of a certain word, how new language constructs are established, the spreading of rumors and opinions, or even how words propagate in social networks. Besides its application in modeling the language formation process for individuals, agents or robots (in particular in the field of artificial intelligence), the naming game is of relevance to understand the consensus dynamics of collaborative tagging systems of web sites like delicious and flickr [5] that have become increasingly popular in recent years. The users of such sites can attach keywords or tags to provide information (e.g., favorite sites on the Internet). In a recent study [6], it is shown how collaborative tagging can lead to both regularities regarding users’ activities, tag frequency and keyword usage, and stabilities concerning relative proportions between tags for a given URL and strings that define the location of programs or files in the Internet. Although it is potentially possible to have a constantly increasing number of tags, these findings indicate convergence of a name descriptor (the collection of tags) and the concept (the contents in the location itself). Different variants of naming games played by humans can help to overcome one of the most challenging problems for search engines: Image labeling. The ESP game [7] aims to use humans’ perceptual abilities in order to create valuable output in the process of image labeling. Two players are shown the same image but they are not able to communicate. They are then asked to describe the image with labels under a given time constraint (e.g. Google Image Labeler uses 2 minutes). As soon as they use a common label, it is saved in the database to index the image, the players earn points accordingly and the next image is shown. The objective is to get as many points as possible. While the ESP game is initially designed for a two- player game, in a broader context, the label consensus dynamics of the naming game can be directly used to improve the description accuracy of the images. Research on the naming game uses mainly the introduced communication model above and focuses on showing its convergence empirically [1,3,8]. The convergence of a deterministic naming game, on is the other hand, mathematically proven in [2]. One common characteristic of
1803.04934	1	1803	2018-03-13T16:53:50	An agent-based evaluation of impacts of transport developments on the modal shift in Tehran, Iran	[ "cs.MA", "cs.AI", "cs.CY" ]	Changes in travel modes used by people, particularly reduction of the private car use, is an important determinant of effectiveness of transportation plans. Because of dependencies between the choices of residential location and travel mode, integrated modelling of these choices has been proposed by some researchers. In this paper, an agent-based microsimulation model has been developed to evaluate impacts of different transport development plans on choices of residential location and commuting mode of tenant households in Tehran, the capital of Iran. In the proposed model, households are considered as agents who select their desired residential location using a constrained NSGA-II algorithm and in a competition with other households. In addition, they choose their commuting mode by applying a multi-criteria decision making method. Afterwards, effects of development of a new highway, subway and bus rapid transit (BRT) line on their residential location and commuting mode choices are evaluated. Results show that despite the residential self-selection effects, these plans result in considerable changes in the commuting mode of different socioeconomic categories of households. Development of the new subway line shows promising results by reducing the private car use among the all socio-economic categories of households. But the new highway development unsatisfactorily results in increase in the private car use. In addition, development of the new BRT line does not show significant effects on the commuting mode change, particularly on decrease in the private car use.	cs.MA	cs	An agent-based evaluation of impacts of transport developments on the modal shift in Tehran, Iran Ali Shirzadi Babakan, Abbas Alimohammadi, and Mohammad Taleai This is an Author's Original Manuscript (Preprint) of an Article Published by Taylor & Francis Group in Journal of Development Effectiveness, 2015, Vol. 7, No. 2, 230-251. DOI: 10.1080/19439342.2014.994656 To link to this article: http://dx.doi.org/10.1080/19439342.2014.994656 Abstract Changes in travel modes used by people, particularly reduction of the private car use, is an important determinant of effectiveness of transportation plans. Because of dependencies between the choices of residential location and travel mode, integrated modeling of these choices has been proposed by some researchers. In this paper, an agent-based microsimulation model has been developed to evaluate impacts of different transport development plans on choices of residential location and commuting mode of tenant households in Tehran, the capital of Iran. In the proposed model, households are considered as agents who select their desired residential location using a constrained NSGA-II algorithm and in a competition with other households. In addition, they choose their commuting mode by applying a multi-criteria decision making method. Afterwards, effects of development of a new highway, subway, and bus rapid transit (BRT) line on their residential location and commuting mode choices are evaluated. Results show that despite the residential self-selection effects, these plans result in considerable changes in the commuting mode of different socio-economic categories of households. Development of the new subway line shows promising results by reducing the private car use among the all socioeconomic categories of households. But the new highway development unsatisfactorily results in increase in the private car use. In addition, development of the new BRT line does not show significant effects on the commuting mode change, particularly on decrease of the private car use. Keywords Transportation impacts; Modal shift; Commuting mode choice; Residential self-selection; Agent-based microsimulation model; NSGA-II algorithm; Multi-criteria decision making. 1 1. Introduction There is a strong and complex relationship between the choices of residential location and travel mode. Necessity of simultaneous modeling of these choices in an integrated framework has been realized by many researchers (e.g. Lerman 1976; Naess 2013; Nurlaela and Curtis 2012; Vega and Reynolds-Feighan 2009). Within this framework, the residential 'self-selection' with respect to travel behavior can be considered. This type of residential self-selection is referred to the tendency of people to choose a particular neighborhood according to their travel abilities, needs, and preferences (Mokhtarian and Cao 2008; Van Wee 2009). For example, people with strong preferences for use of the public transport services usually tend to live in regions with higher accessibilities to these services. On the other hand, those who prefer to use the private car for their daily travels usually tend to live in regions with minimum restrictions on private car use. However, when a car driving lover moves to a region with a good accessibility to public transport services, he/she will be exposed to advantages of using these services and may change his/her travel behavior. Land use and transportation plans can have considerable impacts on the built environment and subsequently on the travel behavior of people (Cao, Mokhtarian, and Handy 2007; Ewing and Cervero 2010; Guo and Chen 2007; Pinjari et al. 2007; Haddad et al. 2011; van de Walle 2009; Rand 2011). Some researchers have found that these impacts are subject to spurious effects of the residential self-selection (e.g. Cao, Mokhtarian, and Handy 2009; Handy, Cao, and Mokhtarian 2005; Mokhtarian and Cao 2008; Van Wee 2009). In other words, land use and transportation plans may not result in change of the travel behavior of people, but they may only result in movement of some people to neighborhoods of their desired transport services. As a result, ignoring the residential self-selection may lead to an overestimation of travel mode changes. For example, after development of a new subway in an area, models ignoring the residential self-selection may overestimate the contribution of subway in people's travel mode choice. Therefore, consideration of the residential self-selection is prerequisite for evaluating the impacts of land use and transportation plans on the travel behavior of people. Majority of researches in the fields of residential location and travel mode choices have used the discrete choice models (e.g. multinomial logit (MNL), nested logit (NL), generalized extra value (GEV), mixed logit, and probit) based on the random utility maximization theory (e.g. Nurlaela and Curtis 2012; Vega and Reynolds-Feighan 2009; Rashidi, Auld, and Mohammadian 2012; Sener, Pendyala, and Bhat 2011). These conventional models are mainly based on the aggregated characteristics of people (Crooks and Heppenstall 2012; Ettema 2011). However, researchers are usually interested to investigate the impacts of land use and transportation plans on the behavior of individual households, persons or subgroups. One of the efficient and most widely used tools in this area is microsimulation. However, microsimulation models (MSMs) have some limitations in modeling of individuals' behavior in terms of their preferences, 2 decisions, plans, and interactions. These limitations can be considerably mitigated by the integration of MSMs with individual-based models such as the agent-based models (ABMs) (Birkin and Wu 2012; Crooks and Heppenstall 2012). ABMs are suitable for the simulation of situations where there are a large number of heterogeneous individuals with different behaviors (Davidsson 2001). ABMs are bottom-up approaches that attempt to model the complex systems at the level of individuals presented as 'agents'. Agents are the fundamental elements of these models who have different characteristics and act and interact with each other based on their perceptions of the environment (Birkin and Wu 2012; Parker et al. 2003). There is no overall agreement on definition of the term 'agent', but most agents have several common features such as the autonomy, heterogeneity, and activity (Crooks and Heppenstall 2012; Wooldridge and Jennings 1995). Behaviors of agents and their relationships with other agents and the surrounding environment are generally modeled by a set of rules. These rules are typically derived from the existing literature, expert knowledge, data surveys or empirical works (Crooks and Heppenstall 2012). Increasingly, several researchers (e.g. Benenson 1998; Devisch et al. 2009; Ettema 2011; Haase, Lautenbach, and Seppelt 2010; Waddell et al. 2003; Gaube and Remesch 2013; Salvini and Miller 2005; Veldhuisen, Timmermans, and Kapoen 2000) have used agent-based and microsimulation models for studying the residential location and travel mode choices. The main purpose of this paper is evaluating the impacts of transport developments on the residential location and commuting mode choices of individual tenant households in Tehran, the capital of Iran. For this purpose, an agent-based microsimulation model is developed to simulate these choices. As far as the authors know, utilizing of ABMs for the evaluation of these impacts has not been considered by previous works. Therefore, the main contribution of this study is the exploration of relationships between transport development, residential self-selection, and modal shift behavior of individual households using an agent- based microsimulation model. Also, in this paper, a novel framework is proposed for choosing the residential location and commuting mode by households. In contrast with previous works, which mainly use the discrete choice models based on the random utility maximization theory, in this framework, an evolutionary multi-objective decision making algorithm (NSGA-II) combined with a competition method is used for residential location choice and a multi-criteria decision making method is used for commuting mode choice. By considering different objectives, criteria and preferences for each agent, the proposed choice mechanism leads to better consideration of the heterogeneity among agents. Due to the lack of efficient public transport services and cheap prices of fuel in Tehran, most people prefer to use the private car in their daily travels. Therefore, different land use and transportation plans are annually prepared in Tehran with the aim of reducing the dependency of people to private car use and encouraging them to use public transport services. For this purpose, urban and transportation planners 3 need a simulation tool to estimate the effectiveness of their plans for reducing the private car use. Although some researchers, such as Jokar Arsanjani, Helbich, and de Noronha Vaz (2013); and Meshkini and Rahimi (2011), attempt to simulate the spatial and/or temporal urban sprawl patterns in Tehran, they do not consider effects of future land use and transportation plans on the residential location and commuting mode choices. Therefore, the proposed microsimulation model in this paper is the first attempt to simulate these effects at the individual level in Tehran. This model considerably helps urban and transportation planners to simulate and predict residential self-selection and commuting mode changes resulting from the future implementation of the transport development plans (TDPs) in Tehran. The remainder of this paper is organized as follows. An overview of the available transportation networks and commuting modes in Tehran is provided in section 2. Section 3 describes the proposed model in details. Results of the implementation and validation of the proposed model are discussed in section 4. In section 5, three transportation projects of the future development plan of Tehran are simulated and resulting changes in the residential location and commuting mode of agents are examined. Finally, major implications of the research findings are presented in the concluding section. 2. Transportation and commuting modes in Tehran This study is conducted in Tehran, the capital of Iran. Tehran with an area of about 750 square kilometers and a population of about 8.3 million is located at 51 8(cid:1) to 51 37(cid:1) longitude and 35 34(cid:1) to 35 50(cid:1) latitude. This metropolis is comprised of extensive transportation networks including about 550 kilometers of highways, a subway network with 4 active intraurban lines and 72 stations with the length of about 125 kilometers, 250 bus lines with the length of about 3000 kilometers, and 6 bus rapid transit (BRT) lines with the length of about 102 kilometers (Tehran Municipality 2013). Tehran is divided into 560 traffic analysis zones (TAZs) which are used as the spatial units in this study. Figure 1 shows the existing transportation networks and TAZs in Tehran. Figure 1: Existing transportation networks and TAZs in Tehran 4 in Tehran, their capacities are not enough to accommodate this problems such as the traffic congestion and air pollution. Although there in urban population as well as the , Tehran has experienced a dramatic growth in urban population as well as Over the past three decades, Tehran coupled with the poor urban planning and inefficient public transport services coupled with the poor urban planning and inefficient public transport car ownership. This growth coupled with the poor urban planning and inefficient public transport have led to significant problems such as the traffic congestion and air pollution. Although there are extensive road networks in Tehran, not enough to accommodate this increasing s its study to the work number of vehicles especially during travel (commuting), because it is one of the important daily travel types, encompassing all transport travel (commuting), because it is one of the , encompassing all transport including the private modes, and results in peak-hours traffic car, subway, bus, BRT, and local taxi car, subway, bus, BRT, and local taxi for commuting of people in Tehran. Shares of these transport people are shown in Figure 2. As illustrated in this Figure, other modes modes in commuting of Tehran's people shown in Figure 2. As illustrated in this Figure, other modes shares in commuting of such as the walking, bicycle, call taxi a such as the walking, bicycle, call taxi and institutional services have minor shares in people. during the peak hours of traffic. This paper limits its study to traffic. There are five common transport modes including 21% 10.5% 9.5% 13.5% 34.5% 11% Private car Subway Bus BRT Taxi Others Tehran Municipality 2013) Figure 2: Shares of different transport modes in commuting of Tehran's people (Tehran Municipality 2013 Figure 2: Shares of different transport modes in commuting of Tehran's people 3. Proposed model of residential location and commuting mode of tenant households considered as agents who select their optimal residential alternatives from based microsimulation model is developed to evaluate the impacts of different developed to evaluate the impacts of different In this paper, an agent-based microsimulation model households in Tehran, TDPs on the joint choices of residential location and commuting mode of tenant . In the proposed model, households are generated using the Monte Carlo simulation. are generated using the Monte Carlo simulation. the capital of Iran. In the proposed model, from the available They are considered as agents who select their optimal residential alternatives constrained NSGA-II algorithm (nondominated sorting genetic algorithm II) (nondominated sorting genetic algorithm II). residential zones using a constrained NSGA is a fast and efficient multi-objective decision making algorithm in which multiple conflicting multiple conflicting NSGA-II is a fast and efficient multi demographic and objectives are considered for residential location choic to consideration of more socioeconomic characteristics and also socioeconomic characteristics and also their criteria and preferences. This leads to consideration of thereby more precise simulation of the residential location choice. heterogeneity among agents and thereby more residential location choice. Afterwards, agents compete with each other in different time periods to select a final residence from their Afterwards, agents compete with each other a final residence from their . After determination of the residence of agents, their employed members employed members are residential alternatives. After determination of the re agents who select the best available commuting mode based on their preferences to considered as the new agents who select the best available commuting mode based on their preferences to different criteria such as the travel time, travel cost, and different criteria such as the travel time, travel cost, and comfortableness using a multi a multi-criteria decision are simulated and resulting changes in the residential self- making method. Finally, different making method. Finally, different TDPs are simulated and resulting changes in the residential residential location choice of agents according to their 5 selection and commuting mode choice of agents are examined. Main components of the proposed model are shown in Figure 3. Population Synthesis Monte Carlo Simulation Agents Residential Alternatives Selection Residential Location Choice Multi-Objective Decision Making (NSGA-II) Appropriate Alternatives Competition Agent-Based Modeling Commuting Mode Choice Commuting Mode Multi Criteria Decision Making Residential Location Figure 3: Main components of the proposed model 3.1. Population synthesis A population of tenant households (agents) is generated using the Monte Carlo simulation such that aggregation of their attributes matches with the available demographic data. Attributes of agents are generated in a sequential approach as used by Miller et al. (2004). In this approach, initially the monthly income, number and age of the members of agents are generated. Then, the former residential zone, the number of employed members and their professional category, the number of owned cars, and the required residential area of agents are generated based on the previously determined attributes. Afterwards, the workplace of employed members of agents is randomly allocated to a zone by using the available rates of different employment categories in TAZs. Finally, criteria and preferences of agents and their employed members for choosing their residential location and commuting mode are generated according to the previously determined attributes and a survey sample consisting of stated preferences of 1580 tenant households in Tehran. This survey was conducted by filling the questionnaires in February 2012, where households with different socioeconomic characteristics stated their criteria and preferences for choosing their residential location and commuting mode in the scale range of 0 (unimportant) to 9 (very important). Table 1 presents the major criteria and the percentage of households in different socioeconomic categories who assigned a preference number greater than 4 to each criterion. 6 Table 1: A summary of the residential and commuting criteria and preferences of 1580 surveyed tenant households in Tehran Attribute Category Percentage Housing Rent to to Accessibility Accessibility Educational Locations Commercial Locations Residential location choice criteria (%) Commuting mode choice criteria (%) Accessibility to Green and Recreational Locations Accessibility to Cultural Locations Accessibility to Remedial Locations Accessibility to Highways Accessibility to Subway Stations Accessibility to Bus Stops Air and Noise Pollutions Distance from the Workplace Distance from the Former Residence Without Traffic Restrictions Commuting Cost In- vehicle time Out-of- vehicle time Comfortability Security Reliability Size Monthly income (million IRR)* Number of Cars Total Single Couple 3-4 > 4 < 10 10-25 > 25 0 1 > 1 5.6 28.1 54.8 11.5 17.6 63.1 19.3 10.8 68.8 20.4 100 100 100 100 100 100 100 100 100 100 100 100 29.5 8.6 66.9 92.9 45.0 59.9 29.5 66.7 56.1 27.3 51.4 11.4 38.3 35.6 31.3 17.6 36.6 43.0 45.0 34.8 28.0 34.5 9.1 41.7 59.9 63.2 36.0 52.3 67.5 51.5 55.5 42.2 52.3 9.1 6.3 9.4 5.5 5.0 8.3 9.8 10.5 7.8 7.5 8.0 5.7 22.3 35.8 44.5 20.5 34.1 32.1 45.6 32.7 19.3 31.3 70.5 82.7 83.4 82.4 67.6 84.5 88.9 22.8 86.5 100 82.3 * Note: IRR (Iranian Rial), At the time of this study 1$ equivalent to 31500 IRR. 39.8 48.6 55.3 59.3 57.9 53.8 46.2 83.6 54.6 31.4 53.0 12.5 18.0 21.0 21.4 32.7 18.7 11.5 62.6 17.6 4.3 19.7 34.1 46.6 43.2 43.4 23.4 42.2 66.9 30.4 40.7 60.9 43.7 92.0 89.9 85.5 82.4 93.9 86.7 80.3 95.3 86.3 83.5 86.7 37.5 50.7 59.8 59.3 65.5 54.7 51.5 64.9 54.1 57.5 55.9 60.2 61.3 55.5 54.9 43.2 54.9 78.4 5.3 53.4 98.4 57.3 100 99.1 99.3 100 100 100 96.7 100 99.6 98.1 99.4 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 96.6 93.0 90.4 90.7 80.2 92.1 100.0 78.9 91.1 99.7 91.5 25.0 27.7 26.2 26.9 19.1 26.2 35.1 18.1 26.1 32.9 26.6 56.8 46.8 45.2 49.5 37.4 46.9 54.8 45.0 45.5 51.9 46.8 3.2. Residential alternatives selection In the real world, households do not search all zones to select their residence, but they limit their search space to few desired zones according to their characteristics, requirements and preferred criteria. This fact is followed for the residential location choice of agents in the proposed model. Thus, using a constrained NSGA-II algorithm, the residential choice set of agents is limited to up to 10 optimal residential alternatives among the all TAZs. The NSGA-II algorithm, developed by Deb et al. (2002), is a fast and efficient multi-objective decision making method for the selection of the best solutions among the many available solutions. Details of this algorithm can be found in the study by Deb et al. (2002). According to the survey sample of stated preferences of tenant households in Tehran, the following rules are used as different objective functions in the NSGA-II algorithm. Depending on their criteria and preferences, agents may consider all or some of these rules in the process of their residential location choice. rule1: Agents select zones as their residential alternatives in which the housing rent per square meter is compatible with their affordability and required residential area. For this purpose, maximum and minimum percentages of the monthly income which agents are willing to pay for renting a residence are simulated by the Monte Carlo simulation approach as described in section 3.1. Therefore, the search space of agents is limited to zones in which rent of their required residential area is between the specified minimum and maximum percentages of their monthly income. It should be noted that the main reason for considering the minimum limit is that many agents, particularly those with high incomes, usually prefer to reside in neighborhoods with qualified cultural and social characteristics. rule2: Agents prefer to select zones in closer distances to the workplaces of their employed members. rule3: Agents prefer to select zones in closer distance from their former residential area. rule4: Agents prefer to select zones with minimum environmental pollutions. A constraint is considered for agents with very important preferences for air and noise pollutions, as they only search among zones in which pollution levels are medium to clean. 7 rule5: Agents prefer to select zones with no traffic restrictions on the private car use. A constraint is considered for agents with very important preferences for traffic restrictions, as they only search among the non-restricted zones. rule6: Agents prefer to select zones with higher accessibilities to public facilities including the commercial, educational, green, recreational, remedial, and cultural facilities. Accessibility of zones to the public facilities for each agent is calculated by Eq. (1) (Tsou, Hung, and Chang 2005). (cid:1)(cid:2)(cid:3) (cid:4) (cid:5)(cid:5)(cid:6)(cid:7)(cid:3) (cid:8) (cid:9)(cid:10)(cid:11) (cid:8) (cid:12)(cid:2)(cid:10)(cid:11) (cid:13)(cid:14) (cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:16)(cid:17)(cid:18)(cid:19)(cid:20)(cid:21) (cid:7) (cid:10)(cid:11) where: (cid:1)(cid:2)(cid:3) is the accessibility of zone i to the public facilities for agent a; f is the type of public facility; (cid:22)(cid:7) is the jth case of the fth type of public facility; (cid:6)(cid:7)(cid:3) is the preference of agent a to the type of public facility f which is simulated by the Monte Carlo simulation method, where (cid:23)(cid:6)(cid:7)(cid:3) (cid:4) (cid:20); (cid:9)(cid:10)(cid:11) is the relative effect of (cid:22)(cid:7) which is calculated by (cid:9)(cid:10)(cid:11) (cid:4) (cid:1)(cid:10)(cid:11)(cid:24)(cid:25)(cid:26)(cid:27)(cid:15)(cid:19)(cid:1)(cid:10)(cid:11)(cid:21), where (cid:1)(cid:10)(cid:11) is the area of (cid:22)(cid:7); (cid:12)(cid:2)(cid:10)(cid:11) is the distance between zone i and the public facility (cid:22)(cid:7). rule7: Agents prefer to select zones with higher accessibilities to transportation services including the highways and/or public transport services (bus and subway stations). Accessibility of zones to the transportation services for each agent is calculated by Eq. (2) (Currie 2010; Delbosc and Currie 2011). (cid:1)(cid:28)(cid:2)(cid:3) (cid:4) (cid:5)(cid:5)(cid:6)(cid:29)(cid:3) (cid:8) (cid:1)(cid:10)(cid:30) (cid:1)(cid:2) (cid:29) (cid:10)(cid:30) (cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:16)(cid:17)(cid:18)(cid:19)(cid:31)(cid:21) where: (cid:1)(cid:28)(cid:2)(cid:3) is the accessibility of zone i to transportation services for agent a; t is the type of transportation service; (cid:22)(cid:29) is the jth case of the tth type of transportation service; (cid:6)(cid:29)(cid:3) is the preference of agent a to the type of transportation service t which is simulated by the Monte Carlo simulation method, where (cid:23)(cid:6)(cid:29)(cid:3) (cid:4) (cid:20); Ai is the area of zone i; (cid:1)(cid:10)(cid:30) is the service area of (cid:22)(cid:29) which is inside the zone i; the service ranges of highways, subway stations, BRT, and bus stops are set to 2, 1.9, 1.7 and 1.2 km, respectively. These distances, obtained from the comprehensive transportation and traffic studies of Tehran, are slightly higher than those used in other 8 cities, because besides walking, Tehran's people usually use cheap local taxis for access to the transport services. 3.3. Competition Agents compete with each other in different time periods to choose a residence from their residential alternatives. For this purpose, they are randomly distributed in monthly time periods according to the volume of relocation in different months of the year in Tehran. In addition, the residential capacity of each zone is estimated by multiplying the ratio of residential area in that zone to all zones, the number of agents, and an equilibrium parameter in each period. The equilibrium parameter is derived from the volume of demand and supply for renting residences in different months of the year in Tehran. Agents search their residential alternatives according to the distance from their former residence and reside in the first alternative which has the residential capacity. If the residential capacity of a zone is not enough to satisfy demands of all agents, they compete with each other for residing in that zone. In this competition, agents with fewer members (except singles), with higher incomes, and without a child would succeed, respectively. If demandants are equal with respect to the mentioned criteria, winners are randomly selected and defeated agents try to reside in their next alternatives. This process is continued until all agents reside in one of their residential alternatives or all alternatives of the defeated agents are evaluated. If a number of agents cannot reside in any of their alternatives, they are moved to the competition process in the next time period. If they still cannot reside in any zone, no residence is allocated for them. Generally, they are agents with low incomes who cannot afford to rent a residence in any zone and they probably have to move to suburbs. 3.4. Commuting mode choice After determination of the residential location of agents, the employed members of agents are considered as the new agents who select the best available commuting mode based on their criteria and preferences using a multi-criteria decision making method. The most important criteria for the commuting mode choice are derived from the survey sample of stated preferences of commuting mode choice in Tehran. These criteria include the travel cost, out-of-vehicle time, in-vehicle time, comfortability, security, and reliability. It is requested from a number of transportation experts to evaluate the transport modes including the walking, private car, bus, BRT, subway, and taxi with respect to these criteria at the daily peak hours of traffic in Tehran. They relatively compare the transport modes with respect to each criterion and assign a score in the scale range of 1 to 10. The average of normalized scores of experts is used as the score of each transport mode with respect to each criterion. Agents have their own preferences with respect to these criteria. These preferences are normalized and used for the calculation of the suitability of each transport mode for each agent as follows. 9 !(cid:3) (cid:4) (cid:5)(cid:6)"(cid:3) (cid:8) &!" (cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:16)(cid:17)(cid:18)(cid:19)'(cid:21) # "$% where: !(cid:3)(cid:15)is the suitability of the transport mode m for agent a; (cid:6)"(cid:3) is the preference of agent a to the criterion c, where (cid:23)(cid:6)"(cid:3) (cid:4) (cid:20); &!" is the score of transport mode m with respect to the criterion c. Finally, agents choose the transport mode with the maximum suitability for their commuting. It should be noted that the level-of-service (LOS) attributes including the travel cost, in-vehicle and out-of-vehicle times for different transport modes are obtained from the shortest path between the residence and workplace of agents on the network models in GIS. 4. Implementation and validation The proposed model was implemented by the MATLAB software version 7.12. Using the Monte Carlo simulation, 50000 tenant households with different socio-economic characteristics were generated as agents. They searched among TAZs and selected their optimal residential alternatives using the NSGA-II algorithm. Then, they chose a final residence from their alternatives by the aforementioned competition mechanism. Figure 4 shows the distribution of residential locations of agents in TAZs. Finally, 58941 employed members of the simulated households selected their commuting mode according to Equation (3). Selected modes by different socioeconomic categories of agents are presented in Table 2. Figure 4: Distribution of residential zones of the simulated agents in Tehran 10 Table 2: Commuting modes of different socioeconomic categories of the simulated agents in Tehran Attribute Category Number Percentage Private Car Gender Household Size Average Monthly Income of Household (million IRR) Number of Private Cars of Household Commuting Distance (km) Total Female Male Single Couple 3-4 > 4 < 10 10-25 > 25 0 1 > 1 < 5 5 - 15 > 15 18104 40837 2888 17394 31636 7023 10597 37331 11013 6532 40050 12359 9607 37782 11552 58941 30.7 69.3 4.9 29.5 53.7 11.9 18.0 63.3 18.7 11.1 67.9 21.0 16.3 64.1 19.6 100 39.1 38.5 47.5 40.1 37.9 35.1 29.2 37.7 51.2 0.0 34.8 71.6 21.7 36.7 59.4 38.7 Commuting mode (%) Subway 12.9 12.2 10.1 11.8 12.7 13.6 17.8 11.9 9.0 25.7 12.2 6.1 11.7 10.8 18.2 12.4 Bus 13.7 14.1 9.8 13.0 14.5 15.9 19.1 14.7 6.5 23.4 15.7 3.5 15.7 17.1 2.3 14.0 BRT 10.6 10.9 7.8 10.0 11.3 11.8 15.7 10.8 6.1 20.6 11.7 2.9 12.0 12.7 3.5 10.8 Taxi 22.5 22.4 23.7 23.6 21.8 22.0 16.8 23.1 25.6 28.5 23.8 14.7 28.6 22.6 16.6 22.4 Walking 1.2 1.9 1.1 1.5 1.9 1.6 1.4 1.8 1.6 1.8 1.8 1.2 10.3 0.0 0.0 1.7 One of the greatest challenges of utilizing the ABMs is their validation (Crooks and Heppenstall 2012). In this research, a survey sample of 1485 actual tenant households was used for validation of the proposed model. These data were not used in the population synthesis of agents. Residence and commuting mode of these households were simulated by the proposed model and results were compared with their actual residence and commuting mode. Validation results for different socioeconomic categories of sample households are presented in Table 3. As shown in this table, the proposed model correctly simulates the residential zone of 62.6% of households and the commuting mode of 69.6% of their employed members. Also, the model properly simulates the commuting mode of 78.0% of females and 76.8% of all households who use the private car for their commuting. Therefore, the proposed model seems to show a good and promising performance. Table 3: Validation results of the proposed model Attribute Category Number Percentage Gender Household Size Average Monthly Income of Household (million IRR) Number of Private Cars of Household Commuting Distance (km) Total Female Male Single Couple 3-4 > 4 < 10 10-25 > 25 0 1 > 1 < 5 5 - 15 > 15 391 1094 58 475 820 132 247 937 301 147 1059 279 205 968 312 1485 26.3 73.7 3.9 32.0 55.2 8.9 16.6 63.1 20.3 9.9 71.3 18.8 13.8 65.2 21.0 100 Correctly Simulated Residential Zone (%) 59.8 63.5 70.7 64.2 62.2 55.3 64.4 60.3 68.1 62.6 62.7 62.0 64.9 62.5 61.2 62.6 Correctly Simulated Residential Neighborhood (%) 78.8 82.4 87.9 83.2 80.5 78.8 86.2 80.7 80.1 78.9 82.7 78.1 86.8 81.5 77.9 81.5 Correctly Simulated Commuting Mode (%) Subway Bus BRT Taxi Walking 67.3 68.3 60.0 65.0 69.8 70.0 68.0 68.6 65.2 76.7 66.7 56.3 72.2 55.2 86.8 68.1 61.7 62.9 50.0 61.5 64.2 58.3 62.7 62.8 60.0 70.0 61.8 40.0 71.4 61.4 42.9 62.6 61.1 63.5 50.0 60.4 65.2 62.5 63.9 62.6 63.2 62.5 63.2 60.0 74.1 61.4 50.0 63.0 65.6 66.0 66.7 70.9 63.8 62.5 63.4 67.2 63.0 67.6 65.9 64.4 78.1 65.6 47.4 65.9 75.0 66.7 0.0 62.5 75.0 66.7 50.0 70.0 75.0 60.0 68.2 100 67.9 100 100 67.9 Private Car 78.0 76.3 79.3 77.4 76.6 73.5 70.8 75.3 82.0 100 72.0 85.0 57.6 74.4 84.3 76.8 * Note: A neighborhood with Euclidean distance of 5 kilometers 11 5. Impact assessment of the TDPs In this section, impacts of three TDPs including construction of a new highway, BRT, and subway on the residential self-selection and commuting mode choice of tenant households are evaluated. These TDPs are the major under construction components of the future development plan of Tehran and will be operational in the near future (Figure 5). In this research, these TDPs are separately simulated and their effects on the accessibility and housing rent are estimated. Then, the simulated agents reselect their residence and commuting mode under the new conditions. Since all other conditions are assumed to be constant and only the transportation system is changed, the observed changes in the residential location and commuting mode of agents can be attributed to the impacts of TDPs. Figure 5: Location of the new transport development plans (TDPs) in Tehran TDPs influence many components of a city including the accessibility, travel behavior, traffic, environmental pollutions, land uses, residential location, and housing price in the long- and/or short-time periods. However, in this research, only the impacts of TDPs on the accessibility, travel behavior, housing rent, and residential location choice are investigated. In other words, it is assumed that these TDPs do not have any significant impacts on other determinants of the residential location and commuting mode choices. TDPs significantly improve the accessibility of their neighborhoods to the transport services. The accessibility of TAZs to the transport services is updated using Eq. (2). As a result, due to improvement of the accessibility, the housing rent is also changed in these neighborhoods. Consultations with real- estate agencies besides the experiences obtained from the construction of previous TDPs in Tehran show that the highway and subway developments can have significant impacts on the housing rent, but the BRT developments have not considerable effects on the housing rent. Therefore, the housing rents within the neighborhoods of the highway and subway developments are updated using the following equation: 12 ((cid:2)# (cid:4) )(cid:1)(cid:2) * (cid:23)(cid:1)(cid:2)++,-((cid:2) . (cid:23)(cid:19)(cid:1)(cid:2)++, /++,((cid:2)(cid:21) (cid:1)(cid:2) (cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:15)(cid:16)(cid:17)(cid:18)(cid:19)0(cid:21)(cid:15) where: ((cid:2)# and Ri are the new and previous housing rents per square meter in zone i, respectively; r and 12 are the beginning and ending radiuses of a ring buffer around the new highway or subway; Ai and (cid:1)(cid:2)++, are the areas of zone i and the ring buffer with radiuses r and 12, respectively; /++, is the change rate of housing rent in the buffer area. The values of this parameter are empirically estimated by consultation with the real-estate agencies and by experiences obtained from the construction of previous TDPs in Tehran. For studying how the residential location and commuting mode of agents are changed by TDPs, the proposed model was repeatedly run and changes of the residential location and commuting mode of agents were observed. A summary of these changes is presented in Table 4. Generally, considerable changes in the commuting mode of agents are observed in neighborhoods of TDPs. However, a part of these changes is due to residential self-selection of some agents with respect to their commuting preferences. In other words, some agents with high interests to commute by the private car, subway or BRT move to vicinities of the new highway, subway or BRT, respectively. Therefore, increase in the ridership of subway or BRT in neighborhoods of the new developments does not necessarily indicate the effectiveness of these plans. But what is important is that which and how many agents change their commuting mode to the developed modes. Results of these investigations are shown in Table 4. Interesting results are obtained from simulation of the residential location and commuting mode choices of agents after implementation of the new highway (TDP 1). Private car use significantly increases in the neighboring zones of the new highway (Figure 6). This increase is greater in the northern neighboring zones and indicates that residents of these zones are more interested to use the private car, because the mean car ownership in these zones is more than those of the other neighboring zones. However, private car use also shows a decreasing trend in some zones, because a number of agents with high interests to private car use move from these zones to the neighborhoods of the new highway (residential self- selection effects). Although increase in the private car use in the neighboring zones of the new highway is partly due to the residential self-selection, some agents also change their commuting mode to the private car. As presented in Table 4, development of the new highway generally results in 1.1% increase in the private car use. On the other hand, the greatest decreases of the private car use are observed in the shares of subway, bus and BRT, respectively. Figure 7 shows increase in the private car use by different socioeconomic categories of agents. As can be seen, after implementation of the new highway, agents with fewer members, higher incomes, more owned cars, and longer commuting distances show more 13 interests to change their commuting mode to the private car. As a result, development of the new highway encourages people to use their private cars and may result in increase in the traffic congestion and environmental pollutions in the long term. This completely conflicts with the main goal of urban and transportation planners in Tehran which is decreasing the private car use. Figure 6: Changes of the private car use in TAZs after implementation of the new highway e t a v i r p e h t n i e s a e r c n I ) % ( e s u r a c (a) e t a v i r p e h t n i e s a e r c n I ) % ( e s u r a c 1.5 1.2 0.9 0.6 0.3 0 1.8 1.5 1.2 0.9 0.6 0.3 0 e t a v i r p e h t n i e s a e r c n I ) % ( e s u r a c Single Couple 3-4 > 4 Household size (b) e t a v i r p e h t n i e s a e r c n I ) % ( e s u r a c 1.5 1.2 0.9 0.6 0.3 0 1.8 1.5 1.2 0.9 0.6 0.3 0 < 10 10-25 > 25 Monthly income (c) 1 0 > 1 Number of owned cars (d) < 5 5-15 > 15 Commuting distance Figure 7: Increase in the private car use by agents with different household sizes (a); monthly incomes (b); number of owned cars (c); and commuting distances (d) after implementation of the new highway 14 By development of the new subway line (TDP 2), accessibility of many deprived areas of the southwest of Tehran to the subway network is significantly improved. This plan results in substantial changes in the commuting mode of residents in the southwest zones of Tehran. As illustrated in Figure 8, share of the subway in commuting of people increases up to 17.5% in some southwest zones. Also, due to interconnection of the subway lines, use of the subway shows slight increases in the neighborhoods of the previous stations. However, because of the residential self-selection effects, use of the subway decreases in some zones which are mainly located in north of the new line and have poor accessibilities to the subway network. In other words, a number of agents with high interests to use the subway mode move from these zones to neighborhoods of the new subway stations. In addition to residential self-selection which leads to increase in subway use in the neighboring zones of the new subway line, some agents change their commuting mode to the subway. As shown in Table 4, implementation of the new subway line results in 2.8% increase in the share of subway and 1.07% decrease in the private car use. Figure 9 shows increase in the subway use by different socioeconomic categories of agents. Although after implementation of the new subway line, all categories of agents show considerable willingness to change their commuting mode to the subway, agents with more members, lower incomes, fewer owned cars, and longer commuting distances show higher interests to change their commuting mode to the subway. Results show that development of the new subway line is significantly effective in decrease in the private car use even among the agents with high incomes and more than one owned car. Figure 8: Changes of the subway use in TAZs after implementation of the new subway line 15 y a w b u s e h t n i e s a e r c n I ) % ( e s u (a) y a w b u s e h t n i e s a e r c n I ) % ( e s u 3 2.5 2 1.5 1 0.5 0 5 4 3 2 1 0 Single Couple 3-4 > 4 Household size y a w b u s e h t n i e s a e r c n I ) % ( e s u (b) y a w b u s e h t n i e s a e r c n I ) % ( e s u (c) 0 1 > 1 Number of owned cars (d) 3.5 3 2.5 2 1.5 1 0.5 0 5 4 3 2 1 0 10-25 < 10 Monthly income > 25 < 5 5-15 > 15 Commuting distance Figure 9: Increase in the subway use by agents with different household sizes (a); monthly incomes (b); number of owned cars (c); and commuting distances (d) after implementation of the new subway line Figure 10 shows changes of the BRT use in TAZs after implementation of the new BRT line (TDP 3). These changes are greater in the southern neighboring zones of the new line. This shows that residents of these zones have more willingness to use BRT for their commuting. The main reason for this may be that the mean income and car ownership in these zones are less than those of the other neighboring zones and generally agents with low incomes and fewer owned cars show higher interests to use BRT. In contrast to increase in BRT use in the neighboring zones of the new line, its use decreases in a significant number of zones. In other words, some agents with high interests to use BRT for their commuting move from these zones to neighborhoods of the new line. Therefore, share of the BRT decreases in these zones and increases in neighborhoods of the new line. This indicates that a major part of impacts of the new BRT line on the commuting mode changes is due to residential self-selection. As a result, total share of the BRT mode does not show a significant increase. As presented in Table 4, after implementation of the new BRT line, only 0.36% of agents change their commuting mode to the BRT. Also, share of the private car does not show a meaningful decrease. As illustrated in Figure 11, agents with more members, lower incomes, fewer owned cars, and shorter commuting distances show higher interests to change their commuting mode to BRT. However, agents with high incomes and more than one owned car do not show considerable interests to change their commuting mode to BRT. As a result, the new BRT line does not show significant efficiency to encourage agents especially those with high incomes and a high number of owned cars to change their commuting mode form the private car to the BRT mode. 16 Figure 10: Changes of the BRT use in TAZs after implementation of the new BRT line 0.4 0.3 0.2 0.1 0 0.6 0.5 0.4 0.3 0.2 0.1 0 e s u T R B e h t n ) % ( i e s a e r c n I (a) e s u T R B e h t n i e s a e r c n I ) % ( (c) e s u T R B e h t n ) % ( Single Couple 3-4 > 4 Household size i e s a e r c n I (b) e s u T R B e h t n i e s a e r c n I ) % ( (d) 1 > 1 0 Number of owned cars 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.6 0.5 0.4 0.3 0.2 0.1 0 < 10 10-25 > 25 Monthly income < 5 5-15 > 15 Commuting distance Figure 11: Increase in the BRT use by agents with different household sizes (a); monthly incomes (b); number of owned cars (c); and commuting distances (d) after implementation of the new BRT line 17 Table 4: Changes in the residential location and commuting mode of agents after implementation of the three transport development plans in Tehran Changes in the Residential Zone and Share of the Commuting Modes (%) Development of the new highway (TDP 1) Private Subway Bus BRT Taxi Walking Residential Development of the new subway line (TDP 2) Private Subway Bus BRT Taxi Walking Residential Development of the new BRT line (TDP 3) Private Subway Bus BRT Taxi Walking Attribute Category Number Percentage Monthly Income Gender Average Female 18104 40837 Male Household Size Single 2888 17394 Couple 31636 7023 10597 37331 11013 6532 40050 12359 9607 37782 11552 58941 3-4 > 4 < 10 10-25 > 25 0 1 > 1 < 5 5 - 15 > 15 Commuting Distance (km) of Household (million IRR) Private Cars of Number of Household Total Residential Zone 3.92 3.93 4.47 4.15 3.80 3.73 3.58 4.28 3.10 1.97 4.64 2.65 3.45 4.00 4.11 3.93 30.7 69.3 4.9 29.5 53.7 11.9 18.0 63.3 18.7 11.1 67.9 21.0 16.3 64.1 19.6 100 Car +1.08 +1.11 +1.25 +1.14 +1.08 +1.03 +0.92 +1.09 +1.30 0.00 +1.09 +1.72 +0.59 +1.08 +1.58 +1.10 -0.39 -0.41 -0.48 -0.42 -0.40 -0.34 -0.33 -0.44 -0.35 -0.06 -0.37 -0.71 -0.07 -0.29 -1.06 -0.41 -0.30 -0.31 -0.35 -0.33 -0.28 -0.33 -0.25 -0.28 -0.44 0.00 -0.30 -0.49 -0.20 -0.40 -0.09 -0.31 -0.10 -0.28 -0.09 -0.29 -0.10 -0.31 -0.10 -0.28 -0.09 -0.30 -0.10 -0.24 -0.14 -0.20 -0.07 -0.28 -0.37 -0.13 0.00 +0.06 -0.12 -0.28 -0.45 -0.08 -0.09 -0.17 -0.03 -0.36 -0.30 -0.14 -0.29 -0.10 -0.01 -0.01 0.00 -0.01 -0.01 -0.01 -0.01 -0.01 -0.01 0.00 -0.01 0.00 -0.06 0.00 0.00 -0.01 Zone 4.08 4.08 4.40 4.04 4.11 3.92 4.63 4.23 3.04 6.87 4.38 1.63 2.07 4.28 5.10 4.08 Car -1.01 -1.10 -0.76 -0.98 -1.11 -1.27 -1.43 -1.01 -0.93 0.00 -1.14 -1.41 -0.09 -0.65 -3.28 -1.07 +2.73 +2.84 +2.35 +2.66 +2.89 +2.96 +3.17 +2.84 +2.32 +4.55 +2.76 +2.03 +0.66 +2.68 +4.99 +2.80 -0.49 -0.45 -0.42 -0.42 -0.48 -0.53 -0.43 -0.50 -0.37 -1.26 -0.46 -0.06 -0.08 -0.69 -0.04 -0.46 -0.56 -0.58 -0.45 -0.53 -0.63 -0.50 -0.45 -0.66 -0.39 -1.35 -0.61 -0.05 -0.07 -0.85 -0.07 -0.57 -0.67 -0.71 -0.73 -0.74 -0.67 -0.67 -0.85 -0.67 -0.64 -1.94 -0.54 -0.52 -0.40 -0.49 -1.60 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -0.01 0.00 0.00 0.00 Zone 0.65 0.69 0.76 0.64 0.69 0.68 0.82 0.79 0.16 1.91 0.61 0.27 0.44 0.84 0.36 0.68 Car -0.07 -0.08 -0.03 -0.08 -0.07 -0.10 -0.14 -0.07 -0.02 0.00 -0.10 -0.02 -0.10 -0.08 -0.02 -0.07 -0.08 -0.07 -0.07 -0.07 -0.08 -0.07 -0.13 -0.08 -0.03 -0.18 -0.08 -0.02 -0.08 -0.09 -0.02 -0.08 -0.04 -0.03 -0.03 -0.02 -0.04 -0.04 -0.08 -0.02 -0.03 -0.09 -0.02 -0.03 -0.04 -0.03 -0.03 -0.03 +0.37 +0.36 +0.31 +0.34 +0.37 +0.38 +0.62 +0.35 +0.15 +0.58 +0.40 +0.12 +0.52 +0.39 +0.15 +0.36 -0.17 -0.17 -0.17 -0.16 -0.18 -0.16 -0.25 -0.17 -0.06 -0.29 -0.19 -0.04 -0.23 -0.18 -0.08 -0.17 -0.01 -0.01 0.00 -0.01 -0.01 -0.01 -0.01 -0.01 -0.02 -0.02 -0.01 -0.01 -0.06 0.00 0.00 -0.01 6. Conclusion In this paper, the residential location choice of tenant households and the commuting mode choice of their employed members were modeled using an agent-based microsimulation model. Then, impacts of the three TDPs including development of the new highway, subway and BRT line on these choices were evaluated by repeatedly running the model. Although microsimulation models have some limitations such as data requirements and computational complexities, recent advances in computational capacities and integration with new technologies such as ABMs have reduced these limitations. Agent-based microsimulation models enable us to study the residential location choice and travel behavior of individuals. Also, it provides a robust and flexible framework for implementation of different rules, assumptions, interactions, and levels of description and aggregation. This framework allows us to better study the behavioral and causal relationships in the urban system. Results of the present study indicate that TDPs lead to considerable effects on the commuting mode choice of households. A major part of these effects is due to the residential self-selection of households with respect to their travel preferences. After implementation of each TDP, some households who prefer to commute by the developed mode attempt to move to its neighborhood. Therefore, use of the developed mode is increased in neighborhoods of the development. As a result, it should be noted that the increase in use of the developed mode in its neighboring areas does not necessarily indicate its efficiency, because a significant part of this increase is due to the residential self-selection. In addition to residential self- selection effects, these plans lead to considerable changes in commuting mode of households. This is very important for evaluating the efficiency of different transportation plans. Results show that the new highway leads to increase in the private car use in commuting of households. In other words, after implementation of the new highway, private car use becomes more attractive for some households who 18 previously used the public transport services. This is exactly the opposite of what urban and transportation planners desire and may result in more traffic, air and noise pollutions in the long term. But, results of the new subway line development are very promising, because it leads to decrease in the private car use. Some households from all socioeconomic categories, even those with high incomes and a high number of owned cars, change their commuting mode from the private car to the subway. Furthermore, development of the new BRT line results in change of the commuting mode of some households. However, these changes are more observable among households with low incomes and without the private car. In other words, households with high incomes and a high number of owned cars do not show interest to change their commuting mode from the private car to BRT. Therefore, it seems that the new subway line is more popular than the new BRT line, and it is more efficient to decrease the private cars use. It should be emphasized that results of this research are exposed to assumptions, rules and conditions of the simulation and can be regarded as the scenario-based results. Various aspects of the proposed model can be improved in the future. First, the rules and conditions used in this study to model the behavior of agents and their relationships with other agents and the surrounding environment can be enhanced to improve the results. Second, use of the TAZs as the spatial units leads to modifiable areal unit problem (MAUP) and errors in measurement of the network-based attributes of the transport services. This means that changes in size or configuration of zones can affect the results. In order to reduce these effects and also to enhance the efficiency of the proposed model in the large-scale planning, smaller spatial units such as the census blocks, cells, and parcels can be used. In addition, effects of varying sizes and configurations of spatial units on results can be examined. Third, the proposed model can be extended to evaluate impacts of TDPs on the residential location and commuting mode choices of not only tenant households but also all households. Fourth, the proposed model can be used to evaluate impacts of different transport policies (e.g. changes in the fuel price, changes in the parking space and cost, and changes in the existing transportation networks) on residential location and commuting mode choices of households. Fifth, although commuting mode choice has been modeled in a typical day of the year in this paper, the model can be developed to consider time-related criteria including season, weather conditions, and school holidays in selection of an appropriate commuting mode. Finally, a price bidding framework and a housing supply module can be added to the proposed model in the future. 19 References Benenson, I. 1998. "Multi-agent simulations of residential dynamics in the city." Computers, Environment and Urban Systems 22:25-42. Birkin, M., and B. Wu. 2012. "A Review of Microsimulation and Hybrid Agent-Based Approaches." In Agent- Based Models of Geographical Systems, edited by A.J. Heppenstall, A.T. Crooks, L.M. See and M. Batty, 51-68. Dordrecht: Springer. Cao, X., P.L. Mokhtarian, and S. Handy. 2009. "Examining the impacts of residential self-selection on travel behavior: a focus on empirical findings." Transport Reviews 29:359-95. Cao, X., P.L. Mokhtarian, and S.L. Handy. 2007. "Do changes in neighborhood characteristics lead to changes in travel behavior? A structural equations modeling approach." Transportation 34:535-56. Crooks, A.T., and A.J. Heppenstall. 2012. "Introduction to Agent-Based Modelling." In Agent-Based Models of Geographical Systems, edited by A.J. Heppenstall, A.T. Crooks, L.M. See and M. Batty, 85-105. Dordrecht: Springer. Currie, G. 2010. "Quantifying spatial gaps in public transport supply based on social needs." Journal of Transport Geography 18:31-41. Davidsson, P. 2001. "Multi agent based simulation: Beyond social simulation." In Multi agent based simulation, edited by S. Moss and P. Davidsson. Heidelberg: Springer. Deb, K., A. Pratap, S. Agarwal, and T. Meyarivan. 2002. "A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II." IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION 6 (2):182-97. Delbosc, A., and G. Currie. 2011. "Using Lorenz curves to assess public transport equity." Journal of Transport Geography 19:1252-9. Devisch, O.T.J., H.J.P. Timmermans, T.A. Arentze, and A.W.J. Borgers. 2009. "An agent-based model of residential choice dynamics in nonstationary housing markets." Environment and Planning A 41(8):1997- 2013. Ettema, D. 2011. "A multi-agent model of urban processes: Modelling relocation processes and price setting in housing markets." Computers, Environment and Urban Systems 35:1-11. Ewing, R., and R. Cervero. 2010. "Travel and the built environment:a meta-analysis." Journal of the American Planning Association 76(3):265-294. Gaube, V., and A. Remesch. 2013. "Impact of urban planning on household's residential decisions: An agent-based simulation model for Vienna." Environmental Modelling & Software 45:92-103. Guo, J.Y., and C. Chen. 2007. "The built environment and travel behavior: making the connection." Transportation 34:529-33. Haase, D., S. Lautenbach, and R. Seppelt. 2010. "Modeling and simulating residential mobility in a shrinking city using an agent-based approach." Environmental Modelling & Software 25:1225-40. Haddad, E.A., F.S. Perobelli, E.P. Domingues, and M. Aguiar. 2011. "Assessing the ex ante economic impacts of transportation infrastructure policies in Brazil." Journal of Development Effectiveness 3 (1):44-61. 20 Handy, S., X. Cao, and P. Mokhtarian. 2005. "Correlation or causality between the built environment and travel behavior? Evidence from Northern California." Transportation Research Part D 10:427-44. Jokar Arsanjani, J., M. Helbich, and E. de Noronha Vaz. 2013. "Spatiotemporal simulation of urban growth patterns using agent-based modeling: The case of Tehran." Cities 32:33-42. Lerman, S.R. 1976. "Location, housing, automobile ownership, and mode to work: a joint choice model." Transportation Research Record 610:6-11. Meshkini, A., and H. Rahimi. 2011. "Changes in population settlement pattern in urban system of Tehran province (1966 to 2006)." Journal of Geography and Regional Planning 4(7):371-82. Miller, E.J., J.D. Hunt, J.E. Abraham, and P.A. Salvini. 2004. "Microsimulating urban systems." Computers, Environment and Urban Systems 28:9-44. Mokhtarian, P.L., and X. Cao. 2008. "Examining the impacts of residential self-selection on travel behavior: a focus on methodologies." Transportation Research Part B 42:204-28. Naess, P. 2013. "Residential location, transport rationales and daily-life travel behaviour: The case of Hangzhou Metropolitan Area, China." Progress in Planning 79:1-50. Nurlaela, S., and C. Curtis. 2012. "Modeling household residential location choice and travel behavior and its relationship with public transport accessibility." Procedia - Social and Behavioral Sciences 54:56-64. Parker, D.C., S.M. Manson, M.A. Janssen, M.J. Hoffmann, and P. Deadman. 2003. "Multi-agent systems for the simulation of land-use and land-cover change: A review." Annals of the Association of American Geographers 93 (2):314-37. Pinjari, A., R. Pendyala, C. Bhat, and P. Waddell. 2007. "Modeling residential sorting effects to understand the impact of the built environment on commute mode choice." Transportation 34:557-73. Rand, J. 2011. "Evaluating the employment-generating impact of rural roads in Nicaragua." Journal of Development Effectiveness 3(1):28-43. Rashidi, T.H., J. Auld, and A. Mohammadian. 2012. "A behavioral housing search model: Two-stage hazard-based and multinomial logit approach to choice-set formation and location selection." Transportation Research Part A 46:1097–107. Salvini, P.A., and E.J. Miller. 2005. "ILUTE: An Operational Prototype of a Comprehensive Microsimulation Model of Urban Systems." Networks and Spatial Economics 5:217-34. Sener, I.N. , R.M. Pendyala, and C.R. Bhat. 2011. "Accommodating spatial correlation across choice alternatives in discrete choice models: an application to modeling residential location choice behavior." Journal of Transport Geography 19:294-303. Tehran Municipality. 2013. Tehran statistical yearbook 2012-2013. Tehran, Iran: ICT Organization of Tehran Municipality. Tsou, K.-W., Y.-T. Hung, and Y.-L. Chang. 2005. "An accessibility-based integrated measure of relative spatial equity in urban public facilities." Cities 22 (6):424-35. van de Walle, D. 2009. "Impact evaluation of rural road projects." Journal of Development Effectiveness 1(1):15-36. 21 Van Wee, B. 2009. "Self-Selection: A Key to a Better Understanding of Location Choices, Travel Behaviour and Transport Externalities?" Transport Reviews 29(3):279-92. Vega, Amaya, and Aisling Reynolds-Feighan. 2009. "A methodological framework for the study of residential location and travel-to-work mode choice under central and suburban employment destination patterns." Transportation Research Part A: Policy and Practice 43 (4):401-19. Veldhuisen, K.J., H.J.P. Timmermans, and L.L. Kapoen. 2000. "RAMBLAS: a regional planning model based on the microsimulation of daily activity travel patterns." Environment and Planning A 32:427-43. Waddell, P., A. Borning, M. Noth, N. Freier, M. Becke, and G.F. Ulfarsson. 2003. "Microsimulation of urban development and location choices: Design and implementation of UrbanSim." Networks and Spatial Economics 3(1):43-67. Wooldridge, M., and N.R. Jennings. 1995. "Intelligent agents: Theory and practice." The Knowledge Engineering Review 10 (2):115-52. 22
1711.10283	1	1711	2017-11-28T13:30:45	Data Backup Network Formation with Heterogeneous Agents	[ "cs.MA" ]	Social storage systems are becoming increasingly popular compared to the existing data backup systems like local, centralized and P2P systems. An endogenously built symmetric social storage model and its aspects like the utility of each agent, bilateral stability, contentment, and efficiency have been extensively discussed in Mane et. al. (2017). We include heterogeneity in this model by using the concept of Social Range Matrix from Kuznetsov et. al (2010). Now, each agent is concerned about its perceived utility, which is a linear combination of its utility as well as others utilities (depending upon whether the pair are friends, enemies or do not care about each other). We derive conditions when two agents may want to add or delete a link, and provide an algorithm that checks if a bilaterally stable network is possible or not. Finally, we take some special Social Range Matrices and prove that under certain conditions on network parameters, a bilaterally stable network is unique.	cs.MA	cs	Data Backup Network Formation with Heterogeneous Agents (Extended Abstract) Harshit Jaina, Guduru Sai Tejaa, Pramod Manea, Kapil Ahujaa and Nagarajan Krishnamurthyb 7 1 0 2 v o N 8 2 ] A M . s c [ 1 v 3 8 2 0 1 . 1 1 7 1 : v i X r a Abstract -- Social storage systems [1], [2], [3] are becoming increasingly popular compared to the existing data backup systems like local, centralized and P2P systems. An endogenously built symmetric social storage model and its aspects like the utility of each agent, bilateral stability, contentment, and efficiency have been extensively discussed in [4]. We include heterogeneity in this model by using the concept of Social Range Matrix from [5]. Now, each agent is concerned about its perceived utility, which is a linear combination of its utility as well as others utilities (depending upon whether the pair are friends, enemies or do not care about each other). We derive conditions when two agents may want to add or delete a link, and provide an algorithm that checks if a bilaterally stable network is possible or not. Finally, we take some special Social Range Matrices and prove that under certain conditions on network parameters, a bilaterally stable network is unique. Index Terms -- Social Storage, F2F Backup System, Endoge- nous and Heterogeneous Network Formation, Bilateral Stability. I. INTRODUCTION AND MOTIVATION In recent years, social storage (friend-to-friend) is emerging as an alternative to local (e.g., hard disc), online (e.g., cloud), and Peer-to-Peer data backup systems. Social storage allows users to store their data on their social relatives' (e.g., friends, colleagues, family, etc.) storage devices. Researchers believe that social relationships (which are at the core of social storage) help to mitigate issues like data availability, reliability, and security. Initial work in this field has primarily focused on developing techniques to exogenously build social storage systems, and performing Quality of Services (QoS) analysis in terms of data reliability and availability in these systems. A most recent study in [4] focuses on explicit data backup partner selection, where agents themselves select their partners. This selection is studied in a strategic setting, and eventually builds a social storage network. They model a utility function, which reveals the benefit an individual receives in a social storage network. Further, they analyze the network by using bilateral stability as a solution concept, where no pair of agents add or delete a link without their mutual consent. There are several advantages of this approach. First, this approach makes it possible to incorporate user's strategic behavior (in terms of with whom user wants to form social connections and with whom it does not). Second, this approach aAuthors are with the IIT Indore. Corresponding author, Kapil Ahuja's, email: {kahuja}@iiti.ac.in bAuthor is with the IIM Indore. helps us to predict the following: which network is likely to emerge; which one is stable (in which no individual has incentives to alter the structure of the network either by forming new or deleting existing social connections); and which one is the best (contended and/ or efficient) from all the participants point of view. However, the above strategic setting does not consider het- erogeneous behavior (e.g., selfless, selfish, etc.) in the network formation. In this paper, we focus on this aspect. This way, we make the model close to a real-world scenario. Although, doing so makes it challenging to deal with the model and as well as predict its outcome. To achieve the above, we incorporate the concept of Social Range Matrix introduced in [5] while exploring different social relationships between agents. We modify the utility of agents in the social storage model discussed in [4]. Further, we revisit the results regarding bilateral stability of such a model. II. MODEL Our model is composed of four components. First is sym- metric social storage. Second is the utility (as derived in [4]) of each agent. Third is Social Range Matrix, which captures the social relationship between these agents. Fourth is the perceived utility obtained by combining the second and the third components. A symmetric social network g is a data backup network consisting of N number of agents and a set of links connecting these agents. A link (cid:104)ij(cid:105) ∈ g represents that the agents i and j are data backup partners, who store their data on each other's shared storage space. In g, pairs of agents share an equal amount of storage space. Agents perform two actions, link addition (represented by g + (cid:104)ij(cid:105)) and deletion (represented by g − (cid:104)ij(cid:105)). The utility function in [4] reveals the cost and benefit that each agent i receives in g and is given by ui(g) = β ∗ (1 − λni(g)) − c ∗ ni(g). This utility function ui(g) consists of the following param- eters: the agent i's neighborhood size ni(g); the benefit β associated with data; the cost c that the agent i incurs to maintain its neighbours; and the probability of disk failure λ. Note that c, β, and λ lie between 0 and 1. The utility function is a combination of two objectives for each agent i. The first is to minimize the total cost of the links, which is c ∗ ni(g). The second is to maximize the expected data backup benefit, which is β ∗ (1 − λni(g)). Note that from now onwards we will use ni to represent neighbourhood size of agent i in g. In our model, we consider three type of agents. First, where an agent helps to maximize other's utility (friend). Second, where an agent aims to decrease other's utility (enemy). Third, where an agent does not care about other's utility (neutral). Based upon [5], we represent the above kind of social relationships between pairs of agents in a Social Range Matrix F . Here, each element fij denotes the above discussed social relationship between i and j. That is, > 0 < 0 = 0 fij i and j are friend of each other, i and j are enemy of each other, and i and j do not care about each other. Also, j ∈ the set of all agents including i and fii > 0. Many such interesting matrices are possible, for example, a matrix of all ones mean that all pair of agents are friends, and a matrix of all zeroes except the diagonal elements means weak social links. In our paper, for the sake of simplicity, we use only three values for fij. That is, 1, -1, and 0 indicating friend, enemy, neutral, respectively. We also consider that agents give more importance to other agents utilities than their own utility, i.e. fii < fij for all j. Now, agents have different social relationships with others, and hence, their utility not only depends on the structure of the network but also their social relationships with others. We define the new utility (of agent i in g) as the perceived utility, which is defined as follows: fijuj(g). (1) j In this setting, each agent's objective is to maximize its perceived utility (which takes care of utilities of other agents). Thus, the optimization problem is max ((cid:101)ui(g)). (cid:88) (cid:101)ui(g) = III. CONDITION FOR LINK ADDITION AND DELETION Link addition between an agent i and an agent j occurs only when the perceived utilities of both the agents increases, i.e. (cid:101)ui(g + (cid:104)ij(cid:105)) > (cid:101)ui(g) (cid:101)uj(g + (cid:104)ij(cid:105)) > (cid:101)uj(g). and Similarly, link deletion happens when (cid:101)ui(g − (cid:104)ij(cid:105)) > (cid:101)ui(g) (cid:101)uj(g − (cid:104)ij(cid:105)) > (cid:101)uj(g). and Thus, equivalent to (4) and (5) we get link deletion conditions as below. fiiλni + fijλnj < fiiλnj + fijλni < (fii + fij) ∗ c ∗ λ (fii + fij) ∗ c ∗ λ (1 − λ) ∗ β (1 − λ) ∗ β and (6) (7) . (cid:12)(cid:12)ln(cid:0) IV. SUFFICIENCY CONDITION FOR LINK ADDITION Theorem 1. For any two agents i and j, let t1 = max (ni, , the nj), t2 = min (ni, nj) and fij > 0. If t1 < link addition conditions (4) and (5) will be true, and hence, a link will be formed between i and j. Proof. Case 1: If ni > nj, then t1 = ni and t2 = nj. Thus, (1−λ)∗β ln λ c (cid:1)(cid:12)(cid:12) λni fiiλni fijλni or Also, < λnj < fiiλnj < fijλnj (since 0 < λ < 1) (since fii > 0). (since fij > 0). Combining above two we get (fii + fij)λni < fiiλni + fijλnj < (fii + fij)λnj (fii + fij)λni < fiiλnj + fijλni < (fii + fij)λnj . and (8) (9) Using (4) -- (5) in (8) -- (9) we get sufficiency condition for link addition as below. (fii + fij)λni > . (10) (fii + fij) ∗ c (1 − λ) ∗ β On solving (10), we get that link between i, j will happen if (cid:16) (cid:12)(cid:12)(cid:12)ln ni < c (1−λ)∗β ln λ (cid:17)(cid:12)(cid:12)(cid:12) . Case 2: If ni ≤ nj Since (4) and (5) are symmetric, interchange i and j. (2) (3) Similarly, we can derive sufficiency conditions when fij < 0 above as well as for link deletion. Algorithm 1 lists the steps for reaching a bilaterally stable network. That is, when no agent has any incentive to add or delete a link. For all agents k, except i and j, the neighbourhood size (nk) remains constant so their utility is the same and cancels out. Simplifying (2) and (3) with (1), we get that link addition happens when fiiλni + fijλnj > fiiλnj + fijλni > (fii + fij) ∗ c (1 − λ) ∗ β (fii + fij) ∗ c (1 − λ) ∗ β . and (4) (5) V. CASE STUDY Consider five agents as follows: a, b, c, d, and e. Assume that social relationships between these agents is captured in the Social Range Matrix F given in Table I. As an example, the way to read this matrix is as follows: a is a friend of b and d, while it is an enemy of c and e. Let us also assume that c = 0.01, λ = 0.2, = 0.1, and β = 0.1. Initially all agents are isolated (i.e., the starting Algorithm 1: Pseudo code to arrive at a bilaterally stable network. Input Comment: i and j are agents, flag = 1 means network is not bilaterally : c, λ, β, F, starting network, f lag = 1 if link is absent between < i, j > then Check link addition conditions (4) and (5) for i, j and add link if they are true. flag = 1 end if link is present between < i, j > then Check link deletion conditions (6) and (7) for i, j and delete link if they are true. flag = 1 stable flag = 0 for i = 1 to n do 1 while flag == 1 do 2 3 4 5 6 7 for j = 1 to n do if i (cid:54)= j then 8 9 10 11 12 13 14 15 16 17 end end end end end TABLE I: Social Range Matrix F a 1 -1 1 -1 b 1 1 -1 -1 c -1 1 -1 1 d 1 -1 -1 1 e -1 -1 1 1 a b c d e network is a null network). We follow the procedure described in Algorithm 1 and obtain the bilaterally stable network g(cid:48) shown in Fig. 1. With the same set of parameters and F , by starting from a random non-null network, we obtain a different bilaterally stable network than in Fig. 1. This implies that bilaterally stable networks are not necessarily unique. By looking at the ratio of c β , where 0 < c, β < 1, λ = 0.2, and above F , we found that for 0.044 < c β < 0.089, Algorithm 1 runs into infinite loop implying that no bilaterally stable network is possible. Interestingly, for all other values of β we found that at least one bilaterally stable network exists. c If c > β, λ = 0.2, and above F , then we get a unique bilaterally stable network in which all pair of agents are enemies. VI. UNIQUE BILATERALLY STABLE NETWORKS β < (1 − λ)λN−2 and if all agents are friends Lemma 1. If c of each other (i.e. fij = 1), then a complete network is the unique bilaterally stable network. Proof. First, consider the link addition conditions stated in (4) and (5). We can observe that, as ni and nj increases, the L.H.S of (4) and (5) decreases. This implies that agents have an incentive to increase their neighborhood size. If the link addition conditions (4) and (5) are true for ni = nj = N − 2 (both agents have the neighborhood size one less than the maximum possible size)1, then they will be true for all values of ni, nj < N-2 (if the smaller value of the L.H.S is greater than the R.H.S., bigger values will be greater). Solving (4) -- (5) with ni = nj = N − 2, i.e. (fii + fij) ∗ c (1 − λ) ∗ β fiiλN−2 + fijλN−2 > gives < (1 − λ)λN−2. c β A similar analysis can be done with the deletion conditions. β > (1 − λ) and if all agents are friends of Lemma 2. If c each other (i.e. fij = 1), then the empty network is the unique bilaterally stable network. Proof. Similar to Lemma 1. (1− λ) and if fij is either 1 or -1, Corollary 1. If c then the bilaterally stable network will be the one where all the pairs of enemies form links. β > 1 1−fii VII. FUTURE WORK In this work, we have extended the social storage model proposed in [4] to include heterogeneous behavior (variety of social relationships). After that, we have analyzed this model using their solution concept of bilateral stability. The preliminary results here give tremendous insight into how a endogenously built social storage system would emerge. Future work involves further analyzing the stability of such networks as well as studying contentment (when everyone has maximized their utility) and efficiency (where the total utility of the network is maximized) in this new heterogeneous agent scenario. REFERENCES [1] N. Tran, F. Chiang, and J. Li, "Efficient cooperative backup with decentralized trust management," ACM Transactions on Storage, vol. 8, no. 3, pp. 8:1 -- 8:25, 2012. [2] R. Gracia-Tinedo, M. S´anchez-Artigas, A. Moreno-Mart´ınez, and P. Garcia-Lopez, "Friendbox: A hybrid F2F personal storage application," in 12th IEEE International Conference on Peer-to-Peer Computing (P2P). IEEE, 2012, pp. 131 -- 138. [3] R. Sharma, A. Datta, M. DeH'Amico, and P. Michiardi, "An empirical study of availability in friend-to-friend storage systems," in IEEE Interna- tional Conference on Peer-to-Peer Computing (P2P), 2011, pp. 348 -- 351. [4] P. Mane, K. Ahuja, and N. Krishnamurthy, "Unique stability point [Online]. Available: in social storage networks," 2017, 33 Pages. https://arxiv.org/abs/1603.07689 [5] P. Kuznetsov and S. Schmid, "Towards network games with social preferences," in 17th International Colloquium on Structural Information and Communication Complexity (SIROCCO), B. Patt-Shamir and T. Ekim, Eds. Springer: Berlin Heidelberg, 2010, pp. 14 -- 28. Fig. 1: Bilaterally Stable Social Storage Network g(cid:48). 1Neighborhood size of N − 1 is a complete network, and hence, is not included in the analysis here.
1508.02685	1	1508	2015-08-11T18:48:39	Augmenting Agent Platforms to Facilitate Conversation Reasoning	[ "cs.MA" ]	Within Multi Agent Systems, communication by means of Agent Communication Languages (ACLs) has a key role to play in the co-operation, co-ordination and knowledge-sharing between agents. Despite this, complex reasoning about agent messaging, and specifically about conversations between agents, tends not to have widespread support amongst general-purpose agent programming languages. ACRE (Agent Communication Reasoning Engine) aims to complement the existing logical reasoning capabilities of agent programming languages with the capability of reasoning about complex interaction protocols in order to facilitate conversations between agents. This paper outlines the aims of the ACRE project and gives details of the functioning of a prototype implementation within the Agent Factory multi agent framework.	cs.MA	cs	Augmenting Agent Platforms to Facilitate Conversation Reasoning David Lillis and Rem W. Collier School of Computer Science and Informatics University College Dublin david.lillis,[email protected] Abstract. Within Multi Agent Systems, communication by means of Agent Communication Languages (ACLs) has a key role to play in the co-operation, co-ordination and knowledge-sharing between agents. De- spite this, complex reasoning about agent messaging, and specifically about conversations between agents, tends not to have widespread sup- port amongst general-purpose agent programming languages. ACRE (Agent Communication Reasoning Engine) aims to complement the existing logical reasoning capabilities of agent programming lan- guages with the capability of reasoning about complex interaction pro- tocols in order to facilitate conversations between agents. This paper outlines the aims of the ACRE project and gives details of the function- ing of a prototype implementation within the Agent Factory multi agent framework. 1 Introduction Communication is a vital part of a Multi Agent System (MAS). Agents make use of communication in order to aid mutual cooperation towards the achievement of their individual or shared objectives. The sharing of knowledge, objectives and ideas amongst agents is facilitated by the use of Agent Communication Lan- guages (ACLs). The importance of ACLs is reflected by the widespread support for them in agent programming languages and toolkits, many of which have ACL support built-in as core features. In many MASs, communication takes place by way of individual messages without formal links between them. An alternative approach is to group related messages into conversations: "task-oriented, shared sequences of messages that they observe, in order to accomplish specific tasks, such as a negotiation or an auction" [1]. This paper presents the Agent Conversation Reasoning Engine (ACRE). The principal aim of the ACRE project is to integrate interaction protocols into the core of existing agent programming languages. This is done by augmenting their existing reasoning capabilities and support for inter-agent communication by adding the ability to track and reason about conversations. Currently at the stage of an initial prototype, ACRE has been integrated with several agent programming languages running as part of the Common Language Framework 5 1 0 2 g u A 1 1 ] A M . s c [ 1 v 5 8 6 2 0 . 8 0 5 1 : v i X r a of the Agent Factory platform [2]. The longer-term goals of ACRE include its use within other mainstream agent frameworks and languages. The principal aim of this paper is to outline the goals of the ACRE project and to discuss its integration into Agent Factory. This paper is laid out as follows: Section 2 outlines some related work on agent interaction. Section 3 then provides an overview of the aims and scope of the ACRE project. The model used to reason about conversations is presented in Section 4. ACRE protocols are defined in an XML format that is outlined in Section 5, followed by an example of a conversation in execution in Section 6. Details of the integration of ACRE into the Agent Factory framework are given in Section 7. Finally, Section 8 outlines some conclusions along with ideas for future work. 2 Related Work In the context of Agent Communication Languages, two standards have found widespread adoption. The first widely-adopted format for agent communication was the Knowledge Query and Manipulation Language (KQML) [3]. An alterna- tive agent communication standard was later developed by the Foundation for Intelligent Physical Agents (FIPA). FIPA ACL utilises what it considers to be a minimal set of English verbs that are necessary for agent communication. These are used to define a set of performatives that can be used in ACL messages [4]. These performatives, along with their associated semantics, are defined in [5]. Recognising that one-off messages are limited in their power to be used in more complex interactions, FIPA also defined a set of interaction protocols [6]. These are designed to cover a set of common interactions such as one agent requesting information from another, an agent informing others of some event and auction protocols. Support for either KQML or FIPA ACL communication is frequently in- cluded as a core feature in many agent toolkits and frameworks, native support for interaction protocols is less common. The JADE toolkit provides specific im- plementations of a number of the FIPA interaction protocols [7]. It also provides a Finite State Machine (FSM) behaviour to allow interaction protocols to be defined [?]. Jason includes native support for communicative acts, but does not provide specific tools for the development of agent conversations using interac- tion protocols. This is left to the agent programmer [8, p. 130]. A similar level of support is present within the Agent Factory framework [9]. There do exist a number of toolkits, however, that do include support for conversations. For example, the COOrdination Language (COOL) uses FSMs to represent conversations [10]. Here, a conversation is always in some state, with messages causing transitions between conversation states. Jackal [11] and KaOS [12] are other examples of agent systems making use of FSMs to model communications amongst agents. Alternative representations of Interaction Pro- tocols include Coloured Petri Nets [13] and Dooley Graphs [14]. 3 ACRE Overview ACRE is aimed at providing a comprehensive system for modelling, managing and reasoning about complex interactions using protocols and conversations. Here, we distinguish between protocols and conversations. A protocol is defined as a set of rules that dictate the format and ordering of messages that should be passed between agents that are involved in prolonged communication (beyond the passing of a single message). A conversation is defined as a single instance of multiple agents following a protocol in order to engage in communication. It is possible for two agents to engage in multiple conversations that follow the same protocol. Such an aim can only be realised effectively if a number of features are already available. These include: -- Protocol definitions understandable by agents: Interaction protocols must be declared in a language that all agents must be able to understand and share. This also has the advantage that the protocol definition is sep- arated from its implementation in the agent, thus providing a programmer with a greater understanding of the format the communication is expected to take. ACRE uses an XML representation of a finite state machine for this purpose. This representation is further discussed in Section 5. The separation of protocol definitions from agent behaviours also facilitates the development of external tools to monitor communication between agents. -- Shared ontologies: A shared vocabulary is essential to agents understand- ing each other's communications. A shared ontology defines concepts about which agents need to be capable of reasoning. -- Plan repository: With the two above features in place, an agent may reason about the sequence of messages being exchanged, as well as the content of those messages. This reasoning will typically result in an agent deciding to perform some action as a consequence of receiving certain communications. In this case, it is useful to have available a shareable repository of plans that agents may perform so that new capabilities may be learned from others. Clearly, the use of shared plans will be dependent on the agent programming language(s) being used. The presence of these features aid greatly in the realisation of ACRE's aims. The principal aims are as follows: -- External Monitoring of Interaction Protocols: At its simplest level, conversation matching and recognition of interaction protocols allows for a relatively simple tool operating externally to any of the agents. This can intercept and read messages at the middleware level and is suitable for an open MAS in which agents communicate via FIPA ACL. This is a useful tool for debugging purposes, allowing developers to monitor communication to ensure that agents are following protocols correctly. This is particularly important where conversation management has been implemented in an ad- hoc way, with incoming and outgoing messages being treated independently and without a strong notion of conversations. In this case, the protocol def- initions can be formalised after the implementation of the agents without interfering with the agent code itself until errors are identified. -- Internal Conversation Reasoning: On receipt of a FIPA ACL message, it should be possible for an agent to identify the protocol being followed by means of the protocol parameter defined in the message (for the specifica- tion of the parameters available in a FIPA ACL message see [15]). Similarly, the initiator of a conversation should also set the conversation-id param- eter, which is a unique identifier for a conversation. By referring to the the protocol identifier, an agent can make decisions about its response by con- sulting the protocol specification. Similarly, the conversation identifier may be matched against the stored history of ongoing conversations. ACRE provides an agent with access to information about the conversations to which it is a party. This allows the agent to reason about this according to the capabilities of the agent programming language being used. One example of this is the use of this information to analyse the status of conversations and generate appropriate goals for the agent to successfully continue the conversation along the appropriate lines for the protocol that is specified. This has previously been done with the AFAPL2 agent programming lan- guage [16], following the use of goals in [17]. Goals represent the motivations of the participants in a conversation. Thus the agents' engagement in a par- ticular conversation is decoupled from the individual messages that are being exchanged, allowing greater flexibility in reasoning about their reactions and responses. -- Organisation of Incoming Messages: It is possible that an agent com- municating with agents in another system may receive messages that do not specify their protocol and/or conversation identifier. In this case, it is useful for the agent to have access to definitions of the protocols in which it is capable of engaging so as to match these with incoming messages so as to categorise the messages. In this situation, message fields such as the sender, receiver, message content and performative can be compared against currently active conversations to ascertain if it matches the expectations of the next step of the underlying protocol. -- Agent Code Verification: The ultimate aim of ACRE is to facilitate the verification of certain aspects of agent code. In particular, given integration of conversation reasoning into a programming language, it should be possible to verify whether or not an agent is capable of engaging in a conversation following a particular protocol. 4 Conversation Management ACRE models protocols as FSMs, with the transitions between states triggered by the exchange of messages between agents participating in the conversation. Messages, protocols and conversations are represented by tuples. As a FSM, each protocol is made up of states and transitions, which are also represented by tuples. This section presents these representations and also provides an informal description of the conversation management algorithm used within ACRE. A message is represented by the tuple (s, r, c, φ, p, x), where s is the agent identifier of the message's sender, r is the agent identifier of the recipient, c is the conversation identifier, φ identifies the protocol, p is the message performative and x is the message content. Each protocol is represented by a tuple (φ, S, T, i, F ) where φ is the protocol's unique identifier, S and T are sets of states and transitions respectively, i is the name of the initial state and F is a set of names of final (terminal) states. Within these conversations, each state is represented by the tuple (n, s, φ) where n is the name of the state, s is the status of the state (whether it is a start, end or intermediate state) and φ is the identifier of the protocol it belongs to. A transition is represented by (σ, , s, r, p, x). Here, σ and are the names of the start and end states respectively, s and r are the agent identifiers of the sending and receiving agents respectively, p is the performative of the message triggering the transition and x is the message content. As a FSM, a protocol can easily visualised as shown in Figure 1. This figure shows a FSM for a simple, one-shot Vickrey-style auction. It shows the states and transitions associated with this protocol. Transitions are triggered by comparison with messages exchanged between the participating agents. Fig. 1. FSM representation of the Vickrey Auction protocol Finally, a conversation may be represented by (φ, A, s, c, B, ψ) where φ is the protocol identifier, A is the set of participating agents, s is the name of the conversation's current state, c is the conversation identifier, B is the current set of variable/value bindings and ψ is the conversation status (active, completed or failed). The values permitted in the tuples shown here are based on first-order logic, meaning that all values are constants, variables or functions. When considering whether a message is capable of advancing a conversation, its fields are compared with the corresponding elements of the conversation's available transitions. When comparing values, the following rules apply: -- Constant values match against other identical constant values (e.g. in Fig- ure 1, the first transition can only be triggered by a message with the per- formative cfp). -- Variables match against any value. -- Functions match other functions that have the same functor, have the same number of arguments and whose arguments in turn match. In the pseudocode that follows in Figures 2, 3 and 4, this is encapsulated by the function matches(a,b). The bindings associated with the conversation (B) is a set of key/value pairs that binds variables to constants or functions against which that they have been matched in triggering a transition. Any variables that have been matched against a constant or function in a triggering message are given a binding that is stored in B. In the example from Figure 1, the sender of the initial message will have their agent identifier bound to the ?initiator variable, so any further messages must be sent by/to that same agent, whenever the ?initiator variable is used. This is an example of a variable being used in immutable context. Once the variable has been bound to a value, that value may not change for the duration of the conversation. An alternative approach is to use a variable in a mutable context. In this situation, a variable may acquire a binding to a new value regardless of whether it has been previously bound. Further explanation (and examples) of the different variable contexts is presented in Section 5.2. One special-case variable also exists. The anonymous variable (denoted by "?") may not acquire any binding. Thus it acts as a wildcard match that will match against any values. The following sections outline the three key stages of the conversation man- agement algorithm. By convention, elements of tuples are denoted by using sub- scripts (e.g. the initial state (i) of a protocol (p) is shown as pi). 4.1 Identifying Candidate Conversations The first stage of the conversation management algorithm is carried out whenever a message is exchanged and is shown in Figure 2. This identifies any active conversations that may be advanced by a message that has been exchanged. If the message contains a defined conversation identifier (which are unique), then only a conversation bearing that identifier may be advanced by the message. In the event that a message is exchanged without a conversation identifier being present, any conversation with an available transition that may be triggered by the message will be considered a candidate. A conversation can be advanced by a message if the elements of the message match against the corresponding elements of any available transitions (i.e. that begin at the current state of the conversation). If the message contains a de- fined conversation identifier, but that conversation cannot be advanced by the message, the status of the conversation must be changed to failed. The apply(B,a) function is used to apply a set of bindings (B) to a term (a). If a is a variable used in an immutable context for which a binding exists in B, then the bound value is returned. Otherwise, a is returned unaltered. C ← ∅ to store candidate conversations m ← message sent/received for each active conversation (c) do if mc = cc or mc = ⊥ then for each transition (t) where tσ = cs do if matches(ms, apply(cB, ts)) and matches(mr, apply(cB, tr)) and matches(mx, apply(cB, tx)) and matches(mp, tp) then Add c to C end if end for end if if mc = cc and c /∈ C then cψ ← f ailed end if end for Fig. 2. Identifying candidate conversations 4.2 Identifying Candidates for New Conversations If no active conversations may be advanced by the given message, the second stage is to identify whether the message is capable of initiating a conversation using a known protocol. This procedure is shown in Figure 3. if C = 0 then for each protocol (p) do if mφ = pφ or mφ = ⊥ then for each transition (t) where tσ = pi do if matches(ms, ts) and matches(mr, tr) and matches(mx, tx) then if mc = ⊥ then Add (pφ,{ms, mr}, pi, nextid(),∅, active) to C Add (pφ,{ms, mr}, pi, mc,∅, active) to C else end if end if end for end if end for end if Fig. 3. Identifying candidate protocols for new conversations If the message contains a protocol identifier, then only the protocol with that identifier is considered. Otherwise, the message is compared against the initial transition of each available protocol. On finding a suitable protocol, a new conversation is created and added to the set of candidate conversations (C). If the message contained a conversation identifier, this is used as the identifier for the new conversation. Otherwise, a new unique conversation identifier is generated (by means of the nextid() function). 4.3 Advancing the Conversation Having identified conversations that match against the given message, the system must advance a conversation, as appropriate. This is shown in Figure 4. At this stage, events are raised to the agent layer to inform the agent of the outcome of the process. If the message was not capable of advancing or initiating any conversation, an "unmatched" event is raised. If there were multiple candidate conversations (which cannot be the case if conversation identifiers are defined for all messages), an "ambiguous" event is raised. If one candidate conversation was identified, this is advanced to the next appropriate state. Its bindings must be updated (using the getBindings(m,t) function) to include bindings for variables in the transition that were matched against values in the message. The anonymous variable may not acquire a bind- ing. This function does not discriminate between variables based on the context in which they are used. Both mutable and unbound immutable variables are free to acquire new bindings. If an immutable context variable has previously been bound to a value, it is this value that is used in matching the message to the transition (by means of the apply(B,a) function shown in Figure 2). As the message content is frequently a function of first order logic, any variables within that function that match against corresponding parts of the message content will also acquire bindings in the same way as standalone variables. if C = 1 then c ← the matched conversation in C t ← the transition matched by the message m cs ← t cB ← cB ∪ getBindings(m, t) if cs is an end state then cψ ← completed raiseEvent(completed, c) else raiseEvent(advanced, c) end if else if C = 0 then raiseEvent(unmatched, m) else raiseEvent(ambiguous, m) end if Fig. 4. Advancing the conversation 5 The ACRE XML Format In ACRE, interaction protocols are modelled using an XML file that follows the ACRE XML protocol schema definition 1. A sample of an XML representation of a Vickrey Auction Interaction Protocol is given in Figure 5 (this is the same protocol as the FSM in Figure 1). Each protocol is identified by a name, a name and a version number (contained in the <namespace>, <name> and <version> tags respectively). The version number is intended to prevent multiple agents attempting to communicate using different protocol implementations (e.g. if an error is discovered in an earlier attempt at modelling a particular protocol). The use of a namespace helps to avoid conflicts whereby various developers implement different models of similar protocols using the same name. Each protocol is represented by a number of states and transitions, defined us- ing <state> and <transition> tags respectively. Each state has only a "name" attribute, so that it can be referred to in the transitions. The type of state each represents (i.e. terminal, initial or other) can be found on the fly when the pro- tocol is loaded. A state at which no transition ends is considered a start state. States at which no transitions begin are terminal states. The reason these are not expressly marked in the protocol definition is because of the ability to import other protocols, which is discussed in Section 5.1. Transitions are more complex, as these are required to match messages so as to trigger a change in the state of a conversation. Each <transition> tag contains up to six attributes, many of which attempt to match against one field of a FIPA message. The attributes in this file correspond with the values in the tuple representing a transition in Section 4. In addition to these message fields, the ACRE conversation manager will also examine the protocol-id and conversation-id fields to match messages to particular conversations. The attributes allowable in a <transition> tag are as follows: -- Performative: This is a mandatory field that specifies the performative that a message must have in order to trigger this transition. The attribute value must be exactly equal to the performative contained in the message for this transition to be triggered. Variables are not permitted in this field. -- From State: Another mandatory field, this indicates the state from which this transition may be triggered. If the conversation is in another state then this rule cannot match. In the majority of cases, the attribute value must be the same as the name of a state that is contained either in the protocol itself or in an imported protocol. In addition to exact state names, regular expression matching is also permit- ted. If a regular expression is provided (indicated by beginning and ending the value with a forward slash), then this transition will be triggerable from any state that matches this regular expression. In practice, the protocol in- terpreter will duplicate this transition for each state name that matches, thus fitting with the model outlined in Section 4. 1 http://acre.lill.is/protocol.xsd -- To State: This is another mandatory field and is used to indicate the state that the conversation will be in upon successful triggering of this transition. As with the "From State" attribute, this should match the name of a state that is either part of the protocol or is imported. However, it may not contain a regular expression, as the conversation state after the sending of a message must be clearly defined. -- Sender: This indicates which agent should be the sender of the message. Although it is allowable to use a constant value for this attribute, this is unusual as it specifically restricts the protocol to an agent with a a particular identifier. Generally, this will use a variable to refer to particular agents. The same variable may be used throughout the protocol to indicate that particular messages should be sent by the same agent, as it will have acquired a bound value the first time it matches against an agent identifier. -- Receiver: This attribute functions in a similar way to "Sender", with the exception that it is the recipient of the message that is being matched. -- Content: This attribute relates to the actual content of the message. It may be a constant, a variable or a function that possibly combines the two. Figure 5 illustrates the use of a function in the content field in each of the transitions. The "Sender", "Receiver" and "Content" attributes are optional in a protocol definition. In each case, the default value if one is not supplied is the anonymous "?" variable that matches any value. 5.1 Importing Protocols One other feature of the ACRE XML format is the ability to import from other protocols. When this occurs, all of the states and transitions from the imported protocol are added to those of the protocol containing the <import> tag. This means that transitions in a protocol may refer to states that are not in the protocol itself but rather are in the imported protocol. One example of a use for this is the "Cancel" meta-protocol that is included in all of the standard FIPA Interaction Protocols. This protocol always works in an identical way, regardless of what the main protocol being followed is: at any non-terminal stage of the conversation, the initiator of the original conversation may terminate the interaction by means of a cancel message. This meta-protocol can be extracted into a separate ACRE protocol that is imported by all other protocols that support it. 5.2 Variable Bindings As mentioned in Section 4, the definition of protocols in ACRE allows the use of three types of variable. The anonymous variable "?" is an unnamed variable that is capable of matching against any value. As such, it can be considered to be a wildcard match. A transition whose content attribute is set to "?" can <?xml version="1.0"?> <protocol xmlns="http://acre.lill.is" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://acre.lill.is \protect\vrule width0pt\protect\href{http://acre.lill.is/proto.xsd}{http://acre.lill.is/proto.xsd}"> <namespace>is.lill.acre</name> <name>acre-vickreyauction</name> <version>0.1</version> <states> <state name="start"/> <state name="awaiting_bid" /> <state name="bid" /> <state name="nobid"/> <state name="accepted"/> <state name="rejected"/> </states> <transitions> <transition performative="cfp" from-state="start" to-state="awaiting_bid" sender="?initiator" receiver="?bidder" content="bidfor(?item)" /> <transition performative="propose" from-state="awaiting_bid" to-state="bid" sender="?bidder" receiver="?initiator" content="bid(?item,?amount)" /> <transition performative="propose" from-state="awaiting_bid" to-state="nobid" sender="?bidder" receiver="?initiator" content="nobid(?item)" /> <transition performative="accept-proposal" from-state="bid" to-state="accepted" sender="?initiator" receiver="?bidder" content="bid(?item,?amount)" /> <transition performative="reject-proposal" from-state="bid" to-state="accepted" sender="?initiator" receiver="?bidder" content="bid(?item,?amount)" /> </transitions> </protocol> Fig. 5. ACRE XML Representation of the Vickrey Auction Protocol be triggered by a message with any content (assuming the other fields in the message match the specified transition). Two types of named variable are permitted: immutable named variables, which have "?" as a prefix followed by the variable name (e.g. ?item) and mu- table named variables that are prefixed with "??" (e.g. ??amount). Each named variable is in scope for the duration of a conversation and is associated with values as it is matched against the actual fields in messages that trigger transitions. Whenever a named variable is used in an immutable context, it may match any content if has not already acquired a value. However, once it has been matched to a value, it may only match that value for the duration of the conversation. For example, Figure 5 uses ?initiator to denote the agent that begins the Vickrey Auction with the sending of the initial call for proposals. Initially this variable does not have a value associated with it, so it may match the name of the relevant agent (e.g. "agent1"). However, once this has been done, the variable ?initiator may only match the value "agent1" for the remainder of the conversation. In some situations it is desirable to have variables whose values may change as the conversation progresses. For that reason, mutable named variables are also facilitated. The difference between a mutable and immutable named variable is that a mutable named variable can match against any content, regardless of whether it previously has a value associated with it. When this occurs, the existing value is overwritten with the new value that has been matched. The motivation behind the use of mutable variables can be seen by exam- ining the Vickrey Auction protocol shown above. This is a one-shot auction, so immutable variables are sufficient. However, implementing an iterated auction is made far more complex if all variables are immutable. The second transi- tion shown in Figure 5 would be unsuitable for this, as the ?amount variable is restricted to match only whatever the first value it is matched against. By changing the content field of this transition to bid(?item,??amount), the vari- able relating to the amount is used in mutable context and so its value may change (although the item being bid for must remain the same). Thus each time this transition is triggered, the amount variable acquires a new value: namely the amount of the latest bid that was submitted. This can later be referred to using texttt?amount by the other agent that wishes to accept or reject the bid. A similar usage can be seen in the example presented in Section 6. 6 ACRE Conversation Example This section presents an example of how an ACRE conversation may progress. The example is based on the "Process Documents" protocol shown in Figure 6. This protocol is designed for a system where text documents must undergo some form of processing. The Initiator of the protocol is capable of performing this processing, though it must be made aware of which documents to work on by the Respondent. Fig. 6. Process Documents Protocol Initially, the Initiator informs the Respondent that it is ready to process doc- uments, to which the Respondent replies with a request to process a particular document. The Initiator may either process the document and inform the Re- spondent of this, or refuse to carry out this processing. In the former case, the Respondent will send the next document for processing and continue to do this until the Initiator eventually refuses. Here, the Agent UML description shown to the left of Figure 6 is converted to the FSM shown on the right. The dashed line surrounding the "Start" state indicate that this is the current state initially. The first transition contains constant values for both the performative and the content. This means that any message matching that transition must have those exact values in those fields. The message sender and receiver are variables, and since there are not yet any bindings associated with the conversation, these will match any agent identifiers in those fields in the message. A message that will trigger this initial transition may look as follows (some unimportant fields have been omitted for clarity): (inform processor :sender :receiver manager :content ready ) This results in the state of the conversation changing to "Waiting", as shown in Figure 7. Because the ?initiator and ?respondent matched against the Fig. 7. Process Documents protocol in the "Waiting" state Fig. 8. Process Documents protocol in the "Requested" state constants "processor" and "manager" respectively in the message, these bindings are associated with the conversation. As these variables are used in immutable context throughout, they must match their exact bound values for the remainder of the conversation. This is indicated in Figure 7 by replacing the variables with these values. Angle brackets have been placed around each of the replacement values in order to emphasise this. The transition from the "Waiting" state can now only be triggered by a message sent by the manager agent to the processor agent with the "request" performative. The content must also match the transition, though with the use of the ??docid variable, there is some flexibility in the values that can be matched. In the next stage, the manager asks the processor to process the document with the identifier "doc123". This is done by means of the following message: (request :sender :receiver :content manager processor process(doc123) ) As this message matches the transition, the conversation moves from the "Waiting" state to the "Requested" state, as shown in Figure 8. Fig. 9. Process Documents protocol in the "Waiting" state for the second time. Fig. 10. Process Documents protocol in the "Requested" state for the second time. At this point, the ?docid variable has also acquired a binding, so this is re- placed in all transitions using it in an immutable context. Again, this is indicated by the value being contained within angle brackets in Figure 8. At this point, the processor agent must either inform the manager that document "doc123" has been processed (thus returning the conversation to the "Waiting" state) or refusing to process that document. In each case, it is only the document identifier that has previously been bound to the ?docid variable that may be used. Re- fusing to process a different document identifier would result in the conversation failing as the message could not match an available transition. If the processor agrees to process the document and informs the manager of its completion, the conversation returns to the "Waiting" state, as shown in Figure 9. Unlike the first time the conversation was in this state, this time the ?docid variable has got a binding associated with it. However, the transition between "Waiting" and "Requested" uses this variable in a mutable context, meaning that it can match against any value contained in the next message. Thus the manager can ask the processor to process any document it wishes. In contrast, as noted previously, when moving from the "Requested" state, the processor is restricted to only discussing the identifier of the last document it was asked to process. If the manager requests that the processor processes document "doc124", the conversation returns to the "Requested" state, as shown in Figure 10. The?docid variable has now acquired an updated binding that is now applied to all the tran- sitions using that variable in an immutable context. At this stage, the processor agent may repeat the cycle by processing the document and informing the man- ager of this, or it may end the conversation by refusing to perform the processing, thus entering the "End" state. 7 Language Integration in Agent Factory Agent Factory is an extensible, modular and open framework for the develop- ment of multi agent systems [2]. It supports a number of agent programming languages, including AFAPL/ALPHA [18,19], AFAPL2 [20], AF-TeleoReactive (based on [21]) and AF-AgentSpeak (an implementation of AgentSpeak(L) [8]). The specific use of ACRE from within AFAPL2 agents has been discussed in a previous paper [16]. Agent Factory's implementation of these agent programming languages are based on its Common Language Framework (CLF) whereby the way in which sensors, actions and modules are implemented have been standardised across the various languages. This greatly facilitates the integration of additional services into each language, as the core components will be shared. This is the case for ACRE, where minimal effort is required to add support for further languages once one integration has been completed. The ACRE Architecture consists of a number of components, some of which are platform-independent and some of which require some work to be ported to other platforms and agent programming languages. 7.1 Protocol Manager The Protocol Manager (PM) is a platform-independent component that is tasked with making protocols available to agents. When an agent identifies the URL of an ACRE protocol definition it will send this to the PM, which will load the protocol, verify it against the appropriate schema and make it available for interested agents to use. It is also capable of accessing online ACRE reposito- ries that may contain multiple protocol definitions in a centralised location. A repository definition file lists the protocols that it has available. Typically, one PM will exist on an agent platform, so any protocol located by any agent will be shared amongst all agents on the platform (within Agent Factory, the PM is a shared Platform Service). However, there is no technical barrier to individual agents having their own PMs if desired. Previously loaded protocols are also stored locally so that they can be recovered in the event of a platform failure or restart. 7.2 Conversation Manager Whereas the PM is shared amongst agents, each agent has its own Conversation Manager (CM), which is used to keep track of the conversations the agent is involved in. The CM monitors both incoming and outgoing communication and matches each message to an appropriate conversation, following the algorithm outlined in Section 4. By monitoring the CM, an agent can gain data that can be used to reason about ongoing conversations and the messages it sends and receives. The CM is also platform-independent. 7.3 Agent/ACRE Interface The Agent/ACRE Interface (AAI) is specific to the platform and agent pro- gramming language being used. This is designed to facilitate the interaction between the agents and the ACRE components mentioned above. The AAI has two distinct principal roles: -- To enable an agent to inform the PM and CM of information it holds. -- To provide the agent with information about the status and activity of the CM and PM. In the former case, an agent must be capable of informing the PM of the location of any protocols that it wishes to use. This information may originally come from another agent with which it wishes to communicate. The CM requires access to the inbox and outbox of the agent also, so the AAI must provide this service also. The key role of the AAI is making information about its own communication available to agents. Within Agent Factory, this is done in two complementary ways: knowledge sensors and event sensors. A knowledge sensor is a sensor that runs on each interpreter cycle of the agent, and provides information on the current state of conversations and pro- tocols. This information currently consists of: -- What protocols are already loaded (PM). -- For each conversation in which the agent is a participant: -- The protocol each conversation is following (CM) -- The identity of the other participating agent in each conversation (CM) -- The current state of each conversation (CM) -- The current status of each conversation (CM) In addition to these, event sensors inform the agent whenever events are raised by the PM or the CM. Events currently handled include: -- A new protocol has been loaded (PM) -- A new conversation has begun (CM) -- A conversation has advanced (CM) -- A conversation has ended (CM) -- An error has occurred in a conversation (CM) In addition to the basic role of information passing, an AAI may augment the capabilities of a language by leveraging the data available from the CM or the PM. For instance, the AAI built for the Agent Factory CLF provides an action of the form advance(conversation-id,performative,content) whereby an agent can advance a specific conversation while providing minimal information. The details about the other participating agent (including its address) are taken from the CM, along with the protocol identifier and content language. Further features in this vein are left for future work. 8 Conclusions and Future Work This paper presents an outline of the ACRE conversation reasoning system, how it models protocols and conversations, and how it is integrated into the Agent Factory multi agent framework. Although currently only used with Agent Fac- tory, it is intended that ACRE will be used in conjunction with several other agent programming frameworks and the languages they support. ACRE has been designed to be as language-independent and platform-independent as pos- sible. Despite this, it will be necessary to adapt the system to frameworks and languages other than Agent Factory and the agent programming languages it supports. Aside from the integration of ACRE into other platforms, future focus will be on the reasoning about conversations at the agent deliberative level and the information that ACRE will need to provide in order to facilitate this. One are of focus will be to allow the grouping of conversations into related groups. As ACRE protocols are limited to two participants, it is necessary to allow agents to relate conversations to each other. One such example will be in a situation where an agent is conducting an auction and has issued a call for proposals to multiple agents. At present, each of these initiates a separate conversation that must be managed by the agent using its own existing capabilities. However, grouping conversations at the Conversation Manager level would allow events to be raised to inform the agent that all conversations in the group had reached the state where a proposal had been received, or left the state where a proposal had been solicited. The key point is the provision of sufficient information for agents to use their deliberative reasoning capabilities in conjunction with the information emanating from the Conversation and Protocol Managers. In addition, it is intended to explore the ways in which having access to an ACRE layer will add to the native messaging capabilities of other programming languages. This includes the ability to directly advance a conversation, as alluded to in Section 7, as well as specifically initiating conversations (rather than relying on message matching on the part of the Conversation Manager). The possibilities are likely to vary with different agent programming languages and it is intended to add these features as appropriate. The availability of cross-platform communication tools such as ACRE can only aid interoperability between distinct agent platforms, toolkits and program- ming languages. References 1. Labrou, Y.: Standardizing agent communication. Multi-Agents Systems and Ap- plications (Advanced Course on Artificial Intelligence) (2001) 74 -- 97 2. Collier, R.W., O'Hare, G.M.P., Lowen, T., Rooney, C.: Beyond Prototyping in the Factory of Agents. In: Multi-Agent Systems and Application III: 3rd International Central and Eastern European Conference on Multi-Agent Systems (CEEMAS 2003), Prague, Czech Republic (2003) 3. Finin, T., Fritzson, R., McKay, D., McEntire, R.: KQML as an Agent Com- munication Language. In: Proceedings of the Third International Conference on Information and Knowledge Management, Gaithersburg, MD (1994) 456 -- 463 4. Poslad, S., Buckle, P., Hadingham, R.: The FIPA-OS Agent Platform: Open Source for Open Standards. In: Proceedings of the 5th International Conference and Exhibition on the Practical Application of Intelligent Agents and Multi-Agents (PAAM2000), Manchester (2000) 368 5. Foundation for Intelligent Physical Agents: FIPA Communicative Act Library Specification (2002) 6. Foundation for Intelligent Physical Agents: FIPA Interaction Protocol Library Specification (2000) 7. Bellifemine, F., Caire, G., Trucco, T., Rimass, G.: Jade Programmer's Guide (2007) 8. Bordini, R.H., Hubner, J.F., Wooldridge, M.: Programming multi-agent systems in AgentSpeak using Jason. Wiley-Interscience (2007) 9. Collier, R.W., Ross, R., O'Hare, G.M.P.: A Role-Based Approach to Reuse in Agent-Oriented Programming. In: AAAI Fall Symposium on Roles, an Interdisci- plinary Perspective (Roles 2005), Arlington, VA, USA (2005) 10. Barbuceanu, M., Fox, M.S.: COOL: A language for describing coordination in In: Proceedings of the First International Conference on multi agent systems. Multi-Agent Systems (ICMAS-95). (1995) 17 -- 24 11. Cost, R.S., Finin, T., Labrou, Y., Luan, X., Peng, Y., Soboroff, I., Mayfield, J., Boughannam, A.: Jackal: a Java-based Tool for Agent Development. In: Working Papers of the AAAI-98 Workshop on Software Tools for Developing Agents, AAAI Press (1998) 12. Bradshaw, J.M., Dutfield, S., Benoit, P., Woolley, J.D.: KAoS: Toward an industrial-strength open agent architecture. Software Agents (1997) 375 -- 418 13. Cost, R.S., Chen, Y., Finin, T., Labrou, Y., Peng, Y.: Modeling agent conversations with colored petri nets. In: In: Workshop on Specifying and Implementing Con- versation Policies, Third International Conference on Autonomous Agents (Agents '99), Seattle. (1999) 59 -- 66 14. Parunak, H.: Visualizing Agent Conversations: Using Enhanced Dooley Graphs for Agent Design and Analysis. In: Proceedings of the Second International Conference on Multi-Agent Systems (ICMAS). (1996) 15. Foundation for Intelligent Physical Agents: FIPA ACL Message Structure Specifi- cation (2002) 16. Lillis, D., Collier, R.W.: ACRE: Agent Communication Reasoning Engine. In: 3rd International Workshop on LAnguages, Methodologies and Development Tools for Multi Agent SystemS (LADS'010), Lyon (2010) 17. Braubach, L., Pokahr, A.: Goal-Oriented Interaction Protocols. In: MATES '07: Proceedings of the 5th German Conference on Multiagent System Technologies. Volume 4687., Leipzig, Germany, Springer (2007) 85 -- 97 18. Collier, R.W.: Agent Factory: A Framework for the Engineering of Agent-Oriented Applications. Phd thesis, University College Dublin (2001) 19. Collier, R.W., Ross, R., O'Hare, G.M.P.: Realising Reusable Agent Behaviours In: Proceedings of the 3rd German Conference on Multi-Agent with ALPHA. System Technologies (MATES 05), Koblenz, Germany (2005) 210 -- 215 20. Muldoon, C., O'Hare, G.M.P., Collier, R.W., O'Grady, M.J.: 6. In: Towards Per- vasive Intelligence: Reflections on the Evolution of the Agent Factory Framework. Springer US, Boston, MA (2009) 187 -- 212 21. Nilsson, N.J.: Teleo-Reactive Programs for Agent Control. Journal of Artificial Intelligence Research 1 (1994) 139 -- 158
cs/0610113	1	0610	2006-10-19T10:41:16	CHAC. A MOACO Algorithm for Computation of Bi-Criteria Military Unit Path in the Battlefield	[ "cs.MA", "cs.CC" ]	In this paper we propose a Multi-Objective Ant Colony Optimization (MOACO) algorithm called CHAC, which has been designed to solve the problem of finding the path on a map (corresponding to a simulated battlefield) that minimizes resources while maximizing safety. CHAC has been tested with two different state transition rules: an aggregative function that combines the heuristic and pheromone information of both objectives and a second one that is based on the dominance concept of multiobjective optimization problems. These rules have been evaluated in several different situations (maps with different degree of difficulty), and we have found that they yield better results than a greedy algorithm (taken as baseline) in addition to a military behaviour that is also better in the tactical sense. The aggregative function, in general, yields better results than the one based on dominance.	cs.MA	cs	CHAC. A MOACO Algorithm for Computation of Bi-Criteria Military Unit Path in the Battlefield A.M. Mora1, J.J. Merelo1, C. Millan2, J. Torrecillas2, and J.L.J. Laredo1 1 Departamento de Arquitectura y Tecnolog´ıa de Computadores University of Granada (Spain) {amorag,jmerelo,juanlu}@geneura.ugr.es 2 Mando de Adiestramiento y Doctrina Spanish Army {cmillanm, jtorrelo}@et.mde.es Abstract. In this paper we propose a Multi-Objective Ant Colony Optimization (MOACO) algo- rithm called CHAC, which has been designed to solve the problem of finding the path on a map (corresponding to a simulated battlefield) that minimizes resources while maximizing safety. CHAC has been tested with two different state transition rules: an aggregative function that combines the heuristic and pheromone information of both objectives and a second one that is based on the dom- inance concept of multiobjective optimization problems. These rules have been evaluated in several different situations (maps with different degree of difficulty), and we have found that they yield better results than a greedy algorithm (taken as baseline) in addition to a military behaviour that is also better in the tactical sense. The aggregative function, in general, yields better results than the one based on dominance. 1 Introduction Prior to any combat manoeuvre, the unit commander must plan the best path to get to an position that is advantageous over the enemy in the tactical sense. This decision is conditioned by two criteria, speed and safety, which he must evaluate. The choice of a safe path is made when the enemy forces situation is not known, so the unit must move through hidden zones in order to avoid detection, which may correspond to a very long, and thus slower, path. On the other hand, the choice of a fast path is made when the unit is going to attack the enemy or if there are few hidden (safe) zones in the terrain and going through it may produce a lot of casualties. In any case, the computation of the best itinerary for the unit to reach the objective point is a very important tactical decision. We will try to make this decision in the paper by solving the what we have called the military unit path finding problem. The military unit path finding problem is similar to the common path finding problem with some additional features and restrictions. It intends to find the best path from an origin to a destination point in the battlefield but keeping a balance between route speed and safety. The unit has an energy (health) and a resource level which are consumed when the unit moves through the path depending on the kind of terrain and difference of height, so the problem objectives are adapted to minimize the consumption of resources (which usually means walking as short/fast as possible) and the consumption of energy. In addition, there might be some enemies in the map (battlefield) which could shoot their weapons against the unit. To solve the problem an Ant Colony System (ACS) algorithm has been adapted. It is a type of ACO [1,2], which offers more control over the exploration and exploitation parts of search. In addition, some features to deal with the two objectives of the problem have been added (in [3] it can be found a survey of MO algorithms), so it is a MOACO algorithm (paper [4] presents a review of some of them). This problem had been solved so far using classical techniques like branch and bound or dynamic programming, which do not usually scale well with size, but to the best of our knowledge find no one that treat the problem as a multiobjective and solve it with a MOACO. The rest of the paper is organized as follows: next section is devoted to a presentation of the problem we want to solve, followed in section 3 by a presentation of the CHAC algorithm. Experiments and results are presented in section 4, and conclusions are drawn in section 5, along with the future lines of development of this work. 2 The Problem Our objective in this paper is to give the unit commander, or to a simulated unit in a combat training simulator, a tactical decision capability so that it will be able to calculate the best path to its target point in the battlefield considering the same factors that a commander would. The battlefield has been modeled as a square cell grid, every cell corresponding to a 500x500 meter zone in the real world. The speed in a path is associated with the unit resources consumption because we have assigned a penalization to every cell related to the 'difficulty' of going through it and we have called this penalization resource cost of the cell. So, going through cells with more resource cost is more difficult and so more slow. Due to this justification, we refers to fast paths or paths with small resource cost. Thus, the problem unit has two properties, a number of energy (which represent global health of unit soldiers or status of the unit vehicles) points and a number of resource (which represent unit supplies such as fuel, food and even moral) points. Its objective is to get to a target point with the maximum level in both properties. The unit energy level is decremented in every cell of the path (the company depletes its human resources or vehicles suffer damage), which have a penalization called no combat casualties, depending on its type. In addition there is an extra penalization due to the impact of an enemy weapon in the cell. This value is calculated as combination of three other, the probability of enemy shoots, the probability of impact in the unit, and the damage it would produce to the unit. Moreover every cell has an assigned resources penalty depending on the type of terrain. There is an extra penalty when the unit goes from a cell to other with different height (if it goes down, the penalty is small than if it goes up). Besides cost in energy, and cost in resources, cells have the following properties: • Type: there are four kinds of cells: normal (flat terrain), forest, water and obstacle, three different types of terrain and obstacle which means a cell that the unit cannot going through. There are several subtypes: it can be an enemy unit position, problem unit position, or the target point; it can also be affected by enemy fire (with one hundred levels of impact) or be lethal for the problem unit. • Height : an integer number (between -3 and 3) which represents a level of depth or height of a cell. We have implemented an application in Delphi 7 for Windows XP in order to create the problem scenarios (battlefields) and also to visualize the solutions obtained by CHAC. This application is available under request. Figure 1 shows an example of battlefield. It includes all typical features: enemy unit, obstacles, forest and water filled terrain accidents. Fig. 1. Map Example. It is a 30x30 map which shows some of the types and subtypes above mentioned. In left, and explanation of cells types is showed, because it is difficult to differ between different kinds of cells due to the black and white image. The example includes the problem unit (circle with black border), an enemy unit (square with gray border and mark with 'X'), the zone affected by enemy weapons (in different shades of light gray depending on the associated damage) surrounding the enemy and the objective point for the unit (circle with light gray border). The different shades in the same color models height (light color) and depth (dark color). This implementation includes some constraints: problem and enemy units fill up exactly one cell (which corresponds with their size in real world). The problem unit as well as the target point must be located on the map; the presence of the enemy is optional. The unit can go only once through a cell and cannot go through cells occupied by an enemy or an obstacle, or cells with a cost in resources or energy bigger than what the unit has available. Some rules also apply to the line of sight, so that it behaves as realistically as possible. 3 The CHAC Algorithm CHAC means Compan´ıa de Hormigas Acorazadas or Armoured Ant Company; we have chosen this (admittedly a bit tongue-in-cheek) name to relate the ACO algorithms with the military scope of this application. It is an ACS adapted to deal with several objectives, that is, a MOACO or Multi-objective Ant Colony Optimization algorithm. In order to approach it via an ACO algorithm, the problem must be transformed into a graph with weighted edges. Every cell in the map is considered a node in the graph with 8 edges connect it to every one of its neighbours (except in border cells). To deal with two objectives, there are two weights in every edge related to the cost of energy and resources, that is, a cost related to the energy expenses (cost assigned to the destination node of the edge) and other related to the consumption of resources if the unit moves from one cell to its neighbour following that edge (which depends on the types of both nodes and the height difference between them). The algorithm implemented within CHAC is constructive which means that in every iteration, every ant builds a complete solution, if possible (there are some constraints, for example the unit cannot go through one node twice in the same path), by travelling through the graph. In order to guide this movement, the algorithm uses information of two kinds: pheromone and heuristic, that will be combined. Since it is a multiobjective problem, CHAC uses two pheromone matrices, one per ob- jective, and two heuristics functions (also matrices), following the BicriterionAnt algorithm designed by Iredi et al. in [5]. However, we decided to use an Ant Colony System (ACS) instead of an Ant System to have better control in the balance between exploration and exploitation by using the parameter q0. We have implemented two state transition rules (which means two different algorithms), first one similar to the Iredi's proposal and second one based on dominance of neighbours. The local and global updating formulas are based in the MACS-VRPTW algorithm proposed by Bar´an et al. in [6], with some changes due to the use of two pheromone matrices. The objectives are named f , minimization the resources consumed in the path (speed maximization) and s, minimization the energy consumed in the path (safety). The Heuristic Functions try to guide search between the start and the target point consid- ering the most important factors for every objective. So, for edge (i,j) they are: ηf (i, j) = ηs(i, j) = ωf r Cr(i, j) ωse Ce(i, j) + + ωf d Dist(j, T ) ωsd Dist(j, T ) + (ωf o · ZO(j)) + (ωso · ZO(j)) (1) (2) In Equation 1, Cr is the resource cost when moving from node i to node j, Dist is the Euclidean distance between two nodes (T is the target node of the problem) and ZO is a score (between 0 and 1) to a cell being 1 when the cell is hidden to all the enemies (or to all the cells in a radius when there are no enemies) and decreasing exponentially when it is seen. ωf r , ωf d and ωf o are weights to assign relative importance to the terms in the formula. In this case, the most important term is the distance to target point because it searches for the fastest path. In second place is important the resources cost and only a little the hidden of the cell (it almost does not mind in a fast path). In Equation 2, Ce is the energy cost of moving to node j, Dist and ZO is the same as the previous formula. ωse , ωsd and ωso are again weights to assign relative importance to the terms in the formula, but in this case the most important are energy cost and hidden (both are to be considered in a safe path) and a little the distance to target point. The Combined State Transition Rule (CSTR) is a formula used to decided which node j is the next in the construction of a solution (path) when the ant is at the node i, it is the pseudo-random-proportional rule used in ACS, but adapted to deal with a two objectives problem by combining the heuristic and pheromone information of both of them: If (q ≤ q0) j = arg max j∈Ni nτf (i, j)α·λ · τs(i, j)α·(1−λ) · ηf (i, j)β·λ · ηs(i, j)β·(1−λ)o (3) (4) Else P (i, j) =   τf (i, j)α·λ · τs(i, j)α·(1−λ) · ηf (i, j)β·λ · ηs(i, j)β·(1−λ) τf (i, u)α·λ · τs(i, u)α·(1−λ) · ηf (i, u)β·λ · ηs(i, u)β·(1−λ) Xu∈Ni 0 if j ∈ Ni otherwise Where q 0 ∈ [0,1] is the standard ACS parameter, q is a random value in [0,1]. τf and τs are the pheromone matrices and ηf and ηs are the heuristic functions for the objectives (Equations 1 and 2). All these matrices have a value for every edge (i,j). α and β are the usual weighting parameters and Ni is the current feasible neighbourhood for the node i. λ ∈ (0,1) is a user-defined parameter which sets the importance of the objectives in the search (this is an application created for a military user who decides which objective has more priority), so for instance, if the user decides to search for the fastest path, λ will take a value close to 1 if he wants the safest path, it has to be close to 0. This value is kept during all the algorithm and for all the ants, unlike other bi-criterion implementations in which the parameter takes value 0 for first ant and it was growing for every ant until it takes a value 1 for the last one. When an ant is building a solution path and it is placed at one node i, if q ≤ q 0 the best neighbour j is selected as the next (Equation 3). Otherwise, the algorithm decides which node is the next by using a roulette considering P(i,j) as probability for every feasible neighbour j (Equation 4). The Dominance State Transition Rule (DSTR) is based on the dominance concept (see reference [3]), which is defined as follows (a dominates b): a ≻ b if : ∀i ∈ 1, 2, ..., k Ci(a) ≤ Ci(b) ∧ ∃j ∈ 1, 2, ...kCj(a) < Cj (b) (5) Where a and b are two different vectors of k values (one per objective) and C is a cost function for every component in the vector. If it intends to minimize the cost and Equation 5 is true, then b is dominated by a. So, in our problem there are two cost functions to evaluate the dominance between nodes. Actually these functions consider edges because they have assigned pheromone and heuristic information and combine them: Cf (i, j) = τf (i, j)α·λ · ηf (i, j)β·λ Cs(i, j) = τs(i, j)α·(1−λ) · ηs(i, j)β·(1−λ) (6) (7) In addition there is a function which uses Equations 6 and 7: D(i, j, u) =  1 0 if (i, j) ≻ (i, u) otherwise At last, the dominance state transition rule is as follows: If (q ≤ q0) Else P (i, j) =   (8) (9) (10) j = arg max j∈Ni (Xu∈Ni D(i, j, u) ∀j 6= u) Xu∈Ni Xk∈Ni Xu∈Ni D(i, j, u)! + 1 D(i, k, u)! + 1! if j ∈ Ni ∧ j 6= u ∧ k 6= u 0 otherwise where all the parameters are the same as in Equations 3 and 4. This rule chooses the next node j in the path (when an ant is placed at node i) considering the number of neighbours dominated for every one. So if q ≤ q0, the node which dominates more of the other neighbours is chosen, otherwise the probability roulette wheel is used. In Equation 10 we add 1 to avoid a 0 probability if no one dominates other neighbour. There are two Evaluation Functions (one per objective, again): Ff (P sol) = Xn∈P sol Fs(P sol) = Xn∈P sol [Cr(n − 1, n) + ωF f o · (1 − ZO(n))] [Ce(n − 1, n) + ωF so · (1 − ZO(n))] (11) (12) Where Psol is the solution path to evaluate and ωF f o and ωF so are weights related to the importance of visibility of the cells in the path. In Equation 11 its importance will be small, it is less important to hide in a fast path and it will be high in Equation 12 for the opposite reason. The other terms are the same that in Equations 1 and 2. Since CHAC is an ACS, there are two levels of pheromone updating, local and global, which update two matrices at each level. The equations for Local Pheromone Updating (performed when a new node j is added to the path an ant is building) are: τf (i, j) = (1 − ρ) · τf (i, j) + ρ · τ0f τs(i, j) = (1 − ρ) · τs(i, j) + ρ · τ0s (13) (14) Where ρ in [0,1] is the common evaporation factor and τ0f , τ0s are the initial amount of pheromone in every edge for every objective, respectively: τ0f = τ0s = 1 (numc · M AXR) 1 (numc · M AXE) (15) (16) With numc as the number of cells in the map to solve, MAXR is the maximum amount of resources going through a cell may require, and MAXE is the maximum energy cost going through a cell may produce (in the worst case). The equations for Global Pheromone Updating are: τf (i, j) = (1 − ρ) · τf (i, j) + ρ/Ff τs(i, j) = (1 − ρ) · τs(i, j) + ρ/Fs (17) (18) Only the solutions inside the Pareto set will make the global pheromone updating once all the ants have finished of building paths in every iteration. The pseudocode for CHAC is as follows: Initialization of pheromone matrices with T0f and T0s For i=1 to NUM_iterations For a=1 to NUM_ants ps=build_path(a) /* Equations 1,2, [(3,4) or (6,7,8,9,10)] , 13,14 / evaluate(ps) / Equations 11,12 / if ps is non-dominated insert ps in Pareto_Set remove from Pareto_Set dominated solutions endif EndFor global_pheromone_updating / Equation 17,18 */ EndFor 4 Experiments and Results We would like to first emphasize that the values for the parameters and weights affect the results because parameters guide the search and balance the exploration and exploitation of the algorithm, and weights set the importance of every term in the heuristics and evaluation function. We fine-tuned both in order to obtain a good behaviour in almost all the maps; the values found were α = 1, β = 2, ρ = 0.1 and q0 = 0.4, in the three experiments we show in this section. The last value establishes a balance between exploration and exploitation, tending to exploration (but without abandoning exploitation altogether). The weights described in Equations 1, 2, 11 and 12 are set to give more importance to minimizing distance to target point and consumption of resources in the speed objective, and to give more importance to minimizing visibility and consumption of energy in the safety objective. The user can only decide the value for λ parameter, which gives relative importance to one objective over the other, so if it is near 1, finding fastest path would be more important and if it is near 0, the other way round. As every MO algorithm, CHAC yields a set of non-dominated solutions from which the user chooses using his own criteria, since, usually he only wants one solution. But in this algorithm the resulting Pareto set is small (about five to ten different solutions on average, depending on the map size) because it only searches in the region of the ideal Pareto front delimited by the λ parameter. We are going to represent in the results only 'the best' solution inside all the Pareto sets of 30 executions, considering better the solution with the smallest cost in the most significant criteria (depending on λ value). In order to clarify this concept, there is not a best path, but paths enough good from the military point of view. These paths have been selected by the military participation of the project taking into account tactical considerations and the features of every battlefield. CHAC have been implemented using the two state transition rules as two different al- gorithms. We are going to compare the results between both CHAC implementations (with the same parameter values), and both of them with a greedy approach which uses the same heuristic functions as cost functions (the pheromones has no equivalent in this algorithm), but using a dominance criterion to select the next node of the actual cell neighbourhood (it selects as next the node which dominates most of the others neighbours considering their cost in both objectives). This algorithm is quite simple, and sometimes does not even reach a solution because it gets into a loop situation (back and forth between a set of cells consuming all the resources or energy). We have performed experiments on three different maps; two of them have cells of a single type of terrain (with different heights) in order to avoid problems of visualization due to the black and white figures. We executed every CHAC implementation 30 times with each extreme value for λ parameter (0.9 and 0.1) in order to search for the fastest and the safest path (they cannot be 1 and 0 because always the two objectives must be considered). In every execution we chose the best solution (enough good from a military point of view) considering the objective we are minimizing and we make mean and standard deviation of them. 4.1 River and Forest Map The first scenario is a 15x15 map with a river flowing between the start and the target point and a patch of forest. There is no enemy. The best of fast paths found and the best of safe paths found are shown in Figure 2, marked by circles. As it can be seen in Table 1, the cost of the solutions for the fastest path have a small standard deviation which means possibly CHAC has reach a quasi-optimal solution (small search space and enough iterations and ants to solve it). Mean and standard deviation in the objective which is not being minimized are logically worse, because it has little importance. Fig. 2. River and Forest map (forest cells in dark gray and water cells in light gray). Fastest (left) and safest (right) paths found. Both use forest cells to avoid being seen, but the latter goes through them to optimize safety. Combined State Transition Rule Dominance State Transition Rule Fastest (λ = 0.9) Safest (λ = 0.1) Fastest (λ = 0.9) Safest (λ = 0.1) Greedy Ff Fs Ff Fs Ff Fs Ff Fs Best 22.198 85.768 25.704 61.189 22.214 80.075 27.274 61.833 Mean 22.207 85.941 28.505 67.296 22.595 88.405 28.724 66.365 ±0.022 ±0.436 ±1.993 ±3.360 ±0.555 ±3.127 ±2.637 ±2.881 NO SOLUTION Table 1. Results for River and Forest map. (500 iterations, 20 ants) DSTR implementation yields similar results even improving sometimes mean and standard deviation values in main objectives so, its solutions are more similar between executions (the algorithm has robustness). Figure 2 (left) shows the fastest path found, which goes zigzagging, because the algorithm usually moves to the most hidden cells when there are no enemies (in a radius that can be set up by the user and which is 10 in this experiment), even in the search for the fastest path, but with less relevance. Forest cells obstruct the line of sight so the unit tends to move near them while goes rather directly to the target point. Figure 2 (right) shows the safest path found which moves in a less straightforward way and hides by moving near, and even inside forest cells (hidden is very important in safe paths) although it mean bigger resource cost (forest cells have more resource cost assigned than flat terrain). 4.2 Flat Terrain with Walls Map The second map contains 30x30 cells, and represents a flat terrain with some 'walls', there is one enemy unit protecting the target cell (there are some cells affected by its weapons fire). The best of fast paths found and the best of safe paths found are shown in Figure 3. Cells with light gray border in the path are hidden to enemy. Fig. 3. Flat Terrain with Walls, and one enemy shooting in a zone near the target point (the enemy is marked with 'X', the cells surrounding it in light shades of gray are the zone of weapons impact and the other cells in light gray are the walls). Fastest (left) and safest (right) paths found by CHAC. Combined State Transition Rule Dominance State Transition Rule Fastest (λ = 0.9) Safest (λ = 0.1) Fastest (λ = 0.9) Safest (λ = 0.1) Greedy Ff Fs Ff Fs Ff Fs Ff Fs Ff Fs Best 36.500 133.000 38.000 112.700 39.000 133.100 Mean 37.220 141.973 45.650 121.210 43.350 240.660 48.500 113.400 59.530 123.980 ±1.216 ±2.904 ±3.681 ±10.425 ±1.592 ±85.254 ±10.174 ±8.119 46.7 322.9 Table 2. Results for the Flat Terrain with Walls map. (1000 iterations, 30 ants) As it can be seen in Table 2, CHAC (in the two implementations) outperforms the greedy algorithm by almost 25% in resource cost and 75% in safety costs. The reason for this is that the greedy path is rather straight, which means low resource expenses; but, at the same time, it means the enemy sees the unit all the time (it moves through uncovered cells) and the safety cost is dramatically increased (it depends too much on cell visibility). The cost for fast path, which also depends on visibility, is increased too. The standard deviation is small in CSTR, but it grows a little in values of safest paths which could mean a bigger exploitation factor is needed. Again, in the DSTR implementation, results are similar to those obtained with CSTR, but with better standard deviation; however, its performance is poorer (even quite bad in one case) for the objectives it is not minimizing, which supports the theory that a higher exploitation level is needed in this scenario. In Figure 3 (left), we can see the fastest path found; the unit moves in a curve mode in order to hide from enemy behind the walls (it consider the hidden of cells from the enemy cell). It surrounds this terrain elevations because resource cost of going through them is big. On the other hand, in Figure 3 (right) we can see the safest path found which moves surrounding the left side walls but remaining during more cells hide for the enemy. In both cases, the unit moves inside the zone affected by weapons choosing the less damaging cells, even in the safe case it arrives by a flank of the enemy unit where they have less fire power and this inflict less damage. It represents a good behaviour, very tactical in the case of safest path because attack to an enemy by the flank is one principle of military tactics. 4.3 Valleys and Mountains Map This 45x45 map represents some big terrain accidents, with clear zones representing peaks (the more higher the more clearer) and dark zones representing valleys or cliffs. There are two enemies located in both peaks and the target point is behind one of them. The best of fast paths found and the best of safe paths found are shown in Figure 4. Circle cells mark the paths, those with light gray border in the path are hidden to both enemies. Fig. 4. Valleys and Mountains, with 2 enemy units (marked with 'X') on watch. Fastest (left) and safest (right) path found. Combined State Transition Rule Dominance State Transition Rule Fastest (λ = 0.9) Safest (λ = 0.1) Fastest (λ = 0.9) Safest (λ = 0.1) Greedy Ff Fs Ff Fs Ff Fs Ff Fs Best 70.500 334.600 84.500 285.800 73.500 374.300 76.000 354.600 Mean 75.133 357.800 105.517 311.390 77.950 397.280 88.780 371.890 ±2.206 ±9.726 ±17.138 ±19.541 ±1.724 ±11.591 ±13.865 ±7.522 NO SOLUTION Table 3. Results for Valleys and Mountains map. (2000 iterations, 70 ants) As Table 3 shows, both implementations yield a low standard deviation, which means the algorithm is robust with respecto to initialization. CSTR implementation results may be nearer an optimal path, but in the DSTR case, the solutions have big costs (compared with the others), which means it needs more iterations or a greater exploitation factor to improve them. As in previous experiments, mean and standard deviation in the objectives that the algorithm is not minimizing are worse, because they have little importance. In this case, the differences between the cost of an objective when search minimizing this objective and when search minimizing the other are bigger than in the previous maps because the search space is bigger too. In Figure 4 (left) we can see the fastest path found, the unit goes in a rather straight way to the target point, but without considering the hidden of cells (from the position of both enemies) so the safety cost increases dramatically. On the other hand, in Figure 4 (right) we can see the safest path found which represents a curve (distance to target point has little importance) which increases speed cost, but the unit goes through many hidden cells. This behaviour is excellent from military tactical point of view. 5 Conclusions and Future Work In this paper we have described a MOACO algorithm called CHAC which tries to find the fastest and safest path, whose relative importance is set by the user, for a simulated military unit. This algorithm can use different state transition rules, and, in this papeer, two of them have been presented and tested. The first one combines heuristic and pheromone information of two objectives (Combined State Transition Rule, CSTR) and the second one is based on dominance over neighbours (Dominance State Transition Rule, DSTR). The algorithm using both state transition rules has been tested in several different sce- narios yielding very good results in a subjective assessment by the military staff of the project (Mr. Mill´an and Torrecillas) and being perfectly compatible with military tactics. They even offer good solutions in complicated maps in less time than a human expert would need. Moreover it is possible to observe an inherent emergent behaviour studying the solu- tions because it tends to be similar to those a real commander would, in many cases, take. In addition CHAC (using any state transition rule of the ones tested so far) outperforms a greedy algorithm. In the comparison between them, CHAC with CSTR yields better results in the same conditions, but CHAC with DSTR is more robust, yielding solutions that per- form similarly, independently of the random initial conditions. If we increase exploitation level or iterations of algorithm, DSTR approach offers similar results to CSTR. As future work, we will compare CHAC with path finding algorithms better than the simple greedy we have used here. From the algorithmic point of view, we will try to ap- proach more systematically parameter setting, and investigate its performance in dynamic environments where, for instance, the enemy can move and shoot on sight. We will also try to evaluate its performance in environments with a hard time constraint, that is, in envi- ronments where the algorithm cannot run for an unlimited amount of time, but a limited and small one, which is usually the case in real combat situations. From the implementa- tion point of view, it would be interesting to implement it within a real combat training simulator, that includes realistic values for most variables in this algorithm, including fuel consumption and casualties caused by projectile impact. We will also try to approach scenario design more systematically, trying to describe its difficulty by looking at different aspects. This will allow us to assess different aspects of the algorithm and relate them to scenario difficulty, finding out which parameter combination is better for each scenario. Acknowledgements This work has been developed within the SIMAUTAVA Project, which is supported by Universidad de Granada and MADOC-JCISAT of Ej´ercito de Tierra de Espana. Mr. Mill´an is a Lieutenant Colonel and Mr. Torrecillas is a Major of the Spanish Army Infantry Corps. References 1. M. Dorigo and G. Di Caro. The ant colony optimization meta-heuristic. In D. Corne, M. Dorigo, and F. Glover, editors, New Ideas in Optimization, pages 11 -- 32. McGraw-Hill, 1999. 2. M. Dorigo and T. Stutzle. The ant colony optimization metaheuristic: Algorithms, applications, and advances. In G.A. Kochenberger F. Glover, editor, Handbook of Metaheuristics, pages 251 -- 285. Kluwer, 2002. 3. Carlos A. Coello Coello, David A. Van Veldhuizen, and Gary B. Lamont. Evolutionary Algorithms for Solving Multi-Objective Problems. Kluwer Academic Publishers, 2002. 4. C. Garc´ıa-Mart´ınez, O. Cord´on, and F.Herrera. An empirical analysis of multiple objective ant colony optimization algorithms for the bi-criteria TSP. In ANTS 2004. Fourth International Workshop on. Ant Colony Optimization and Swarm Intelligence, number 3172 in LNCS, pages 61 -- 72. Springer, 2004. 5. Steffen Iredi, Daniel Merkle, and Martin Middendorf. Bi-criterion optimization with multi colony ant algorithms. In E. Zitzler, K. Deb, L. Thiele, C. A. Coello Coello, and D. Corne, editors, Proceedings of the First International Conference on Evolutionary Multi-Criterion Optimization (EMO 2001), volume 1993 of Lecture Notes in Computer Science, pages 359 -- 372, Berlin, 2001. Springer-Verlag. 6. B. Bar´an and M. Schaerer. ing on Applied http://www.scopus.com/scopus/inward/record.url?eid=2-s2.0-1442302509&partner=40&rel=R4.5.0. A multiobjective ant colony system for vehicle rout- IASTED International Multi-Conference 2003. problem with time windows. IASTED IMCAI, 97 -- 102, In Informatics, number 21 in pages
1810.06918	1	1810	2018-10-16T10:28:21	MoCaNA, un agent de n{\'e}gociation automatique utilisant la recherche arborescente de Monte-Carlo	[ "cs.MA" ]	Automated negotiation is a rising topic in Artificial Intelligence research. Monte Carlo methods have got increasing interest, in particular since they have been used with success on games with high branching factor such as go.In this paper, we describe an Monte Carlo Negotiating Agent (MoCaNA) whose bidding strategy relies on Monte Carlo Tree Search. We provide our agent with opponent modeling tehcniques for bidding strtaegy and utility. MoCaNA can negotiate on continuous negotiating domains and in a context where no bound has been specified. We confront MoCaNA and the finalists of ANAC 2014 and a RandomWalker on different negotiation domains. Our agent ouperforms the RandomWalker in a domain without bound and the majority of the ANAC finalists in a domain with a bound.	cs.MA	cs	MoCaNA, un agent de n´egociation automatique utilisant la recherche arborescente de Monte-Carlo C´edric L.R. Buron∗1,2, Zahia Guessoum†1,3, Sylvain Ductor‡4, and Olivier Roussel§5 1LIP6, Sorbonne Universit´e, Paris, France 2Thales Research and Technology, Palaiseau, France 3CReSTIC, Universit´e de Reims, Reims, France 4Universidade Estadual do Cear´a, Fortaleza, Br´esil 5Kyriba Corp, San Diego, USA 11 octobre 2018 R´esum´e La n´egociation automatique est un sujet qui suscite un int´eret croissant dans la recherche en IA. Les m´ethodes de Monte-Carlo ont quant `a elles v´ecu un grand essor, notamment suite `a leur utilisation sur les jeux `a haut facteur de branchement tel que le go. Dans cet article, nous d´ecrivons un agent de n´egociation automatique, Monte- Carlo Negotiating Agent (MoCaNA) dont la strat´egie d'offre s'appuie sur la recherche arborescente de Monte Carlo. Nous munissons cet agent de m´ethodes de mod´elisation de la strat´egie et de l'utilit´e adverse. MoCaNA est capable de n´egocier sur des do- maines de n´egociation continus et dans un contexte o`u aucune borne n'est sp´ecifi´ee. Nous confrontons MoCaNA aux agents de l'ANAC 2014 et `a un RandomWalker sur des domaines de n´egociation diff´erents. Il se montre capable de surpasser le Random- Walker dans un domaine sans borne et la majorit´e des finalistes de l'ANAC dans un domaine avec borne. Mots-cl´es : Monte-Carlo, N´egociation automatique, Agent ∗[email protected] †[email protected] ‡[email protected] §[email protected] Vingt-sixi`emes Journ´ees Francophones sur les Syst`emes Multi-Agents (JFSMA'18), M´etabief, France, p. 85 -- 94, oct. 2018 1 Abstract Automated negotiation is a rising topic in Artificial Intelligence research. Monte Carlo methods have got increasing interest, in particular since they have been used with success on games with high branching factor such as go. In this paper, we describe an Monte Carlo Negotiating Agent (MoCaNA) whose bidding strategy relies on Monte Carlo Tree Search. We provide our agent with op- ponent modeling tehcniques for bidding strtaegy and utility. MoCaNA can negotiate on continuous negotiating domains and in a context where no bound has been spec- ified. We confront MoCaNA and the finalists of ANAC 2014 and a RandomWalker on different negotiation domains. Our agent ouperforms the RandomWalker in a domain without bound and the majority of the ANAC finalists in a domain with a bound. Keywords: Monte Carlo, Automated Negotiation, Agent 1 Introduction La n´egociation est une forme d'interaction dans laquelle un groupe d'agents ayant des conflits d'int´eret et un d´esir de coop´erer essaient de trouver un accord mutuellement acceptable sur un objet de n´egociation. Ils explorent `a cette fin les solutions selon un protocole pr´ed´efini. La question de l'automatisation de la n´egociation, bien connue dans les domaines ´economiques depuis l'av`enement des applications de e-commerce, a re¸cu une attention toute particuli`ere dans le champ de l'intelligence artificielle et des syst`emes multi-agents. De nombreux frameworks ont ´et´e propos´es [17] : ench`eres, r´eseaux contractuels, ´equilibres g´en´eraux ; ils permettent de n´egocier selon plusieurs modalit´es. Les n´egociations peuvent etre caract´eris´ees selon divers aspects portant notamment sur l'ensemble des participants (bilat´eral ou multilat´eral), les pr´ef´erences des agents (lin´eaires ou non), les attributs sur lesquels la n´egociation porte (discrets ou continus), ou encore les caract´eristiques du proto- cole (born´e ou non par le temps ou le nombre de tours). Chaque agent utilise une strat´egie pour ´evaluer l'information re¸cue et faire des offres. Plusieurs strat´egies ont ´et´e propos´ees comme [8, 5, 19]. Elles peuvent etre fixes ou adaptatives. Cependant, la plupart de ces strat´egies reposent sur une borne connue (en temps ou en tours). Dans les applications quotidiennes, par exemple lorsque nous cherchons `a acqu´erir un bien de consommation non critique, il est commun que le protocole ne propose aucune borne, et que chaque agent ignore le moment o`u il coupera court `a la n´egociation. Dans cet article, nous ´etudions MoCaNA (Monte-Carlo Negotiating Agent), un agent de n´egociation automatique con¸cu pour le protocole de bargaining (ou marchandage). Dans ce protocole, deux agents s'´echangent des offres de mani`ere `a trouver un accord. La n´egociation s'interrompt lorsqu'un des deux adversaires accepte l'offre qui lui est faite par son adversaire ou qu'il la rejette. MoCaNA a pour particularit´e de ne faire aucune pr´esupposition sur : 1) la lin´earit´e des pr´ef´erences, 2) le caract`ere continu ou discret de 2 l'espace de n´egociation ni 3) l'imposition d'une borne par le protocole. Il exploite pour cela des connaissances issues des domaines du General Game Playing et de l'Apprentissage Automatique. Afin de d´ecider de la valeur d'une offre, notre agent s'appuie sur la recherche arborescente de Monte-Carlo (MCTS), une heuristique qui a ´et´e utilis´ee avec succ`es dans de nombreux jeux comme le go, qu'il combine avec de la mod´elisation d'adversaires pour plus d'efficacit´e. Le choix de cette heuristique est guid´e par deux ´el´ements. D'abord, MCTS s'est montr´e efficace dans des jeux tr`es diff´erents, ce qui lui a valu de nombreux succ`es en General Game Playing. Il est aussi adapt´e aux jeux `a haut facteur de branchement. Il constitue donc une bonne alternative pour les domaines de n´egociation complexes dont le facteur de branchement `a chaque proposition est l'ensemble des propositions possibles. Cet article est oorganis´e comme suit : la section 2 introduit les travaux de la litt´erature traitant de n´egociation automatique et des m´ethodes de Monte-Carlo. Nous proposons ensuite une mod´elisation de la n´egociation comme un jeu dans la section 3. La section 4 pr´esente les fonctionnalit´es de MoCaNA. Nous pr´esentons les r´esultats de la confrontation de MoCaNA aux finalistes de l'ANAC 2014 et au RandomWalker dans la section 5. Nous r´esumons ces contributions et donnons des pistes d'am´elioration de notre travail dans la section 6. 2 Travaux connexes MoCaNA se trouve au croisement des domaines de la n´egociation automatique et des techniques de Monte-Carlo appliqu´ees aux jeux. Nous donnons dans cette section une introduction `a ces deux domaines. 2.1 N´egociation automatique Les architectures des agents de n´egociation automatique impliquent trois fonctionnalit´es [1] : la strat´egie d'offres d´efinit les offres que l'agent fait `a son adversaire ; la strat´egie d'acceptation d´efinit si l'agent accepte la proposition qui lui a ´et´e faite ou s'il fait une contre-offre ; la mod´elisation d'adversaires permet de mod´eliser certains aspects de l'adversaire, comme sa strat´egie d'offre, son utilit´e ou sa strat´egie d'acceptation et est utilis´e pour am´eliorer l'efficacit´e de la strat´egie d'offres de l'agent et parfois celle de sa strat´egie d'acceptation. 2.1.1 Strat´egie d'offres Les strat´egies d'offres utilis´ees dans le cadre de n´egociations complexes s'appuient sur deux familles de techniques, que nous pr´esentons par la suite : les heuristiques coupl´ees au domaine et les algorithmes g´en´etiques. Les heuristiques s'appuient g´en´eralement sur des tactiques [8], qui consistent `a faire des concessions selon un ´el´ement de la n´egociation comme le temps ´ecoul´e, la raret´e de 3 la denr´ee n´egoci´ee, ou les concessions faites par l'adversaire. Il est possible d'adapter ces tactiques ou de s´electionner une offre ayant une utilit´e gagnant-gagnant s'il en existe une. La majorit´e de ces strat´egies [8] reposent sur la connaissance de la borne, ce qui les rend inapplicables dans notre contexte. Les algorithmes g´en´etiques [5] s´electionnent les offres qui maximisent une fonction ob- jectif d´etermin´ee, les croisent et les modifient de mani`ere `a les am´eliorer. Ces strat´egies g´en`erent un certain nombre d'offres parmi lesquelles l'agent doit choisir celle qui sera en- voy´ee `a l'adversaire. Elles ont obtenu de bons r´esultats, sans toutefois vaincre les agents utilisant des heuristiques. 2.1.2 Strat´egies d'acceptation On distingue deux principales cat´egories de strat´egies d'acceptation [3] : les strat´egies myopes et les strat´egies optimales. Les strat´egies myopes reposent sur les offres de l'ad- versaire et celles de l'agent. Il peut s'agir d'accepter toute offre d´epassant un certain seuil, toute offre meilleure que la derni`ere faite par l'agent lui-meme ou que la prochaine g´en´er´ee par strat´egie d'offre. Il est aussi possible de combiner ces strat´egies. Les strat´egies opti- males [3] exploitent un mod`ele de la strat´egie adverse afin de d´eterminer la probabilit´e d'obtenir une meilleure offre. 2.1.3 Mod´elisation de l'adversaire La majorit´e des m´ethodes de mod´elisation d'adversaire utilis´ees en n´egociation auto- matique ont ´et´e analys´ees par Baarslag et ses coll`egues [2]. Elles permettent de mod´eliser la strat´egie d'offre, la strat´egie d'acceptation, l'utilit´e de l'adversaire, la borne qu'il s'est fix´e, et son prix de r´eserve le cas ´ech´eant. Deux familles de techniques sont adapt´ees `a la mod´elisation de strat´egies d'offres adaptatives et non born´ees : les r´eseaux de neurones et les techniques reposant sur l'ana- lyse de s´eries temporelles. Les approches bas´ees sur l'analyse de s´eries temporelles sont diverses. Parmi elles, la r´egression de processus gaussien est stochastique. Elle s'appuie sur les mouvements pr´ec´edents pour pr´edire une densit´e de probabilit´e pour le tour suivant. Introduite par Rasmussen et Williams [15], elle a ´et´e utilis´ee avec succ`es par Williams et al. [20]. L'aspect stochastique de cette m´ethode permet de g´en´erer des propositions diff´erentes `a chaque ´etape de simulation de Monte Carlo Tree Search (MCTS), selon la gaussienne pr´edite par la r´egression. C'est cette m´ethode que nous adoptons. L'utilit´e d'un adversaire est g´en´eralement consid´er´ee comme la somme pond´er´ee de fonctions pour chaque attribut `a valeurs dans [0, 1] [2]. Deux familles de m´ethodes sont utilis´ees. La premi`ere s'appuie sur la fr´equence d'apparition de chaque valeur parmi les propositions de l'adversaire. Elle suppose que les valeurs propos´ees le plus r´eguli`erement par l'adversaire sont celles qu'il pr´ef`ere, et que les attributs variant le plus souvent sont ceux qui ont le moins d'importance pour lui. Ces approches ne sont cependant valables que dans le cas o`u le domaine de n´egociation est discret. En effet, l'extension au cas continu requiert la d´efinition d'une fonction de distance d´ependant du domaine. La seconde famille de 4 m´ethodes repose sur l'apprentissage bay´esien. Elle est adapt´ee au cas continu. La m´ethode pr´esent´ee par Hindriks et Tykhonov [10] repose sur la g´en´eration d'hypoth`eses. Chaque hypoth`ese est constitu´ee d'un ordonnancement des attributs, et d'une fonction d'utilit´e simple (lin´eaire ou lin´eaire par morceaux) pour chacun d'entre eux. Le poids associ´e `a chaque attribut est calcul´e en fonction de l'ordre qui lui est attribu´e. La fonction d'utilit´e estim´ee est la somme des hypoth`eses pond´er´ee par leur probabilit´e. La strat´egie d'acceptation d'un adversaire peut etre apprise de deux mani`eres. On peut d'abord faire la supposition que l'adversaire acceptera une offre sous certaines condi- tions d´ependant de sa strat´egie d'offre et/ou de sa fonction d'utilit´e [2]. Dans ce cas, il est possible de la d´eduire des mod`eles ci-dessus sans faire de calcul suppl´ementaire. Dans le cas o`u l'agent ne mod´elise pas les ´el´ements pr´ec´edemment d´ecrits, il est aussi possible d'utiliser des r´eseaux de neurones [7]. Cela demande cependant un calcul suppl´ementaire, potentiellement couteux. 2.2 M´ethodes de Monte Carlo Les m´ethodes de Monte-Carlo sont utilis´ees comme heuristiques pour les jeux. Elles attribuent `a chaque mouvement possible un score calcul´e au terme d'une ou de plusieurs simulations. Kocsis et Szepevsv´ari [12] proposent une m´ethode hybridant la construction d'un arbre de jeu, m´ethode qui a fait ses preuves, et les m´ethodes de Monte-Carlo. Chaque noeud de l'arbre de jeu garde en m´emoire les scores obtenus lors des simulations o`u il a ´et´e jou´e, ainsi que le nombre de simulations faites dans la branche partant du noeud. Pendant l'exploration de cet arbre, les branches privil´egi´ees sont celles qui ont ´et´e peu explor´ees et celles qui ont le plus haut score dans les simulations. Cette m´ethode, nomm´ee recherche arborescente de Monte-Carlo (MCTS 1) s'est vue am´elior´ee au moyen de nombreuses ex- tensions [4]. MCTS se d´ecompose en quatre parties. (1) La premi`ere ´etape consiste `a parcourir l'arbre d´ej`a constitu´e selon une strat´egie pr´ed´efinie. `A chaque noeud, on d´ecide s'il faut s´electionner une branche de niveau inf´erieur ou en explorer une nouvelle. (2) Dans le second cas, l'arbre se voit pourvu d'un nouveau noeud, g´en´er´e selon une strat´egie d'expansion. (3) Ensuite, on simule le jeu jusqu'`a un ´etat final. (4) On calcule les scores et on les r´etropropage sur les noeuds de l'arbre ayant ´et´e explor´es. Les m´ethodes de Monte-Carlo ont gagn´e en popularit´e suite `a leur succ`es dans les jeux `a fort facteur de branchement, notamment le Go. En particulier, AlphaGo [18], qui a vaincu l'un des meilleurs joueurs au monde, utilise les m´ethodes de Monte-Carlo coupl´ees `a de l'apprentissage profond. `A notre connaissance, seuls de Jonge et Zhang [6] ont utilis´e MCTS dans le cadre de la n´egociation automatique. Ils se sont concentr´es sur des domaines de n´egociations petits, et o`u la fonction d'utilit´e adverse est facilement inf´erable, comme la r´epartition d'un dollar entre deux joueurs. Nous nous concentrons sur des domaines plus complexes, o`u la fonction d'utilit´e adverse ne peut etre facilement inf´er´ee et o`u l'espace de n´egociation peut etre grand, voire infini. 1. Monte Carlo Tree Search 5 D´efinition 1 (Bargaining) Un bargaining est un triplet B = H, A, (ui)i∈(cid:74)1,2(cid:75) v´erifiant : (cid:16) (cid:17) 3 Jeu et n´egociation La recherche arborescente de Monte-Carlo est une m´ethode appliqu´ee aux jeux extensifs. Dans cette section, nous montrons comment il est possible d'assimiler la n´egociation `a un tel jeu. Nous commen¸cons par associer chaque aspect de la n´egociation `a un ´el´ement d'un jeu. Nous d´ecrivons ensuite les particularit´es de la n´egociation, qui empechent l'utilisation des strat´egies les plus r´epandues dans MCTS. Nous concluons cette section en donnant d'autres formalismes possibles pour la n´egociation et en expliquant les relations qu'ils entretiennent avec celui que nous avons choisi. 3.1 Notre formalisme Un jeu extensif [14] est compos´e d'un ensemble de joueurs, de la description des histo- riques de jeu possibles, d'une fonction indiquant le tour de chaque joueur et d'un profil de pr´ef´erence. En s'appuyant sur cette d´efinition, il est possible de d´efinir un bargaining B comme un jeu sous forme extensive. 1. joueurs : il y a deux joueurs : un acheteur (joueur 1) et un vendeur (joueur 2), 2. historique : l'historique de la n´egociation est compos´e des messages que les agents s'envoient : les propositions faites par les agents, et les accept et reject. Les historiques terminaux sont les historiques infinis et les historiques se terminant par accept ou par reject. Chaque message est compos´e d'une paire (α, c) o`u α est l'acte de langage (performatif) du message et, c est le contenu du message, i.e. une liste de couples (k, v) o`u k est la cl´e d'un attribut du domaine de n´egociation et v la valeur correspondante. On s´epare tout historique en deux, chacun correspondant aux actions d'un joueur : hi = (α, c)i 3. tour de jeu : la fonction tour fonctionne selon la parit´e. On suppose que c'est l'acheteur (le joueur 1) qui commence. Ainsi, ∀h ∈ H, joueur(h) = 2− (h mod 2) o`u h est la taille de h, 4. pr´ef´erence : l'ordre sur les situations terminales de chaque joueur est induit par une fonction d'utilit´e ui qui associe son utilit´e pour chaque historique terminal. Cette fonction associe une utilit´e `a chaque accord possible trouv´e, ainsi qu'au cas de rejet et au cas d'historique infini 2. La n´egociation n'est pas un jeu combinatoire classique, comme peuvent l'etre les ´echecs ou le go. La premi`ere diff´erence entre ces jeux et la n´egociation tient au fait que cette derni`ere est un jeu `a somme non nulle. Les agents cherchent `a trouver un accord mutuel qui soit profitable `a chacun. Cela est particuli`erement vrai dans des domaines complexes, 2. Notons que ces deux derni`eres valeurs peuvent etre diff´erentes, par exemple si l'agent consid`ere qu'il alloue des ressources `a la n´egociation. 6 ayant de nombreux attributs, o`u une solution peut etre bien meilleure que la solution par d´efaut o`u chaque agent re¸coit son utilit´e de r´eserve, i.e. quand les agents ne trouvent pas d'accord. Ensuite, la n´egociation est un jeu `a information incompl`ete, c'est `a dire que le type des joueurs, ici leur profil de pr´ef´erence, est inconnu et fait meme g´en´eralement l'objet d'une mod´elisation. Ces deux particularit´es rendent inutilisable l'Upper Confidence Tree [12] utilis´e pour les jeux combinatoires comme le go car il est fait pour les jeux combina- toires. Enfin, le domaine de la n´egociation est tr`es particulier. Il peut etre cat´egoriel (e.g. couleur) mais ´egalement num´erique voire continu. Cela a des cons´equences sur l'explora- tion de l'arbre, en particulier sur le crit`ere d´ecidant de l'expansion d'un nouveau noeud. Malgr´e ces difficult´es d'adaptation, les succ`es qu'a remport´e MCTS dans les jeux complexes semblent indiquer qu'il pourrait s'av´erer etre une bonne strat´egie pour la n´egociation dans un contexte complexe. 3.2 Autres formalismes Bien que nous utilisions le formalisme des jeux extensifs pour mod´eliser la n´egociation, d'autres mod´elisations sont possibles, notamment en utilisant les jeux bay´esiens et les jeux stochastiques. Les jeux `a information incompl`ete peuvent etre mod´elis´es en utilisant le formalisme des jeux bay´esiens [16]. Ces jeux sont traditionnellement utilis´es pour mod´eliser les jeux `a information incompl`ete, mais supposent en g´en´eral la connaissance d'une probabilit´e a priori sur les types possibles de l'adversaire. Notons que dans le cadre que nous nous sommes fix´es, il n'existe pas de telle distribution a priori. La prise en compte de l'information r´ev´el´ee par l'adversaire au cours de la n´egociation, qui permet d'´etablir des probabilit´es sur le profil de pr´ef´erence, se fait au cours de la n´egociation en utilisant la mod´elisation de l'adversaire. Un autre formalisme possible est la mod´elisation par un jeu stochastique [11]. Les jeux stochastiques sont une g´en´eralisation des processus markoviens. Le bargaining peut etre vu comme un processus de d´ecision markovien (MDP), o`u chaque action de l'agent g´en`ere une r´eaction de son adversaire, amenant une transition vers un autre ´etat du jeu. Dans ce nouvel ´etat, les actions possibles restent les memes, mais l'utilit´e induite par l'acceptation de la proposition de l'adversaire change. L'exploitation de ce formalisme se fait par une exploration des transitions possibles et une pond´eration par leur probabilit´e. MoCaNA cherche `a ´evaluer ces probabilit´es en utilisant la mod´elisation de l'adversaire et estime l'utilit´e attendue pour les diff´erentes transitions possibles, i.e. les choix de l'adversaire, au moyen de MCTS. Notons que cette m´ethode s'est montr´ee particuli`erement efficace pour la planification dans les MDP, comme le montre notamment [12]. 4 MoCaNA Comme nous l'avons expliqu´e dans la section pr´ec´edente, la n´egociation est un jeu particulier. Il est donc n´ecessaire d'ajuster les heuristiques pour les jeux `a ces particularit´es. 7 Dans cette section, nous pr´esentons notre agent de n´egociation automatique exploitant MCTS. L'architecture g´en´erale des diff´erents modules de notre agent est pr´esent´ee dans la Figure 1. La strat´egie d'offre impl´emente MCTS et utilise le module de mod´elisation d'adversaire, qui comporte deux sous-modules : l'un pour l'utilit´e de l'adversaire, l'autre pour sa strat´egie d'offre. Le dernier module, celui de strat´egie d'acceptation, effectue une comparaison entre la proposition de l'adversaire et la proposition g´en´er´ee par la strat´egie d'offre. Les diff´erents sous-modules de l'agent ainsi que les interactions entre eux sont d´ecrits dans la suite de cette section. Strat´egie d'offres S´election Expansion Simulation R´etropropagation Strategie (R´egression de processus gaussien) Utilit´e (Appren- tissage Bayesien) Mod´elisation d'adversaire S t r a t e g i e d ' a c c e p t a t i o n Figure 1 -- Interaction entre les modules de MoCaNA 4.1 Mod´elisation d'adversaire Afin de d'am´eliorer l'efficacit´e de MCTS, il est n´ecessaire de disposer d'un mod`ele de sa strat´egie d'offre ainsi qu'un mod`ele de son profil de pr´ef´erence. 4.1.1 Mod´elisation de la strat´egie d'offre Le but est de pr´edire une future proposition faites par l'adversaire au tour x∗ de la n´egociation. Nous utilisons la r´egression de processus gaussiens [15] qui permet de g´en´erer une gaussienne centr´ee en la valeur pr´edite par l'algorithme et dont l'´ecart type repr´esente l'incertitude du mod`ele. On commence par calculer la matrice de covariance K, qui repr´esente la proximit´e entre les tours (xi)i∈(cid:74)1,n(cid:75) de la s´equence, en nous reposant sur une fonction de covariance, aussi on calcule ensuite la distance entre le tour dont on veut estimer la proposition x∗ et les tours pr´ec´edents : k(xn, x1) . . . k(xn, xn) K∗ = (k(x∗, x1), . . . , k(x∗, xn)) (2) 8 (1) appel´ee kernel k. Soit K =  k(x1, x1) ...  . . . k(x1, xn) ... La r´egression de processus gaussien fait la supposition que chacune de ces valeurs est la composante d'une gaussienne multivari´ee. On en d´eduit : y∗ = K∗K−1y σ2∗ = Var(y∗) = K∗∗ − K∗K−1K(cid:62) ∗ (3) (4) o`u K∗∗ = k(x∗, x∗). Nous identifions y∗ `a une variable al´eatoire gaussienne de moyenne y∗ et d'´ecart type σ∗. Une des composantes les plus importantes de la r´egression de processus gaussien est la d´etermination du kernel. Les plus communs sont les fonctions `a base radiale, les fonctions rational quadratic, le kernel de Mat´ern et le kernel exponential sine squared. Ces kernels servent `a d´eterminer la distance entre deux tours de n´egociation. Nous avons test´e ces quatre kernels sur plusieurs n´egociations entre agents. La Table 1 donne les r´esultats de la r´egression de processus gaussien pour chacun de ces kernels. Nous avons fait ces tests dans le contexte de l'ANAC 2014, o`u les pr´ef´erences ne sont pas lin´eaires, et avec les finalistes de l'ANAC. Nous avons g´en´er´e al´eatoirement 25 sessions, et avons mod´elis´e les deux agents, obtenant ainsi 50 mod´elisations au total. Chaque offre de chaque s´erie est pr´edite en fonction des pr´ec´edentes et sert `a pr´edire les suivantes. La table montre la distance euclidienne moyenne entre les propositions faites et le r´esultat de l'apprentissage. Plus la valeur est basse, plus la pr´ediction est proche de la v´eritable s´erie. Kernel Fonction `a base radiale Rational quadratic Mat´ern Exp. sine squared Dist. moyenne 43.288 17.766 43.228 22.292 Table 1 -- Distance moyenne entre les propositions et les pr´edictions de la r´egression de processus gaussien selon le kernel Le kernel donnant le meilleur r´esultat est le Rational quadratic. C'est donc celui-ci que notre agent utilise. 4.1.2 Mod`ele du profil de pr´ef´erence L'apprentissage bay´esien d´ecrit dans [10] suppose seulement que l'agent fait une conces- sion `a un rythme constant, une supposition faible par rapport aux autres m´ethodes. laires. Une fonction t de [a, b] ⊂ R dans [0, 1] est dite triangulaire si et seulement si : On approxime l'utilit´e de l'adversaire par une somme pond´er´ee de fonctions triangu- -- t est affine, et t(a) = 0 et t(b) = 1 ou t(a) = 1 et t(b) = 0 ou -- il existe c dans [a, b] tel que t est affine sur [a, c] et sur [c, b], t(a) = 0, t(b) = 0, t(c) = 1. 9 La m´ethode se divise en deux ´etapes. D'abord, un certain nombre d'hypoth`eses sont g´en´er´ees. Ces hypoth`eses sont compos´ees d'une somme pond´er´ee de fonctions triangulaires. Chaque attribut se voit attribuer un poids et une fonction triangulaire. L'utilit´e estim´ee d'un adversaire est la somme de ces hypoth`eses pond´er´ee par la pro- babilit´e de chacune. Cette m´ethode a pour avantage de ne pas faire de supposition sur la strat´egie de l'adversaire, mais elle suppose qu'il fait des concessions `a un rythme globale- ment lin´eaire. 4.1.3 Mod`ele de strat´egie d'acceptation La strat´egie d'acceptation des agents est mod´elis´ee de mani`ere tr`es simple : on consid`ere qu'un agent accepte la derni`ere proposition faite par son adversaire si et seulement si son utilit´e est meilleure que l'utilit´e g´en´er´ee par la strat´egie d'offre. Cette mod´elisation, ´evoqu´ee par [2], est peu couteuse et permet de ne pas alourdir les calculs d´ej`a nombreux de l'agent. 4.2 Strat´egie d'offres bas´ee sur MCTS Les m´ethodes de Monte-Carlo sont tr`es adaptatives, et ont obtenu de bons r´esultats dans diff´erents jeux, y compris des jeux `a grand facteur de branchement. Dans cette section, nous introduisons une strat´egie d'offres bas´ee sur MCTS. 4.2.1 MCTS sans ´elagage Comme nous l'avons expliqu´e, MCTS est un algorithme g´en´eral, qui repose sur un certain nombre de strat´egies. `A chaque fois que l'agent doit prendre une d´ecision, il g´en`ere un nouvel arbre et l'explore en utilisant MCTS. L'impl´ementation la plus commune de MCTS est l'Upper Confidence Tree (UCT). Cette m´ethode a notamment donn´e de tr`es bons r´esultats pour le Go. N´eanmoins, lors de la phase de s´election, cette impl´ementation ´etend un nouveau noeud d`es lors que tous les enfants d'un arbre n'ont pas ´et´e ´etendus. Or, dans le cas d'un domaine de n´egociation contenant un attribut infini, le nombre de fils possibles pour un noeud est infini, puisqu'il peut prendre toutes les valeurs possibles pour cet attribut. L'algorithme ´etendra donc ind´efiniment de nouveaux noeuds sans jamais explorer l'arbre en profondeur. Si l'arbre est tr`es large, si le nombre de noeuds possibles est plus grand que nombre de simulations, la meme situation apparaıtra. Dans le contexte de la n´egociation, il est donc n´ecessaire de d´efinir une impl´ementation de MCTS diff´erente : S´election Pour cette ´etape, nous utilisons le progressive widening. Le crit`ere du pro- gressive widening peut s'exprimer ainsi : un nouveau noeud est ´etendu si et seulement si : p ≥ nc nα (5) o`u np est le nombre de fois o`u le parent a ´et´e simul´e, nc est le nombre de fils qu'il a, et α est un param`etre du mod`ele. S'il n'y a pas d'expansion, nous s´electionnons 10 le noeud i qui maximise : (cid:115) ln(n) ni + 1 Wi = si ni + 1 + C × nα (6) o`u n est le nombre total de simulations faites jusqu'alors, si est le score du noeud i et C est ´egalement un param`etre du mod`ele. Expansion La valeur d'un noeud ´etendu est quant `a elle choisie de mani`ere al´eatoire, avec une distribution uniforme sur le domaine de n´egociation. Simulation Nous utilisons le mod`ele de strat´egie de l'adversaire, afin de rendre les simulations plus pertinentes. Nous utilisons la r´egression de processus gaussien et l'apprentissage bay´esien. R´etropropagation L'´etape de r´etropropagation utilise aussi le mod`ele de l'utilit´e de l'adversaire, et r´epercute sur les noeuds l'utilit´e de l'agent et l'utilit´e de son adver- saire selon la mod´elisation d'utilit´e. Enfin, MCTS fait des simulations et calcule l'utilit´e attendue de toute la n´egociation. Il a tendance `a sous-estimer la valeur de la proposition qu'il ressort, et la probabilit´e que l'adversaire l'accepte. Nous renvoyons donc l'offre b∗ parmi les fils de la racine maximisant : (cid:18) utility(b) + score(b) (cid:19) b∗ = argmaxb∈racine.f ils 4.2.2 ´Elagage 2 (7) Afin de n'explorer que les noeuds donnant une utilit´e importante `a l'agent, il est possible d'utiliser des connaissances sur le jeu pour ´elaguer certaines parties de l'arbre et ne pas les d´evelopper. Nous consid´erons deux cas, un ´elagage dit (cid:28) fixe (cid:29) et un ´elagage dit (cid:28) variable (cid:29). Dans le cas de l'´elagage fixe, nous fixons `a notre agent une utilit´e minimale attendue, i.e. une utilit´e de r´eserve. Au sein de la simulation, lorsque MoCaNA produit une propo- sition, l'utilit´e de cette proposition est calcul´ee en utilisant le profil d'utilit´e de MoCaNA. Si cette utilit´e est inf´erieure `a la valeur seuil, la branche est supprim´ee. Lorsque l'agent n'est pas capable de d´eterminer un prix de r´eserve, il est possible de proc´eder `a un ´elagage variable d´ependant des propositions de l'adversaire. Dans ce cas, l'agent garde en m´emoire la proposition de l'adversaire ayant pour lui la meilleure utilit´e. Toute proposition g´en´er´ee par MoCaNA dans une simulation ayant une utilit´e inf´erieure est ´elagu´ee. Avec cette politique, l'agent ne fait que des propositions meilleures que celles que son adversaire a faites. 4.3 Architecture logicielle Notre agent a ´et´e d´evelopp´e en java. Nous l'avons con¸cu de fa¸con `a ce qu'il soit mo- dulaire, et `a ce qu'il soit d´ecoupl´e du framework sur lequel les exp´erimentations ont ´et´e 11 Mod´elisation ǡ Strat´egie ǈ Connecteur Donn´ees Genius Mod´elisation (cid:255) Utilit´e R Strat´egie 3 d'offre (MCTS) Agent R R Figure 2 -- Architecture logicielle des modules de MoCaNA r´ealis´ees. Nous avons pour cela construit un module permettant `a notre agent de s'in- terfacer avec Genius. L'agent est organis´e sous la forme d'interfaces impl´ement´ees par les classes de chaque module, et qui sont traduites pour Genius le cas ´ech´eant. On re- trouve dans l'architecture de MoCaNA les diff´erents ´el´ements d´ecrits ci-dessus : la strat´egie d'offre incluant MCTS et les concepts allant avec (e.g. notions d'arbre, de noeud), un mo- dule de mod´elisation de la strat´egie adverse et un mod`ele de l'utilit´e. Nous n'avons pas impl´ement´e de module pour la mod´elisation de la strat´egie d'acceptation. Cette architec- ture est repr´esent´ee dans la Figure 2. Les R repr´esentent les messages envoy´es et re¸cus par l'agent. Les simulations de MCTS ´etant tr`es gourmandes en terme de puissance de calcul, nous les avons parall´elis´e. L'aspect stochastique de notre mod´elisation de strat´egie d'offre assure que deus simulations lanc´ees du meme point n'auront pas le meme r´esultat, tout en privil´egiant les valeurs les plus probables. La r´egression de processus gaussien, qui repose sur le calcul matriciel, a ´et´e d´evelopp´ee en utilisant la librairie JAMA. Les param`etres du kernel de cette m´ethodes sont quant `a eux optimis´es en utilisant la librairie Apache Commons Math. 5 Exp´erimentations Pour ´evaluer notre agent, nous utilisons le framework Genius [13], qui permet de cr´eer des sessions de n´egociation et des tournois. Nous confrontons MoCaNA au RandomWalker ainsi qu'aux agents de l'ANAC 2014. 5.1 Protocole exp´erimental Genius permet de n´egocier sur des attributs entiers ou discrets, mais pas encore continus. Nous utilisons le domaine de n´egociation de l'ANAC 2014 d´ecrit par [9]. Les comp´etitions de 2015 et de 2016 se sont concentr´ees respectivement sur la n´egociation mul- tilat´erale, et sur les smart grids. `A notre connaissance, les agents de n´egociation de l'ANAC 2017 n'ont pas encore ´et´e mis `a disposition. Les attributs sont entiers, et varient de 0 `a 10. Plusieurs variantes ont ´et´e propos´ees allant de 10 attributs `a 50 attributs. Nous nous 12 focalisons sur le contexte `a 10 attributs. Le contexte utilis´e n'a aucun taux de r´eduction. L'utilit´e est une somme pond´er´ee de fonctions non lin´eaires sur chaque attribut. L'utilit´e de r´eserve des agents (dans le cas o`u ils ne trouvent pas d'accord) est 0, c'est `a dire la valeur minimale. Nous fixons en outre la borne de chaque session de n´egociation `a 1 heure. Apr`es ce temps, la n´egociation s'interrompt et les agents obtiennent leur utilit´e de r´eserve, c'est-`a- dire 0. Nous calibrons notre mod`ele de mani`ere empirique. Lorsque notre agent peut choisir parmi 200 propositions, il s'en trouve g´en´eralement une qui procure `a la fois une haute utilit´e pour lui et son adversaire. Afin que l'arbre soit ´egalement explor´e en profondeur, nous lui laissons assez de temps pour qu'il g´en`ere en moyenne 50.000 simulations. Ainsi, chaque pour faire une proposition, notre agent prend 3 minutes. Nous obtenons α = 0.489. 5.2 R´esultats Nous distinguons deux contextes de n´egociation : un contexte o`u la borne est pr´esente et connue des adversaires, face aux agents de l'ANAC 2014, et un o`u elle est assez loin pour etre consid´er´ee comme inexistante, face au RandomWalker. 5.2.1 ANAC 2014 La Table 2 montre les r´esultats de notre agent face aux finalistes de l'ANAC 2014, d´ecrits par [9], ordonn´es selon leur r´esultats `a l'ANAC 2014 dans la cat´egorie utilit´e individuelle. Les scores repr´esentent l'utilit´e moyenne de notre agent et de son adversaire moyenn´e sur 20 n´egociations, 10 avec chaque profil. Plus le score est haut, plus l'utilit´e de l'agent est ´elev´ee et mieux l'agent a r´eussi la n´egociation. Le nombre suivant le score pr´ec´ed´e du signe ± est l'´ecart type de la s´erie. L'utilit´e moyenne est calcul´ee en ne prenant en compte que les n´egociations qui ont r´eussi. La derni`ere colonne repr´esente le taux d'accord atteint par les agents, c'est `a dire le nombre de fois o`u les agents parviennent `a un accord divis´e par le nombre total de n´egociations. Le taux d'accord tr`es faible peut etre expliqu´e par notre strat´egie, qui ne s'accorde pas du tout `a la distance `a la borne, puisqu'elle ne la prend pas en compte. Cette caract´eristique l'empeche de ressentir la pression du temps (en anglais time pressure) qui est utilis´ee par les autres agents pour d´ecider s'ils conc`edent beaucoup ou non. Plus de 50% des n´egociations se soldent par un ´echec (contre 45% pour les autres agents). Les deux versions de l'´elagage am`enent une am´elioration significative. Avec l'´elagage par une valeur fixe, notre agent ne fait et n'accepte que des propositions pour lesquelles il recevrait une utilit´e de 0.8 au moins. Cela a un effet important sur le r´esultat des n´egociations. Celles qui se terminent permettent `a notre d'agent d'obtenir une bonne utilit´e, mais elles sont moins nombreuses que dans le cas pr´ec´edent, en particulier en ce qui concerne les agents avec lesquels notre agent avait d´ej`a un taux d'accord faible. L'autre m´ethode d'´elagage est moins dure pour les adversaires, puisque l'´elagage repose sur les propositions qu'ils font tout au long de la n´egociation. L'utilit´e de l'agent, en 13 ) e l b a i r a v e g a g a l ´e / e x fi e g a g a l ´e / e g a g a l ´e s n a s ( 4 1 0 2 C A N A ' l e d s e t s i l a n fi x u a e c a f A N a C o M e d s t a t l u s ´e R -- 2 e l b a T d r o c c a ' d x u a T 0 1 . 0 / 5 1 . 0 / 5 3 . 0 0 1 . 0 / 0 / 5 0 . 0 5 5 . 0 / 5 5 . 0 / 5 4 . 0 5 2 . 0 / 5 4 . 0 / 5 4 . 0 5 4 . 0 / 5 4 . 0 / 5 4 . 0 5 0 . 0 / 0 / 0 2 . 0 5 1 . 0 / 0 1 . 0 / 5 1 . 0 1 / 5 9 . 0 / 5 9 . 0 ) 8 0 . 0 ± ( 6 6 . 0 / ) 2 0 . 0 ± ( 5 8 . 0 / ) 8 0 . 0 ± ) 6 0 . 0 ± ( 5 5 . 0 / A N / ) 0 . 0 ± . ) 0 1 . 0 ± ( 7 7 . 0 / ) 2 0 . 0 ± ) 0 1 . 0 ± ( 9 7 . 0 / ) 4 0 . 0 ± ) 6 1 . 0 ± ( 0 7 . 0 / ) 5 0 . 0 ± ( 3 8 . 0 / ) 1 1 . 0 ± ( 7 8 . 0 / ) 9 1 . 0 ± ( 5 8 . 0 / ) 4 1 . 0 ± ( 7 5 . 0 ( ( ( A N a C o M e r o c S ) 0 ± ( 9 4 . 0 / A N / ) 5 0 . 0 ± . ( ) 1 1 . 0 ± ( 4 6 . 0 / ) 3 0 . 0 ± ( 4 8 . 0 / ) 5 1 . 0 ± ) 2 1 . 0 ± ( 6 6 . 0 / ) 7 0 . 0 ± ( 6 8 . 0 / ) 0 1 . 0 ± ( ( 5 5 . 0 ( 6 6 . 0 4 . 0 0 7 . 0 2 8 . 0 2 5 7 . 0 5 7 . 0 ) 1 1 . 0 ± ) 2 2 . 0 ± ) 9 0 . 0 ± ) 8 0 . 0 ± e r i a s r e v d a e r o c S . ) 0 ± ( 1 / A N / ) 5 8 0 . 0 ± ) 1 0 . 0 ± ( 6 8 . 0 / A N / ) 0 . 0 ± ( 1 . ( 7 7 . 0 / ) 8 0 . 0 ± ( 0 7 . 0 / ) 7 0 . 0 ± ( 6 5 . 0 / ) 9 1 . 0 ± ( 4 6 . 0 / ) 6 0 . 0 ± ( 3 5 . 0 / ) 6 0 . 0 ± ( 3 5 . 0 / ) 3 1 . 0 ± ( 1 7 . 0 / ) 6 1 . 0 ± ( 3 5 . 0 / ) 1 2 . 0 ± 8 9 . 0 ( ( ( ( 1 8 . 0 7 7 . 0 8 6 . 0 ) 3 1 . 0 ± ) 3 2 . 0 ± ( 7 7 . 0 / ) 7 0 . 0 ± ( 1 7 . 0 / ) 7 1 . 0 ± ( 2 5 . 0 / ) 0 2 . 0 ± ( 9 4 . 0 / ) 7 1 . 0 ± ( 2 6 . 0 1 8 . 0 ( ( 8 8 . 0 M t n e g A e r i a s r e v d A A N o D r e t s g n a G e l a h W 2 p u o r G t n e g A G k k Y t n e g A t a C e v a r B 14 revanche, est bien meilleure que dans le cas sans ´elagage. Elle l'est moins, en revanche, que dans le cas avec un ´elagage fixe de 0.8. Notre agent bat la majorit´e de ses comp´etiteurs. Quelle que soit la technique d'´elagage, elle permet aux agents de ne pas explorer les branches inint´eressantes, et donc de rechercher les meilleures solutions parmi les solutions acceptables. Elles sont en cela meilleure que la pond´eration a posteriori. 5.2.2 Contexte sans borne En ce qui concerne la n´egociation avec le Random Walker, MoCaNA semble beaucoup mieux s'en sortir, meme sans ´elagage. Il obtient un score ´elev´e, 0.703(±0.062) contrairement au Random Walker 3, qui obtient un score moyen de 0.378(±0.077). ce qui confirme que notre agent est gen´e par la pr´esence de la borne, et le fait qu'il ne s'appuie pas dessus. Notre agent est donc un bon n´egociateur d`es lors qu'on ne lui impose pas de borne. La pr´esence de cette derni`ere coupe la n´egociation avant qu'il ne puisse arriver `a un accord avantageux pour lui. Les n´egociation r´eussissant (moins de 50%), sont donc celles pour lesquelles il a fait le plus de concessions, ce qui a pour effet de lui conf´erer une utilit´e moindre que celle de son adversaire. 6 Conclusion Dans cet article, nous avons pr´esent´e un agent de n´egociation automatique capable de n´egocier dans un contexte o`u les agents ne se fixent pas de borne, ni temporelle ni en termes de nombre de tours et o`u le domaine de n´egociation peut etre continu. Apr`es avoir d´ecrit la n´egociation sous forme de jeu extensif, nous avons d´ecrit une strat´egie d'offres s'appuyant sur une version de MCTS et sur deux m´ethodes de mod´elisation de l'adversaire. Pour l'une d'entre elles, la r´egression de processus gaussiens, dont nous avons test´e plusieurs kernels et retenu le meilleur. Nous avons propos´e deux variantes `a notre strat´egie d'offre s'appuyant sur des ´elagages des branches. Aucun agent n'est con¸cu pour n´egocier dans un contexte aussi r´ealiste que celui pour lequel MoCaNA a ´et´e con¸cu. Nous l'avons donc compar´e aux finalistes de l'ANAC 2014, dont le contexte est le plus proche de celui pour lequel notre agent est con¸cu et `a un RandomWalker. Les exp´erimentations ont montr´e qu'il est difficile d'obtenir un haut taux d'accord pour un agent n'ayant pas conscience de la time pressure. Dans les cas o`u un accord est trouv´e, l'´elagage am´eliore grandement la performance de MoCaNA et lui permet de battre la majorit´e de ses adversaires, au prix d'une baisse du taux d'acceptation. Les perspectives de ces travaux comprennent l'´elargissement du spectre de domaines sur lequel MoCaNA est capable de n´egocier, notamment `a tester des domaines de n´egociations form´es d'attributs cat´egoriels uniquement, continus uniquement, et mixtes. Un autre type de protocoles `a tester r´eside dans les protocoles multilat´eraux, notamment des protocoles 1:n de type ench`eres ou n:m de type many-to-many bargaining. Une autre perspective consisterait `a concevoir une version de l'ANAC sans borne. Il serait enfin int´eressant de 3. Cet agent est d´ecrit dans [3] et fait des propositions al´eatoires. 15 voir dans quelle mesure MoCaNA pourrait int´egrer des informations suppl´ementaires telles que la borne, des connaissances sur la nature des attributs afin d'en tirer parti lorsque cela est possible. Il est aussi possible d'am´eliorer MoCaNA en se concentrant sur MCTS. Il serait par exemple int´eressant de voir comment un double ´elagage, par une valeur fixe et selon les propositions de l'adversaire, influerait sur l'utilit´e et le taux d'accord de MoCaNA. Une autre am´elioration, qui permettrait d'augmenter le nombre de simulations de MoCaNA et donc de diminuer le temps de calcul ou d'augmenter la performance de notre agent serait d'impl´ementer des algorithmes comme All Moves As First ou la Rapid Action Value Estimation. R´ef´erences [1] Tim Baarslag. Exploring the Strategy Space of Negotiating Agents : A Framework for Bidding, Learning and Accepting in Automated Negotiation. PhD thesis, Delft University of Technology, 2016. [2] Tim Baarslag, Mark J C Hendrikx, Koen V Hindriks, and Catholijn M Jonker. Lear- ning about the opponent in automated bilateral negotiation : a comprehensive sur- vey of opponent modeling techniques. Autonomous Agents and Multi-Agent Systems, 20(1) :1 -- 50, 2015. [3] Tim Baarslag and Koen V. Hindriks. Accepting optimally in automated negotiation In AAMAS '13, pages 715 -- 722, Richland, SC, 2013. with incomplete information. International Foundation for Autonomous Agents and Multiagent Systems. [4] Cameron C Browne, Edward Powley, Daniel Whitehouse, Simon M. Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Taverner, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of Monte Carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1) :1 -- 43, 2012. [5] Dave de Jonge and Carles Sierra. Recent Advances in Agent-based Complex Au- tomated Negotiation, volume 638 of Studies in Computational Intelligence, chapter GANGSTER : An Automated Negotiator Applying Genetic Algorithms, pages 225 -- 234. Springer International Publishing, 2016. [6] Dave de Jonge and Dongmo Zhang. Automated negotiations for general game playing. In AAMAS '17, pages 371 -- 379, Richland, SC, 2017. [7] Fang Fang, Ye Xin, Yun Xia, and Xu Haitao. An opponent's negotiation behavior In 2008 model to facilitate buyer-seller negotiations in supply chain management. International Symposium on Electronic Commerce and Security, 2008. [8] Peyman Faratin, Nicholas R Jennings, and Carles Sierra. Negotiation decision func- tions for autonomous agents. Robotics and Autonomous Systems, 24(3-4) :159 -- 182, 1998. 16 [9] Naoki Fukuta, Takayuki Ito, Minjie Zhang, Katsuhide Fujita, and Valentin Robu, editors. Recent Advances in Agent-based Complex Automated Negotiation, volume 638 of Studies in Computational Intelligence. Springer International Publishing, 2016. [10] Koen Hindriks and Dmytro Tykhonov. Opponent modelling in automated multi-issue In Proceedings of the 7th International Joint negotiation using bayesian learning. Conference on Autonomous Agents and Multiagent Systems, volume 1, pages 331 -- 338, 2008. [11] Anna Ja´skiewicz and Andrzej S. Nowak. Non-Zero-Sum Stochastic Games, pages 1 -- 64. Springer International Publishing, Cham, 2016. [12] Levente Kocsis and Csaba Szepesv´ari. Bandit based monte-carlo planning. In Jo- hannes Furnkranz, Tobias Scheffer, and Myra Spiliopoulou, editors, Machine Lear- ning : ECML 2006, pages 282 -- 293, Berlin, Heidelberg, 2006. Springer Berlin Heidel- berg. [13] Raz Lin, Sarit Kraus, Tim Baarslag, Dmytro Tykhonov, Koen Hindriks, and Catho- lijn M Jonker. Genius : an integrated environment for supporting the design of generic automated negotiators. Computational Intelligence, 30(1) :48 -- 70, 2014. [14] Martin J Osborne and Ariel Rubinstein. A course in game theory. MIT press, 12 edition, 1994. [15] Carl E Rasmussen and Christopher K I Williams. Gaussian processes for machine learning. MIT Press, 2006. [16] Daniel M Reeves and Michael P Wellman. Computing best-response strategies in infinite games of incomplete information. In Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence, UAI '04, pages 470 -- 478, Arlington, Virginia, United States, 2004. AUAI Press. [17] Tuomas W Sandholm. Distributed rational decision making. Multiagent systems : a modern approach to distributed artificial intelligence, pages 201 -- 258, 1999. [18] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, and al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587) :484 -- 489, 2016. [19] B´alint Szollosi-Nagy, David Festen, and Marta M Skarzy´nska. Recent Advances in Agent-based Complex Automated Negotiation, volume 638 of Studies in Computational Intelligence, chapter A Greedy Coordinate Descent Algorithm for High-Dimensional Nonlinear Negotiation, pages 249 -- 260. Springer International Publishing, 2016. [20] Colin R Williams, Valentin Robu, Enrico H Gerding, and Nicholas R Jennings. Using gaussian processes to optimise concession in complex negotiations against unknown opponents. In IJCAI'11, pages 432 -- 438, 2011. 17
1909.02396	1	1909	2019-08-25T18:55:10	Modelling transport provision in a polycentric mega city region	[ "cs.MA" ]	The aim of this paper is to present a model of interaction between transport and land use which aims at endogenously integrates the provision of transportation infrastructure and its effects on land use, with a long term perspective (Lowry, 1964, Wegener, 2004, Levinson, 2011, Bretagnolle, 2014, Mimeur et al., 2015). LUTECIA (Land Use, Transport, Evaluation of Cooperation, Infrastructure provision and Agglomeration effects) model puts emphasis on multiscale processes of urban growth, in the context of the emergence of Mega-City Regions (MCR, Hall & Pain, 2006). It allows us, in this exploratory phase, to characterize via simulation the conditions of development of polycentric metropolises. Nevertheless, we argue that such approaches are all the more necessary that we are in a period of multiple transitions that classical modelling tools have difficulty to capture.	cs.MA	cs	Modelling transport provision in a polycentric mega city region. Florent Le Néchet Université Paris-Est, Laboratoire Ville Mobilité Transport, UPEMLV : 5 boulevard Copernic, Cité Descartes F 77454 Marne-la-Vallée cedex 2 France [email protected] This article was written in the context of the Rochebrune conference (Le Néchet, 2010). Some elements are published in Le Néchet (2017). Abstract The aim of this paper is to present a model of interaction between transport and land use which aims at endogenously integrates the provision of transportation infrastructure and its effects on land use, with a long term perspective (Lowry, 1964, Wegener, 2004, Levinson, 2011, Bretagnolle, 2014, Mimeur et al., 2015). LUTECIA (Land Use, Transport, Evaluation of Cooperation, Infrastructure provision and Agglomeration effects) model puts emphasis on multiscale processes of urban growth, in the context of the emergence of Mega-City Regions (MCR, Hall & Pain, 2006). It allows us, in this exploratory phase, to characterize via simulation the conditions of development of polycentric metropolises. Nevertheless, we argue that such approaches are all the more necessary that we are in a period of multiple transitions that classical modelling tools have difficulty to capture. Keywords Mega-City Regions, agent-based modelling, polycentricity, urban dynamics, LUTI model Introduction The aim of this paper is to present a model of interaction between transport and land use which aims at endogenously integrates the provision of transportation infrastructure and its effects on land use, with a long term perspective (Lowry, 1964, Wegener, 2004, Levinson, 2011, Bretagnolle, 2014, Mimeur et al., 2015). LUTECIA (Land Use, Transport, Evaluation of Cooperation, Infrastructure provision and Agglomeration effects) model puts emphasis on multiscale processes of urban growth, in the context of the emergence of Mega-City Regions (MCR, Hall & Pain, 2006). It allows us, in this exploratory phase, to characterize via simulation the conditions of development of polycentric metropolises. Cities are said to be complex (Benenson & Torrens, 2004; Batty et al., 2012), in the sense that the modification of a component of one of the subsystems would potentially induce changes to all the components of all the subsystems. LUTECIA model focus on the coevolution of location system 1 and transportation system (Raimbault, 2018).As an illustration of such interactions, let us mention the rebound effect of travel speed on trip distances (Litman, 2017a). At metropolitan level, gains in accessibility due to highway construction and thus improved speeds overall resulted in unforeseen urban sprawl (Banister, 2011). Complex imbrication also occurs between multiple spatial scales of planning: decisions of transport provision made at a local scale can have repercussions at a wider scale, and vice versa. For instance, Appert (2004) showed how the Green Belt policy in London (belt of restricted urbanization located circa 30 km from the city center), a measure which was dedicated to limit urban sprawl, created in fact leapfrog sprawl and increased travel distance for new residents located outside the green belt and commuting to London. Conversely, the spatial organization at a metropolitan level can influence the need for transport provision: the proximity of two cities and increase of commuting flows can over time create needs for the development of efficient transportation links linking the two cities (Le Néchet, 2012) This article focuses on the complex coevolution between decision making of transport infrastructure provision operated by local and metropolitan stakeholders and urban dynamics hence changes in the spatial organization of population and activities densities. Land use and transport infrastructure networks come from planning decisions made over a long period of time by a diversity of actors at individual and collective levels (Venables, 2007; Vickerman, 2017). Individuals make location choices based on several criteria (for instance, housing affordability, accessibility to workplaces). At the collective level, urban actors play a role in the development of cities, such as national or supra- national bodies and private actors, capable of making heavy investments. The contribution of this article lies in the exploration of paths of urban dynamics with evolving transportation network, and of the conditions of the emergence of governance at MCR level. We start be a short discussion on the emergence of Mega-City Regions (Hall & Pain, 2006), then raising the question of the level of governance most fitted to transportation issues. In the last section, we introduce the theoretical agent-based model used to explore such questions. Emergence of Mega-City Regions This section describes the emergence of MCR, in Europe and the United States. Many examples are studies in Europe (PolyNet research network for instance), including the well-known Randstad Hol- land (Lambregt & Kloosterman, 2012). During the second part of the twentieth century, the reduc- tion in transport costs has been accompanied by a double dynamic of urban sprawl (Le Néchet, 2015) and hierarchization of the system of cities (Bretagnolle et al., 2007; Pumain et al., 2015). In- tra-urban and inter-urban dynamics tend to overlap: (Champion, 2001; Pumain et al., 2006). As stat- 2 ed by Soja (2011) « The urban, the metropolitan, and the subnational-regional scales seem to be blending together ». Mega City Regions (Hall & Pain, 2006) are geographical objects lying some- where between cities and system of cities, structured by flows and networks of communication. Territories of several thousands of square kilometers of discontinuous urbanization are now func- tionally integrated, with various kind of flows (business & leisure trips, Scott, 1997, logistics supply chains, Heitz & Dablanc, 2015, long-distance commuting trips, Conti, 2015). Several geographical levels are now de facto interrelated in the fabric of cities (what Soja, 2011 call "multilevel urbaniza- tion processes"). In particular, these evolutions the questions of which type of governance are more adequate to such geographical objects to achieve social, economic or environmental goals. The links between the reconfiguration of flows within territories and reconfiguration of governance at the metropolitan or regional level are though at different scales, and constitute an open and difficult question (Lefèvre, 2009; Cowell, 2010; Le Néchet, 2017). The literature does not offer a consensus over the evaluation of the resulting synergies (Meijers, 2004, Davoudi, 2007) of such metropolitan integration policies, but the question of the main scale of decision is little raised by the literature. As an example, let us compare Paris (France) and Rhine- Two research questions are treated at this stage: what is the impact of the spatial organization of densities (polycentric versus monocentric) on the emergence of Mega-City Regions? What is the influence of governance system (centralized ver- sus noncentralized) on the level of polycentricity of the Mega-City Region? Urban sprawl, functional integration and governance levels Ruhr area (Germany), two European metropolitan areas of relatively similar total population (12M) and size (12,000 km²), with very contrasted spatial organization. Figure 1 illustrates how different these networks are, that have emerged over the long term. In Paris, the local network ("métro") is of small spatial scope, which arises from the scale of deci- sion before the construction of the network : in 1900, the municipality has been able to impose its project to the national level (Larroque et al., 2002). Thus, two scales of network emerged, still visi- ble today, which is not the case for example in another monocentric city like London. As the cities of the Rhine-Ruhr region (Köln, Düsseldorf, Essen, Dortmund, Duisburg) have developed in rela- tive independence over time, before a gradual and incomplete functional integration (Blotevogel, 2001, Le Néchet, 2012), it is logical to observe here also two scales in the transportation network, the regional network "S- Bahn" having been organized since 1967. 3 Figure 1: Transit networks, at several spatial scales, in Paris Metropolitan Area "Ile-de-France-Paris", France (left) and in the Rhine-Ruhr region, Germany (right). Research question : multilevel governance and coevolution transport and land use In this article, we start from this very simple question: why are transport networks organized the way they are? Since the time scales involved are long (railway networks back from the beginning of the 19th century, the motorway networks of the early 20th century), this question invites us to think more generally about the interactions between transportation networks on one hand, and demographic evolution on the other hand. Kasraian et al. (2016) provide an extensive meta-analysis which highlights the complexity of these links. Most of the studies in this field use the notion of accessibility (Hansen, 1959) to understand and predict the effects of a transport infrastructure on territorial development, and in a more systemic approach, as an operative concept for studying the coevolution dynamics between transport and territories (Raimbault, 2018). To add to this complexity, transport infrastructures are decided and financed by a multiplicity of territorial stakeholders, since they potentially affect multiple territorial scales, from local to national or even international scale for fast networks such as high-speed railways or airports. Due to the polycentric nature of the MCR, our research question is thus transformed into the following one : what is the impact of a type of governance, hence giving the advantage to the local or to the whole region, on the dynamics of coevolution between transport and spatial distribution of the demography? This question takes place in the current debates about the formation of governance bodies at metropolitan level (Heeg et al., 2003 ; Douay, 2010) or at MCR level (Le Néchet, 2017). In this article, we tackle these questions via simulation. 4 According to our hypotheses, local stakeholders will tend to favor the accessibility of workers living within their governance area. Conversely, a regional stakeholder would aim at better global accessibility of the whole territory, which would often lead to different choices. Overall, one would expect MCR governance integration to occur when regional planning significantly outpaces the sum of local planning outcomes in terms of total accessibility (even if other factors might affect integration of governance, such as cultural differences, Blotevogel, 2001). Within the framework of simulation, to account for the rationality of territorial stakeholders at various spatial scales (cities, region), we are inspired by the paradigm Belief Desire Intention (Taillandier et al. 2011): actions of a territorial stakeholder will be driven by the perception of the elements of the system that he is likely to act on. However, most models of literature do not raise the issue of multiple territorial stakeholders. In most cases, choices are made only at the level of economic agents (individuals, real estate developers, etc.), or they are made according to a global rationality, implicitly assuming that one single actor has ability to take all decisions for a whole region. The reality is more complex: individual rationalities and the concept of global rationality come up against the rationalities of multiple territorial actors operating on multiple scales (Trigalo, 2000). Adopting a modelling framework separating individual mobility and migration choices on one hand (micro scale) and stakeholder long-term investments such as transport provision on the other hand (macro scale), the following section focus on the literature review to understand agents choices at each scale. Transport provision, accessibility, and residential location models Indeed, transport infrastructure provision follows various logics, for instance supporting land development (the Transit-Oriented Development Literature, Litman, 2017b) or improved accessibility for deprived neighborhoods (Desjardins & Drevelle, 2014). Acknowledging this complexity, a part of the literature focus on the possible effects of transport provision on to the spatial organization of densities (Banister, 2001; Litman, 2017b). LUTI (Land Use Transport Interaction) models extend classical four-stage model of transportation demand to take into account residential mobility. A key hypothesis of the model is that an increase in accessibility due to a new transportation link is likely to bring opportunities for new urban developments and/or an increase in residential or employment density around the new modes of transportation (Wegener, 2004). Most of such research are focused at the metropolitan level, only a few studies trying to assess the effects 5 of a variation of boundaries on the residential choice model output (see Thomas, 2018 for a complete analysis on LUTI models). Based on economic hypotheses such as monocentric model by Alonso (1964) and Lowry dynamic model (1964), some models link individual choices, whose determinants are synthesized by sophisticated mathematical functions via utility functions that result from these choices. Based on elements identified as exogenous (transport cost, available infrastructure, existing housing stock), and a finite number of selection criteria, it is possible to study the Pareto equilibrium configurations of the system (Delons et al., 2008). Some other have an agent-based approaches (Tannier et al., 2015). Scenarios can then be made to assess how a system would react in response to exogenous modifications such as an increase in the cost of mobility or a change in transport supply) This type of approach has proven useful in operational contexts to assess the socio-economic benefit of transport provision (Vickerman, 2017). However, data on the supply side are considered as exogenous: the model cannot predict the growth of infrastructure or other major structural changes (Raimbault, 2018). This constitutes the main limitation of this approach in the perspective of modeling the dynamics of cities in a longer term. To understand the long-term evolution of urban systems, it is useful to have other tools. Literature on transport network evolution is scarce (Raimbault, 2018), and existing models often focus on the geometric aspects (for instance percolation models) and are rarely explicitely with the socio- demographic aspects of the territory. Articles by Zhang and Levinson (2007), Xie & Levinson (2009), or Cavailhès et al. (2010) are notable exceptions that endogenously account for the changing transportation networks in relation to the changes of land use and commuting flows. Our approach is comparable but extends to the case of multiple actors. Overview of the LUTECIA model This paper introduces a LUTI model, developed and implemented on the Netlogo platform (Wilenski, 1999) and presented in Le Néchet (2010, 2017) and then Raimbault (2018) aimed at exploring the complex multiscale evolution of transportation networks and spatial structure, in the context of polycentric Mega City-Regions (MCR, Hall & Pain, 2006). The model combines a classical discrete choice LUTI model and a module that predicts endogenously the evolution of transportation network, given a configuration of territorial stakeholders respective powers. In order to explore the formation of governance at MCR level over several decades. Two main 6 hypotheses are used in the formalization of the model: firstly, the development of transport infrastructures affects the spatial organization of population and jobs; secondly, planning stakeholders are in interaction to decide for heavy transport provision investment. A stylized city is proposed which consists of the following elements: - The cells are the elementary zones which support housing and jobs - The metropolis: set of cells constituting the closed environment in which population (only workers in this version) and workplaces take place. - The workers are located on the cells and differentiated by socio-occupational categories. - Jobs are also located on the cells; the types of jobs are associated with socio-professional categories (we used a simple typology) - Planners are agents responsible for providing transport infrastructure. Two type of actors are implemented in the model: local stakeholders called "mayors", who divide the metropolis into M disjointed areas, and a metropolitan "governor". These (M + 1) agents have their own objective function when provisioning transport infrastructures. The model is articulated around three sub-modules, ran successively during a time step: - A transport module, which computes the demand for travel according to a given configuration of the internal distribution of workers and jobs, as well as the supply of transport infrastructure. - A land use module that reallocate workers and jobs based on metropolitan accessibility via transportation network - A governance module, which settles the decision making of transport infrastructure provision according to differentiated development rules at several scales. To account for the hypothesis stated, the underlying principle of the model is as follows: a spatial configuration of transport infrastructures being given, workers and jobs relocate in space, according to the transport possibilities offered by the city. In return, depending on the location of workers and jobs, a new transportation infrastructure is built in order to improve the overall accessibility level. This growth of accessibility is non-linear: some infrastructures simply extend an existing line and reinforce an ongoing process of urban sprawl; conversely, when an infrastructure connects two major centers, a structural change occurs, with possible further integration towards a polycentric metropolis as described by Champion (2001). In the initial configuration, workers and jobs, are spatially distributed exogenously. Workers are distributed according to Clark's (1951) empirical law: the monocentric model. The population density ρ(c) in cell c, distant from from center Ci by a distance di is according to polycentric model 7 of Heikkila et al. (1989): with M centers. This section details the equations used in the model, for each of the three modules presented. The model relies on a classical four-stage transport demand model, discrete choices model of location choices of households and jobs. The originality of the model is the inclusion of a model accounting for the construction of new transport infrastructures. Transport submodel The four-stage model of transportation demand is commonly used in transportation engineering to model the demand for travel, once given location-based attributes (workers and jobs per municipality generating transport demand) and the attributes of transport infrastructure (travel time, capacity) summarizing the transport offer. In detail, the four stages conventionally implemented, with varying degrees of sophistication, are as described as suggested by Bonnel (2001): 1. Generation of the travel demand. The aim is to estimate, trips at origin and at destination, by travel zone, for various trip purposes. In this model, only commuting is taken into account, differentiated by socio-professional category. 2. Distribution of commuting flows, carried out according to the gravity model: If and are respectively the number of workers of zone i, and the number of jobs of zone j, the number of displacements between the two zones is obtained by solving the system of coupled equations (Furness algorithm): under constraints : et given et and are temporary parameters; λ can be interpreted aversion to the distance of individuals; here represents the travel time between the two zones, depending on the transport infrastructure present (if the "as the crow flies" -- AFC - route induces a lower time, it is retained). 8 iidbieA−==M1i(c)iAjEijijdjijiijeEAqpNji−==,..1,===NkjkjENj1,..1===NljilANi1,..1=−==NldlliileEqpNi11,..1=−==NkdklkjkjeApqNj11,..1ipjqijd3. Modal choice. Here, only car use is implemented; therefore, this submodule is disabled. 4. Traffic assignment will compute the frequentation of the different sections of the transport networks. This step is performed by calculations of shortest paths, taking into account congestion, via an implementation of a static Dijkstra algorithm. Note that LUTECIA model differentiate "local" and "regional" roads: local roads are not represented in the model but correspond to AFC distance, they are no subject to congestion. Regional roads are the transportation links that are built throughout the model: they are subject to possible congestion (travel speeds depend on flow- capacity ratio according to the classical BPR 1964 function). Location choice submodel The literature on residential mobility is mainly based on the idea that individuals and firms, in the choice of their location, promote greater accessibility to the desired resources as well as attributes of the place (Tannier et al., 2015). In LUTECIA model, no land auction is implemented but the local distribution of densities are accounted for in the relocation submodule, making it possible to implement the two opposite forces: density aversion and desire for accessibility. This utility function depends on the position of the cell relative to its environment (Accessibility, noted Xc), and its own attributes (Urban Form, noted Fc). The implementation adopted is as follows: a Cobb-Douglas function makes it possible to synthesize the utility function of the workers (and jobs) for each cell c, Uc. In a detailed way, the functions are calculated as follow (for the workers, the formalizations are symmetrical for the jobs): The parameter ν accounts for the cost of energy in the sense that greater distances can be more or less difficult to achieve on a daily basis depending on the city wealth and the technology of vehicles; this parameter is common to all agents of the model. Note that it is a strong assumption which can be relaxed in further versions of the model. A proximity matrix between socioprofessional categories allows to account for the preference 9 −=1cccFXU−=jdsjssicijeEAX',','''',)()())((ssssmsimsisssisicEAEAF=between workers and jobs in location choice model; three possibilities are implemented: categories s and s' are indifferent to their mutual presence (ms, s' = 0), avoid each other (ms, s' <0) or appreciate their mutual presence (ms, s'> 0). Once these utility functions have been determined, the workers and jobs are distributed on the grid, so that the best locations are the most chosen; a discrete choice model has been implemented: the probability for a worker (respectively a job) to move into cell c is as follows, where μ is the sensitivity of workers (respectively jobs) towards a differential of accessibility. NB : il manque toute une partie sur l'évaluation socio-économique des infrastructures Governance submodel Growth of a transportation network through coordination between several territorial stakeholders Finally, a last step is implemented to allow for endogeneous modification of the transportation network. As stated earlier, we wish to include in the model the evaluations of transport provision needs from both local and regional stakeholder. Our modelling choices lie on three main hypothesis : (i) in general stakeholders will seek maximization of accessibility to jobs for its residents given a level of new transport resources ; (ii) due to historical pathways (Paasi, 1986), there is an external and somehow constant level of concentration of governance within the region, between two extreme cases : all decisions are taken at regional level, with no role to local stakeholders and all decision are taken at local city level, with no role for regional stakeholders ; a parameter can account for the proximity towards such extreme configurations ; (iii) : in case of local stakeholder having a say in transportation provision, the relative power of local stakeholder will be a function a their relative economic wealth. Thus, due to hypothesis (iii) a local city has an implicit interest in developing urban amenities. In addition, the territories are in a situation of cooperation: the workers of each territory can use the infrastructures created, independently of the stakeholder that made the investment; this concurrence / cooperation processes appear realistic in a context where there are increasing interactions between spatial scales as mentioned by Parr (2004) and Cowell (2010). The infrastructures are here developed one by one, by investments made 100% by one of the 10 ='')(cUUcceecPadministrative authorities: local mayor or regional autorithy. This hypothesis indeed neglects part of the complexity of the political decision-making process, which often leads, in the case of transport infrastructure, to compromises carried by multiple actors (Ollivier-Trigalo, 2000), and to joint financing by different stakeholders. However, we argue that over the course of several iterations of the model, the various stakeholders will have had chance to develop infrastructure according to their utility function, proportionally to their respective importance. Transport provision is implemented using a two-step approach: (i) The choice of stakeholders that will develop the infrastructure network: local decision (one of the mayors) or metropolitan decision? (ii)The choice of infrastructure: which criteria? which method is chosen? (i) Firstly, randomly chosen decision will be at the local or metropolitan level, according to an exogeneous parameter, the share of local decisions ξ. Note that the determination of mayor is set exogenously, and do not vary during the simulation (there is no feedback loop from land use to the administrative division between stakeholders). If the decision is local, another random step is achieved to determine which mayor will have the opportunity to decide on the construction of the new infrastructure. This random choice is conducted according to a probabilistic model taking into account the relative power of mayors through number of jobs in the mayor's territory (in order to take into account the relative wealth of each territory); In other words, there is a feedback loop of the distribution of employment on the propensity of each mayor to make a development decision. In detail, if Yi is the total number of jobs in Mayor i's territory, the probability that this mayor will make the decision is: . (ii) Once a territory T has been chosen on which the decision to build an infrastructure is based, the choice of infrastructure is made with a view to maximizing the accessibility of the T territory's workers to all metropolitan jobs. This choice is, of course, debatable, but it allows us to give a synthetic account of the desire usually expressed by planners and urban planners to encourage social exchanges between individuals and economic functioning of the territory. The current configuration is associated with a "current" accessibility , where X (c) is the accessibility of the cell c, belonging to T. If a new infrastructure, Z, is built, the accessibility function will, after application of the transport module, be modified: XZ (T). It's about retaining the infrastructure that maximizes new accessibility: 11 ==MjjiiYY1)()(cXTXTc=.. This research is carried out here for all the infrastructures connecting two neighboring zones, which implies substantial computing times, explaining why a small grid is used at this exploratory stage. Choice of stakeholder Local decision Metropolitan level decision ξε1 ξε2 … ξεM 1-ξ Figure 2 : Probability for each territorial stakeholder to obtain the decision at each time step. As such, the metropolitan dynamic is the result of a stochastic process, which we explore in the next section. Results In this last section, first applications of the model are proposed. Note that in this archive version of the paper, the validation roadmap is not detailed ; parameters used here derive from a "expert" validation in the sense that they were chosen to produce realistic urban dynamics (for instance, the speed of sprawl has to be controlled to ensure realistic response to provision of transport infrastructure). The following section describes the dynamics of the transportation network in absence of land use change, but with varying type of governance and different initial configurations in term of land use. Growth of transportation network, no land use change Figure 3 shows the state of a simulation after a small number of steps; Mayor 0 (gray, bottom right) 12 )(*maxargTXZZZ=and Mayor 1 (red, top left) develop the territory. The population density, in green, shows a city more populated than the other. It is a rather monocentric city, with a short distance between the two centers. Hence, Mayor 1 is more likely to apply its utility function to the transport infrastructure provision, compared to Mayor 0. Figure 3 : Simulation outputs after 3 timesteps : exclusively local decisions (left); exclusively metropolitan (right). Each simulation (construction of six infrastructures) is reproduced 30 times to account for the stochastic variability of the model. Figure 4 shows the mean values (and the variation ellipse) of the following indicators: - Total accessibility, - Total travel time of individuals Given the spatial configuration, high accessibility level corresponds to a well-meshed network (facilitating exchanges between the two cities of the metropolis), and low travel times illustrate either a lower territorial integration, or an absence of congestion. The results indicate equivalent accessibility in the case where decisions are taken at the metropolitan level only or by the mayor of greatest importance, the mayor 1. However, when decisions are made by the other, the total time of travel is lower which indicates a lower degree of congestion. Conversely, when decisions are mainly made at Mayor level 0, total accessibility is lower. It is not a question of determining here an "optimal" governance system means, these two indicators having been retained among many others conceivable, but of exploring the articulation between scales in this model transport / land use 13 Figure 4 : Outputs of the simulations : links between accessibility level and average travel times under different governance regimes. Sensitivity analysis Figure 5 shows the average accessibilities at the end of the simulation for four different initial configurations, making it possible to explore the most appropriate decision-making structures for a given urban form. Total accessibility is again taken as an output indicator; on the X-axis, the proportion of decisions taken locally shows, in the case of monocentric cities (the two cities are of unequal demographic importance), a clear tendency for centralized governance systems to offer better accessibility to the city. . Such a result is not reproduced in monocentric metropolises, where the two cities are of the same weight: the variable ξ seems independent of total accessibility even if it probably has an impact on the unequal spatial distribution of this accessibility. 14 Figure 5 : Sensitivity study: influence of urban form and type of governance on total accessibility level. Discussion The results presented constitute a first exploration of a model of interaction between transport and land use, proposing simultaneously to take into account the individual interests of location and displacement, and the collective interests of development. Of course, in its current simplistic configuration, and leave aside important aspects of the complexity of urban dynamics; the criteria with which transport infrastructures are built endogenously leave little room for the complex reality of the terrain. In addition, the evolution of land use has presented here only theoretically, because there is a challenge to adjust the temporalities of joint evolutions of the transport system and the settlement system. So-called "complex systems" approaches aim at taking into account the succession of non- equilibrium states occurring in urban systems, for which neither individual nor global objective functions are available (Batty & Xie, 1999, Xie & Levinson, 2009, Cavailhes et al., 2010, Demare et al., 2017). Organized spatial configurations can emerge from a set of elementary rules, at the local level (as showed by Schelling's segregation model in the urban context (Gauvin et al., 2009). With such models, it is possible to obtain abrupt transitions as a result of the modification of one or a very small number of initial parameters. This is at the same time more realistic and more difficult to use in operational context for technical reasons such as the difficulty to calibrate such a model. Nevertheless we argue that such approaches are all the more necessary that we are in a period of multiple transitions that classical modelling tools have difficulty to capture. 15 References Alonso, William (1964). Location and land use: Towards a general theory of rent. Cambridge, MA: Harvard University Press. Appert, M. (2004). Métropolisation, mobilités quotidiennes et forme urbaine: le cas de Londres. Géocarre- four, 79(2), 109-118. Banister, D., & Berechman, Y. (2001). Transport investment and the promotion of economic growth. Journal of transport geography, 9(3), 209-218. Banister, D. (2011). The trilogy of distance, speed and time. Journal of Transport Geography, 19(4), 950- 959. Batty, M., & Xie, Y. (1999). Self-organized criticality and urban development. Discrete Dynamics in Nature and Society, 3(2-3), 109-124. Batty, M., Axhausen, K. W., Giannotti, F., Pozdnoukhov, A., Bazzani, A., Wachowicz, M., Ouzounis, G., Por- tugali, Y. (2012). Smart cities of the future. The European Physical Journal Special Topics, 214(1), 481-518. Benenson, I. & Torrens, P. (2004), Chapter 4: Modeling Urban Land-Use with Cellular Automata, John Wiley and Sons Ltd, pp. 287. Blotevogel, H. H. (2001). Regionalbewusstsein und Landesidentität am Beispiel von Nordrhein-Westfalen. Gerhard-Mercator Universität. Bonnel, P. (2001), 'Prévision de la demande de transport', Rapport HDR, Décembre 2001, 409. Bretagnolle, A., Pumain, D., & Vacchiani-Marcuzzo, C. (2009). The organization of urban systems. In Com- plexity perspectives in innovation and social change (pp. 197-220). Springer, Dordrecht. Bretagnolle, A. (2014). Les effets structurants des transports, une question d'échelles?. L'Espace géogra- phique, 43(1), 63-65. Bretagnolle, A.; Pumain, D. & Vacchiani-Marcuzzo, C. (2007), Les formes des systèmes de villes dans le monde, Economica, Anthropos, pp. 301-314. Cavailhès, J.; Peeters, D.; Caruso, G.; Frankhauser, P.; Thomas, I. & Vuidel, G. (2010), 'The S-Ghosts' City Self-Generating Housing, Open Space and Transportation in a Sprawling City', Champion, A. (2001), 'A Changing Demographic Regime and Evolving Poly centric Urban Regions: Conse- quences for the Size, Composition and Distribution of City Populations', Urban Studies 38(4), 657-677. Clark, C. (1951), 'Urban Population Densities', Journal of the Royal Statistical Society A 114, 490-496. 16 Conti, B. (2019). Daily inter-urban mobility in France: heterogeneous commuters and commuting. Flux, (1), 14-32. Cowell, M. (2010), 'Polycentric Regions: Comparing Complementarity and Institutional Governance in the San Francisco Bay Area, the Randstad and Emilia-Romagna', Urban Studies 47(5), 945-965. Davoudi, S. (2007), Polycentricity: Panacea or pipedream?, John Libbey Eurotext, pp. 209. Delons, J., Coulombel, N., & Leurent, F. (2008), « an integrated transport and land-use model for the Paris area », HAL archive, https://hal.archives-ouvertes.fr/hal-00319087/ Démare, T., Bertelle, C., Dutot, A., & Lévêque, L. (2017). Modeling logistic systems with an agent-based model and dynamic graphs. Journal of Transport Geography, 62, 51-65. Desjardins, X., & Drevelle, M. (2014). Trends in the social disparities in access to jobs by train in the Paris region since 1975. Town Planning Review, 85(2), 155-170. Douay, N. (2010). Collaborative planning and the challenge of urbanization: Issues, actors and strategies in Marseilles and Montreal metropolitan areas. Canadian Journal of Urban Research, 19(1), 50-69. Gauvin, L., Vannimenus, J., & Nadal, J. P. (2009). Phase diagram of a Schelling segregation model. The Euro- pean Physical Journal B, 70(2), 293-304. Hall, P. G., & Pain, K. (Eds.). (2006). The polycentric metropolis: learning from mega-city regions in Europe. Routledge. Hansen, W. G. (1959). How accessibility shapes land use. Journal of the American Institute of planners, 25(2), 73-76. Heeg, S.; Klagge, B. & Ossenbruuumlgge, J. (2003), 'Metropolitan cooperation in Europe: Theoretical issues and perspectives for urban networking 1', European Planning Studies 11(2), 139-153. Heitz, A., & Dablanc, L. (2015). Logistics spatial patterns in Paris: rise of Paris Basin as Logistics Megaregion. Transportation Research Record, 2477(1), 76-84. Heikkila, E.; Gordon, P.; Kim, J.; Peiser, R. & Richardson, H. (1989), 'What happened to the CBD-Distance Gradient? Land Values in a Policentric City.', Environment and Planning A 21, 221-232. Kasraian, D., Maat, K., Stead, D., & van Wee, B. (2016). Long-term impacts of transport infrastructure net- works on land-use change: An international review of empirical studies. Transport Reviews, 36(6), 772-792. Lambregts, B., & Kloosterman, R. (2012). Randstad Holland: Probing hierarchies and interdependencies in a polycentric world city region. International Handbook of Globalization and World Cities. Cheltenham, UK: Edward Elgar, 476-486. 17 Larroque, D., Margairaz, M., & Zembri, P. (2002). Paris et ses transports: XIXe-XXe siècles, deux siècles de décisions pour la ville et sa région. Recherches/Ipraus. Le Néchet, F. (2010). Approche multiscalaire des liens entre mobilité quotidienne, morphologie et soutena- bilité des métropoles européennes: cas de Paris et de la région Rhin-Ruhr(Doctoral dissertation, Université Paris-Est). Le Néchet, F. (2012). Approche multiscalaire de la mobilité domicile-travail en Île-de-France et dans la ré- gion Rhin-Ruhr. Cahiers de géographie du Québec, 56(158), 405-426. Le Néchet, F. (2015). De la forme urbaine à la structure métropolitaine: une typologie de la configuration interne des densités pour les principales métropoles européennes de l'Audit Urbain. Cybergeo: European Journal of Geography. Le Néchet, F. (2017). De l'étalement urbain aux régions métropolitaines polycentriques: formes de fonc- tionnement et formes de gouvernance. Peupler la terre-De la préhistoire à l'ère des métropoles. Presses Universitaires Francois Rabelais. Lefèvre, C. (2009). Gouverner les métropoles. LGDJ/Lextenso Editions. Levinson, David M (2011). The coevolution of transport and land use : An introduction to the Special Issue and an outline of a research agenda. Journal of Transport and Land Use 4.2. Litman, T. (2017a). Generated traffic and induced travel. Victoria Transport Policy Institute. Litman, T. (2017b). Evaluating accessibility for transport planning. Victoria Transport Policy Institute. Lowry, I. S. (1964). A model of metropolis. Meijers, E. (2004), 'Polycentric urban regions and the quest for synergy: is a network of cities more than the sum of the parts?', Urban Studies 42(4), 765-781. Mimeur, C., Thévenin, T., Vuidel, G., Granjon, L., & Schwartz, R. (2015, May). Géohistoire du lien ré- seau/territoire en France entre 1830 et 1930: une approche géographiquement pondérée. In Douzièmes Rencontres de Théo Quant. Ollivier-Trigalo, M. (2000). Les grands projets de transport transeuropéens: multiplicité des acteurs, conflits et coordination de l'action. Les Cahiers scientifiques du transport, (37), 3-30. Paasi, A. (1986). The institutionalization of regions: a theoretical framework for understanding the emer- gence of regions and the constitution of regional identity. Fennia-International Journal of Geography, 164(1), 105-146. Parr, J. (2004). The polycentric urban region: a closer inspection. Regional studies, 38(3), 231-240. 18 Pumain, D.; Bretagnolle, A. & Glisse, B. (2006), Modelling the future of cities, in 'European Conference of Complex Systems, Oxford University'. Pumain, D., Swerts, E., Cottineau, Céline Vacchiani-Marcuzzo, Cosmo Antonio Ignazzi, Anne Bretagnolle, François Delisle, Robin Cura, Liliane Lizzi et Solène Baffi. (2015). Multilevel comparison of large urban sys- tems. Cybergeo: European Journal of Geography. Raimbault, J. (2018). Caractérisation et modélisation de la co-évolution des réseaux de transport et des territoires (Doctoral dissertation, Université Paris 7 Denis Diderot). Scott, A. J. (1997). The cultural economy of cities. International journal of urban and regional research, 21(2), 323-339. Soja E. (2011), «Beyond Postmetropolis», Urban Geography, 32:4, p. 451-469, DOI: 10.2747/0272- 3638.32.4.451 Taillandier, P., Therond, O., & Gaudou, B. (2012). A new BDI agent architecture based on the belief theory. Application to the modelling of cropping plan decision-making. In International environmental modelling and software society (iEMSs). Thomas, I., Jones, J., Caruso, G., & Gerber, P. (2018). City delineation in European applications of LUTI mod- els: review and tests. Transport Reviews, 38(1), 6-32. Venables, A. J. (2007). Evaluating urban transport improvements: cost -- benefit analysis in the presence of agglomeration and income taxation. Journal of Transport Economics and Policy (JTEP), 41(2), 173-188. Vickerman, R. (2017). Beyond cost-benefit analysis: the search for a comprehensive evaluation of transport investment. Research in Transportation Economics, 63, 5-12. Wegener, M. (2004). Overview of land-use transport models. Handbook of transport geography and spatial systems, 5, 127-146. Wilensky U., (1999), Netlogo, http://ccl.northwestern.edu/netlogo/ Xie, F., & Levinson, D. (2009). Modeling the growth of transportation networks: A comprehensive review. Networks and Spatial Economics, 9(3), 291-307. Zhang, L., & Levinson, D. (2007). The economics of transportation network growth. In Essays on transport economics (pp. 317-339). Physica-Verlag HD. 19
1801.07140	5	1801	2019-05-08T13:41:00	Courtesy as a Means to Coordinate	[ "cs.MA" ]	We investigate the problem of multi-agent coordination under rationality constraints. Specifically, role allocation, task assignment, resource allocation, etc. Inspired by human behavior, we propose a framework (CA^3NONY) that enables fast convergence to efficient and fair allocations based on a simple convention of courtesy. We prove that following such convention induces a strategy which constitutes an $\epsilon$-subgame-perfect equilibrium of the repeated allocation game with discounting. Simulation results highlight the effectiveness of CA^3NONY as compared to state-of-the-art bandit algorithms, since it achieves more than two orders of magnitude faster convergence, higher efficiency, fairness, and average payoff.	cs.MA	cs	Courtesy as a Means to Coordinate Panayiotis Danassis École Polytechnique Fédérale de Lausanne (EPFL) Artificial Intelligence Laboratory Lausanne, Switzerland [email protected] Boi Faltings École Polytechnique Fédérale de Lausanne (EPFL) Artificial Intelligence Laboratory Lausanne, Switzerland [email protected] 9 1 0 2 y a M 8 ] A M . s c [ 5 v 0 4 1 7 0 . 1 0 8 1 : v i X r a ABSTRACT We investigate the problem of multi-agent coordination under ra- tionality constraints. Specifically, role allocation, task assignment, resource allocation, etc. Inspired by human behavior, we propose a framework (CA3NONY) that enables fast convergence to efficient and fair allocations based on a simple convention of courtesy. We prove that following such convention induces a strategy which constitutes an ϵ-subgame-perfect equilibrium of the repeated al- location game with discounting. Simulation results highlight the effectiveness of CA3NONY as compared to state-of-the-art bandit algorithms, since it achieves more than two orders of magnitude faster convergence, higher efficiency, fairness, and average payoff. KEYWORDS [Learning and Adaptation] Multiagent learning; [Economic Paradigms] Noncooperative games: theory & analysis ACM Reference Format: Panayiotis Danassis and Boi Faltings. 2019. Courtesy as a Means to Coor- dinate. In Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), Montreal, Canada, May 13 -- 17, 2019, IFAAMAS, 9 pages. 1 INTRODUCTION In multi-agent systems (MAS), agents are often called upon to im- plement a joint plan in order to maximize their rewards. Typically, coordination in a joint plan incorporates (possibly a combination of) two distinct elements: agents may be required to take the same action [14, 31], or agents may be required to take distinct actions [13, 19]. The latter is ubiquitous in everyday life, e.g. managing common-pool resources, or deciding on car/ship/airplane routes in traffic management. This paper studies coordination in repeated allocation games. Consider for example the problem of channel allocation in wireless networks [36]. In such a problem, N wireless devices contend for R transmission slots (i.e. particular frequency band in a certain period of time), where N ≫ R. If more than one agent access a slot simultaneously, a collision occurs and the col- liding parties incur a cost ζ < 0. The goal then for the agents is to transmit on different slots to minimize collisions over time. Other scenarios include role allocation (e.g. teammates during a game), task assignment (e.g. employees of a factory), resource allocation (e.g. parking spaces and/or charging stations for autonomous ve- hicles), etc. What follows is applicable to any such scenario. For simplicity hereafter we will refer only to indivisible resources (i.e. Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), N. Agmon, M. E. Taylor, E. Elkind, M. Veloso (eds.), May 13 -- 17, 2019, Montreal, Canada. © 2019 International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. resources which can not be shared). Beyond the scope of repeated games, the proposed framework can also be applied as a negoti- ation protocol in one-shot interactions. E.g. self-driving vehicles attempting to get a parking space can utilize such a protocol in a simulated environment with message exchange. Conforming to the relevant literature, in what follows we refer to the coordina- tion problem of selecting distinct actions as the anti-coordination problem [9, 13, 16, 19, 23]. The most straightforward solution would be to have a central coordinator with complete information recommend an action to each agent. This, though, would require the agents to communicate their preferences/plans, which creates high overhead and raises incentive compatibility and truthfulness issues. Moreover, in real- world applications with partial observability, agents might not be willing to trust such recommendations. On the other hand, agents can learn to anti-coordinate their actions in a completely decentral- ized manner. However, in a fully decentralized scheme, agents have an incentive to 'bully' others (stick to your action until you drive out the competition) [25], which makes it impossible to converge to solutions that are both efficient and fair. As such, a mechanism is needed that employs some sort of external authority. In this paper, we employ simple monitoring authorities (MA), with the primary goal to keep track of successful accesses. We do not require any planning capabilities by the MAs, nor any knowledge of the plans and preferences of the participants; only the ability to monitor successful accesses on their respective resource. Such MAs already exist naturally in many domains (e.g. the port authority in maritime traffic management [1], or the access point in wireless networks, etc.), and can be easily introduced in the rest (e.g. an agent mon- itoring each charging / parking station, etc.). At the end, agents learn to anti-coordinate their actions in a decentralized manner, while the introduced MAs allow us to impose quotas to the use of resources and punishments upon violating them to align individual incentives with an efficient and fair correlated equilibrium. A second aspect of anti-coordination problems involves conver- gence speed. Inter-agent interactions often need to take place in an ad-hoc fashion. Typical approaches (e.g. Monte Carlo algorithms, Bayesian learning, bandit algorithms, etc.) tend to require too many rounds to converge to be feasible in dynamic environments and real-life applications. Yet, humans are able to routinely coordinate in such an ad-hoc fashion. One key concept that facilitates human ad-hoc (anti-)coordination is the use of conventions [24]. Behavioral conventions are a fundamental part of human societies, yet they have not appeared meaningfully in empirical modeling of MAS. Inspired by human behavior, we propose the adoption of a simple convention of courtesy. Courtesy arises by the social nature of hu- mans. Society demands that an individual should conduct himself in consideration of others. This allows for fast convergence, albeit it is not game theoretically sound; people adhere to it due to social pressure. Such problems become even more severe in situations with scarcity of resources. Under such conditions, courtesy breaks down and in the name of self-preservation people exhibit urgency and competitive behavior [20]. Thus, to satisfy our rationality con- straint (no incentive to deviate for self-interested agents) in an artificial system, we need a deterrent mechanism. In this paper we present a framework (CA3NONY: Contextual Anti-coordination in Ad-hoc Anonymous games) for repeated al- location games with discounting. CA3NONY reproduces courtesy based on the allocation algorithm of [13], and uses monitoring and punishments as a deterrent mechanism. It exhibits fast convergence, and high efficiency and fairness, while relying only on occupancy feedback which facilitates scalability and does not require any inter- agent communication. The main contributions are: • Introduction of an anti-coordination framework (CA3NONY) which consists of a courteous convention and a monitoring scheme. • Proof that under such a framework, the use of the courteous convention induces strategies that constitute an approximate subgame-perfect equilibrium. • Comparison to state-of-the-art bandit algorithms. 2 RELATED WORK In ad-hoc multi-agent (anti-)coordination the goal is to design au- tonomous agents that achieve high flexibility and efficiency in a setting that admits no prior coordination between the participants [33]. Typical scenarios include the use of Monte Carlo algorithms [6], Bayesian learning [2], or bandit algorithms [7, 11]. Tradition- ally, pure ad-hoc approaches suffer from slow learning [15], which makes pure ad-hoc coordination a very ambitious goal for real- life applications. In this paper we propose a middle-ground ap- proach. Inspired by human ad-hoc coordination, we incorporate prior knowledge in the form of simple conventions. The coordination can still be considered ad-hoc as it is not pre-programmed, rather it involves learning. This allows for faster convergence compared to pure ad-hoc approaches. A convention is defined as a customary, expected and self-enforcing behavioral pattern [24, 37]. In MAS, there are two scopes through which we study conventions. First, a convention can be consid- ered as a behavioral rule, designed and agreed upon ahead of time or decided by a central authority [32, 35]. Second, a convention may emerge from within the system itself through repeated inter- actions [26, 35]. The proposed courteous convention falls on the first category. It is incorporated as prior knowledge, and it is self- enforcing since the induced strategies constitute an approximate subgame-perfect equilibrium of the repeated allocation game. An alternative way to model the anti-coordination problem is as a multi-armed bandit (MAB) problem [3]. In MAB problems an agent is given a number of arms and at each time-step has to decide which arm to pull to get the maximum expected reward. Bandit (or no-regret) algorithms typically minimize the total regret of each agent, which is the difference between the expected received payoff and the payoff of the best strategy in hindsight. As such, they sat- isfy our rationality constraint since they constitute an approximate correlated or coarse correlated equilibrium [21, 27, 30]. However, the studied problem presents many challenges: there is no station- ary distribution (adversarial rewards), all agents are able to learn (similar to recursive modeling) which results to a moving-target problem, and yielding gives a reward of 0 (desirable option for minimizing regret, but not in respect to fairness). Moreover, regret minimization does not necessarily lead to payoff maximization [15]. Nevertheless, due to their ability to learn from partial feedback, bandit algorithms constitute the natural choice for a pure ad-hoc approach. The latter motivates our choice to use them as a baseline, since our agents only receive binary feedback of success or failure upon taking their action. Game theoretic equilibria are desirable (since they satisfy the rationality constraint), but hard to obtain. Deciding whether an anti-coordination (anonymous) game has a pure Nash equilibrium (NE) is NP-complete [10]. Furthermore, allocation games often ad- mit undesirable equilibria: pure NE which are efficient but not fair, or mixed-strategy NE which are fair but not efficient [13]. Hence, iterative best-response algorithms are not satisfactory. On the other hand, correlated equilibria (CE) [5] can be both efficient & fair, while from a practical perspective they constitute perhaps the most relevant non-cooperative solution concept [21]. An optimal CE of an anonymous game may be found in polynomial time [29]. More- over, it is possible to achieve a CE without a correlation device (central agent) [18, 21], using the history of the actions taken by the opponents. However, we are interested in information-restrictive learning rules (i.e. completely uncoupled [34]), where each agent is only aware of his own history of action/reward pairs. Such an approach was applied in [13] to design a decentralized algorithm for reaching efficient & fair CE in wireless channel allocation games. Yet, while the algorithm reaches an equilibrium in a polynomial number of steps, cooperation to achieve this state is not rational. A self-interested agent could keep accessing a resource forever, until everyone else backs off (also known as 'bully' strategy [25]). In this paper, we build upon the ideas of Cigler and Faltings and de- velop an anti-coordination strategy that constitutes an approximate subgame-perfect equilibrium, i.e. cooperation with the algorithm is a best-response strategy at each sub-game of the original stage game, given any history of the play. A generalization of anti-coordination games, called dispersion games, was described in [19]. In a dispersion game, agents are able to choose from several actions, favoring the one that was chosen by the smallest number of agents (analogous to minority games [12]). The authors in [19] define a maximal dispersion outcome as an outcome where no agent can switch to an action chosen by fewer agents. The agents themselves do not have any particular preference for the attained equilibrium. Contrary to that, we are interested in achieving an efficient and fair outcome. Besides, in many real world applications, the agents are indifferent to which role/task/resource they attain, as long as they receive one (e.g. wireless frequencies). Tackling dispersion games, and therefore non-binary utilities, remains open for future research. A similar approach to ours was introduced for weighted match- ing in [17]. The authors propose a heuristic which is decentralized, requires only partial feedback, and has constant in the total problem size running time, under reasonable assumptions on the preference domain of the agents. However, it is applicable in cooperative sce- narios, and not in the presence of strategic agents. 3 THE CA3NONY FRAMEWORK 3.1 The Repeated Allocation Game Let a 'resource' be any element that can be successfully assigned to only one agent at a time. At each time-step, N = {1, . . . , N} agents try to access R = {1, . . . , R} identical and indivisible resources, where possibly N ≫ R. The set of available actions is denoted as A = {Y , A1, . . . , AR}, where Y refers to yielding and Ar refers to accessing resource r. We assume that access to a resource is slotted and of equal duration 1. A successful access yields a positive payoff, while no access has a payoff of 0. If more than one agent access a resource simultaneously, a collision occurs and the colliding parties incur a cost ζ < 0. Thus, the agents only receive a binary feedback of success or failure. Let an denote agent n's action, and a−n = ×∀n′∈N\{n}an′ the joint action for the rest of the agents. The payoff function is defined as: un(an, a−n) = if an = Y if an (cid:44) Y ∧ ai (cid:44) an,∀i (cid:44) n 0, 1, ζ , otherwise (1)  We assume that rewards are discounted by δ ∈ (0, 1), and, con- forming to real-world scenarios, that each agent n is only aware of his own history of action/reward pairs. Finally, we assume that the agents can observe side information from their environment at each time-step t. We call this side infor- mation context (e.g. time, date etc.). The agents utilize this context as a common signal in their decision-making process, a means to learn and anti-coordinate their actions. Let K = {1, . . . , K} de- note the context space. The rationale behind the introduction of the common context is that, under completely uncoupled learning rules, having positive probability mass on undesirable actions (e.g. collisions) is unavoidable. Moreover, from a practical perspective, common environmental signals are amply available to the agents [21]. We do not assume any a priori relation between the context space and the problem. The only constraints are that the values should repeat periodically, and satisfy K = ⌈N/R⌉. The context signals could be produced either on the resource side (e.g. by the port authority in maritime traffic management, or the access point in wireless networks), or in a decentralized manner (e.g. in distributed networks with no authorities the senders can attach identifier signals to data traffic [36]). Finally, in situations where communication is possible, the agents can agree upon the signal themselves by solving the distributed consensus problem. 3.2 Adopted Convention The adopted convention is based on the cooperative allocation algorithm of [13]. Each agent n has a strategy дn : K → A that determines a resource to access at time-step t after having observed context kt . The strategy is initialized uniformly at random in A. If дn(kt) = Ar , then agent n accesses resource r. Otherwise, if дn(kt) = Y, the agent does not access a resource but instead chooses uniformly at random a resource r to monitor for activity. If it is free, then the agent updates дn(kt) ← Ar (see Alg. 1). 1This is done to facilitate the proofs. Real world problems have this property anyway (e.g. access to radio channels, role allocation in a plan, etc.), and the algorithms we compare to all work for slotted resources. else Agent n accesses resource r if Collision(r) then Set дn(kt) ← Y with probability pbackof f > 0 Set accessedn ← True Agents observe context kt if дn(kt) = Ar & accessedn = False then Algorithm 1 Learning rule. Require: Initialize дn u.a.r. in A. Set accessedn ← False. 1: for kt ∈ K do 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: end for 17: Set accessedn ← False end if Agent n monitors random resource r ∈ R if Free(r) then else if дn(kt) = Y then Set дn(kt) ← Ar end if end if In [13], agents back-off probabilistically in case of a collision (set дn(kt) ← Y with probability pnbackof f ). In such a setting, it is possible to reach a symmetric subgame-perfect equilibrium. But in order to actually play it, the agents need to be able to calculate it. It is not always possible to obtain the closed form of the back- off probability distribution of each resource. Furthermore, a self- interested agent could stubbornly keep accessing a resource forever, until everyone else backs off ('bully' strategy [25]). Instead, we adopted a simple convention where agents are being courteous, i.e. if there is a collision, the colliding agents will back-off = p > 0,∀n ∈ N. with some constant positive probability: pnbackof f Being courteous, though, does not satisfy the rationality constraint. However, a uniform distribution of resources is socially optimal (i.e. fair allocations maximize the social welfare). Hence, if we introduce quotas to the resources and punishments upon violating them, courtesy induces rational strategies. In the following sections we introduce a monitoring scheme and prove that the resulting strategy constitutes an ϵ-subgame-perfect equilibrium. 3.3 Rationality In order to ensure the proposed convention's rationality, the agents must be assured that they will eventually be successful, i.e. we must provide safeguards against the monopolization of resources. The proposed framework employs simple Monitoring Authorities (MA). The MAs do not require any planning capabilities, nor any knowl- edge of the agents' plans and preferences. Their primary goal is to keep track of successful accesses. Depending on the domain, MAs may already exist naturally, or can easily be introduced. Examples include centralized MAs like the port authority in maritime traf- fic management, or a set of decentralized MAs (one per resource) like the access points in wireless networks, or agents monitoring a charging / parking station. Their function is twofold. First, they could provide occupancy signals. Agents (e.g. CA3NONY , bandit, or Q-learning agents) must be able to receive some form of feed- back from their environment to inform collisions and detect free resources. This could be achieved (if possible) by the use of various sensors, or by receiving occupancy signals (e.g. 0, 1) from the MAs. Second, (which is their principal purpose) MAs deter agents from monopolizing resources to the point that each agent can access a resource only for one context value out of K. To achieve the latter, MAs must be able to keep track of successful accesses. Upon the violation of the imposed quotas, the framework is responsible to enforce the necessary punishments. Punishments are application specific. They can be individual (e.g. in a wireless scenario, if an agent transmits to some other than the designated channel, then his packets will no longer be relayed), or group punishments (e.g. if quotas are exceeded, access is denied for everyone). The imposed punishments make exceeding the quotas an irrational strategy (sim- ulated in Alg. 1 by the accessedn flag), which in turn aligns the individual incentives with an efficient & fair correlated equilibrium. 3.3.1 Access Monitoring. The primary function of the employed MAs is the tracking of successful accesses by the agents. The latter can be achieved in various ways, depending on the domain. The simplest one would be to maintain a log with successful accesses per episode (period of context values), whether we are dealing with a centralized MA or set of decentralized MAs that are able to communicate to ensure coherence. For more involved scenarios (e.g. if we require agent anonymity per access) we may employ more complex schemes, e.g. using tokens, or artificial currency (solely as an internal mechanism). Note that the latter does not fence the resources (i.e. punishments are still required). In what follows, we present such a self-regulated monitoring scheme. Initially all the agents that 'buy-in' are issued the same amount m ∈ R of artificial cash (AC). This amount also corresponds to the initial fee for every resource fr ← m. To allow access to resource r, the MA of r charges fr units of AC, and monitors the event. If there was a successful access, the MA reimburses the amount of (1 − ξ)fr AC to the accessing agent, where ξ → 0 ∈ R is a commission fee. Otherwise, the MA reimburses the full amount of fr AC to the colliding agents, so that they are able to try again for a different context value. Finally, after each episode, the MAs r ← (1 − ξ)fr ,∀r ∈ R. After every successful lower the fee to f ′ access, the amount of AC that an agent possesses drops below the access fee of a resource. Waiting for the fee to drop to the point that fr = m/2 is not rational since, assuming ξ → 0, the number of iterations required to allow accessing two resources at the same time will reach ∞. At that point the rest of the agents will have reached a correlated equilibrium and the adversarial agent will not have an incentive to access an additional resource, besides the one that corresponds to him, since it would result in a collision. If, due to implementation constraints, we can not select a small ξ, the MAs can change the artificial currency every I episodes, invalidating the old one and again making such strategy irrational. 3.4 Rate of convergence Theorem 3.1. In a repeated allocation game with N agents and R resources, the expected number of steps before Alg. 1 converges to a correlated equilibrium is bounded by: (cid:18) (cid:18) (cid:19) (cid:25) (cid:24) N R O N log + 1 (log N + R) (cid:19) (2) 2Slightly tighter bound on the convergence speed of the adopted learning rule, based on Theorems 12 and 13 of [13]. See the appendix for the proof. (cid:18) Proof. The adopted learning rule is based on the allocation algorithm of [13]. For N , R, K ≥ 1, and back-off probability 0 < p < 1, the expected number of steps before the algorithm converges is bounded by (3) 2, which for a constant back-off probability and K = ⌈N/R⌉ gives the required bound. 2 − p 2(1 − p) (K log K + 2K) R log N + R (cid:18) 1 (cid:19)(cid:19) □ Corollary 3.2. Under a common, constant back-off probability assumption, p∗ = 2 − √2 minimizes the convergence time of Alg. 1 in high congestion scenarios, i.e. N ≫ R. (3) O p Proof. According to bound (3), in high congestion scenarios = K → ∞), the common, constant back-off probability that (i.e. N R minimizes the convergence time is: (cid:19) (cid:18) 2 − p 2(1 − p) 1 p ∗ = arg min p = 2 − √ 2 (4) □ 3.5 Courtesy Pays Off In this section we prove that if the agents back-off with a constant positive probability pbackof f > 0, then Alg. 1 induces a strategy that is an ϵ-equilibrium. Suppose that in a repeated allocation game with discounting (δ ∈ (0, 1)) the agents who collide back-off with a constant prob- ability pbackof f > 0. Let σp denote the aforementioned strategy (courteous strategy), and σ∗ denote the optimal (best-response) strategy under the monitoring authorities (possibly better than a n, at−n) denote the cumulative payoff of agent n following strategy σ, as- suming the rest of the agents follow the strategy σ−n. The following theorem proves that starting at any time-step, agent n does not gain more than ϵ by deviating to the optimal strategy σ∗. stage game NE). Moreover, let Un(σ , σ−n, δ) =∞ t =0 δ tun(at p n , σ E[Un(σ Proof. Note that E[Un(σ∗ Theorem 3.3. Under a high enough discount factor, the courteous strategy σp constitutes an approximate subgame-perfect equilibrium, i.e. ∀ϵ > 0, ∃δ0 ∈ (0, 1) such that ∀δ, δ0 ≤ δ < 1: ∗ p−n, δ)] > (1 − ϵ)E[Un(σ n, σ n, σ∗−n, δ)] ≥ E[Un(σ∗ p−n, δ)], since n, σ p−n the rest of the agents can potentially introduce addi- by playing σ p−n, δ)] > tional collisions. Thus, it suffices to prove that E[Un(σ n, σ∗−n, δ)]. (1 − ϵ)E[Un(σ∗ The introduced monitoring scheme prohibits the monopolization of resources, i.e. each agent can only access a resource for his corresponding context value. Thus, the best-response strategy's (σ∗) payoff for some δ is bounded by: p−n, δ)] p n , σ E[Un(σ ∗ n, σ ∗−n, δ)] ≤ 1 + δ K + δ 2K + · · · = δiK = 1 1 − δ K (5) ∞ i =0 When agents adopt the courteous convention, in each round until the system converges to a correlated equilibrium, the agents receive a payoff between ζ < 0 (collision cost) and 1. Thus, until conver- 1−δ τ K gence, the expected payoff is lower bounded by ζ 1−δ K , where τ is the number of steps to converge. After convergence, their expected payoff is 1−δ K . Hence, the convention induced strategy's payoff is at least: δτ +iK = δ τ τ−1 δiK = ζ i =0 i =0 ∞ ∞ τ =1 (cid:18) ζ(1 − δτ K) + δτ (cid:19) 1 − δ K E[Un(σ p n , σ p−n, δ)] ≥ Pr[conv. in τ steps] · We can define a random variable X such that X = τ if the algorithm converges after exactly τ steps. Since δ x is a convex function we have that E(δ x) ≥ δ E(x), therefore: E[Un(σ p n , σ p−n, δ)] ≥ ζ(1 − δ E(X)K) + δ E(X) 1 − δ K By dividing (6) by (5) we get: (6) p n , σ p−n, δ)] E[Un(σ E[Un(σ∗ n, σ∗−n, δ)] ≥ ζ(1 − δ (7) E(X) does not depend on δ. Moreover, δ E(X) is continuous in δ, δ→1−δ E(X) = 1. Thus, we can take the limit of E(X)K) + δ E(X) monotonous, and lim (7) as δ → 1−, which equals to lim δ→1− E[Un(σ p E[Un(σ ∗ n ,σ p−n,δ)] n,σ ∗−n,δ)] = 1 □ In order to guarantee rationality, the discount factor δ must be close to 1 since, as δ gets closer to 1, the agents do not care whether they access now or in some future round. Since the proposed moni- toring scheme guarantees that every agent will access a resource for his corresponding context value, when δ → 1, the expected payoff for agents who are accessing a resource and for those who have not accessed a resource yet will be the same. In other words, the cost (overhead) of learning the correlated equilibrium decreases. 3.6 Indifference Period In many real world applications, agents are indifferent in claiming a resource in a period of Tind rounds, i.e. δt = 1,∀t ≤ Tind. E.g. data of wireless transmitting devices might remain relevant for a specific time-window, during which the agent is indifferent of transmitting. In such cases, we can use the Markov bound to prove that with high probability the proposed algorithm will converge in under Tind time-steps, thus satisfying the rationality constraint. We assume the agents are willing to accept linear 'delay' with regard to the number of resources R, the number of agents N , and the size of the context space K, specifically: Tind = O (RN K) (8) Theorem 3.4. Under a linear indifference period Tind (i.e. δt = 1,∀t ≤ Tind = O (RN K)) the probability of the system of agents following Alg. 1 not having converged during Tind diminishes as the congestion increases ( N R → ∞). (cid:170)(cid:174)(cid:174)(cid:172) □ Proof. Using the Markov bound, it follows that the probability that the system takes more than the accepted number of steps (Tind) to converge is: Pr[¬conv. after Tind] = O(cid:169)(cid:173)(cid:173)(cid:171) (cid:16)log(cid:6) N R (cid:7) + 1(cid:17) (log N + R) N Taking the limit: lim N R →∞Pr[¬conv. after Tind] = 0. Even though the Markov's inequality generally does not give very good bounds when used directly, Th. 3.4 proves that our al- gorithm converges in the required time with high probability. The latter holds under high congestion, which constitutes the more interesting scenario since for small number of agents the required time to converge is only a few hundreds of time-steps. Moreover, the higher the indifference period, the higher the probability to converge under such time-constraint. For quasilinear indifference period for example, the system converges in the required time- window with high probability even for a small number of agents. We can further strengthen our rationality hypothesis by using a tighter bound (e.g. Chebyshev's inequality), albeit computing the theoretical variance of the convergence time is an arduous task, thus it remains open for future work. 4 EVALUATION In this section we model the resource allocation problem as a multi-armed bandit problem and provide simulation results of CA3NONY's performance in comparison to state-of-the-art, well established bandit algorithms, namely the EXP4 [4], EXP4.P [8], and EXP3 [4]. In every case we report the average value over 128 runs of the same simulation. For the EXP family of algorithms, the input parameters are set to their optimal values, as prescribed in the aforementioned publications. We assume a reward of 1 for a successful access, −1 if there is a collision, and 0 if the agent yielded. 4.0.1 Level of Courtesy. We evaluated different back-off prob- abilities (pbackof f ∈ {0.1, 0.25, 2 − √2, 0.75, 0.9}) for R = K ∈ {2, 4, 8, 16} and N = R × K. There was no significant difference in convergence time, but since pbackof f is directly correlated with the number of collisions, it can have significant impact on the average payoff (see Table 1). It is important to note, though, that agents do not have global nor local knowledge of the level of congestion of the system (they only receive binary occupancy feedback for one resource per time-step). Thus, agents can not select the optimal back-off probability depending on congestion. In what follows, the back-off probability of CA3NONY is set to pbackof f = 2 − √2 of Eq. 4, which is only optimal under high congestion settings (i.e. = K → ∞), yet constitutes a safe option. Note that the provided N R theoretical analysis holds for any constant back-off probability. 4.0.2 Bandits & Monitoring. CA3NONY has three meta actions (Access, Yield, and Monitor {A, Y , M}), while bandit algorithms have only two ({A, Y}), and CA3NONY assumes the existence of monitoring authorities (MAs). For fairness' sake in the reported results, we made the following two modifications. First, we include Table 1: Avg. payoff depending on the level of courtesy (back- off probability), K = R, N = R × K. pbackof f 0.1 0.25 2 − √2 0.75 0.9 R = 2 37.6 43.6 45.7 45.5 44.2 R = 4 -4.7 10.0 15.6 15.7 12.5 R = 8 -39.5 -16.9 -5.2 -3.7 -5.8 R = 16 -63.1 -40.0 -23.0 -20.6 -19.4 a variation of CA3NONY , denoted as CA3NONY, where we as- sume agents incur a cost (equal to collision cost ζ ) every time they monitor a resource (reflected in Table 3). Second, all the employed bandit algorithms make use of the MAs, which indirectly grants them the the ability to monitor resources as well. More specifically, the accumulated payoff is updated only if they are allowed to access a resource. If not, we consider it a monitoring action, which means the agents still receive occupancy feedback. The latter is important, otherwise the bandit algorithms would require significantly longer time to converge. Note that, contrary to CA3NONY, monitoring is free for bandit algorithms with respect to the accumulated payoff, i.e. they do not incur a collision cost. 4.1 Employed Bandit Algorithms The reward of each arm does not follow a fixed probability distribu- tion (adversarial setting). Moreover, the agents are able to observe side-information (context) at each time-step t. The arm that yields the highest expected reward can be different depending on the con- text. Hence we focus on adversarial contextual bandit algorithms (see [38] for a survey). A typical approach is to use expert advice. In this method we assume a set of experts M = {1, . . . , M} who generate a probability distribution on which arm to pull depending on the context. A no-regret algorithm performs asymptotically as well as the best expert. Such algorithms are the EXP4 and EXP4.P, the difference being that EXP4 exhibits high variance [38], while EXP4.P achieves the same regret with high probability by combining the confidence bounds of UCB1 [3] and EXP4. The computational complexity and memory requirements of the above algorithms are linear in M, making them intractable for large number of experts. In order to deal with the increased complexity in larger simulations, we gave an edge to these algorithms by restricting the set of experts M to the same uniform correlated equilibria (CE) that CA3NONY converges to. The latter enabled us to perform larger simulations (the unrestricted versions could not handle N ≥ 64), while resulting in the same order of magnitude convergence time, slightly higher average payoff, and significantly higher fairness (> 35% since by design the experts proposed fair CE). Alternatively, we can use non-contextual adversarial bandit algorithms, such as the EXP3. Moreover, we can convert EXP3 to a contextual algorithm by set- ting up a separate instance for all k ∈ K. We call this 'CEXP3'. This results in a contextual bandit algorithm which has the edge over EXP4 from an implementation viewpoint since its running time at each time-step is O(R) and its memory requirement is O(KR) (CA3NONY's is O(1) and O(K) respectively). Figure 1: Utilization: R = 4, N = 16 (x-axis in log scale). 4.2 Simulation Results 4.2.1 Convergence Speed & Efficiency. We know that CA3NONY converges to a CE which is efficient. If all agents follow the cour- teous convention of Alg. 1, the system converges to a state where no resources remain un-utilized and there are no collisions (Th. 13 of [13]). Furthermore, Th. 3.1 argues for fast convergence. The former are both corroborated by Fig. 1 (similar results (> ×102 faster convergence) were acquired for R ∈ {2, 8, 16} as well.). Fig. 1 depicts the total utilization of resources for a simulation period of T = 106 time-steps. Note that the x-axis is in logarithmic scale. CA3NONY converges significantly (> ×102) faster than the bandit algorithms to a state of 100% efficiency. On the other hand, the ban- dit algorithms exhibit high variance, never achieve 100% efficiency, and are not able to handle efficiently the increase in context space size and number of resources. 4.2.2 the Jain index is given by J(x) = Fairness. The usual predicament of efficient equilibria for allocation games is that they assign the resources only to a fixed sub- set of agents, which leads to an unfair result (e.g. an efficient pure NE (PNE) is for R agents to access and N − R agents to yield). This is not the case for CA3NONY , which converges to an equilibrium that is not just efficient but fair as well. Due to the enforced mon- itoring scheme, all agents acquire the same amount of resources. As a measure of fairness, we will use the Jain index [22]. The Jain index is widely used by economists, and exhibits a lot of desirable properties such as: population size independence, continuity, scale and metric independence, and boundedness. For an allocation game th agent receives an allocation of xn, of N agents, such that the n 2 n. An n∈N x allocation is considered fair, iff J(x) = 1, x = (x1, . . . , xN )⊤. Table 2 presents the expected Jain Index of the evaluated algo- rithms at the end of the time horizon T . CA3NONY converges to a fair equilibrium, achieving a Jain index of 1. The EXP4 and EXP4.P were the fairest amongst the bandit algorithms, achieving a Jain index of close to 1. This is to be expected since the set of experts M is limited to the same set of equilibria that CA3NONY converges to. On the other hand, EXP3 and CEXP3 performed considerably worse, with the EXP3 exhibiting the worst performance in terms of fairness, equal to a PNE's: JP N E(x) = R2 n∈N xn2 / N = 1 K = JEX P3(x). N R 4.2.3 Average Payoff. The average payoff corresponds to the total discounted payoff an agent would receive in the time horizon T . Table 3 presents the average payoff for the studied algorithms. The clustered pairs of bandit algorithms exhibited < 1% difference, hence we included the average value of the pair. The discount factor was set to δ = 0.99. At the top half we do not assume any Table 2: Fairness (Jain Index), K = R, N = R × K. CA3NONY EXP3 CEXP3 EXP4 EXP4.P R = 2 1.0000 0.5000 0.7018 1.0000 1.0000 R = 4 1.0000 0.2500 0.5865 0.9999 0.9999 R = 8 1.0000 0.1250 0.5341 0.9959 0.9798 Table 3: Average Payoff, K = R, N = R × K. CA3NONY CA3NONY* (C)EXP3 EXP4(.P) CA3NONY CA3NONY* (C)EXP3 EXP4(.P) R = 2 45.4 20.0 -72.4 -59.0 53.3 23.8 -84.0 -68.4 R = 4 15.7 -22.4 -93.1 -81.0 79.0 -55.5 -331.0 -288.4 R = 8 -5.3 -50.6 -99.8 -90.7 502.5 -1337.1 -4175.7 -3807.1 R = 16 1.0000 0.0625 0.9638 0.9198 0.7503 R = 16 -23.0 -72.6 -100.0 -95.4 4057.5 -26723.9 -60595.1 -62631.5 indifference period, while at the bottom half we assume linear indifference period (δt = 1,∀t ≤ Tind, where Tind is given by Eq. 8). Recall that CA3NONY and the bandit algorithms do not incur any cost for monitoring resources, while CA3NONY* incurs a cost equal to the collision cost ζ . Once more, CA3NONY (and even CA3NONY*) significantly out- performs all the bandit algorithms. The latter have relatively similar performance, with EXP4 and EXP4.P being the best amongst them. It is worth noting that adding an indifference period has a dramatic effect on the results. Comparing CA3NONY to bandit algorithms we observe that CA3NONY achieves a large increase on average payoff, while the opposite happens for the bandit algorithms. This is because the learning rule of Alg. 1 prohibits from accessing an al- ready claimed resource, thus minimizing collisions. On the contrary, bandit algorithms constantly explore (they assign a positive proba- bility mass to every arm) which leads to collisions. In a multi-agent system where every agent learns this can have a cascading effect. The latter becomes apparent when fixing δ = 1 for Tind steps. The collision cost remains high for longer which, as seen by Table 3, has a significant impact on the bandit algorithms' performance. 4.2.4 Large Scale Systems. The innovation of CA3NONY stems from the adoption of a simple convention, which allows its applica- bility to large scale MAS. To evaluate the latter, Fig. 2 and 3 depict the convergence time for increasing number of resources R, and increasing system congestion ( N = K) respectively. Both graphs R are in a double logarithmic scale, and the error bars represent one standard deviation of uncertainty. The total number of agents is given by N = R × K. Thus, the largest simulations involve 16.384 agents. Along with CA3NONY, we depict the fastest (based on the previous simulations) of the bandit algorithms, namely EXP3. In both cases we acquire ×103 − ×105 faster convergence. The above validate CA3NONY's performance in both scenarios with abun- dance (N ≈ R or small K), and scarcity of resources (N ≫ R or large K). As depicted, CA3NONY is significantly faster than the Figure 2: Convergence time: increasing #resources R, vary- ing context space size K, N = R × K (double log scale). Figure 3: Convergence time: increasing congestion ( N R varying #resources R, N = R × K (double log scale). = K), EXP3 and can gracefully handle increasing number of resources, and high congestion. Finally note that, in several of the simulations, EXP3 was unable to reach its convergence goal of 90% efficiency (utilization of resources) in a reasonable amount of computation time (1.5× 108 time-steps), hence the resulting gaps in EXP3's lines in Fig. 2 and 3. Especially in situations with scarcity of resources the utilization was significantly lower. 5 CONCLUSION In this paper we proposed CA3NONY, an anti-coordination frame- work under rationality constraints. It is based on a simple, human- inspired convention of courtesy which prescribes a positive back- off probability in case of a collision. Coupled with a monitoring scheme which deters the monopolization of resources, we proved that the induced strategy constitutes an ϵ-subgame-perfect equilib- rium. We compared CA3NONY to state-of-the-art bandit algorithms, namely EXP3, EXP4, and EXP4.P. Simulation results demonstrated that CA3NONY outperforms these algorithms by achieving more than two orders of magnitude faster convergence, a fair allocation, and higher average payoff. The aforementioned gains suggest that human-inspired conventions may prove beneficial in other ad-hoc coordination scenarios as well. A APPENDIX A.1 Proof of Bound (3) In this section we provide a formal proof of bound (3). The proof is an adaptation of the convergence proof of [13]. We will prove the following theorem: Theorem A.1. For N agents and R ≥ 1, K ≥ 1, 0 < p < 1 the expected number of steps before the learning algorithm converges to an efficient correlated equilibrium of the allocation game for every k ∈ K is O(cid:16)(K log K + 2K) R + 1(cid:17) (cid:17) log N + R . (cid:16) 1 p 2−p 2(1−p) We begin with the case of having multiple agents but only a sin- gle resource (R = 1). We will describe the execution of the proposed learning rule as a discrete time Markov chain (DTMC) 3. In every time-step, each agent performs a Bernoulli trial with probability of 'success' 1 − p (remain in the competition), and failure p (back- off). When N agents compete for a single resource, a state of the system is a vector {0, 1}N denoting the individual agents that still compete for that resource. But, since the back-off probability is the same for everyone, we are only interested in how many agents are competing and not which ones. Thus, in the single resource case (R = 1), we can describe the execution of the proposed algorithm using the following chain: Definition A.2. Let {Xt }t ≥0 be a DTMC on state space S = {0, 1, . . . , N} denoting the number of agents still competing for the resource. The transition probabilities are as follows: Pr(Xt +1 = N Xt = 0) = 1 Pr(Xt +1 = 1Xt = 1) = 1 Pr(Xt +1 = j Xt = i) = (cid:18)i (cid:19) j restart absorbing i > 1, j ≤ i pi−j(1 − p)j (all the other transition probabilities are zero) Intuitively, this Markov chain describes the number of individ- uals in a decreasing population, but with two caveats: The goal (absorbing state) is to reach a point where only one individual remains, and if we reach zero, we restart. Lemma A.3. [13] Let A = {0, 1}. The expected hitting time of the set of states A in the Markov chain described in Definition A.2 is O(cid:16) 1 (cid:17) log N . p Definition A.4. Let {Yt }t ≥0 be a DTMC on state space S = {0, 1, . . . , N} with the following transition probabilities (two ab- sorbing states, 0 and 1): Pr(Yt +1 = 0Yt = 0) = 1 Pr(Yt +1 = 1Yt = 1) = 1 Pr(Yt +1 = j Yt = i) = (cid:18)i j (cid:19) pi−j(1 − p)j absorbing absorbing i > 1, j ≤ i (all the other transition probabilities are zero) Let hA i denote the hitting probability of a set of states A, starting from state i. We will prove the following lemma. Lemma A.5. The hitting probability of the absorbing state {1}, starting from any state i ≥ 1, of the DTMC {Yt } of Definition A.4 is given by Eq. 9. This is a tight lower bound. (9) (cid:18) 2(1 − p) (cid:19) 2 − p {1} i h = Ω , ∀i ≥ 1 {1} i ∆= h Proof. For simplicity we denote hi for p ∈ (0, 1), hi ≥ λ = 3For an introduction on Markov chains see [28] . We will show that 2−p ,∀i ≥ 1 using induction. First note 2(1−p) that since state {0} is an absorbing state, h0 = 0, h1 = 1 ≥ λ and that λ ∈ (0, 1). for a set of states A is the minimal non-negative solution to the system of linear equations 10: The vector of hitting probabilities hA = (hA i : i ∈ S = {0, 1, . . . , N}) = = 1, j∈S i hA i hA i = 1, = j =0 hA i hA i if i ∈ A if i (cid:60) A pi jhA j , (10) By replacing pij with the probabilities of Definition A.4, the system of equations 10 becomes: (cid:0)i j (cid:1)pi−j(1 − p)jhA j , if i ∈ A if i (cid:60) A (11) Base case: (cid:18)i i−1 (cid:18)i i−1 j =2 j (cid:19) (cid:19) 2p(1 − p) 1 − (1 − p)2 = 2(1 − p) 2 − p h2 = (1 − p)2h2 + 2p(1 − p)h1 + p2h0 = ≥ λ Inductive step: We assume that ∀j ≤ i − 1 ⇒ hj ≥ λ. We will prove that hi ≥ λ,∀i > 2. (cid:19) hi = pi−j(1 − p)jhj i (cid:18)i j j =0 = pi h0 + ipi−1(1 − p)h1 + pi−j(1 − p)jhj + (1 − p)i hi ≥ pi h0 + ipi−1(1 − p)h1 + = ipi−1(1 − p) + [1 − pi − (1 − p)i − ipi−1(1 − p)]λ + (1 − p)i hi pi−j(1 − p)j λ + (1 − p)i hi j =2 j ⇒ hi = λ − pi 1 − (1 − p)i λ + ipi−1(1 − p) 1 − (1 − p)i (1 − λ) We want to prove that hi ≥ λ: pi ipi−1(1 − p) 1 − (1 − p)i (1 − λ) ≥ 1 − (1 − p)i λ ⇒ ipi−1(1 − p) + pi − pi ≥ λ ⇒ pi + ipi−1(1 − p) pi + ipi−1(1 − p) ≥ 2(1 − p) 1 − 2 − p pi + ipi−1(1 − p) ≤ p ⇒ 2 − p ⇒ pi pi pi(2 − p) ≤ p[pi + ipi−1(1 − p)] ⇒ 2 − 2p − i + ip ≤ 0 ⇒ 2 − i − p(2 − i) ≤ 0 ⇒ (2 − i)(1 − p) ≤ 0 ⇒ 2 − i ≤ 0 which holds since i > 2. The above bound is also tight since ∃i ∈ S : hi = λ, specifically h2 = λ. □ Plugging the above hitting probability bound to the convergence proof of [13], results on the required bound. REFERENCES [1] Lucas Agussurja, Akshat Kumar, and Hoong Chuin Lau. 2018. Resource- Constrained Scheduling for Maritime Traffic Management. https://www.aaai. org/ocs/index.php/AAAI/AAAI18/paper/view/17116 [2] Stefano V Albrecht, Jacob W Crandall, and Subramanian Ramamoorthy. 2016. Belief and truth in hypothesised behaviours. Artificial Intelligence 235 (2016), 63 -- 94. [3] Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer. 2002. Finite-time Analysis of the Multiarmed Bandit Problem. Machine Learning 47, 2 (2002), 235 -- 256. https://doi.org/10.1023/A:1013689704352 [4] Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert E Schapire. 2002. The nonstochastic multiarmed bandit problem. SIAM journal on computing 32, 1 (2002), 48 -- 77. [5] Robert J. Aumann. 1974. Subjectivity and correlation in randomized strategies. Journal of Mathematical Economics 1, 1 (1974), 67 -- 96. https://doi.org/10.1016/ 0304-4068(74)90037-8 [6] Samuel Barrett, Avi Rosenfeld, Sarit Kraus, and Peter Stone. 2017. Making friends on the fly: Cooperating with new teammates. Artificial Intelligence 242 (2017), 132 -- 171. [7] Samuel Barrett and Peter Stone. 2011. Ad hoc teamwork modeled with multi- armed bandits: An extension to discounted infinite rewards. In Proceedings of 2011 AAMAS Workshop on Adaptive and Learning Agents. 9 -- 14. [8] Alina Beygelzimer, John Langford, Lihong Li, Lev Reyzin, and Robert Schapire. 2011. Contextual bandit algorithms with supervised learning guarantees. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics. 19 -- 26. [9] Yann Bramoullé, Dunia López-Pintado, Sanjeev Goyal, and Fernando Vega- Redondo. 2004. Network formation and anti-coordination games. International Journal of Game Theory 33, 1 (2004), 1 -- 19. [10] Felix Brandt, Felix Fischer, and Markus Holzer. 2009. Symmetries and the Com- plexity of Pure Nash Equilibrium. J. Comput. Syst. Sci. 75, 3 (May 2009), 163 -- 177. https://doi.org/10.1016/j.jcss.2008.09.001 [11] Mithun Chakraborty, Kai Yee Phoebe Chua, Sanmay Das, and Brendan Juba. 2017. Coordinated Versus Decentralized Exploration In Multi-Agent Multi-Armed Bandits. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17. 164 -- 170. https://doi.org/10.24963/ijcai.2017/24 [12] Damien Challet, Matteo Marsili, Yi-Cheng Zhang, et al. 2013. Minority games: interacting agents in financial markets. OUP Catalogue (2013). [13] Ludek Cigler and Boi Faltings. 2013. Decentralized anti-coordination through multi-agent learning. Journal of Artificial Intelligence Research 47 (2013), 441 -- 473. [14] Russell Cooper. 1999. Coordination Games. Cambridge University Press. https: //doi.org/10.1017/CBO9780511609428 [15] Jacob W Crandall. 2014. Towards minimizing disappointment in repeated games. Journal of Artificial Intelligence Research 49 (2014), 111 -- 142. [16] Panayiotis Danassis and Boi Faltings. 2018. Learning in Ad-hoc Anti-coordination Scenarios. AAAI Spring Symposium Series (2018). https://www.aaai.org/ocs/ index.php/SSS/SSS18/paper/view/17485 [17] Panayiotis Danassis, Aris Filos-Ratsikas, and Boi Faltings. 2019. Anytime Heuris- tic for Weighted Matching Through Altruism-Inspired Behavior. arXiv:1902.09359 (2019). https://arxiv.org/abs/1902.09359 [18] Dean P Foster and Rakesh V Vohra. 1997. Calibrated learning and correlated equilibrium. Games and Economic Behavior 21, 1-2 (1997), 40. [19] Trond Grenager, Rob Powers, and Yoav Shoham. 2002. Dispersion games: general definitions and some specific learning results. In AAAI/IAAI. 398 -- 403. [20] Shipra Gupta and James W. Gentry. 2016. The behavioral responses to perceived scarcity âĂŞ the case of fast fashion. The International Review of Retail, Distribution and Consumer Research 26, 3 (2016), 260 -- 271. https://doi.org/10.1080/09593969. 2016.1147476 arXiv:https://doi.org/10.1080/09593969.2016.1147476 [21] Sergiu Hart and Andreu Mas-Colell. 2000. A simple adaptive procedure leading to correlated equilibrium. Econometrica 68, 5 (2000), 1127 -- 1150. [22] Raj Jain, Dah-Ming Chiu, and W. Hawe. 1998. A Quantitative Measure Of Fairness And Discrimination For Resource Allocation In Shared Computer Systems. CoRR cs.NI/9809099 (1998). http://arxiv.org/abs/cs.NI/9809099 [23] Jeremy Kun, Brian Powers, and Lev Reyzin. 2013. Anti-coordination games and stable graph colorings. In International Symposium on Algorithmic Game Theory. Springer, 122 -- 133. [24] David Lewis. 2008. Convention: A philosophical study. John Wiley & Sons. [25] Michael L. Littman and Peter Stone. 2002. Implicit Negotiation in Repeated Games. Springer Berlin Heidelberg, Berlin, Heidelberg, 393 -- 404. https://doi.org/10.1007/ 3-540-45448-9_29 [26] Mihail Mihaylov, Karl Tuyls, and Ann Nowé. 2014. A decentralized approach for convention emergence in multi-agent systems. Autonomous Agents and Multi-Agent Systems 28, 5 (2014), 749 -- 778. [27] Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V Vazirani. 2007. Algo- rithmic game theory. Vol. 1. Cambridge University Press Cambridge. [28] James R Norris. 1998. Markov chains. Number 2. Cambridge university press. [29] Christos H. Papadimitriou and Tim Roughgarden. 2008. Computing Correlated Equilibria in Multi-player Games. J. ACM 55, 3, Article 14 (Aug. 2008), 29 pages. https://doi.org/10.1145/1379759.1379762 [30] Tim Roughgarden. 2016. Twenty Lectures on Algorithmic Game Theory (1st ed.). [31] Thomas C Schelling. 1960. The strategy of conflict. Cambridge, Mass (1960). [32] Yoav Shoham and Moshe Tennenholtz. 1995. On social laws for artificial agent societies: off-line design. Artificial Intelligence 73, 1 (1995), 231 -- 252. https: //doi.org/10.1016/0004-3702(94)00007-N Computational Research on Interaction and Agency, Part 2. Cambridge University Press, New York, NY, USA. [33] Peter Stone, Gal A. Kaminka, Sarit Kraus, and Jeffrey S. Rosenschein. 2010. Ad Hoc Autonomous Agent Teams: Collaboration without Pre-Coordination. In Proceedings of the Twenty-Fourth Conference on Artificial Intelligence. [34] Mohammad Sadegh Talebi. 2013. Uncoupled Learning Rules for Seeking Equilibria in Repeated Plays: An Overview. CoRR abs/1310.5660 (2013). http://arxiv.org/ abs/1310.5660 [35] A. Walker and M. J. Wooldridge. 1995. Understanding the Emergence of Conven- tions in Multi-Agent Systems. In ICMAS95. San Francisco, CA, 384 -- 389. http: //groups.lis.illinois.edu/amag/langev/paper/walker95understandingThe.html [36] L. Wang, K. Wu, M. Hamdi, and L. M. Ni. 2013. Attachment-Learning for Multi- Channel Allocation in Distributed OFDMA-Based Networks. IEEE Transactions on Wireless Communications 12, 4 (April 2013), 1712 -- 1721. https://doi.org/10. 1109/TWC.2013.022013.120627 [37] H Peyton Young. 1996. The economics of convention. The Journal of Economic [38] Li Zhou. 2015. A survey on contextual multi-armed bandits. arXiv preprint Perspectives 10, 2 (1996), 105 -- 122. arXiv:1508.03326 (2015).
1912.01629	1	1912	2019-12-03T19:03:57	A Simulation Model for Pedestrian Crowd Evacuation Based on Various AI Techniques	[ "cs.MA", "cs.AI", "physics.soc-ph" ]	This paper attempts to design an intelligent simulation model for pedestrian crowd evacuation. For this purpose, the cellular automata(CA) was fully integrated with fuzzy logic, the kth nearest neighbors (KNN), and some statistical equations. In this model, each pedestrian was assigned a specific speed, according to his/her physical, biological and emotional features. The emergency behavior and evacuation efficiency of each pedestrian were evaluated by coupling his or her speed with various elements, such as environment, pedestrian distribution and familiarity with the exits. These elements all have great impacts on the evacuation process. Several experiments were carried out to verify the performance of the model in different emergency scenarios. The results show that the proposed model can predict the evacuation time and emergency behavior in various types of building interiors and pedestrian distributions. The research provides a good reference to the design of building evacuation systems.	cs.MA	cs	Cite as: Muhammed, D.A., Saeed, S.A.M., Rashid, T.A. (2019). A simulation model for pedestrian crowd evacuation based on various AI techniques. Revue d'Intelligence Artificielle, Vol. 33, No. 4, pp. 283-292. https://doi.org/10.18280/ria.330404 A Simulation Model for Pedestrian Crowd Evacuation Based on Various AI Techniques Danial A. Muhammed1, Soran A.M. Saeed2, Tarik A. Rashid3 1 Computer Department, College of Science, University of Sulaimani, Sulaymaniyah 46001, Kurdistan, Iraq 2 Sulaimania Polytechnic University, Sulaymaniyah 46001, Kurdistan, Iraq 3 Computer Science and Engineering, University of Kurdistan Hewler, Erbil 44001, Kurdistan, Iraq Corresponding Author Email: [email protected] Keywords: computational evacuation models, modeling, participants' emergency behavior, evacuation time, environment, engineering applications simulation, ABSTRACT This paper attempts to design an intelligent simulation model for pedestrian crowd evacuation. For this purpose, the cellular automata (CA) was fully integrated with fuzzy logic, the kth nearest neighbors (KNN), and some statistical equations. In this model, each pedestrian was assigned a specific speed, according to his/her physical, biological and emotional features. The emergency behavior and evacuation efficiency of each pedestrian were evaluated by coupling his/her speed with various elements, such as environment, pedestrian distribution and familiarity with the exits. These elements all have great impacts on the evacuation process. Several experiments were carried out to verify the performance of the model in different emergency scenarios. The results show that the proposed model can predict the evacuation time and emergency behavior in various types of building interiors and pedestrian distributions. The research provides a good reference to the design of building evacuation systems. 1. INTRODUCTION related to it such as protective, preventive, the physical movement of evacuees; Occurring emergencies in big places makes the evacuation operation difficult due to adding confusion, fear and even vagueness and anxiety to mass residents [1]. Constructed environment, evacuees' communication, and the plans and measures in place to direct the reactions of participants to environmental conditions and characteristics can influence the evacuation system. Therefore, various evacuation methods appear, rescue, and reconstructive evacuations [2]. The problem of evacuation is not solely is multifaceted and the physical and social circumstances, such as the high possibility for hazard, great stage of pressure, and inadequate data. These circumstances illustrate robust communication between environment, danger, egress process, population demographics, and participant behavior [3]. A crowd is a communication of a set of people [4]. Various features belong to crowds; therefore, during an emergency, a number of behaviors would be exposed in community areas [5], and simulation is a means of expecting behaviors via responding to the "what-if" situations [6]. Hence, Crowd evacuation simulation is a method for mimicking the flow of participants for the same goal of evacuation from the same place [4]. In the last two decades, rising dedicated attempts via researchers clearly have been seen to explore crowd evacuation during an emergency. Evacuation dynamics have been widely considered to eliminate deaths and injuries during emergency pedestrian incidences [7-13]. A vast number of models is developed to know individual behaviors, such as how individual react a specific condition and how to deal with varied kind of scenarios in the crowd during an emergency; the entire investigations have been there for creating a foundation of enhancement in managing emergencies [14-20]. Modeling is a way of answering difficulties, thus, a newly developed framework after analyzing and controlling via basic pro-test shows the authenticity of its procedures to be mentioned as a model [4]. Existing models can be separated into three main groups: the first one is classical models, the second one is hybridized models, and the third one is a generic model. Each group comprises some models, the recent past; various applications to investigate crowd evacuation in both normal and emergency situations were done due to these models. in individuals and Microscopic is one of the classical models and the details of the individual behaviors are precisely identified. However, it has a problem in dealing with the large size of participants. In microscopic, cellular automata, lattice gas, social force, agent-based, game theory, and experimental approaches were designed. In the last two decades, cellular automata models have been created to consider an evacuating group of individuals under different circumstances. These models can be categorized into two groups. The first depends on the associations among situations and pedestrians. For example, in 2007, Varas et al. utilized a two-dimensional cellular automaton model to imitate the process of leaving from a single and double door classroom with complete ability. In this study, to each grid, the structure of the room, obstructions spreading, floor field were measured. Moreover, the effect of panic as a counted dimension was calculated which % 5 possibilities of not moving. The model applied random selection to cope with the collision issue. Therefore, the proposed model changed into non-deterministic via these characteristics. From the simulation result, it was clearly observed that the best locations of the door and evacuation efficiency were not enhanced by substituting a double door with two distinct doors. Finally, for the evacuation time, a number of persons and way out width were considered due to 283 suggesting simple scaling law [21]. In 2011, Alizadeh proposed a CA model to analyze process of evacuation in a room contained obstructions and had different configuration of the room, such as exit and obstruction locations, width of the exit, light of the area, evacuee's psychology, and spreading of the crowd, which had a significant impact on the evacuation process. The case study of this model was the restaurant and a classroom. The impact of the distribution of the evacuees, location, and width of the door on time of the evacuation argued and output of the model was com-pared with some static models [22]. In 2014, Guo, Ren-Yong created a model based on CA with better separation of the area and advanced speed of walking to demonstrate moving away of pedestrians from a room with a single exit door. When experiments were simulated, near the exit they noticed the shape of the crowd was affected by both the separation of the area and advanced speed of walking, duration of the individuals at various positions, and efficiency of the evacuees expressed via two-time indicators. Besides, the association between width and flow of the exit was expressed via this model [23]. In 2015, Li and Han recommended a model for simulating pedestrian evacuation based on extended cellular automata to configure different behavioral propensities in people. Familiarity and aggressive were the two chosen behavioral propensities to be examined via this model. During the experimentations of this simulation, behavioral parameters and pedestrian flow arrangements were verified. The output of this study demonstrated that evacuation time de-creases with increasing individuals' familiarity and increases with the increasing individuals' aggressiveness. Thus, it is clearly seen that the best evacuation was based on the familiarity of the individual with avoiding aggressive behavior [24]. In 2018, Kontou et al. created a model of crowd evacuation based on cellular automata (CA) parallel computing tool to simulate and evaluate behaviors and distinct characteristics of the people within the area of evacuation, which comprises disable people. A secondary school in the region of Xanthi, which contained disable children, was chosen for the simulation process. With noticing and existing earthquake, the school controlled safety exercise; the overall time of the evacuation was recorded. Finally, the proposed model via the empirical data authenticated and there was a convenience inference to the special area [25]. Table 1 shows information for future use. The main goals of this paper are: 1) Collecting most of the research works related to various applications applied to pedestrian evacuation using a cellular automata approach in the microscopic classical model. 2) Investigating features, techniques, and implications of the different applications. 3) Learning from 1 and 2 to make a decision to design a new intelligent and reliable model to simulate participants' appearing emergency behaviors and evacuation efficiency during an emergency evacuation. The model is designed based on an amalgamation of cellular automata, fuzzy logic technique, KNN algorithm, and some statistical equations. Consequently, this paper provides details and specifics in the area of pedestrian evacuation and provides opportunities to the researchers to reach the related information easily and determine their future research directions. This research study's contributions are: 1) Utilizing participants' physical, biological, and emotional elements to apply different speeds for each participant. 2) Incorporating the speed with different elements, such as environment, 284 participants' distributions, and familiarity of the participant to the exit doors to the evacuation model to improve their appearing emergency behaviors and evacuation efficiency. The remainder of this paper is organized as follows: Section 2 presents the research method. Section 3 shows the result and analysis of the simulations. Finally, Section 4 concludes the main points and suggests future research studies. 2. RESEARCH METHOD The following structure, which is presented in Figure 1 define subsections of the suggested framework of pedestrian evacuation through the first floor during the existing emergency to record each of the pedestrian evacuation time and emergency behaviors. Figure 1. Proposed model methodology framework 2.1 Gathering data involved The methodology of our framework started with gathering data. Essentially, different procedures are in gathering data. It is obvious direct observing of the situations appears within the evacuation process is vital to acquire reliable data for different forms of environments. However, due to difficulty in performing this method, new methods were applied, such as applying a general database to hold whole parameters' values with lacking information about provided agents [26]. In the crowd evacuation, the interior of the building plays a great role, besides; there isn't sufficient investigation for influences of the geometry inside the researches [27]. During the evacuation, obstructions' locations and size of the area have a great role [28]. Gathering data for our model is done via two common approaches; interviewing and questionnaire. In the interviewing, the participants faced some physical, biological, and emotional properties structured questions. Also, they asked a number of dichotomous questions of the type of closed-ended questions. 2.2 Gathered data specification The collected data consists of 81 participants 35 males and 46 females. Furthermore, age and weight for the male participants were 18 to 50 years old and 68 to 95 k/g. Age and weight for the female participants were 18 to 43 years old and 57 to 83 k/g. This part of the gathering data utilized in defining the agent's speed while the second part of the collecting data was via the questionnaire was utilized in implementing the agent's behaviors, such as failing, waiting, helping, jumping, and others. Results of the 81 participants for the first part and second part with their analyzing results were briefed in Tables 2 and 3. The percentage value in tables 2 and 3 represents the number of people with the behaviour over the number of participants, for example, in Table 2, the wait behaviour is 20, which means 20 people are waiting and the number of participants is 35, thus, 20/35 = %57.14. Table 1. Highlighting the CA method and its features, techniques, and implications of current simulation models References Methods models occupants State of simulation Investigated occurrences and behaviors Cellular Automata Model Applications Homogenous Microscopic Emergency Effect of obstacles Emergency Impact of distribution of the evacuees, location, and width of the door on time of the evacuation argued The shape of the crowd, duration of the individuals at various positions, the Normal efficiency of the evacuees expressed via Heterogeneous Normal two-time indicators, the association between width and flow of the exit Familiarity and aggressive, evacuation time Emergency Disable children, evacuation time proposed New Model Microscopic heterogeneous Emergency Determining heterogeneous speed to evacuees based on their properties, effect of environment, evacuees distribution, number of room exit doors and main exit doors, and familiarity of evacuees to the exit door on evacuation time and appeared emergency behaviors within the evacuation process Type of models Classical Models CAM CAM [21] [22] [23] CAM [24] [25] CAM CAM proposed model proposed model CA (combined with the idea of fuzzy logic technique, KNN algorithm, and some statistical equations) Table 2. Questionnaire of 81 participants for their behaviors during an emergency evacuation Gender Male Female Number of participants 35 46 Behaviors during an emergency evacuation Wait 20 (%57.14) 18 (%39.13) Aside 24 (%68.57) 40 (%86.96) Jump over 13 (%37.14) 5 (%10.86) Help Wai to fail the person 10 (%28.57) 3 (%6.52) 18 (%51.43) 6 (%13.04) Table 3. Interview of 81 participants' for their physical biological and emotional properties Physical and Biological Properties Gender Number of participants Male 35 Age(yrs.) Weight(kg) 65 -- 95 18 -- 50 Female 46 18 -- 43 57 -- 83 Levels V. Low Low Medium High V. High V. Low Low Medium High V. High Disease 25 (%71.43) 5 (%14.29) 2 (%5.71) 2 (%5.71) 1 (%2.86) 35 (%76.09) 5 (%10.87) 3 (%6.52) 2 (%4.35) 1 (%2.17) Emotional Properties Shock Collaboration 18 (%51.43) 8 (%22.86) 4 (%11.43) 3 (%8.57) 2 (%5.71) 5 (%10.87) 5 (%10.87) 18 (%39.13) 15 (%32.61) 3 (%6.52) 26 (%74.29) 5 (%14.29) 2 (%5.71) 1 (%2.86) 1 (%2.86) 21 (%45.65) 13 (%28.26) 7 (%15.22) 3 (%6.52) 2 (%4.35) 2.3 Layout of the first floor First floor layout for our presented framework was the first floor of one of the institutes' in Sulaimaniyah city in the Kurdistan region of Iraq, which was a polytechnic institute. The first floor of this building contains a cafeteria, which different people visit this part of the building such as, employees, students, visitors, and cafeteria staff with various health, weight, age, and different response or behavior throughout the emergency case. Inside the first floor, 540m2 is provided, for the cafeteria with a width of 36 meters and length 15 meters. This area is partitioned into three different subareas; students' area, employees' area, and kitchen and services areas. The area of employees is 10 meters width and 15 meters length, 285 inside the employees' area there are 6 tables, and one exit doors with 2.5 meters width. The area of students is 17 meters width and 15 meters length, inside the students' area, there are 8 tables and two exit doors with 2.5 meters width for each. Area of kitchen and services is 9 meters width and 15 meters length; this part contains some components for cooking, refrigerator, and cooler and it has two exit doors with 2 meters width, one is on the students' area and the other one from the back end of the area. Figure 2 illustrates the outline of a part of the first-floor polytechnic institute building. 2.4 Structure of the model A structure is designed for this framework and within the design, different characteristics, such as environment, individuals, group behaviors, and emergencies are mentioned. For the first characteristic, the size of the area, the boundary of the floor, the size of the exit doors with their locations, and obstacles are determined. For the second characteristic, agent position, group behaviors such as fail, wait, help, and jump is determined and individual proper-ties such as age, weight, gender, and others are specified to define various speeds of agents. Furthermore, the third characteristic is emergency situations such as fire, smoke, and others are mentioned and participated in defining agents speed with the individual properties these determinations, environment components, agent speed, and behavior patterns together are employed to execute the evacuation process and then evacuation time and various appeared behaviors of each agent during the evacuation process would be recorded. Figure 3 shows the structure of the presented model. second characteristic. After in the Figure 2. Illustrate the outline of a part of first-floor polytechnic university's precedency building Figure 3. Illustrates the proposed model structure 2.5 Definition of the model In the definition of this framework, we demonstrate the individuals, individuals' behaviors, case study's geometry, and spreading of individuals throughout the geometry. This framework for the pedestrian evacuation during emergency built based on cellular automata (CA) approach, which is one of the approaches of the microscopic classical model, and combining it with the idea of fuzzy logic technique, some statistical equations, and K-Nearest Neighbor (KNN) algorithm. This floor is divided into grids, while the CA method has many characteristics such as discrete, continues, and dynamic [7]. Size of each cell set with 0.50.5m2. Naturally, each cell might or might not be available due to the individual, exit doors, walls or obstacles. A two-dimensional array with the dimensions of the floor is used to separate the area into grids. Inside this framework, males and females with various properties, for instance, emotional, biological, and individuals. Different components physical properties are applied. Spreading of individuals within this floor of the case study inside this framework is either randomly or manually. Expressing an emergency in this study is not visible for the agents through the evacuation, but it is due to a coefficient, which puts an impact on the speed of the to construct this framework, such as individuals with yellow color, obstacles with gray color, and exit doors with dark blue color, rooms' wall with pink color, and main walls with the cyan color, which determine the area of the evacuation. Figure 4 (a) illuminates various implemented components inside the proposed framework. Individuals can move with 8 directions to evacuate the building. Figure 4 (b) is an individuals' movement with 8 directions. implemented During the evacuation process, individuals evacuate with determined speed and due to the strike with other individual dilation occurs for some while. In this framework, some behavior is implemented, for instance, when an individual 286 reaches an obstacle, distance from both ends of the obstacle would be checked by the individual, and then decides to pass the obstacle from the nearest one. Figure 4 (c) is the participant's behavior when faced with an obstacle. Colliding individuals in this framework causes to collider to make a decision to go to another empty location and pass the other collider or wait to the other collider to move from the occupied location. For the passing situation, the framework changes the individual's color from yellow into light green, while for the waiting situation the framework changes color from yellow into dark green. Figure 4 (d) is the participant's behavior when faced with other participants. Figure 4. Model definition: (a) various implemented components inside the proposed framework, (b) Individuals' movement with 8 directions, (c) participant's behavior when a faced obstacle, (d) participant's behavior when faced other participants Figure 5. Model definition: (a) Age membership function, (b) Weight membership function, (c) Disease, shock, collaborate membership function, (d) Speed membership function 2.6 Participant speed determination Inside this framework, determination of individual speed with considering the mentioned individual properties such as physical, biological, and emotional was due to taking benefits from the idea of fuzzy logic technique. By means of accepting 287 the fuzzy logic idea to the fuzziness of each parameter, this framework reaches a realistic solution. Many-valued logic is the output of reality, which managed in fuzzy logic [29]. This framework is designed based on the fuzzy logic idea, thus it examines the individual properties that found from the gathered data and creates fuzzy linguistic variables represent the qualities spanning a particular range. For example, disease {very low, low, medium, high, very high}, weight {very slim, slim, heavy, very heavy}, age {adult, very young, young, old, very old}, collaboration {very low, low, medium, high, very high} and shock {very low, low, medium, high, very high}. Then the framework creates a membership in each assortment. Figures 5 (a, b, c, d) present the membership functions of the age, weight, disease, shock, collaborate separately. The following example clarifies how evacuees speed is determined in our model: After collecting and analyzing individual properties, the designed membership functions (See Figures 5 (b), 5 (c), 5 (d) and 7) are applied to identify the degree of the properties of an individual with its speed range. The degree comprises of two values; lower value and upper value, and then weighted mean equation (See Eq. (1)) is used to work on these values [30]. 𝑤𝑒𝑖𝑔ℎ𝑡𝑝𝑟𝑜𝑝 = 𝑛 ∑ 𝑖=1 𝑝𝑤𝑖∗𝑠𝑟𝑑𝑖 (1) ∑ 𝑛 𝑖=1 𝑝𝑤𝑖 𝑝𝑤𝑖 is the degree of the given property's weight and 𝑠𝑟𝑑𝑖 is the speed range for the degree. In this study, equation (1) has different forms (see Eq. (2) and Eq. (3)) to make heterogeneity inside a single class interval. An individual with age and weight properties, their properties' weights defined separately with applying Eq. (2) and Eq. (3). 𝑤𝑒𝑖𝑔ℎ𝑡𝑝𝑟𝑜𝑝 = 𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑚𝑒𝑎𝑛 = (𝑙𝑜 ∗ 𝑚𝑖𝑛𝑖𝑠𝑟𝑑 + 𝑢𝑝 ∗ (2) 𝑚𝑎𝑥𝑖𝑠𝑟𝑑)/(𝑙𝑜 + 𝑢𝑝) 𝑤𝑒𝑖𝑔ℎ𝑡𝑝𝑟𝑜𝑝 = 𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑚𝑒𝑎𝑛 = (𝑢𝑝 ∗ 𝑚𝑖𝑛𝑖𝑠𝑟𝑑 + 𝑙𝑜 ∗ (3) 𝑚𝑎𝑥𝑖𝑠𝑟𝑑)/(𝑙𝑜 + 𝑢𝑝) From Eq. (2) and Eq. (3), 𝑙𝑜 is lower value and 𝑢𝑝 is upper value, 𝑚𝑖𝑛𝑖𝑠𝑟𝑑 is minimum interval speed range, and 𝑚𝑎𝑥𝑖𝑠𝑟𝑑 is maximum interval speed range of the mentioned properties. Before using these two equations, designed memberships (See Figure 5 (a) and 5 (b)) need to be used to find the degree of the given properties. For example, if the individual is 38 years old, designed membership for age property is used to find the degree of that given age property (lower value and upper value of the age). To find the right class interval, the model checks the intervals and finds the right one (See Figure 7). Hence, the framework determines the desired value for 38 years old is 0.35 adult and 0.65 very young. On the other hand, to find the property at which speed range is, the designed membership function of speed is used (See Figure 5 (d)). The specified range for speed in this framework is between 2k/h and 7k/h. This framework breaks down the speed range into several class intervals. This separation is for making logical heterogeneity in values of property's weight for various individuals within the speed range. For example, for an individual with 38 years old class interval of speed is 5k/h - 6k/h, whereas for an individual with 46 years old is 4k/h - 5k/h (See Figure 7). The aim of this diversity in the class interval for speed is associated with logic; younger persons are faster than elder persons. Additionally, finding the middle of 288 the chosen class interval is necessary for finding mid value, Midvalue equation (See Eq. (4)) [30] would be utilized. Midvalue = lov + upv/2 (4) From Eq. (4), for the chosen class, the lower limit is denoted with 𝑙𝑜𝑣 and the upper limit is denoted with𝑢𝑝𝑣. The class interval would be divided into two distinct parts via Midvalue; first half of the interval and second half of the interval (See Figure 7). This division makes the values of given properties' weight different in the two distinct halves of the interval, which leads to remove weight redundancy and arrange different properties weights in both halves of the interval. The result of the Midvalue equation and the given property's value make the framework decide to use Eq. (2) or Eq. (3). For example, when the given property value is greater than the Midvalue result, the framework applies Eq. (3), conversely, when the given property value smaller than the Midvalue result, the framework applies Eq. (2). Both Eq. (2) and Eq. (3) could be used when given value for the property is equal to the Midvalue result. The same operations are done to find the weight of the property of weight. Results of Eq. (2) and Eq. (3) are used to calculate properties' weights as in Eq. (5), which is the mean equation [30] after combination the equation with gender factor 𝑔𝑒𝑛𝑖 and emergency coefficient𝑒𝑚𝑖. desiredSpeed = ( i=1 ∑n pwi ) ∗ geni ∗ emi (5) n Figure 6. Result of determining the desired speed for 81 participants based their properties Figure 7. How weights of participant's properties are identified Finally, gender and emergency factors were counted for the desired speed. From this, the desired speed is affected by those factors, while speeds for male and female is different even both have the same properties. The female speed in our expectation for the simulation was decreased by 0.5. Thus, if the speed of males is 6 kph, it is for female agents decreased to 3 kph. On the other hand, during an emergency, evacuees run as fast as they could. Hence, the speed of the evacuees during an emergency is more than the normal situation. Figure 6 illustrates the conducted desired Speed in kph for the agents. 3. RESULTS AND ANALYSIS In this research work, the model is based on various scenarios, which have been written and executed through various experiments. Concise of experiments' results for the scenarios via the developed intelligent model were briefed inside Table 4. For this simulation process, data gathered for 81individuals. As mentioned in the previous sections within the gathering data, individual properties were recorded for each of them. Therefore, these properties were entered into the model, based on these properties individuals were generated and desired speeds were specified for them. These individuals arbitrarily set through the cafeteria. Figure 8 illustrates the participants' distribution through the first floor. The above-mentioned scenarios could be briefed in the following points: as mentioned in subsection 4.2, 43.21 percent was male and 56.79 percent was female with (1) One main exit for the first floor, two exit doors for students part, only one exit door to each employee part and staff part, and distribution of the individuals in small and large area of the cafeteria. (2) Two main exit doors for the first floor, only one exit door for each of the student part, employee part and staff part, and distribution of the individuals in a small and large area of the cafeteria. (3) Two main exit doors for the first floor, two exit doors for student part, one exit door for employee part and staff part, and distribution of the individuals in a small or large area of the cafeteria. (4) Three main exit doors, two exit or one exit doors for students' part, one exit door for each employee part and staff part, and distribution of the individuals through a large area of the cafeteria. Individuals' familiarity is considered with the points (2), (3) and (4). Figure 8. Illustrates the participants' distribution through the cafeteria Table 4. Experimentations' results from the developed model Experimentations Scenarios Two exit doors for the student part, three main exit doors and evacuees were no familiar with the exits 25.926 Same condition No4A, evacuees familiar with the exits 22.222 7.407 7.407 Aside, wait 0:40:51 Aside, wait 0:31:307 The simulation results of the above-mentioned points via our developed intelligent model were briefed in Table 4. Some abbreviations were used, such as C.W.A is a collision with agents, C.W.O is a collision with obstacles, and O.B.D.E is occurring different behaviors during evacuation. Inside the developed our new model, it was appeared that increasing number of room exit doors, main exit doors with familiarity of the evacuees to the exit doors and distribution in a larger area 289 Evacuees distributed in an average small area in the cafeteria, two main exits, each of the employee part, student part, and staff part has one exit door, evacuees were no familiar with the exits Same as the No1A but evacuees were familiar with the exits Same as the No1A but evacuees distributed in a larger area of the cafeteria Evacuees distributed in an average small area in the cafeteria, two main exits, two exit doors for students part, and each of the employee part, and staff part has one exit door, evacuees were familiar with the exits. Only one main exit door was used, evacuees distributed in a larger area of the cafeteria Same as No2A, but all agents distributed through a very small portion of the floor Same as No2A but agents positions were near the exit door Only one main exit door was used with two exit doors of student part and employees' part and staff part has only one exit door, evacuees distributed in a small area of the cafeteria More than two main exit doors and increase evacuees' distribution area Same condition No3A, they are all familiar with the exit doors Rate of C.W.A (%) Rate of C.W.O (%) Evacuation O.B.D.E time min:sec:ms 61.728 30.864 Aside, wait 1:0:941 51.852 65.432 Aside, wait 0:47:261 27.16 20.988 Aside, wait, help 0:46:711 17.284 33.333 Aside, wait 0:47:130 32.099 13.58 Aside, wait 0:41:635 54.321 20.988 Aside, wait 0:50:682 49.383 4.938 Aside, wait, help, jump over 0:35:373 44.444 17.284 Aside, wait 0:43:487 23.457 13.58 Aside, wait 0:41:241 16.49 16.049 Aside, wait 0:37:976 No1A No1B No1C No1D No2A No2B No2C No2D No3A No3B No4A No4B of the floor are the key factors to minimize the congestion, collision and appearing agent's emergency behaviors that directly have a great influence on improving evacuation efficiency. In the remaining of this section, the result of our model simulation is discussed. For details of simulation results, refer to Supplementary File comprises 36 pages is available in the supplementary Files. for Obviously specifying heterogeneous speed the participants creates jamming due to colliding agents with each other and with obstructions. This situation was clearly seen in the simulation results. Furthermore, changing numbers of main exit doors and cafeteria's distinct parts exit doors create variation in evacuation time and appearing emergency behaviors, such as C.W.A, C.W.O, and O.B.D.E. For example, in experimentation of No1C, No2A, they had different numbers of main exit doors, but they had the same distributions, C.W.A, C.W.O, and O.B.D.E were recorded as 27.66 percent, 33.33 percent, and (aside, wait) for No1C respectively, 45.679 percent, 50.617 per-cent, and (aside, wait, help) for No2A respectively. Besides, it was found that with adding a number of main exit doors C.W.A and C.W.O were decreased and different O.B.D.E appeared. Furthermore, agents' evacuation times were reduced and recorded 0:46:74 for No1C, 0:47:805 for No2A, the time managed as a format of min.sec.ms. When participants distributed through a larger area of the cafeteria great effect would be created on the occurring behaviors and also on the evacuation time. For instance, in the experimentation of No1A and No1C behaviors, such as C.W.A, C.W.O, OBOE were 61.78, 30.64, and (aside, wait) for No1A respectively, while these behaviors were 27.859, 20.988, and (aside, wait, help) for No1C, respectively. Because of the larger distribution in No1C collision behaviors significantly decreased between participants and participants with obstacles. Besides that, evacuation time considerably improved and recorded as 1:0:941 for No1A and 0:46:711 for No1C. On the other hand, participant familiarity with the exits doors had a great impact on changing the participants' evacuation behaviors and evacuation time, for instance, in the experimentations of No1A vs No1B, No3A vs No3B, and No4A vs no4B behaviors and evacuation time, such as, C.W.A and C.W.O changed from 61.728 to 51.852, 65.432 to 30.864, 1:0:941 to 0:47:261 for No1A and No1B, respectively, 23.457 to 16.49, 13.58 to 16.49, 0:41:241 to 0:37:976 for No3A and No3B receptively , and 25.926 to 22.222, 7.407 to 7.407 ,0:40:51 to 0:31:307, for No4A &No4B, respectively. Furthermore, changing the number of exit doors made a great change in participant's evacuation behaviors and evacuation time, for example, in the experimentation of No3A vs No1A could be clearly noticed. In No1A which had two main exits and participants distributed through a small portion of the cafeteria, the appeared behaviors, such as, C.W.A, C.W.O, and O.B.D.E significantly more occurred than in No3A which had three main exits and participants distributed through larger area of the cafeteria, the result of this experimentation for behaviors, such as, C.W.A, C.W.O, OBDE, and evacuation time was recorded as 61.728, 30.864, (aside, wait), and 1:0:941 for No1A, receptively, 23.457, 13.58, (aside, wait), and 0:41:241 for No3A, respectively. The result shows appearing behaviors decreased and also evacuation time considerably improved due to an increase in the number of main exits and area of distribution. From the experimentation result of No1D vs No2D, it appeared that increasing the number of exits is not the only reason to decrease appearing emergency behaviors and improving evacuation efficiency. For example, the experimentation of No1D, there were two main exit doors, and in the experimentation of No2D, there was only one main exit door. However, the C.W.A is better to be compared to C.W.A in No2D, but C.W.O and evacuation time was better in No2D due to larger distribution. in 4. CONCLUSION Despite the fact that developed models based on pedestrian evacuation approaches have been extensively used and confirmed, there is still essential to create models that carefully simulate people's evacuation time and appeared behaviors during an emergency evacuation. It is vital to review many previous works and notice different factors that made an impact on participant behaviors during the evacuation process and also on participants' evacuation time. In this study, an intelligent model built based on analyzing the previous applications of the cellular automata approach which is one of the microscopic model's approaches. Inside this model, cellular automata (CA) employed with a combination of the fuzzy logic idea to define the value of the participant's properties, such as physical, biological, and emotional and then defined values of the participant's properties were used within some statistical equation to define desired speed for that participant. Additionally, the KNN algorithm was applied to find the nearest exit door during the evacuation process. familiarity, environment, This study discusses the impact of the defined desired heterogeneous speed for participants with a combination of participants' behavior, and participants' distribution through the distinct parts of the cafeteria on the appearing emergency behaviors and evocation efficiency. From simulation results appeared different properties of participants caused different speeds for participants. Hence, a collision between participants C.W.A occurred during the evacuation process. Different behaviors, such as C.W.O and O.B.D.E appeared and different evacuation time was recorded for each participant. More experiments presented that environment, such as obstacles, size and number of main and distinct parts of the cafeteria exit doors had a great impact on changing participants' emergency behaviors and evacuation time. This research confirms that participants with different properties, various distributions through the cafeteria, and the presence of obstacles are the key to appearing in various behaviors and congestion. Conversely, it is approved that various designs may decrease the impact of those factors. Firstly, instead of one main exit use two or more main exit doors. Secondly, the distinct part of the cafeteria increase the number of exit doors, meanwhile, changes the distribution of the participants inside the cafeteria into larger distribution. Thirdly, introduce the exit doors to the participants to choose the closest exit to evacuate as these situations simulated with the proposed model. In the future, designing and implementing fire will be examined. Other features, such as: how the fire spread out through the building according to its environment will be added. Additionally, the effect of the fire on the agents' behaviors would be analyzed and discussed. References [31] is very close to our future work while in our work numbers of 290 dead, injured or suffocated agents as a result of fire and smoke will be recorded. Additionally, this model will be enhanced to make a simulation for the second floor and above. Also, it can be seen as a real-world application and will be optimized by different optimizer algorithms, such as fitness dependent optimizer (FDO) [32], WOA-BAT optimization [33], donkey and smuggler optimization [34], and modified grey wolf optimizer [35] to find best location of the main exit door through the area of evacuation in a building. Finally, interested readers can read references [36-38] for the possibility of obtaining future directions. ACKNOWLEDGEMENTS The study is fully funded by the University of Sulaimani (UOS). The authors would like to thank the UOS for providing facilities and equipment for this review work. REFERENCES [1] Emílio Almeida, J., Kokkinogenis, Z., Rossetti, R.J. (2013). NetLogo implementation of an evacuation scenario. arXiv preprint arXiv:1303.4695. [2] Christensen, K.M., Sharifi, M.S., Chen, A. (2013). Considering individuals with disabilities in a building evacuation: An agent-based simulation study. In 92nd Annual Meeting of the Transportation Research Board, Washington, DC, (pp. 11-13). [3] Duives, D.C., Mahmassani, H.S. (2012). Exit choice decisions during pedestrian evacuations of buildings. Transportation Research Record, 2316(1): 84-94. https://doi.org/10.3141/2316-10 [4] Bakar, N.A.A., Majid, M.A., Ismail, K.A. (2017). An overview of crowd evacuation simulation. Advanced Science https://doi.org/10.1166/asl.2017.10298 23(11): 11428-11431. Letters, [5] Liu, H., Liu, B., Zhang, H., Li, L., Qin, X., Zhang, G. (2018). Crowd evacuation simulation approach based on navigation knowledge and two-layer control mechanism. Information 436: 247-267. https://doi.org/10.1016/j.ins.2018.01.023 Sciences, [6] Majid, M.A., Siebers, P.O., Aickelin, U. (2013). Modelling reactive and proactive behaviour in simulation: A case study in a university organisation. arXiv preprint arXiv:1307.1073. http://dx.doi.org/10.2139/ssrn.2829269 [7] Burstedde, C., Klauck, K., Schadschneider, A., Zittartz, J. (2001). Simulation of pedestrian dynamics using a two- dimensional cellular automaton. Physica A: Statistical Mechanics and its Applications, 295(3-4): 507-525. https://doi.org/10.1016/S0378-4371(01)00141- 8 [8] Tajima, Y., Nagatani, T. (2001). Scaling behavior of crowd flow outside a hall. Physica A: Statistical Mechanics and its Applications, 292(1-4): 545-554. https://doi.org/10.1016/S0378-4371(00)00630-0 [9] Kirchner, A., Schadschneider, A. (2002). Simulation of evacuation processes using a bionics-inspired cellular automaton model for pedestrian dynamics. Physica A: Statistical Mechanics and Its Applications, 312(1-2): 260-276. https://doi.org/10.1016/S0378-4371(02)00857- 9 291 [10] Klüpfel , H., Meyer-Konig, M., Wahle, J., Schreckenberg, M. (2000). Microscopic simulation of evacuation processes on passenger ships. Proceedings of the Fourth International Conference on CA for Research and Industry, pp. 63-71. https://doi.org/10.1007/978-1-4471- 0709-5_8 [11] Helbing, D., Farkas, I., Vicsek, T. (2000). Simulating dynamical features of escape panic. Nature, 407(6803): 487. https://doi.org/10.1038/35035023 [12] Helbing, D., Farkas, I.J., Vicsek, T. (2000). Freezing by heating in a driven mesoscopic system. Physical Review Letters, 1240. https://doi.org/10.1103/PhysRevLett.84.1240 84(6): [13] Helbing, D., Isobe, M., Nagatani, T., Takimoto, K. (2003). Lattice gas simulation of experimentally studied evacuation dynamics. Physical Review E, 67(6): 067101. https://doi.org/10.1103/PhysRevE.67.067101 [14] Gwynne, S., Galea, E.R., Owen, M., Lawrence, P.J., Filippidis, L. (1999). A review of the methodologies used in the computer simulation of evacuation from the built environment. Building and Environment, 34(6): 741-749. https://doi.org/10.1016/S0360-1323(98)00057-2 [15] Santos, G., Aguirre, B.E. (2004). A critical review of emergency evacuation simulation models. Disaster Research Center. http://udspace.udel.edu/handle/19716/299 [16] Zhan, B., Monekosso, D.N., Remagnino, P., Velastin, S.A., Xu, L.Q. (2008). Crowd analysis: A survey. Machine Vision and Applications, 19(5-6): 345-357. https://doi.org/10.1007/s00138-008-0132-4 [17] Zheng, X., Zhong, T., Liu, M. (2009). Modeling crowd evacuation of a building based on seven methodological approaches. Building and Environment, 44(3): 437-445. https://doi.org/10.1016/j.buildenv.2008.04.002 [18] Kobes, M., Helsloot, I., De Vries, B., Post, J.G. (2010). Building safety and human behaviour in fire: A literature review. Fire Safety Journal, 45(1): 1-11. https://doi.org/10.1016/j.firesaf.2009.08.005 [19] Zhou, S.P., Chen, D., Cai, W.T., Luo, L.B., Low, M.Y.H., Tian, F., Tay, V.S., Ong, D.W.S., Hamilton, B.D. (2010). Crowd modeling and simulation technologies. ACM Transactions on Modeling and Computer Simulation (TOMACS), http://doi.acm.org/10.1145/1842722.1842725 20(4). [20] Radianti, J., Granmo, O.C., Bouhmala, N., Sarshar, P., Yazidi, A., Gonzalez, J. (2013). Crowd models for emergency evacuation: A review targeting human- centered sensing. In 2013 46th Hawaii International Conference on System Sciences, pp. 156-165. http://doi.org/10.1109/HICSS.2013.155 [21] Varas, A., Cornejo, M.D., Mainemer, D., Toledo, B., Rogan, J., Munoz, V., Valdivia, J.A. (2007). Cellular automaton model for evacuation process with obstacles. Physica A: Statistical Mechanics and its Applications, 382(2): 631-642. https://doi.org/10.1016/j.physa.2007.04.006 [22] Alizadeh, R. (2011). A dynamic cellular automaton model for evacuation process with obstacles. Safety Science, 315-323. https://doi.org/10.1016/j.ssci.2010.09.006 49(2): [23] Guo, R.Y. (2014). New insights into discretization effects in cellular automata models for pedestrian evacuation. Physica A: Statistical Mechanics and its Applications, 1-11. 400: modelling and simulation for evacuation of people from a building in case of fire. Procedia Computer Science, 130: 10-17. https://doi.org/10.1016/j.procs.2018.04.006 [32] Abdullah, J.M., Ahmed, T. (2019). Fitness dependent optimizer: Inspired by the bee swarming reproductive process. 43473-43486. https://doi.org/10.1109/ACCESS.2019.2907012 Access, IEEE 7: [33] Mohammed, H.M., Umar, S.U., Rashid, T.A. (2019). A systematic and meta-analysis survey of whale optimization algorithm. Computational Intelligence and Neuroscience, 25pages. https://doi.org/10.1155/2019/8718571 2019: [34] Shamsaldin, A.S., Rashid, T.A., Agha, R.A.A.R., Al- Salihi, N.K., Mohammadi, M. (2019). Donkey and smuggler optimization algorithm: A collaborative working approach Journal of Computational Design and Engineering, 6(4): 562-583. https://doi.org/10.1016/j.jcde.2019.04.004 to path finding. [35] Rashid, T.A., Abbas, D.K., Turel, Y.K. (2019). A multi hidden recurrent neural network with a modified grey wolf e0213237. https://doi.org/10.1371/journal.pone.0213237 optimizer. PloS One, 14(3): [36] Muhammed, D.A., Saeed, S., Rashid, T.A. (2019). A comprehensive study on pedestrians' evacuation. arXiv preprint arXiv:1911.01165. https://doi.org/10.3991/ijes.v7i4.11767 [37] Arji, G., Ahmadi, H., Nilashi, M., Rashid, T.A., Ahmed, O.H., Aljojo, N., Zainol, A. (2019). Fuzzy logic approach for infectious disease diagnosis: A methodical evaluation, literature and Biomedical Engineering, 39(4): 937-935. https://doi.org/10.1016/j.bbe.2019.09.004 classification. Biocybernetics and [38] Nilashi, M., Samad, S., Manaf, A.A, Ahmadi, H., Rashid, T.A., Munshi, A., Almukadi, W., Ibrahim, O., Ahmed, O.H. (2019). Factors tourism adoption in Malaysia: A DEMATEL-Fuzzy TOPSIS approach. Computers & Industrial Engineering, 137: 106005. https://doi.org/10.1016/j.cie.2019.106005 influencing medical https://doi.org/10.1016/j.physa.2014.01.001 [24] Li, D., Han, B. (2015). Behavioral effect on pedestrian evacuation simulation using cellular automata. Safety Science, 41-55. https://doi.org/10.1016/j.ssci.2015.07.003 80: [25] Kontou, P., Georgoudas, I.G., Trunfio, G.A., Sirakoulis, G.C. (2018). Cellular automata modelling of the movement of people with disabilities during building evacuation. In 2018 26th Euromicro International Conference on Parallel, Distributed and Network-based Processing 550-557. http://doi.org/10.1109/PDP2018.2018.00093 (PDP), pp. [26] Shi, L., Xie, Q., Cheng, X., Chen, L., Zhou, Y., Zhang, R. (2009). Developing a database for emergency evacuation model. Building and Environment, 44(8): 1724-1729. https://doi.org/10.1016/j.buildenv.2008.11.008 [27] Zhang, Q., Zhao, G., Liu, J. (2009). Performance-based design for large crowd venue control using a multi-agent model. Tsinghua Science and Technology, 14(3): 352- 359. http://doi.org/10.1016/S1007-0214(09)70051-3 [28] Pan, X., Han, C.S., Dauber, K., Law, K.H. (2006). Human and social behavior in computational modeling and analysis of egress. Automation in Construction, 15(4): 448-461. https://doi.org/10.1016/j.autcon.2005.06.006 [29] Suganthi, L., Iniyan, S., Samuel, A.A. (2015). Applications of fuzzy in renewable energy systems -- a review. Renewable and Sustainable Energy Reviews, 585-607. https://doi.org/10.1016/j.rser.2015.04.037 logic 48: [30] Varalakshmi, V., Suseela, T.N., Sundaram, T.G.G., Ezhilarasi, T.S., Indrani, T.B. (2004). Statistics Higher Secondary -- First Year. https://docplayer.net/3467441- Statistics-higher-secondary-first-year-untouchability-is- a-sin-untouchability-is-a-crime-untouchability-is- inhuman.html, accessed on 2 May 2019. [31] Kasereka, S., Kasoro, N., Kyamakya, K., Goufo, E.F.D., Chokki, A.P., Yengo, M.V. (2018). Agent-based 292
1503.09066	1	1503	2015-03-31T14:39:27	MORE: Merged Opinions Reputation Model	[ "cs.MA" ]	Reputation is generally defined as the opinion of a group on an aspect of a thing. This paper presents a reputation model that follows a probabilistic modelling of opinions based on three main concepts: (1) the value of an opinion decays with time, (2) the reputation of the opinion source impacts the reliability of the opinion, and (3) the certainty of the opinion impacts its weight with respect to other opinions. Furthermore, the model is flexible with its opinion sources: it may use explicit opinions or implicit opinions that can be extracted from agent behavior in domains where explicit opinions are sparse. We illustrate the latter with an approach to extract opinions from behavioral information in the sports domain, focusing on football in particular. One of the uses of a reputation model is predicting behavior. We take up the challenge of predicting the behavior of football teams in football matches, which we argue is a very interesting yet difficult approach for evaluating the model.	cs.MA	cs	MORE: Merged Opinions Reputation Model Nardine Osman1, Alessandro Provetti2 Valerio Riggi2, and Carles Sierra1 1 Artificial Intelligence Research Institute (IIIA-CSIC), Barcelona, Spain 2 Department of Math & Informatics, University of Messina, Italy Abstract. Reputation is generally defined as the opinion of a group on an aspect of a thing. This paper presents a reputation model that fol- lows a probabilistic modelling of opinions based on three main concepts: (1) the value of an opinion decays with time, (2) the reputation of the opinion source impacts the reliability of the opinion, and (3) the cer- tainty of the opinion impacts its weight with respect to other opinions. Furthermore, the model is flexible with its opinion sources: it may use explicit opinions or implicit opinions that can be extracted from agent behaviour in domains where explicit opinions are sparse. We illustrate the latter with an approach to extract opinions from behavioural in- formation in the sports domain, focusing on football in particular. One of the uses of a reputation model is predicting behaviour. We take up the challenge of predicting the behaviour of football teams in football matches, which we argue is a very interesting yet difficult approach for evaluating the model. Keywords: Trust, reliability and reputation 1 Introduction This paper is concerned with the classic, yet crucial, issue of reputation. We propose MORE, the Merged Opinions REputation model, to compute reputation on the basis of opinions collected over time. MORE uses a probabilistic modelling of reputation; adopts the notion of information decay; considers the reliability of an opinion as a function of the reputation of the opinion holder; and assesses the weight of an opinion based on its certainty. This latter feature constitutes the most novel feature of our algorithm. Furthermore, MORE may be applied to fields with varying abundancy of explicit opinions available. In other words, if explicit opinions are available, as it is the case with so-called eMarkets, then those opinions may directly be used by MORE. In other cases, where such opinions are sparse, behavioural infor- mation can be translated into opinions that MORE can then use. For example, if Barcelona beats Real Madrid at football, then this may be translated into mutual opinions where Barcelona expresses Real Madrid's inadequate skills and Real Madrid acknowledges Barcelona's superior skills. This paper also proposes an approach for extracting opinions from behavioural information in the sports domain. 2 MORE's calculated reputation measures may then be used for different ob- jectives, from ranking performance to predicting behaviour and sports results. Evaluating reputation is a notoriously tricky task, since there seldom is an objective measure to compare to. For instance, how can we prove which opinion is correct and which is biased? In this paper, we present an extensive valida- tion effort that has sought to assess MORE's predictive abilities in the football domain, where accurate predictions are notoriously hard to make [4]. The rest of this paper is divided as follows: Section 2 presents the MORE model, Section 3 introduces the necessary approximations, Section 4 summarises the MORE algorithm; Sections 5, 6, and 7 presents our evaluation, before con- cluding with Section 8. 2 The MORE Model condition that(cid:80) We define the opinion that agent β may form about agent α at time t as: ot β(α) = {e1 (cid:55)→ v1, . . . , en (cid:55)→ vn}, where G = {α, β, . . .} is a set of agents; t ∈ T and T represents calendar time; E = {e1, . . . , en} is an ordered evaluation space where the terms ei may account for terms such as bad, good, very good and so on; and vi ∈ [0, 1] represents the value assigned to each element ei ∈ E under the β(α) is specified as a discrete probability distribution over the evaluation space E. We note that the opinion one holds with respect to another may change with time, hence various instances of ot β(α) may exist for the same agents α and β but for distinct time instants t. i∈[1,E] vi = 1. In other words, the opinion ot Now assume that at time t, agent β forms an opinion ot β(α) about agent α. To be able to properly interpret the opinion, we need to consider how reliable β is in giving opinions. We reckon that the overall reliability of any opinion is the reliability of the person holding this opinion, which changes along time. That is the more reliable an opinion is, the closer its reviewed value is to the original one; inversely, the less reliable an opinion is, the closer its reviewed value is to the flat (or uniform) probability distribution F, which represents complete ignorance and is defined as ∀ ei ∈ E · F(ei) = 1/E. This reliability value R is defined later on in Section 2.3. However, in this section, we use this value to assess the reviewed value Ot β(α), which we define accordingly: β(α) of the expressed opinion ot β) × F β(α) + (1 − Rt β(α) = Rt Ot β × ot (1) 2.1 Opinion Decay Information loses its value with time. Opinions are no exception, and their in- tegrity decreases with time as well. Based on the work of [9], we say the value of an opinion should tend to ignorance, which may be represented by the flat created at time t(cid:48), we distribution F. In other words, given a distribution Ot(cid:48) say at time t > t(cid:48), Ot(cid:48) ), where Λ is the decay function satisfying the property limt(cid:48)→∞ Ot(cid:48) would have decayed to Ot = Λ(t, F, Ot(cid:48) = F. One possible definition, used by MORE, for Λ is the following: where ν ∈ [0, 1] is the decay rate, and: Ot(cid:48)→t = ν∆t Ot(cid:48) + (1 − ν∆t) F (cid:40) ∆t = 0 , if t − t(cid:48) < κ 1 + t−t(cid:48) κ , otherwise 3 (2) (3) ∆t serves the purpose of establishing a minimum grace period during which the information does not decay and that once reached the information starts decaying. This period of grace is determined by the parameter κ, which is also used to control the pace of decay. 2.2 Certainty and its Impact on Group Opinion A group opinion on something at some moment is based on the aggregation of all the previously-expressed individual opinions. However, the certainty of each of these individual opinions has a crucial impact on the aggregation. This is a concept that, to our knowledge, has not been used in existing aggregation methods for reputation. We say, the more uncertain an opinion is then the smaller its effect on the final group opinion is. The maximum uncertainty is defined in terms of the flat distribution F. Hence, we define this certainty measure, which we refer to as the opinion's value of information, as follows: I(Ot β(α)) = H(Ot β(α)) − H(F) (4) where, H represents the entropy of a probability distribution, or the value of information of a probability distribution. In other words, the certainty of an opinion is the difference in entropies of the opinion and the flat distribution. Then, when computing the group opinion, we say that any agent can give opinions about another at different moments in time. We define Tβ(α) ⊆ T to describe the set of time points at which β has given opinions about α. The group opinion about α at time t, Ot G(α), is then calculated as follows: (cid:88) (cid:88) (cid:88) β (cid:88) β∈G Ot G(α) = t(cid:48)∈Tβ (α) β∈G t(cid:48)∈Tβ (α) Ot(cid:48)→t (α) · I(Ot(cid:48)→t β (α)) I(Ot(cid:48)→t β (α)) (5) β This equation states that the group opinion is an aggregation of all the de- cayed individual opinions Ot(cid:48)→t (α) that represent the view of every agent β that has expressed an opinion about α at some point t(cid:48) in the past. However, different views are given different weights, depending on the value of their information I(Ot(cid:48)→t β Note that in the proposed approach, one's latest opinion does not override previous opinions. This choice to override previous opinions or not is definitely (α)). 4 context dependent. For example, consider one providing an opinion about a cer- tain product on the market, then changing his opinion after using the product for some time. In such a case, only the latest opinion should be considered and it should override the previous opinion. However, in our experiments, we use the sports domain, where winning football matches are interpreted as opinions formed by the teams about each others strength in football. In such a case, the opinions obtained from the latest match's score should not override opin- ions obtained from previous matches. In such a context, past opinions resulting from previous matches will still need to be considered when assessing a team's reputation. G (α) = F. In other words, in the absence of any information, the group opinion is equivalent to the flat distribution accounting for maximum ignorance. As individual opinions are expressed, the group opinion starts changing following Equation 5. Finally, we note that initially, at time t0, we have ∀ α ∈ G · Ot0 2.3 Reliability and Reputation An essential point in evaluating the opinions held by someone is considering how reliable they are. This is used in the interpretation of the opinions issued by agents (Equation 1). The idea behind the notion of reliability is very simple. A person who is considered very good at solving a certain task, i.e. has a high reputation with respect to that task, is usually considered an expert in assessing issues related to that task. This is a kind of ex-cathedra argument. An example of current practice supported by this argument is the selection of members of committees or advisory boards. But how is reputation calculated? First, given an evaluation space E, it is easy to see what could be the best opinion about someone: the 'ideal' distribution, or the 'target', which is defined as T = {en (cid:55)→ 1}, where en is the top term in the evaluation space. Then, the reputation of β within a group G at time t may be defined as the distance between the current aggregated opinion of the group Ot G(β) and the ideal distribution T, as follows: Rt β = 1 − emd(Ot G(β), T) (6) where emd is the earth movers distance that measures the distance between two probability distributions [7] (although other distance measurements may also be used). The range of the emd function is [0,1], where 0 represents the minimum distance (i.e. both distributions are identical) and 1 represents the maximum distance possible between the two distributions. As time passes and opinions are formed, the reputation measure evolves along with the group opinion. Furthermore, at any moment in time, the measure Rt can be used to rank the different agents as well as assess their reliability. 3 Necessary Approximation As Equation 5 illustrates, the group opinion is calculated by aggregating the decayed individual opinions and normalising the final aggregated distribution by 5 considering the value of the information of each decayed opinion (I(Ot(cid:48)→t (α))). This approach imposes severe efficiency constraints as it demands exceptional computing power: each time the group opinion needs to be calculated, all past opinions need to decay to the time of the request, and the value of the information of these decayed opinions should be recomputed. β We suggest an approximation to Equation 5 that allows us to apply the algorithm over a much longer history of opinions. To achieve this, when a group opinion is requested, its value is calculated by obtaining the latest group opinion and decaying it accordingly. In other words, we assume the group opinion to decay just like any other source of information. Instead of recalculating them over and over again, we simply decay the latest calculated value following Equation 2 as follows: G(α) = ν∆t Ot(cid:48) Ot G(α) + (1 − ν∆t) F When a new opinion is added, the new group opinion is then updated by adding the new opinion to the decayed group opinion. In this case, normalisation is still achieved by considering the value of the information of the opinions being aggregated; however, it also considers the number of opinions used to calculate the latest group opinion. This is because one new opinion should not have the exact weight as all the previous opinions combined. In other words, more weight should be given to the group opinion, and this weight should be based on the number of individual opinions contributing to that group opinion. As such, when a new opinion ot β(α) is added, Equation 5 is replaced with Equation 7: Ot G(α) = nα Ot(cid:48)→t G (α) · I(Ot(cid:48)→t nα I(Ot(cid:48)→t G (α)) + Ot G (α)) + I(Ot β(α) · I(Ot β(α)) β(α)) (7) where nα represents the number of opinions used to calculate the group opinion about α. Of course, this approach provides an approximation that is not equivalent to the exact group opinion calculated following Equation 5. This is mainly because the chosen decay function (Equation 2) is not a linear function since the decay parameter ν is raised to the exponent of ∆t, which is time dependent. In other words, decaying the group opinion as a whole results in a different probability distribution than decaying all the individual opinions separately and aggregating the results following Equation 5. Hence, there is a need to know how close is the approximate group opinion to the exact one. In what follows, we introduce the test used for comparing the two, along with the results of this test. 3.1 The Approximation Test To test the proposed approximation, we generate a number of random opinions Ot (α) over a number of years, where α is fixed, βi is an irrelevant variable βi (although we do count the number of opinion sources every year, the identity of the source itself is irrelevant in this specific experiment), and t varies according to 6 the constraints set by each experiment. For example, if 4 opinions were generated every year for a period of 15 years, then the following is the set of opinion sets that will be generated over the years: {{O1 β1 (α), ..., O1 β4 (α)}, ...,{O15 β1 (α), ..., O15 β4 (α)}} With every generated opinion, the group opinion is calculated following both the exact model (Equation 5) and the approximate model (Equation 7). We then plot the distance between the exact group opinion and the approximate one. The distance between those two distributions is calculated using the earth mover's distance method outlined earlier. We note that a good approximation is an approximation where the earth mover's distance (EMD) is close to 0. Two different experiments were executed. In the first, 10 opinions were being generated every year over a period of 6 years. In the second, 4 opinions where being generated every year over a period of 15 years. Each of these experiments were repeated several times to test a variety of decay parameters. The final results of these experiments are presented in the following section. 3.2 Results of the Approximation Test Figure 1 presents the results of the first experiment introduced above. The results show that the approximation error increases to around 11% in the first few rounds, and after 12 opinions have been introduced. The approximation error then starts to decrease steadily until it reaches 0.3% when 60 opinions have been added. Experiment 2 has the exact same results, although spanning over 15 years instead of 6. For this reason, as well as well as lack of space, we do not present the second experiment's results here. However, we point that both experiments illustrate that it is the number of opinions that affect the increase/decrease in the EMD distance, rather than the number of years and the decay parameters. In fact, undocumented results illustrate that the results of Figure 1 provide a good estimate of the worst case scenarios, since the earth mover's distance does not grow much larger for smaller ν values, but starts decreasing towards 0. When ν = 0 and the decay is maximal (i.e. opinions decay to the flat distribution at every timestep), the EMD distance is 0. However, when the decay is minimal (i.e. opinions never decay), then the results are very close to the case of ν = 0.98 and κ = 5. Fig. 1. Distance between the exact and approximate Ot G 0123456780.020.040.060.080.10.12EMDYearsυ=0.80, κ=1υ=0.90, κ=1υ=0.98, κ=5xDecay Parameters7 We conclude that the larger the available number of opinions, then the more precise the approximation is. This makes this approximation suitable for appli- cations where more and more opinions are available. 4 The MORE Algorithm The merged opinions reputation model, MORE, is implemented using the ap- proximation of Section 3 and formalised by Algorithm 1. rameter and κ ∈ N is the pace of decay Algorithm 1 The MORE Algorithm Require: E = {e1, . . . , en} to be an evaluation space Require: G = {α, β, . . .} to be a group of agents Require: t ∈ N to be a point in time Require: odb to describe the database of all opinions Require: ot(cid:48) δ(γ) = {(cid:62), if t(cid:48) (cid:22) t;⊥, otherwise} Require: Ot(cid:48)→t β (α) (cid:22) ot X (α) = (Ot(cid:48) P(E) × 2 that calculates the distance between two probability distributions ∀ ei ∈ E · F(ei) = 1/n ∀ ei ∈ E · (i < E ⇒ T(ei) = 0) ∧ (i = E ⇒ T(ei) = 1) H(F) = − log(1/n) ∀ α ∈ G · Ot0 ∀ α ∈ G · nα = 0 while ∃ ot Require: emd : 2 G (α) = F β(α) (cid:22) o) do X (α) − F)ν1+(t−t(cid:48))/κ + F, where ν ∈ [0, 1] is the decay pa- P(E) → [0, 1] to represent the earth mover's distance function β(α) ∈ odb · (∀ o ∈ odb · ot β) × F β(α)(ei) · log Ot Ot Ot(cid:48)→t G (α)(ei) · log Ot(cid:48)→t β = 1 − emd(Ot(cid:48)→t Rt β(α) = Rt Ot I(Ot I(Ot(cid:48)→t β(α)) = − (cid:88) G (α)) =−(cid:88) G (β), T) β(α) + (1 − Rt β × ot ei∈E β(α)(ei) − H(F) G (α)(ei) − H(F) nα Ot(cid:48)→t Ot G(α) = Rt α = 1 − emd(Ot nα = nα + 1 ei∈E G (α)·I(Ot(cid:48)→t nα I(Ot(cid:48)→t G(α), T) G (α)) + Ot G (α)) + I(Ot β (α)·I(Ot β (α)) β (α)) end while In summary, the algorithm is called with a predefined set of opinions, or the opinions database odb. For each opinion in odb, the reviewed value is calculated following Equation 1, the informational value of the opinion as well as that of the decayed latest group opinion are calculated following Equation 4, the updated group opinion is then calculated following Equation 7, and the reputation of the agent is calculated via Equation 6. These steps are repeated for all opinions in odb in an ascending order of time, starting from the earliest given opinion and moving towards the latest given opinion. 8 We note that the complexity of this algorithm is constant (O(1)). Whereas if we were using Equation 5 as opposed to the proposed approximation, then the complexity would have been linear w.r.t. the number of opinions n (O(n)). For very large datasets, such as those used in the experiment of Section 7, the approximation does provide a great advantage. 5 From Raw Scores to Opinions This section describes the extraction of opinions from behavioural information. While we focus on football, we note that these methods may easily be applied to other domains. We say the possible outcomes of a match between teams α and β are as follows: (i) α wins, (ii) α loses, or (iii) the match ends up in a draw. We denote as ng(α) (resp., ng(β)) the number of goals scored by α (resp., β). We then define three methods to convert match results into opinions. Generated opinions belongs to a binary evaluation space consisting of two outcomes, namely bad (B) and good (G): E = {B, G}. 5.1 The Naive Conversion In this first strategy, we simply look for the winner. If α wins, then it receives β(α) = {B (cid:55)→ 0, G(cid:55)→ 1}, and β will get an opinion an opinion from β equal to ot α(β) = {B (cid:55)→ 1, G (cid:55)→ 0}. In case of a draw, they both get the from α equal to ot α(β) = {B (cid:55)→ 0.5, G(cid:55)→ 0.5}. The method is quite simple same opinion: ot and it does not take into account important aspects such as the final score of the match. For instance, losing 0 to 3 is equivalent to losing 2 to 3. β(α) = ot 5.2 Margin-of-Victory Conversion A second strategy we consider is called Margin of Victory -- MV. The margin of victory of a match involving clubs α and β is defined as the difference of goals M = ng(α) − ng(β) scored by α and β. Of course M > 0 if α wins. The main idea here is this: if we know α beats β, this tells us something about the relative strength of α against β. If we know α scored more than 3 goals against β (which is rather unusual in many professional leagues), we could probably have a better picture of the relative strength of the two clubs. We believe that including more data in the process of generating opinions should produce more accurate results and, ultimately, this should help us in better predicting the outcome of a football match. The rules we used to include the number of goals scored by each club are as follows: (cid:40){B(cid:55)→ 0.5, G(cid:55)→ 0.5} (cid:110) ot α(β) = B(cid:55)→ ng(α) ng(α)+ng(β) , G(cid:55)→ ng(β) ng(α)+ng(β) (cid:111) , for a 0-0 tie , otherwise (8) In analogous fashion we can compute the opinion of β on α. Equation 8 tells us that if the margin of victory ng(α)− ng(β) is large, then ng(α) is higher than 9 ng(α) ng(β) and the ratio ng(α)+ng(β) will be closer to 1. As a consequence, the larger the margin of victory between α and β, the more likely α will get an evaluation biased towards good. In case of a 0-0 tie, the terms ng(α)+ng(β) are undefined. To manage such a configuration, we assume that the probability that α (resp., β) gets the evaluation good is equal to the probability it gets the evaluation bad. ng(α)+ng(β) and ng(α) ng(β) A potential drawback of the MV strategy is that different scores may be translated into the same distribution. This happens every time one of the clubs does not score any goal. For instance, the winners in two matches that end with the scores 1−0 and 4−0 would received an opinion {B(cid:55)→ 0, G(cid:55)→ 1}, as calculated by the MV strategy. 5.3 Gifted Margin of Victory The third strategy we propose is called the Gifted Margin of Victory -- GMV. It has been designed to efficiently handle the case of football matches in which one of the clubs does not score any goal. The GMV strategy computes opinions accordingly: (cid:110) (cid:110) ot β(α) = ot α(β) = (cid:111) (cid:111) B(cid:55)→ ng(α)+X B(cid:55)→ ng(β)+X ng(α)+ng(β)+2X , G(cid:55)→ ng(β)+X ng(α)+ng(β)+2X , G(cid:55)→ ng(α)+X ng(α)+ng(β)+2X ng(α)+ng(β)+2X (9) (10) ng(β)+X In other words, we give as a gift both clubs with a bonus of X > 0 goals in order to manage all matches in which one (or possibly both) of the two clubs does not score any goal. Here X is any positive real number. If X → 0, then the GMV strategy would collapse to the MV strategy. On the other hand, if X is extremely large then the constant X would dominate over both ng(α) and ng(β) and the terms ng(α)+ng(β)+2X would converge to 0.5. This result is potentially negative because the probability that any team is evaluated as good is substantially equivalent to the probability that it is evaluated as bad and, therefore, all the opinions would be intrinsically uncertain. An experimental analysis was carried out to identify the value of X guaranteeing the highest prediction accuracy. Due to space limitations we omit the discussion on the experimental tuning of the X parameter and we suffice with the results of our experiment that show that the best value found for X was 1. ng(α)+ng(β)+2X and ng(α)+X A further improvement of the GMV strategy comes from normalization. Nor- malization is motivated by the observation that, since X > 0, term ng(α) + X (resp., ng(β) + X) is strictly less than ng(α) + ng(β) + 2X. Hence, α (resp., β) will never get an opinion where the probability of good comes close to 1, even if it has scored much more goals than β (resp., α). At the same time, since ng(α) + X > 0, there is no chance that α will get an opinion where the probability of bad is close to 0. Let pGM V α,β (G) be the probability of α being evaluated good by β, according to the GM V strategy. We then normalize pGM V (G) to the [0,1] range by con- sidering, for a given set S of teams, the highest and lowest probabilities of being α,β 10 evaluated good according to the calculations of the GMV strategy, which we de- note as M (S) and m(S), respectively. We then define the normalized probability pG(α) of team α being evaluated good by β as follows: (G) − m(S) pGM V α,β M (S) − m(S) pG(α) = (11) And the probability of team α being evaluation bad by β becomes: pB(α) = 1 − pG(α). 6 From Reputation to Predictions This section illustrates how we can use MORE to predict the outcome of a football match. We note that a football match may be depicted as an ordered pair (cid:104)α, β(cid:105), where α and β are opponent clubs. We will follow this convention: we will let α be the 'home club' whereas β will be the 'visiting club'. To compute the reputation of teams α and β, we define the relative strength of α w.r.t. β at time t as follows: rα,β(t) = Rt αRt α + Rt β (12) In what follows, and for simplification, we omit the reference to time t and we use the simplified notation rα,β. Notice that 0 ≤ rα,β ≤ 1 and the higher (resp., lower) rα,β is, the stronger (resp., weaker) the club α is at playing and winning a football match. We shall adopt the following rules to predict the outcome of a match:3 1. If rα,β (cid:39) 1 2. If rα,β ≈ 1 3. If rα,β (cid:47) 1 2 , then the winner will be α. 2 , then the match will end up in a draw. 2 , then the winner will be β. 7 Experimental Results In this section, we test the effectiveness of our approach. In detail, we designed our experiments to answer the following questions: Q1. What is the accuracy of the MORE algorithm in correctly predicting the outcome of a football match? Q2. Which score-to-opinion strategy is reliably the most accurate? Q3. To what extent does information decay impact the accuracy of MORE? 3 We note that we look for values that are approximately greater ((cid:39)), approximately less than ((cid:47)), or approximately equal (≈) to 1 2 . In practice, this is achieved by defining three different intervals to describe this. 11 7.1 Datasets and Experimental Procedure To answer questions Q1 -- Q3, we ran several experiments, drawn on a large dataset of match scores that we collected from public sources.4 Our dataset contains the complete scores of several seasons of the Spanish Primera Divisi´on (Liga), the top football league in Spain. At the moment of writing, 20 clubs play in the Liga. Each club plays every other club twice, once at home and once when visiting the other club. Points are assigned according to the 3/1/0 schema: 3 for win, 1 for draw and 0 for loss. Clubs are ranked by the total number of points they accumulate and the highest-ranked club at the end of the season is crowned champion. The dataset consists of 8182 matches from the 1928-29 season until the 2011-12 season. Overall, the home club won 3920 times and lost 2043 times, and the number of ties amounted to 2119. For the football domain, a major goal of our experimental tests was to check MORE's predicting accuracy. For each match in our database involving clubs α (home club) and β (visiting club) we separately applied the Naive, MV and GMV strategies to convert the outcomes of a football match into opinions. We then applied the MORE algorithm and computed the relative strength rα,β of α against β. We tried various configurations of the decay parameter ν in order to study how the tuning of this parameter influences the overall predictive per- formance of MORE. The usual 3/1/0 scoring system for football rankings (and other games) provided us with a baseline to study the predictive accuracy of MORE. The experimental procedure we followed to compare the predictive accu- racy of MORE and 3/1/0 was as follows. We partitioned the dataset containing football matches into 10 intervals, I1,I2, . . . ,I10, on the basis of the relative strength of opponent clubs. In detail, for an arbitrary pair of clubs α and β, the first interval I1 was formed by the matches such that 0 ≤ rα,β < 0.1, the second interval I2 contained the matches for which 0.1 ≤ rα,β < 0.2 and so on until the tenth interval I10 (consisting of the matches in which 0.9 ≤ rα,β ≤ 1). Observe that the intervals may have different sizes (because, for instance, the number of matches in I1 could differ from those in I2). Given an interval Ik, we have that the larger k, the better the skills of α are and, then, the more likely α should be able to beat β. For different strategies and parameter settings, we computed the percent- age of times (FH (k)) that MORE accurately predicted the outcome of matches in the Ik interval that ended with the victory of the home club. Accordingly, we refer to FH (k) as the home success frequency. In an analogous fashion, we computed the percentage of times (FA(k)) that MORE accurately predicted the outcome of matches in the Ik interval that ended with the victory of the visiting club. Accordingly, we refer to FA(k) as the visiting success frequency. We would expected that the higher rα,β the higher FH (k). In fact, as rα,β → 1 MORE be- comes more and more confident on the ability of α of beating β and, therefore, we expect that FH (k) is consequently large. The situation for rα,β is similar: 4 Data were extracted from http://www.lfp.es/LigaBBVA/Liga_BBVA_Resultados. aspx 12 its increase corresponds to a decrease of rβ,α and, therefore, an increase of rβ,α should correspond to a decrease in the frequency of (home club) α wins. In the following, when it does not generate confusion, we shall use the sim- plified notation FH (resp., FA) in place of FH (k) (resp., FA(k)) because, for a fixed match (cid:104)α, β(cid:105) we can immediately identify the interval Ik to which (cid:104)α, β(cid:105) belongs to and, therefore, the FH (k) (resp., FA(k)) becomes redundant. 7.2 Assessing the Quality of Predictions The first series of experiments we performed aimed at assessing the accuracy of the predictions with respect to the different strategies. The results are plotted in Figures 2 -- 4 for the Naive, MV, and GMV strategies, respectively. In each figure, the plot on the left represents the frequency of successful predictions for the home team winning, and that on the right represents the frequency of successful predictions for the visiting team winning. From the analysis of these results we can draw some relevant conclusions. The Naive strategy, despite its simplicity and independence from the final outcome Fig. 2. Naive strategy: success frequencies for FH and FA (resp.) over relative strength. Fig. 3. MV strategy: success frequencies for FH and FA (resp.) over relative strength. Fig. 4. GMV strategy: success frequencies for FH and FA (resp.) over relative strength. 13 of the match, is able to generate accurate predictions. In fact, Naive is ofter a better forecaster than ranking-based prediction (i.e., using the 3/1/0 point system). For home victories, the maximum value of FH is around 66% whereas the 3/1/0 algorithm peaks at around 59%. For away victories, the values of FA range between 25% and 50% whereas the 3/1/0 algorithm has its success frequency flat around 0.3. It is also interesting to observe that the decay factor ν has little impact on the values of both FH and FA. In particular, the peak value of FH is obtained when ν = 0.7 but the value ν = 0.5 provides more stable results. In contrast, setting ν = 0.5 is the best option for visiting victories, even if the curves describing the evolution of FA tend to coincide when the relative strength (depicted as rs in Figures 2 -- 4) is greater than 0.5. Let us now consider the MV strategy, whose results are reported in Fig- ure 3. This second experiment provides evidence of an increase in the accuracy of MORE, as the highest value of FH is now equal to 64% and the highest value of FA is equal to 46%. This suggests that including the number of goals scored by each team in the process of generating opinions is effective in better com- puting the strength of each club and, ultimately, in producing more accurate predictions. From these figures we can also conclude that for both home and visiting victories, FH and FA achieve their peak when ν = 0.6. But the trends of the curves depicted in Figure3 are quite similar. This implies that information decay has little impact when the MV strategy is chosen. Finally, we consider the GMV strategy. Once again, we computed FH and FA for different values of ν and the corresponding results are graphically reported in Figure 4. This last experiment illustrates that the GMV strategy (with X = 1) provides the highest values of FH and FA. The best value of FH is around 78% (while FH associated with the 3/1/0 algorithm does not exceed 59%). Analogously, in case of visiting victories, the best value of FA is equal to 68% (while the 3/1/0 algorithm is not able to go beyond 37%). The value of ν providing the peak values of FH and FA was 0.6 even though the information decay has little impact, as in the case of MV strategy. We conclude this section by observing that when rα,β is less than 0.3, the value of FH is around 0.5, independently of the adopted strategy. This result is clearly superior to a merely guess-and-check strategy, where choices are chosen uniformly at random and the probability of guessing the correct result is 1 3 (as there are three possible outcomes: α winning, β winning, or neither - having a draw). 8 Conclusion This paper proposed a reputation model based on a probabilistic modelling of opinions, a notion of information decay, an understanding that the reputation of an opinion holder provides an insight on how reliable his/her opinions are, 14 as well as an understanding that the more certain an opinion is, the more its weight, or impact. An interesting aspect of this model is that it may be used in domains rich with explicit opinions, as well as in domains where explicit opinions are sparse. In the latter case, implicit opinions are extracted from the behavioural infor- mation. This paper has also proposed an approach for extracting opinions from behavioural information in the sports domain, focusing on football in particular. In the literature, several ranking algorithms exist that are also based on the notion of implicit opinions. For instance, PageRank [1] and HITS [6] compute the reputation of entities based on the links between these entities. Indirectly, their approach assumes that a link describes a positive opinion: one links to the "good" entities. Both have been applied successfully in the context of web search. In [11], ranking algorithms like PageRank and HITS were applied to the social network to find experts in the network based on who is replying to the posts of whom. In [2], HITS has been used in a similar manner to help find experts based on who is replying to the emails of whom. EigenTrust [5] calculates the reputation of peers in P2P networks by relying on the number of downloads that one peer downloads files from another. In [3], a personalised version of PageRank that also relies on the download history is used to find trustworthy peers in P2P networks. Also, CiteRank [10] and SARA [8] are algorithms that rank research work by interpreting a citation as a positive opinion about the cited work. In comparison, we note that MORE is more generic than existing ranking algorithms, since it has the power to incorporate both explicit and implicit opin- ions in one system. Although built upon previous work, MORE also introduces the novel idea of considering the certainty of an opinion as a measure of its weight, or impact, when aggregating the group members' opinions. Finally, the model is validated by evaluating its performance in predicting the scores of foot- ball matches. We consider the football league scenario particularly interesting because it describes well the opportunities and limitations of the mechanisms by which we would like to evaluate reputation, and thus estimate the true strength of agents in general. Furthermore, we note that unlike the sophisticated predictive models in use today, (e.g., Goldman Sachs' model that was used for World Cup 2014, and relied on around a dozen statistical/historical parameters), MORE relies solely on game scores. In other words, it requires no tuning of complex parameters, and yet its predictions are reasonably accurate. Acknowledgements This work is supported by the Agreement Technologies project (CONSOLIDER CSD2007-0022, INGENIO 2010) and the PRAISE project (EU FP7 grant num- ber 388770). References 1. Brin, S., Page, L.: The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems 30(1-7), 107 -- 117 (Apr 1998), http://dx. 15 doi.org/10.1016/S0169-7552(98)00110-X 2. Campbell, C.S., Maglio, P.P., Cozzi, A., Dom, B.: Expertise identification using email communications. In: Proceedings of the twelfth international conference on Information and knowledge management. pp. 528 -- 531. CIKM '03, ACM, New York, NY, USA (2003), http://doi.acm.org/10.1145/956863.956965 3. Chirita, P.A., Nejdl, W., Schlosser, M.T., Scurtu, O.: Personalized reputation man- agement in p2p networks. In: Golbeck, J., Bonatti, P.A., Nejdl, W., Olmedilla, D., Winslett, M. (eds.) ISWC Workshop on Trust, Security, and Reputation on the Semantic Web. CEUR Workshop Proceedings, vol. 127. CEUR-WS.org (2004) 4. Hill, I.: Association football and statistical inference. Applied statistics pp. 203 -- 208 (1974) 5. Kamvar, S.D., Schlosser, M.T., Garcia-Molina, H.: The eigentrust algorithm for reputation management in p2p networks. In: Proceedings of the 12th international conference on World Wide Web. pp. 640 -- 651. WWW '03, ACM, New York, NY, USA (2003), http://doi.acm.org/10.1145/775152.775242 6. Kleinberg, J.M.: Authoritative sources in a hyperlinked environment. Journal of the ACM 46(5), 604 -- 632 (Sep 1999), http://doi.acm.org/10.1145/324133.324140 7. Peleg, S., Werman, M., Rom, H.: A unified approach to the change of resolution: Space and gray-level. IEEE Transactions on Pattern Analysis and Machine Intel- ligence 11(7), 739 -- 742 (Jul 1989), http://dx.doi.org/10.1109/34.192468 8. Radicchi, F., Fortunato, S., Markines, B., Vespignani, A.: Diffusion of scientific credits and the ranking of scientists. Physical Review E 80, 056103 (Nov 2009), http://link.aps.org/doi/10.1103/PhysRevE.80.056103 9. Sierra, C., Debenham, J.: Information-based reputation. In: First International Conference on Reputation: Theory and Technology (2009) 10. Walker, D., Xie, H., Yan, K.K., Maslov, S.: Ranking scientific publications using a model of network traffic. Journal of Statistical Mechanics: Theory and Exper- iment 2007(06), P06010 (2007), http://stacks.iop.org/1742-5468/2007/i=06/ a=P06010 11. Zhang, J., Ackerman, M.S., Adamic, L.: Expertise networks in online communities: structure and algorithms. In: Proceedings of the 16th international conference on World Wide Web. pp. 221 -- 230. WWW '07, ACM, New York, NY, USA (2007), http://doi.acm.org/10.1145/1242572.1242603
1307.4477	1	1307	2013-07-17T01:43:11	Modularity and Openness in Modeling Multi-Agent Systems	[ "cs.MA", "cs.LO" ]	We revisit the formalism of modular interpreted systems (MIS) which encourages modular and open modeling of synchronous multi-agent systems. The original formulation of MIS did not live entirely up to its promise. In this paper, we propose how to improve modularity and openness of MIS by changing the structure of interference functions. These relatively small changes allow for surprisingly high flexibility when modeling actual multi-agent systems. We demonstrate this on two well-known examples, namely the trains, tunnel and controller, and the dining cryptographers. Perhaps more importantly, we propose how the notions of multi-agency and openness, crucial for multi-agent systems, can be precisely defined based on their MIS representations.	cs.MA	cs	Modularity and Openness in Modeling Multi-Agent Systems Wojciech Jamroga Artur Me¸ski Computer Science and Communication, and Interdisciplinary Centre on Security, Reliability and Trust University of Luxembourg [email protected] Institute of Computer Science Polish Academy of Sciences, Warsaw, Poland, and FMCS, University of Ł´od´z, Poland [email protected] Maciej Szreter Institute of Computer Science Polish Academy of Sciences, Warsaw, Poland [email protected] We revisit the formalism of modular interpreted systems (MIS) which encourages modular and open modeling of synchronous multi-agent systems. The original formulation of MIS did not live entirely up to its promise. In this paper, we propose how to improve modularity and openness of MIS by changing the structure of interference functions. These relatively small changes allow for surprisingly high flexibility when modeling actual multi-agent systems. We demonstrate this on two well-known examples, namely the trains, tunnel and controller, and the dining cryptographers. Perhaps more importantly, we propose how the notions of multi-agency and openness, crucial for multi-agent systems, can be precisely defined based on their MIS representations. 1 Introduction The paradigm of multi-agent systems (MAS) focuses on systems consisting of autonomous entities acting in a common environment. Regardless of whether we deem the entities to be intelligent or not, proactive or reactive, etc., there are two design-level properties that a multi-agent system should satisfy. First, it should be modular in the sense that it is inhabited by loosely coupled components. That is, interaction between agents is crucial for the system, but it should be relatively scarce compared to the intensity of local computation within agents (otherwise the system is in fact a single-agent system in disguise). Secondly, it should be open in the sense that an agent should be able join or leave the system without changing the design of the other components. Models and representations of MAS can be roughly divided into two classes. On one hand, there are models of various agent logics, most notably modal logics of knowledge, action, time, and strategic ability [7, 8, 2]. These models are well suited for theoretical analysis of general properties of agent systems. However, they are too abstract in the sense that: (a) they are based on abstract notions of global state and global transition so the structure of a model does not reflect the structure of a MAS at all, and (b) they come with neither explicit nor implicit methodology for design and analysis of actual agent systems. At the other extreme there are practical-purpose high-level representation languages like Promela [11], Estelle [6], and Reactive Modules (RM) [1]. They are application-oriented, and usually include too many features to be convenient for theoretical analysis. The middle ground consists of formalisms that originate from abstract logical models but try to encapsulate a particular modeling methodology. For instance, interpreted systems [8] support local design of the state space; however, transitions are still global, i.e., they are defined between global rather than local states. Synchronous automata networks [10] and ISPL specifications [17, 19] push the idea further: they are based on local states and semi-local transitions, Gabriele Puppis, Tiziano Villa (Eds.): Fourth International Symposium on Games, Automata, Logics and Formal Verification EPTCS 119, 2013, pp. 224 -- 239, doi:10.4204/EPTCS.119.19 c(cid:13) W. Jamroga, A. Me¸ski & M. Szreter This work is licensed under the Creative Commons Attribution License. W. Jamroga, A. Me¸ski & M. Szreter 225 i.e., the outcome of a transition is local, but its domain global. This makes agents hard to separate from one another in a model, which hampers its modularity. On the other hand, concurrent programs [16] and asynchronous automata networks [10] are fairly modular but they support only systems whose execution can be appropriately modeled by interleaving of local actions and/or events. Modular Interpreted Systems (MIS) are a class of models proposed in [13] to achieve separation of the interference between agents from the local processing within agents. The main idea behind MIS was to encapsulate the way agents' actions interfere by so called interaction tokens from a given alphabet In, together with functions outi,ini that define the interface of agent i. That is, outi specifies how i's actions influence the evolution of the other agents, whereas ini specifies how what rest of the world influences the local transition of i. Modular interpreted systems received relatively little attention, though some work was done on studying computational properties of the related verification problem [12], facilitating verification by abstraction [14], and using MIS to analyze homogeneous multi-agent systems [4]. This possibly stems from the fact that, in their original incarnation, MIS are not as modular and open as one would expect. More precisely, the types of functions used to define interference fix the number of agents in the MIS. Moreover, the assumption that all the functions used in a model are deterministic limit the practical applicability, as modeling of many natural scenarios becomes cumbersome. In this paper, we try to revive MIS as an interesting formalism for modeling multi-agent systems. We propose how to improve modularity and openness of the original class by changing the structure of interference functions out,in. The idea is to use multisets of interference tokens instead of k-tuples. This way, we do not need to "hardwire" information about other modules inside a module. Additionally, we assume that the "manifestation" function out can be nondeterministic. These relatively small changes allow for surprisingly high flexibility when modeling MAS. We demonstrate that on two well-known benchmark examples: trains, tunnel and controller, and the dining cryptographers. Perhaps more importantly, we propose how two important features of multi-agent systems can be formally defined, based on MIS representations. First, we show how to decide if a system is designed in a proper multi-agent way by looking at the relation between the complexity of its interference layer to the complexity of its global unfolding. Moreover, we define the degree of openness of a MIS as the complexity of the minimal transformation that the model must undergo in order to add a new agent to the system, or remove an existing one. We apply the definitions to our benchmark models, and show that different variants of cryptographers grossly differ in the amount of openness that they offer. The paper has the following structure. In Section 2, the new variant of MIS is defined, along with its execution semantics. Section 3 presents MIS representations for two benchmarks: Tunnel, Trains and Controller (TTC) and Dining Cryptographers (DC). A graphical notation is provided to make the exam- ples easier to read. In Sections 4 and 5, we propose formal definitions of multi-agency and openness, respectively, and apply them to several variants of the benchmarks. Section 6 concludes the paper. 1.1 Related Work The modeling structures discussed in this paper share many similarities with existing modeling frame- works, in particular with Reactive Modules [1]. Still, MIS and RMs have different perspectives: Reactive Modules is an application-oriented language, while the focus of modular interpreted systems is more the- oretic. This results in a higher abstraction level of MIS which are based on abstract states and interaction tokens. MIS aim at separating internal activities of modules and interactions between modules, what is not (explicitly) featured in RM. Modularity in models and model checking has been the focus of many papers. Most notably, Hi- erarchical State Machines of Alur et al. [3, 20] and the approach of hierarchical module checking by 226 Modularity and Openness in Modeling Multi-Agent Systems Murano et al. [18] feature both "horizontal" and "vertical" modularity, i.e., a system can be constructed by means of parallel composition as well as nesting of modules. Similarly, dynamic modifications and "true openness" of models has been advocated in [9]. In that paper, Dynamic Reactive Modules (DRM) were proposed, which allow for dynamic reconfiguration of the structure of the system (including adding and removing modules). Our approach differs from the ones cited above in two ways. On one hand, we focus on an abstract formulation of the separation of concerns between modules (and agents), rather than providing concrete mechanisms that implement the separation. On the other, we define indicators that show how good the resulting models is. That is, our measures of agentivity and openness are meant to assess the model "from the outside". In particular, the focus of the DRM is on providing a mechanism for adding and removing agents in the RM representation. We implement these operations on the meta-level, as a basis of the mathematical measure of openness. Our work could in principle be applied to DRMs and other formalisms, but it would require defining the appropriate multi-agent mechanisms which are already present in Interpreted Systems. 2 Modular Interpreted Systems Revisited Modular interpreted systems were proposed in [13] to encourage modular and open design of syn- chronous agent systems. Below, we present an update on the formalism. The new version of MIS differs from the original one [13] as follows. First, a single agent can be now modeled by more than one module to allow for compact design of agents' local state spaces and transition functions. Secondly, the type of function ini is now independent from the structure and cardinality of the set of agents, thus removing the main obstacle to modularity and openness of representation in the previous version. Thirdly, the interac- tion functions ini,outi are nondeterministic in order to enable nondeterministic choice and randomization (needed, e.g., to obtain fair scheduling or secure exchange of information). Fourthly, we separate agents from their names. This way, agents that are not present in the "current" MIS can be referenced in order to facilitate possible future expansion of the MIS. 2.1 New Definition of MIS Let a bag (multiset) over set X be any function X → N. The set of all bags over X will be denoted by B(X), and the union of bags by (cid:93). Definition 1 (Modular interpreted system) We define a modular interpreted system (MIS) as a tuple S = (Agtnames,Act, In,Agt), where Agtnames is a finite set of agent names, Act is a finite set of action names, In is a finite interaction alphabet, and Agt = {a1, . . . ,ak} is a finite set of agents (whose structure is defined in the following paragraph). A set of directed tokens, used to specify the recipients of interactions, is defined as Tok = In× (Agtnames∪{ε}), where ε denotes that the interaction needs to be broadcasted to all the agents in the system. Each agent a j = (id,{m1, . . . ,mn}) consists of a unique name id ∈ Agtnames (also denoted with name(a j)), and one or more modules m j = (St j,Init j,d j,out j,in j,o j,Π j,π j), where: • St j is a set of local states, • Init j ⊆ St j is the set of initial states, • d j : St j → P(Act) defines local availability of actions; for convenience of the notation, we addi- tionally define the set of situated actions as D j = {(q j,α) q j ∈ St j,α ∈ d j(q j)}, W. Jamroga, A. Me¸ski & M. Szreter 227 • out j, in j are interaction (or interference) functions: -- out j : D j → P(P(Tok)) refers to the set of influences (chosen nondeterministically) that a given situated action (of module m j) may possibly have on the recipients of the embedded interaction symbols, and -- in j : St j × B(In) → P(In) translates external manifestations from the other modules into the (nondeterministically chosen) "impression" that they make on module m j depending on the local state of m j; we assume in j(·) (cid:54)= /0; • o j : D j × In → P(St j) is a local transition function (possibly nondeterministic), • Π j is a set of local propositions of module m j (we require that Π j and Πm are disjoint when j (cid:54)= m), • π j : Π j → P(St j) is a valuation of these propositions. Typically, each agent in a MIS consists of exactly one module, and we will use the terms interchange- Additionally, we define the cardinality of S (denoted card(S)) as the number of agents in S. ably. Also, we will omit Init j from the description of a module whenever Init j = St j. Note that function in j is in general infinite. For practical purposes, finite representation of in j is needed. We use decision lists similarly to [15, 19]. Thus, ini will be described as an ordered list of pairs of the form condition (cid:55)→ value. The first pair on the list with a matching condition decides on the value of the function. The conditions are boolean combinations of membership and cardinality tests, and are defined over the variable s for the conditions defined on states, and over H for the conditions on multisets of received interferences. We require that the last condition on the list is (cid:62), so that the function is total. Several examples of MIS's are presented in Sections 3 and 5. 2.2 Execution Semantics for MIS Definition 2 (Explicit models) A nondeterministic concurrent epistemic game structure (NCEGS) is a tuple C = (A , St, St0,PV , V , Act, d,t,∼1, . . . ,∼k), where: A = {1, . . . ,k} is a nonempty set of agents, St is a nonempty set of states, St0 ⊆ St is the set of initial states, PV is a set of atomic propositions, V : PV → P(St) is a valuation function, d : A × St → P(Act) assigns nonempty sets of actions available at each state, and t is a (nondeterministic) transition function that assigns a nonempty set Q = t(q,α1, . . . ,αk) of outcome states to a state q, and a tuple of actions (α1, . . . ,αk) that can be executed in q. We define the semantics of MIS through an unfolding to NCEGS. Definition 3 (Unfolding of MIS) Unfolding of the modular interpreted system S from Definition 1 to a nondeterministic concurrent epistemic game structure NCEGS(S) = (A (cid:48), St(cid:48), St(cid:48)0,PV (cid:48), V (cid:48), Act(cid:48), d(cid:48),t(cid:48)) is defined as follows: i=1 Sti, • PV (cid:48) =(cid:83)k • A (cid:48) = {1, . . . ,k}, and Act(cid:48) = Act, • St(cid:48) = ∏k • St(cid:48)0 = {(q1, . . . ,qk) (∀i ∈ {1, . . . ,k}) qi ∈ Initi}, i=1 Πi, and V (cid:48)(p) = πi(p) when p ∈ Πi, • d(cid:48)(i,q) = di(qi) for global state q = (q1, . . . ,qk), and i ∈ A (cid:48), • The transition function t(cid:48) is constructed as follows. Let q = (q1, . . . ,qk) be a state, and α = (α1, . . . ,αk) be a joint action. We define an auxiliary function oii(qi,αi) of all the possible interfer- ences of agent i, for qi, and αi: γ(cid:48) ∈ oii(qi,αi) iff there exist T1, . . . , Tk such that Tj ∈ out j((q j,α j)), and γ(cid:48) ∈ ini(qi, I1 (cid:93) . . .(cid:93) Ik), where I j = {γ j (∃r ∈ {name(a j),ε}) (γ j,r) ∈ Tj} for all j ∈ A (cid:48). Then (q(cid:48)1, . . . ,q(cid:48)k) ∈ t(q,α1, . . . ,αk) iff q(cid:48)i ∈ oi((qi,αi),γ), where γ ∈ oii(qi,αi); 228 Modularity and Openness in Modeling Multi-Agent Systems • q ∼i q(cid:48) iff q and q(cid:48) agree on the local states of all the modules in agent ai. Definition 3 immediately provides some important logics (such as CTL, LTL, ATL, epistemic logic, and their combinations) with semantics over modular interpreted systems. By the same token, the model checking and satisfiability problems for those logics are well defined in MIS. 3 Modeling with MIS We argue that the revised definition of MIS achieves a high level of separation between components in a model. The interaction between an agent and the rest of the world is encapsulated in the agent's interference functions outi,ini. Of course, the design of the agent must take into account the tokens that can be sent from modules with which the agent is supposed to interact. For instance, the out,in functions of two communicating agents must be prepared to receive communication tokens from the other party. However, the interference functions can be oblivious to the modules with which the agent does not interact. In this section, we demonstrate the advantages on two benchmark scenarios: Trains, Tunnel, and Controller (TTC), and Dining Cryptographers (DC). 3.1 Tunnel, Trains, and Controller (TTC) TTC is a variant of classical mutual exclusion, and models n trains moving over cyclic tracks sharing a single tunnel. Because only one train can be in the tunnel at a time, trains need to get a permission from the controller before entering the tunnel. We model the scenario by MIS T TCn = (Agt,Act,In), where: • Agtnames = {tr1, . . . ,trn,ctrl}, • Act = {nop,approach,request,enter,leave}, • In = {idle, appr, try1, . . . , tryn, retry, granted, left, enter, aw reqs, grant, grant1, . . . , grantn, no reqs, infd, ack release, aw leave}. • Agt = {tr1, . . . ,trn,ctrl}, The system includes n trains tri =(cid:0)tri,{(Sti,Initi,di,outi,ini,oi,Πi,πi)} (cid:1) for i ∈ {0, . . . ,n} such that: Sti = {out,tun needed,granted,in}, and Initi = {out}. di is defined as: outi is defined as: • out (cid:55)→ {nop,approach}, • tun needed (cid:55)→ {request}, • granted (cid:55)→ {enter}, • in (cid:55)→ {nop,leave} oi is defined as: • ((out,nop), idle) (cid:55)→ {out} • ((out,approach), appr) (cid:55)→ {tun needed}, • ((tun needed,request), retry) (cid:55)→ {tun needed}, • ((tun needed,request), granted) (cid:55)→ {granted}, • ((granted,enter), enter) (cid:55)→ {in}, • ((in,nop), idle) (cid:55)→ {in}, • ((in,leave), leave) (cid:55)→ {out} Πi = {in tunnel} • (out,nop) (cid:55)→ {{(idle,tri)}}, • (out,approach) (cid:55)→ {{(appr,tri)}}, • (tun needed,request) (cid:55)→ {{(tryi,ctrl)}}, • (granted,enter) (cid:55)→ {{(enter,tri)}}, • (in,nop) (cid:55)→ {{(idle,tri)}}, • (in,leave) (cid:55)→ {{(left,ctrl), (left,tri)}} ini is defined as: • s = out ∧ appr ∈ H (cid:55)→ {appr}, • s = tun needed ∧ grant ∈ H (cid:55)→ {granted}, • s = tun needed (cid:55)→ {retry}, • s = granted ∧ enter ∈ H (cid:55)→ {granted}, • s = in∧ left ∈ H (cid:55)→ {left}, • (cid:62) (cid:55)→ {idle} πi = {in (cid:55)→ in tunnel} W. Jamroga, A. Me¸ski & M. Szreter Moreover, the agent ctrl =(cid:0)ctrl,{(Stc,Initc,dc,outc,inc,oc,Πc,πc)} as follows: Stc = {tun f ree,in f d,tr1granted, . . . ,trngranted}, and Initc = {tun f ree}. dc is defined as: • tun f ree (cid:55)→ {accepting}, • in f d (cid:55)→ {waiting}, • tr1granted (cid:55)→ {in f orm}, • . . . • trngranted (cid:55)→ {in f orm} oc is defined as: • ((tun f ree,accepting), no reqs) (cid:55)→ {tun f ree}, • ((in f d,waiting), aw leave) (cid:55)→ {in f d}, • ((in f d,waiting), ack release) (cid:55)→ {tun f ree}, • ((tun f ree,accepting), grant1) (cid:55)→ {tr1granted}, • . . . • ((tun f ree,accepting), grantn) (cid:55)→ {trngranted}, • ((tr1granted,in f orm), infd) (cid:55)→ {in f d}, • . . . • ((trngranted,in f orm), infd) (cid:55)→ {in f d} (cid:1) modeling the controller is defined 229 outc is defined as: • (tun f ree,accepting) (cid:55)→ {{(aw reqs,ε)}}, • (in f d,waiting) (cid:55)→ {{(aw leave,ctrl)}}, • (tr1granted,in f orm) (cid:55)→ {{(grant,tr1)}}, • . . . • (trngranted,in f orm) (cid:55)→ {{(grant,trn)}}, inc is defined as: • s = tun f ree∧ try1 ∈ H (cid:55)→ {grant1}, • . . . • s = tun f ree∧ tryn ∈ H (cid:55)→ {grantn}, • s = tun f ree (cid:55)→ {no reqs}, • s = tr1granted ∨ ...∨ s = trngranted (cid:55)→ {infd}, • s = in f d ∧ left ∈ H (cid:55)→ {ack release}, • s = in f d (cid:55)→ {aw leave}, • (cid:62) (cid:55)→ {idle} Πc = {tunnel busy} πc = {in f d (cid:55)→ tunnel busy} The model is illustrated in Figure 1 using the notation introduced in Section 3.2. The protocol focuses on the procedure of gaining a permission to access the tunnel. Before requesting the permission, a train approaches the tunnel, and its state changes to tun needed. In this state it requests the permission from the controller. When the controller grants the permission to one of the nondeterministically chosen trains (trigranted) it informs the train that got access to the tunnel about this fact, and moves to the state in f d. The train enters the tunnel in the next step of the protocol, and changes its state to in, whereas the remaining trains may continue requesting the access (they remain in tun needed). When the train leaves the tunnel, it changes its state to tun f ree. 3.2 Graphical Representation As the definitions of MIS tend to be verbose, we introduce a simple graphical notation, based on networks of communicating automata. Let us explain it, based on Figure 1, which is a graphical representation of the tunnel, trains, and controller model from Section 3.1: • Modules defining different agents and belonging to the same agent are separated by solid and dashed lines, respectively, • Circles correspond to local states. An arrow with loose end pointing into a circle denotes an initial state, • Boxes define local actions associated with a state, • For a local action, dashed lines going out of it define emitted influences, specified with the receiver and the influence at the left and right side of an harpoon arrow pointed left, respectively. When no receiver is specified, the influence is broadcasted, 230 Modularity and Openness in Modeling Multi-Agent Systems Figure 1: Tunnel, trains and controller (TTC) in the graphical representation • Solid lines with arrows, connecting an action with a local state, correspond to a local transition function, • For a local state, guarded commands (possibly in a box) define the translation of external manifes- tations received by an agent into local impressions. A harpoon arrow pointed right corresponds to a sender at the left side and the message at the right side, and if the sender is not specified it means receiving from anyone. For a transition, dotted arrows pointing at it correspond to application of those impressions. The number of interactions x received by an agent is denoted with n(x). The notation ∗ labeling a transition means that it is executed when none of the remaining transitions are enabled. For example, it could be used instead of directly specifying the generation and application of aw leave manifestation in the controller. Some parts can be skipped or abstracted away if it does not lead to confusion. For example, if no influence is emitted and only one transition is associated with an action, this action needs not be In Figure 1, the self-loop from the in state is not accompanied by the associated directly specified. local impression nor the impression. Similarly, a single influence addressed to the very module that issued it can be omitted. For example, we do not show the manifestation idle in the graph. Valuations of approachLEFTrequestenterleavenopnopacceptingwaitinginfdtunfreetrngrantedinformtr1)grantGRANT1+try1NOREQUESTSGRANT1GRANTnACKRELEASEACKRELEASEAWLEAVEinforminformtunneededgrantedinoutctrl)tryitri)enterRETRYapprGRANTEDENTERctrl)lefttri)lefttri)apprtrictrl)awreqsctrl)awleavetrn)grant+trynGRANTn......NOREQUESTS+left>>AWLEAVEtri)idle+grantGRANTED>RETRY+apprappr+leftLEFT+enterENTERW. Jamroga, A. Me¸ski & M. Szreter 231 Figure 2: MIS for dining cryptographers (DC1) propositions can be depicted in a similar way as for networks of automata. We omit them in our examples throughout, as they do not play a role in this paper. 3.3 Dining Cryptographers: Standard Version (DC1) Dining Cryptographers is a well-known benchmark proposed by Chaum [5]. n cryptographers are having dinner, and the bill is to be paid anonymously, either by one of them or by their employer. In order to learn which option is the case without disclosing which cryptographer is paying (if any), they run a two- stage protocol. First, every cryptographer is paired with precisely one other participant (they sit around the table), thus forming a cycle. Every pair shares a one-bit secret, say by tossing a coin behind a menu. In the second stage, each cryptographer publicly announces whether he sees an odd or an even number of coin heads, saying the opposite if being the payer. In the simplest case (DC1) the number of cryptographers is fixed, and each cryptographer is directly CounterevenoddheadtailsaydifferentsayequalCiWhopaysCi−)tailCi−)headCounter)sayequali=falseCounter)sayequali=true∨i∈[1,n]Ci+sayequali=falseαi:((Ci++tail∧n(Whopays+payi)=0)αiβiβiαiSAYEQUALSAYDIFFSAYEQUALSAYDIFFSAYDIFFSAYEQUALISODDISEVENISODDISEVENsaysayC1)pay1)paydecided...pay1paynpaynone∨i∈[1,n]Ci+sayequali=trueCn)payn∨(Ci++head∧Whopays+payi)))paydecided)paydecidedβi:((Ci++head∧n(Whopays+payi)=0)∨(Ci++tail∧Whopays+payi))∧Whopays+paydecided)∧Whopays+paydecided)start232 Modularity and Openness in Modeling Multi-Agent Systems bound with its neighbours. Cryptographers announce their utterances by broadcasting them. A modular interpreted system modeling this setting is presented in Figure 2. For n cryptographers, the ith cryptogra- pher is modeled by agent Ci (0 ≤ i ≤ n). We introduce notation i+ and i− to refer to the right and the left neighbour of cryptographer i, respectively. The system includes also two additional agents. W ho pays initializes the system by determining who is the payer, and communicating it to the cryptographers. Ac- cording to the protocol definition, either one of the cryptographers is chosen, or none of the participants pays. Agent Counter counts the utterances of the cryptographers, computes the XOR operation (denoted by ∨ and assuming that utterances different and equal correspond to true and false values, respectively, thus the result is true iff the number of different utterances is odd), and determines the outcome of the protocol. Figure 2 shows the modular interpreted system for DC1. 4 How to Measure Multi-Agency In this section, we present our preliminary attempt at defining what it means for a design to be multi- agent. Intuitively, separate agents should have only limited coordination and/or communication capabil- ities. Otherwise, the whole system can be seen as a single agent in disguise. The idea is to measure the complexity of interference between different agents, and relate it to the complexity of the system. The former factor will be captured by the number of directed interaction tokens that a given agent can gener- ate; the latter by the number of global transitions that can occur. We say that the agent is well designed if its interference complexity is reasonably smaller than overall complexity of the system. Definition 4 (Interaction complexity) The interaction complexity of agent i in modular interpreted sys- tem M, denoted IC(i), is defined as follows. Let #outi(qi) be the the maximal number of directed to- kens generated by function outi to modules of other agents in state qi. Furthermore, let #ini(qi) be the maximal number of tokens admitted by function ini from modules of other agents in state qi. Now, IC(i) = ∑qi∈Sti(#outi(qi) + #ini(qi)). The interaction complexity of M is defined as IC(M) = ∑i∈Agt IC(i). Definition 5 (Global complexity) The global complexity of MIS M, denoted GC(M), is the number of transitions in the NCEGS unfolding of M. How can we express that IC(M) is "reasonably smaller" than GC(M)? Such a requirement is rela- tively easy to specify for classes of models, parameterized with values of some parameter (for instance, the number of identical trains in the tunnel-controller scenario). Definition 6 (C -sparse interaction, multi-agent design) Let M be a class of MIS and C a class of complexity functions f : N → R+∪{0}. We say that M is characterized by C -sparse interaction iff there is a function f ∈ C such that IC(M) ≤ f (GC(M)) for every M ∈ M . Furthermore, we say that M has multi-agent design iff M has LOGTIME-sparse interaction, and card(M) ≥ 2 for every M ∈ M . Proposition 1 Classes TTC and DC1 have multi-agent design. The proof is straightforward. It is easy to see that the other variants of Dining Cryptographers, discussed in Section 5, also have multi-agent design. W. Jamroga, A. Me¸ski & M. Szreter 233 5 How Open is an Open System? The idea of open systems is important for several communities: not only MAS, but also verification, software engineering, etc. It is becoming even more important now, with modern technologies enabling dynamic networks of devices, users and services whose nodes can be created and removed according to current needs. Traditionally, the term open system is understood as a process coupled with the environ- ment, which is rather disappointing given the highly distributed nature of MAS nowadays. One would rather like "openness" to mean that components (agents in our case) can freely join and leave the system without the need to redesign the rest of it. Perfectly open systems are seldom in practice; usually, adding/removing components requires some transformation of the remaining part (for instance, if a server is to send personalized information to an arbitrary number of clients then it must add the name of each new client to the appropriate distribution lists). So, it is rather the degree of openness that should be captured. We try to answer the question How open is the system? (or, to be more precise, its model) in the next subsection. 5.1 A Measure of Openness We base the measure on the following intuition: openness of a system is simplicity of adding and remov- ing agents to and from the model. That is, we consider two natural transformations of models: expansion (adding agents) and reduction (removing agents). We note that the simplicity of a transformation is best measured by its algorithmic complexity, i.e., the number of steps needed to complete the transformation. A perfectly open system requires no transformation at all (0 steps) to accommodate new components, whereas at the other extreme we have systems that require redesigning of the model from scratch when- ever a new agent arrives. Note that the openness of a model depends on which agents want to join or leave. For instance, the system with trains and controllers should be able to easily accommodate additional trains, but not necessarily additional controllers. Likewise, departure of a train should be straightforward, but not nec- essarily that of the controller. No less importantly, the context matters. We are usually not interested in an arbitrary expansion or reduction (which are obviously trivial). We want to add or remove agents while keeping the "essence" of the system's behavior intact. The following definitions formalize the idea. Definition 7 (Expansion and reduction of a MIS) Let M = (Agtnames,Act, In,Agt) be a MIS, and aaa an agent (in the sense of Definition 1). By agt(aaa) (resp. act(aaa), in(aaa)) we denote the set of agent names (resp. action symbols, interaction symbols) occurring in aaa. Moreover, ns(aaa,M) will denote the set of aaa's namesakes in M.1 Note that ns(aaa,M) can contain at most 1 agent. where: Agtnames(cid:48) = Agtnames∪agt(aaa), Act(cid:48) = Act∪act(aaa), In(cid:48) = In∪in(aaa), and Agt(cid:48) = Agt\ns(aaa,M)∪ {aaa}. The reduction of M by aaa is defined as M (cid:9)aaa = (Agtnames,Act, In,Agt(cid:48)) where Agt(cid:48) = Agt\{aaa}. Thus, expansion corresponds to "dumb" pasting an agent into a MIS, and reduction corresponds to The expansion of M by aaa is defined as the modular interpreted system M⊕aaa = (Agtnames(cid:48),Act(cid:48), In(cid:48),Agt(cid:48)) simple removal of the agent. The operations are well defined in the following sense. Proposition 2 Expansion/reduction of a MIS is always a MIS.2 It is easy to see that removing an agent and pasting it in again does not change the MIS. The reverse sequence of operations does change the MIS. However, both structures have the same unfoldings: 1 That is, agents in M that have the same id as aaa. 2 The proofs of results in Section 5 are straightforward from the construction of MIS, and we leave them to the reader. 234 Modularity and Openness in Modeling Multi-Agent Systems Proposition 3 Let aaa be an agent in M. Then, (M (cid:9)aaa)⊕aaa = M. Moreover, let aaa be an agent with no namesake in M. Then, NCEGS((M ⊕aaa)(cid:9)aaa) = NCEGS(M). Now we can make our first attempt at a measure of openness. Definition 8 (Degree of openness) Let θ be a property of models,3 M a modular interpreted system, and aaa an agent. The degree of openness of M wrt expansion by aaa under constraint θ is defined as the minimal number of steps that transform M ⊕aaa into a MIS M(cid:48) such that card(M(cid:48)) = card(M ⊕aaa) and M(cid:48) satisfies θ. Likewise, the degree of openness of modular interpreted system M wrt reduction by agent aaa under constraint θ is the minimal number of steps that transform M (cid:9) aaa into an M(cid:48) such that card(M(cid:48)) = card(M (cid:9)aaa) and M(cid:48) satisfies θ. The constraint θ can for example refer to liveness of the system or some of its components, fairness in access to some resources, and/or safety of critical sections. Note that the cardinality check is essential in the definition -- otherwise, a possible transformation would be to simply delete the newly added agent from M ⊕aaa (respectively, to restore aaa in M (cid:9)aaa). Definition 9 (Openness of a class of models) Let M be a class of MIS, aaa an agent, and θ a property of models. Moreover, let C be a class of complexity functions f : N → R+ ∪{0}. M is C -open wrt expansion (resp. reduction) by aaa under constraint θ iff there is a complexity function f ∈ C such that for every M ∈ M the degree of openness of M wrt expansion (resp. reduction) by aaa under θ is no greater than f (M). The most cumbersome part of the above definitions is the constraint θ. How can one capture the "essence" of acceptable expansions and reductions? Note that, semantically, θ can be seen as a subclass of models. We postulate that in most scenarios the class that defines acceptable expansions/reductions is the very class whose openness we want to measure. This leads to the following refinement of the previous definitions. Definition 10 (Openness in a class) The degree of openness of M wrt expansion (resp. reduction) by aaa in class M is the minimal number of steps that transform M ⊕aaa (resp. M (cid:9)aaa) into a MIS M(cid:48) ∈ M such that card(M(cid:48)) = card(M ⊕aaa). Moreover, M is C -open wrt expansion (resp. reduction) by aaa iff there is a complexity function f ∈ C such that for every M ∈ M the degree of openness of M wrt expansion (resp. reduction) by aaa in M is no greater than f (M). We explain the measure in greater detail in the remainder of Section 5. It is important to note that (in contrast to the measure of multi-agentivity proposed in Section 4) our measure of openness is not specific to MIS, and can be applied to other modeling frameworks. Remark 4 Alternatively, we could define the openness of M wrt aaa and θ by the Kolmogorov complexity of an appropriate expansion/reduction, i.e., by the size of the shortest algorithm that transforms M in an appropriate way. We chose time complexity instead, for two reasons. First, Kolmogorov complexity often obscures the level of difficulty of a process (e.g., a two-line algorithm with an infinite while loop can implement infinitely many changes, which gives the same complexity as changing the names of two communication channels for a controller). Secondly, computing Kolmogorov complexity can be cumbersome as it is Turing-equivalent to answering the halting problem. 3 We do not restrict the language in which θ is specified. It can be propositional logic, first-order temporal logic, or even the general language of mathematics. The only requirement is that, for every MIS M, the truth of θ in M is well defined. W. Jamroga, A. Me¸ski & M. Szreter 235 Figure 3: Dining cryptographers version DC2: direct channels instead of broadcast We observe, however, that a Kolmogorov-style measure of openness can be a good alternative for infinite models, especially ones that require infinitely many steps to accommodate changes in the config- uration of components. 5.2 How to Open Up Cryptographers In Section 3.3 we modeled the standard version of the Dining Cryptographers protocol as a modular interpreted system (class DC1). In this section, we will determine the openness of DC1, plus two other classes of MIS modeling other versions of the protocol. To comply with classical rules of composition, we begin with the least open variant. CounterievenoddHEADTAILsaydifferentsayequalCiCi−)tailCi−)headCounterj)sayequali=falseCounterj)sayequali=true∨j∈[1,n],j6=iCj+sayequalj=falseαiβiβiαiSAYEQUALSAYDIFFSAYEQUALSAYDIFFSAYDIFFSAYEQUALISODDISEVENISODDISEVENsaysay∨j∈[1,n],j6=iCj+sayequalj=trueforj∈[1,n],j6=iforj∈[1,n],j6=iWhopaysC1)pay1)paydecided...pay1paynpaynoneCn)payn)paydecided)paydecidedαi:((Ci++tail∧n(Whopays+payi)=0)∨(Ci++head∧Whopays+payi))βi:((Ci++head∧n(Whopays+payi)=0)∨(Ci++tail∧Whopays+payi))∧Whopays+paydecided∧Whopays+paydecidedstart236 Modularity and Openness in Modeling Multi-Agent Systems 5.2.1 DC-Net, Direct Channels, No Broadcasting (DC2) Let us assume that no broadcast channel is available, or it is too faulty (or insecure) to be of use in multi- party computation. In such case, every pair of cryptographers must use a direct secured channel for communicating the final utterance. The result of the computation is calculated independently by every cryptographer. We denote this class of models by DC2, and construct it as follows. Each cryptographer i is modeled by agent Ci, similar to the cryptographer agents in DC1. Instead of a single global counter of utterances, there is one counter per every cryptographer (Counteri). The final utterance is sent by direct point-to-point channels to the counters of all other participants. The resulting MIS is shown in Figure 3. Adding a new cryptographer Ci to DC1n requires the following changes. First, modifying links among the new neighbours of Ci yields 10 changes. Secondly, every agent Cj in DC1n must be modified in order to establish a communication channel with Ci. This requires 2· 5 changes per cryptographer, thus 10n changes are needed. Thirdly, for the agent W ho pays, we add the state payi(cid:48) with corresponding transitions: a single non-deterministic transition from start to payi(cid:48) (17 steps: 2 for di + 4 for outi + 8 for ini + 3 for oi), and the loop sending payment information (19 steps: 4 for di + 4 for outi + 4 for ini + 3 for oi). Finally, Counter needs to be updated to take into account the new participant. A XOR argument is added with receiving a manifestation, yielding 2· 4 = 8 changes. Thus, the overall openness complexity for DC2n is 10n + 54. Proposition 5 Class DC2 is O(n)-open wrt expansion by a cryptographer. 5.2.2 Dining Cryptographers: Standard Version with Broadcast (DC1) Let us now go back to the standard version of the protocol, presented in Section 3.3 Adding a new cryptographer Ci requires the following changes. First, modifying links for the new neighbors of Ci requires 10 changes. Secondly, changes in W ho pays and Counter are the same as for DC2n, yielding 44 steps. Thus, 54 changes are needed to accommodate the new cryptographer, regardless of the number of agents already present in the system. Proposition 6 Class DC1 is O(1)-open wrt expansion by a cryptographer. 5.2.3 Fully Open System, Cryptographers without Identifiers (DC0) In our most radical variant, cryptographers are not arbitrarily assigned as neighbors. Instead, they es- tablish their neighborhood relation on their own before starting the protocol. Every cryptographer is modeled by two modules Ci and Payi, and there are two additional agents Oracle and Counter, cf. Fig- ure 4. The system proceeds as follows: Setting up the payer. Every cryptographer sends the oracle his declaration whether he is going to pay or not (chosen nondeterministically). This is performed by module Payi. If Oracle receives at most one statement want pay, it confirms to all cryptographers. If more than one statements want pay is sent, the round is repeated until the payment issue becomes resolved. Establishing the neighbourhood relation and tossing coins. Each cryptographer either nondetermin- istically tosses a coin and announces the outcome, or listens to such announcements from the other agents. If there is exactly one cryptographer announcing and one listening, they become paired. They register the value of the announcement, and proceed further. A cryptographer who started with announcing will now listen, and vice versa. This takes several rounds, and completes when every cryptographer has been paired with two neighbors (one to whom he listened, and one to whom he announced). W. Jamroga, A. Me¸ski & M. Szreter 237 Figure 4: Cryptographers without identifiers (DC0) Counter)¬sayequalCounter)sayequalLheadLtailsayequalsaydiff)wait)head)tail)tail)headRtailRheadn(+tail)=0∧n(+head)=1→RHEADn(+wait)=0∧n(+tail)=1∧Payi+ipay→HSEHSEn(+wait)=0∧n(+head)=1∧Payi+inotpay→HSDHSDn(+wait)=1∧n(+head)=n(+tail)=0∧Payi+inotpay→SEn(+wait)=1∧n(+head)=n(+tail)=0∧Payi+ipay→SDSDSEn(tail)=1∧n(head)=0→RTAILRTAILRHEAD∗∗)wait∗CiOraclePayinotdecideddecidednotdecidedwantpayipaynotwantpayinotpayn(+wantpayi)≤1→ACCEPT)CONFIRMEDACCEPTCi)ipayCi)inotpaySDOracle)wantpayOracle)notwantpayRWAITSn(+head)=n(+tail)=0→RWAITSCounter∨(+sayequal)=false∧n(+oo6=sayequal)=0→NOTEQ∨(+sayequal)=true∧n(+oo6=sayequal)=0→EQequalnotequalEQNOTEQ∗+CONFIRMED+CONFIRMED∗∗∗∗)tail)head∗∗∗∗∗n(+wait)=1∧n(+wait)=0∧n(+wait)=0∧238 Modularity and Openness in Modeling Multi-Agent Systems Computation. A broadcast channel is used for sending around the utterances (say equal or ¬say equal). Counter counts the utterances and computes their XOR on the spot, in the way described before. DC0 is fully open, as adding a new cryptographer requires no adaptation of DC0n. Proposition 7 Class DC0 is O(0)-open wrt expansion by a cryptographer. By comparing their classes of openness, it is clear that DC1 is significantly more open wrt expansion than DC2 (constant vs. linear openness). On the other hand, it seems that the gap between DC1 and DC0 is rather slight (O(1) vs. O(0)). Is that really the case? We believe that the difference between O(1)- openness and O(0)-openness is larger than one is used to in complexity of algorithms. First, constant openness means that, when expanding the MIS by a set of new agents, the required transformation can be linear in the size of the set. More importantly, non-zero openness signifies the need to come up with a correct procedure of expansion. In contrast, zero openness means zero hassle: the new agents can join the system as they come. There is no need for "maintenance" of the system so that it stays compliant with its (usually implicit) specification. 6 Conclusions In this paper, we propose a new version of modular interpreted systems. The aim is to let modeling and analysis of multi-agent systems benefit from true separation of interference between agents and the "internals" of their processes that go on in a system. Thanks to that, one can strive for a more modular and open design. Even more importantly, one can use the MIS representation of a system to assess its agentivity and openness through application of simple mathematical measures. We emphasize that it was not our aim to create yet another agent programming language or represen- tations that will be used as input to cutting-edge model checkers. Instead, we propose a class of models which enables to expose the internal structure of a multi-agent system, and to define the concepts of openness and multi-agentivity in a precise mathematical sense. While our definition of multi-agentivity is specific to MIS, the measure of openness is in fact generic, and can be applied to models defined in other formalisms (such as Reactive Modules). We plan to look closer at the degree of openness provided by different representation frameworks in the future. We would also like to stress that the focus of this paper regarding the measures of agentivity and openness is on formalizing the concepts and showing how they work on benchmarks. An formal study of the measures and their properties is a matter of future work. Acknowledgements. The authors thank Andrzej Tarlecki for his suggestion to improve modularity of MIS by using multisets, and Thomas Agotnes for discussions. Wojciech Jamroga acknowledges the support of the FNR (National Research Fund) Luxembourg under project GALOT -- INTER/DFG/12/06. Artur Me¸ski acknowledges the support of the European Union, European Social Fund. Project PO KL "Information technologies: Research and their interdisciplinary applications" (UDA-POKL.04.01.01- 00-051/10-00). References [1] R. Alur & T. A. Henzinger (1999): Reactive Modules. Formal Methods in System Design 15(1), pp. 7 -- 48, doi:10.1023/A:1008739929481. W. Jamroga, A. Me¸ski & M. Szreter 239 [2] R. Alur, T. A. Henzinger & O. Kupferman (2002): Alternating-Time Temporal Logic. Journal of the ACM 49, pp. 672 -- 713, doi:10.1145/585265.585270. [3] R. Alur, S. Kannan & M. Yannakakis (1999): Communicating Hierarchical State Machines. In: Proceedings of ICALP, pp. 169 -- 178, doi:10.1007/3-540-48523-6 14. [4] J. Calta (2012): Synthesis of Strategies for Multi-Agent Systems. Ph.D. thesis, Humboldt University Berlin. [5] D. Chaum (1988): The Dining Cryptographers Problem: Unconditional Sender and Recipient Untraceabil- ity. Journal of Cryptology 1(1), pp. 65 -- 75, doi:10.1007/BF00206326. [6] P. Dembi´nski, A. Janowska, P. Janowski, W. Penczek, A. P´ołrola, M. Szreter, B. Wo´zna & A. Zbrzezny (2003): Verics: A Tool for Verifying Timed Automata and Estelle Specifications. In: Proceedings of the of the 9th Int. Conf. on Tools and Algorithms for Construction and Analysis of Systems (TACAS'03), LNCS 2619, Springer, pp. 278 -- 283, doi:10.1007/3-540-36577-X 20. [7] E. A. Emerson (1990): Temporal and Modal Logic. In J. van Leeuwen, editor: Handbook of Theoretical Computer Science, B, Elsevier Science Publishers, pp. 995 -- 1072. [8] R. Fagin, J. Y. Halpern, Y. Moses & M. Y. Vardi (1995): Reasoning about Knowledge. MIT Press. [9] J. Fisher, T. A. Henzinger, D. Nickovic, N. Piterman, A. V. Singh & M. Y. Vardi (2011): Dynamic Reactive Modules. In: Proceedings of CONCUR, pp. 404 -- 418, doi:10.1007/978-3-642-23217-6 27. [10] F. Gecseg (1986): Products of Automata. doi:10.1007/978-3-642-61611-2. EATCS Monographs on Theor. Comput. Sci., Springer, [11] G. J. Holzmannn (1997): The Model Checker SPIN. IEEE Transactions on Software Engineering 23(5), pp. 279 -- 295, doi:10.1109/32.588521. [12] W. Jamroga & T. Agotnes (2006): Modular Interpreted Systems: A Preliminary Report. Technical Report IfI-06-15, Clausthal University of Technology. [13] W. Jamroga & T. Agotnes (2007): Modular Interpreted Systems. In: Proceedings of AAMAS'07, pp. 892 -- 899, doi:10.1145/1329125.1329286. [14] M. Koster & P. Lohmann (2011): Abstraction for model checking modular interpreted systems over ATL. In: Proceedings of AAMAS, pp. 1129 -- 1130. [15] F. Laroussinie, N. Markey & G. Oreiby (2008): On the Expressiveness and Complexity of ATL. Logical Methods in Computer Science 4, p. 7, doi:10.2168/LMCS-4(2:7)2008. [16] O. Lichtenstein & A. Pnueli (1985): Checking that finite state concurrent programs satisfy their linear spec- In: POPL '85: Proceedings of the 12th ACM SIGACT-SIGPLAN symposium on Principles of ification. programming languages, ACM, New York, NY, USA, pp. 97 -- 107, doi:10.1145/318593.318622. [17] A. Lomuscio & F. Raimondi (2006): MCMAS : A Model Checker for Multi-agent Systems. In: Proceedings of TACAS, Lecture Notes in Computer Science 4314, pp. 450 -- 454, doi:10.1007/11691372 31. [18] A. Murano, M. Napoli & M. Parente (2008): Program Complexity in Hierarchical Module Checking. In: Proceedings of LPAR, pp. 318 -- 332, doi:10.1007/978-3-540-89439-1 23. [19] F. Raimondi (2006): Model Checking Multi-Agent Systems. Ph.D. thesis, University College London. [20] S. La Torre, M. Napoli, M. Parente & G. Parlato (2008): Verification of scope-dependent hierarchical state machines. Information and Computation 206(9-10), pp. 1161 -- 1177, doi:10.1016/j.ic.2008.03.017.
1606.00799	1	1606	2015-10-02T17:01:17	Multi-Agent Modeling of Dynamical Systems: A Self-organized, Emergent, Homeostatic and Autopoietic Approach	[ "cs.MA", "nlin.AO" ]	This thesis presents the theoretical, conceptual and methodological aspects that support the modeling of dynamical systems (DS) by using several agents. The modeling approach permits the assessment of properties representing order, change, equilibrium, adaptability, and autonomy, in DS. The modeling processes were supported by a conceptual corpus regarding systems dynamics, multi-agent systems, graph theory, and, particularly, the information theory. Besides to the specification of the dynamical systems as a computational network of agents, metrics that allow characterizing and assessing the inherent complexity of such systems were defined. As a result, properties associated with emergence, self-organization, complexity, homeostasis and autopoiesis were defined, formalized and measured. The validation of the underlying DS model was carried out on discrete systems (boolean networks and cellular automata) and ecological systems. The central contribution of this thesis was the development of a methodological approach for DS modeling. This approach includes a larger set of properties than in traditional studies, what allows us to deepen in questioning essential issues associated with the DS field. All this was achieved from a simple base of calculation and interpretation, which does not require advanced mathematical knowledge, and facilitates their application in different fields of science.	cs.MA	cs	Modelado Multi-agentes de Sistemas Dinámicos -Un enfoque Auto-organizado, Emergente, Homeostático y Autopoiético- Nelson Josué Fernández Parada Universidad de los Andes Facultad de Ingeniería, Doctorado en Ciencias Aplicadas Mérida, Venezuela 2015 Modelado Multi-agentes de Sistemas Dinámicos -Un enfoque Auto-organizado, Emergente, Homeostático y Autopoiético- Nelson Josué Fernández Parada Tesis presentada como requisito parcial para optar al título de: Ph.D. en Ciencias Aplicadas Director: Ph.D. José Lizandro Aguilar Co-director: Ph.D. Oswaldo Terán Centro de Microelectrónica y Sistemas Distribuidos Universidad de los Andes Facultad de Ingeniería, Doctorado en Ciencias Aplicadas Mérida, Venezuela 2015 IV Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨- Dedicatoria A quienes conforman mi bello hogar: Yois, Tere y Samito. A quienes conforman mi más amorosa familia: Benji, Chechi, Alix, Dora, Oswaldo, Sonia, Javier, Yamile; Jairo, Andrés, Diego, Santi; Alejandro, Jessica y Camilo Iván. A mis amigos y colegas más cercanos: Carlos Gershenson, Alberto Ramírez, Diego Lizcano, Silvia Álvarez, Carlos Ernesto Maldonado, Adriana Maldonado, Andrés Gonzáles, Alí Coronel, Christopher Stephens, Tom Froese y Mario Cosenza. Y finalmente, a la computadora y a la computación universal… Contenido Agradecimientos V Han pasado cuatro años desde que inicié el gran reto de obtener mi doctorado en una facultad de ingeniería, donde fui el primer biólogo marino en ser admitido. Por ello, quiero agradecer a quienes confiaron en mí la tarea de tender puentes, adicionales a los ya logrados, entre la computación y la biología. A mis mentores: José Aguilar, quien fue mi mayor motivador para no abandonar y mi mayor crítico a la hora de las ideas. A Oswaldo Terán, por su confianza. Agradezco muy especialmente a Carlos Gershenson, quien ha sido mi ejemplo más preciado de dedicación, sapiencia y brillantez. Así lo viví en cada estancia en el Instituto de Matemáticas Aplicadas y en Sistemas, así como en el Centro de Ciencias de la Complejidad de la UNAM. Muchos de mis logros y de la superación de grandes obstáculos, los debo a las interacciones relevantes que se generaron con Carlos. En la UNAM también quiero agradecer al Dr. Domínguez, director del IIMAS, por sus 3 invitaciones como investigador visitante. A Martha Flórez en comunicaciones del IIMAS y a Suyin en la biblioteca. A todos los investigadores del C3 por sus enriquecedoras charlas en los diversos seminarios. Agradezco, en la ULA, a quien me llamó hombre del renacimiento: Mario Cosenza, un físico con el poder de ver siempre "The Big Picture". Un valioso ser humano que siempre me motivó a pensar más allá, a adaptarme, a ser creativo. A Niriaska Perozo en la UCLA, a Tom Froese en el IIMAS-UNAM y Junior Altamiranda en la ULA, a Eduard Puerto en la UFPS por la camaradería, sus aportes y retroalimentación. VI Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨- Resumen y Abstract VII RESUMEN Esta tesis presenta una propuesta de modelado de sistemas dinámicos (SD) usando múltiples agentes. El modelado se desarrolla desde una perspectiva que permite evaluar las propiedades que confiere a los SDC orden, cambio, adaptabilidad, equilibrio y autonomía. La base teórica la constituyen la teoría de sistemas multi-agentes, la teoría de grafos, y en especial, la teoría de la información. Además de la especificación de un SD como una red computacional de agentes, se logró el establecimiento de métricas que permiten caracterizar y evaluar su complejidad inherente. El resultado fue la definición, formalización y medición de las propiedades de emergencia, auto-organización, complejidad, homeostasis y autopoiesis. La valoración del modelo de SD propuesto se hizo en sistemas discretos (Redes Booleanas Aleatorias y Autómatas Celulares) y sistemas ecológicos. En particular, el aporte de esta tesis fue el desarrollo de un enfoque metodológico de modelado de SD con un conjunto mayor de propiedades. Estas propiedades fueron la emergencia, auto-organización, complejidad, homeostasis y autopoiesis, a partir de las cuales se ahondó en los aspectos de cambio, regularidad, adaptabilidad, equilibrio dinámico y autonomía en SD. Lo anterior se logró desde una base sencilla de cálculo e interpretación basada en la teoría de la información, que no requiere de conocimiento matemático avanzado, lo cual hizo más práctica su aplicación en diversos campos como las ciencias computacionales, el urbanismo, y en especial, la ecología. Palabras clave: Ciencias de la complejidad, teoría de la información, caos, criticalidad, autonomía, sistemas ecológicos. VIII Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨- ABSTRACT This thesis presents the theoretical, conceptual and methodological aspects that support the modeling of dynamical systems (DS) by using several agents. The modeling approach permits the assessment of properties representing order, change, equilibrium, adaptability, and autonomy, in DS. The modeling processes were supported by a conceptual corpus regarding systems dynamics, multi-agent systems, graph theory, and, particularly, the information theory. Besides to the specification of the dynamical systems as a computational network of agents, metrics that allow characterizing and assessing the inherent complexity of such systems were defined. As a result, properties associated with emergence, self-organization, complexity, homeostasis and autopoiesis were defined, formalized and measured. The validation of the underlying DS model was carried out on discrete systems (boolean networks and cellular automata) and ecological systems. The central contribution of this thesis was the development of a methodological approach for DS modeling. This approach includes a larger set of properties than in traditional studies, what allows us to go deepen in questioning essential issues associated with the DS field. All this was achieved from a simple base of calculation and interpretation, which does not require advanced mathematical knowledge, and facilitates their application in different fields of science. Keywords: Complex sciences, information theory, chaos, criticality, autonomy, ecological systems. Contenido Contenido IX Pág. Dedicatoria ......................................................................................................................... IV RESUMEN ........................................................................................................................... VII ABSTRACT ......................................................................................................................... VIII Lista de figuras ................................................................................................................. XIII Lista de tablas .................................................................................................................. XVII Introducción ........................................................................................................................ 1 Motivación ........................................................................................................................... 3 Objetivos .............................................................................................................................. 4 Antecedentes ....................................................................................................................... 5 Organización de la tesis ..................................................................................................... 8 1. CAPITULO 1: SISTEMAS DINÁMICOS; AUTO-ORGANIZACIÓN, EMERGENCIA, COMPLEJIDAD, HOMESTASIS Y AUTOPOIESIS. ......................................................... 11 Resumen ............................................................................................................................ 11 Teoría de los Sistemas Dinámicos (TDS) ........................................................... 11 Visión General ........................................................................................ 11 Aspectos Históricos ................................................................................ 12 La Incertidumbre de la Previsibilidad .................................................... 12 Implicaciones Matemáticas .................................................................... 13 Noción General de SD ....................................................................................... 13 La Complejidad en Sistemas Dinámicos: Propiedades de Auto-organización, 1.2 1.3 Emergencia, Complejidad, Homeostasis y Autopoiesis. ............................................... 15 Emergencia ............................................................................................ 16 Auto-organización ................................................................................. 17 Complejidad ........................................................................................... 18 Homeostasis ........................................................................................... 19 Autopoiesis............................................................................................. 20 Síntesis .............................................................................................................. 21 1.4 1.1 1.1.1 1.1.2 1.1.3 1.1.4 1.3.1 1.3.2 1.3.3 1.3.4 1.3.5 X Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 2. CAPÍTULO 2: ESPECIFICACIÓN DE SISTEMAS DINÁMICOS COMO REDES COMPLEJAS DE AGENTES ............................................................................................. 23 2.1 Interdependencia Funcional de la Auto-organización, Homeostasis y Resumen ............................................................................................................................ 23 Nociones de SD: de lo Tradicional a la Inclusión de lo Complejo. .................... 23 Noción Tradicional de SD ...................................................................... 24 Una Definición de SD que Considera la Complejidad o Sistemas 2.1.1 2.1.2 Dinámicos Complejos (SDC)................................................................................. 24 2.1.3 Autopoiesis, y la Generación del Espacio Estructural del SDC . ............................ 27 Los SDC cómo Redes Computacionales de Agentes. ........................................ 29 Aportes de la Teoría de SDs ................................................................... 29 Aportes de la Teoría de agentes y de los Sistemas Multi-agentes-SMA .. 31 Aportes de las Teorías de Grafos y Redes ............................................... 32 Aspectos Formales de la Red SDC-Ag. ................................................... 33 La Especificación de Nodo Tipo Ag y SMA ............................................. 35 Síntesis .............................................................................................................. 37 2.2.1 2.2.2 2.2.3 2.2.4 2.2.5 2.2 2.3 3. CAPITULO 3: FORMALISMOS MATEMÁTICOS PARA LA MEDICIÓN DE LA EMERGENCIA, AUTO-ORGANIZACIÓN, COMPLEJIDAD, HOMEOSTASIS Y AUTOPOIESIS EN SISTEMAS DINÁMICOS.................................................................... 39 Resumen ............................................................................................................................ 39 Teoría de la Información cómo Fundamento .................................................... 39 3.1 3.2 Medidas ............................................................................................................. 42 Emergencia ............................................................................................ 42 Auto-organización ................................................................................. 44 Complejidad ........................................................................................... 45 Homeostasis ........................................................................................... 46 Autopoiesis............................................................................................. 54 Síntesis .............................................................................................................. 57 3.2.1 3.2.2 3.2.3 3.2.4 3.2.5 3.3 4. Capítulo 4: APLICACIONES EN SISTEMAS DISCRETOS Y SISTEMAS URBANOS .......................................................................................................................... 61 4.1 Resumen ............................................................................................................................ 61 Sistemas Discretos............................................................................................. 61 Las Redes Booleanas Aleatorias-RBA...................................................... 61 Midiendo la Complejidad en Redes Booleanas con Diferentes Regímenes 4.1.1 4.1.2 y Grados de Conectividad. .................................................................................... 62 Resultados para Redes Booleanas ........................................................... 63 4.1.3 Autómatas Celulares Elementales (ACE) .......................................................... 67 Experimentos con ACE. .......................................................................... 68 Resultados ACE ...................................................................................... 72 4.2.1 4.2.2 4.2 4.3 Medición de la Complejidad en Sistemas Urbanos-Semáforos Auto- organizantes ................................................................................................................. 76 La Movilidad Vehicular en las Ciudades Modernas y los Métodos Auto- 4.3.1 Organizantes para su Solución ............................................................................. 76 4.3.2 Modelo de Tráfico .................................................................................. 77 Contenido XI 4.3.3 Resultados .............................................................................................. 80 4.4 Discusión........................................................................................................... 84 La Aplicación de Las Medidas ................................................................ 84 4.4.1 4.4.2 La Complejidad cómo Balance ............................................................... 84 4.4.3 Múltiples Escalas y Perfiles..................................................................... 85 La Complejidad su Diferencia con el Caos ............................................. 86 4.4.4 Homeostasis ........................................................................................... 87 4.4.5 4.4.6 Autopoiesis y el Requisito de Variedad de Ashby. .................................. 88 Síntesis .............................................................................................................. 89 4.5 5. Capítulo 5: APLICACIONES EN SISTEMAS ECOLÓGICOS .................................. 91 5.1 5.1.1 5.1.2 5.1.3 5.1.4 Resumen ............................................................................................................................ 91 Visiones Tradicionales y Complejidad Ecosistémica ......................................... 91 Fundamentos de Limnología .................................................................. 92 Tipos de Lagos Estudiados ..................................................................... 96 Simulaciones y Cálculo de las Medidas .................................................. 96 Resultados .............................................................................................. 97 Complejidad comparativa en el Gradiente Altitudinal Ártico-Trópico (Ar-T). 103 Complejidad del Número Incremental de Especies de Mamíferos .................. 105 5.2 5.3 5.4 Modelado y Especificación de la Emergencia Acuática a través del uso de programación genética ............................................................................................... 107 5.5 Discusión......................................................................................................... 113 Descripción de los Ecosistemas y la Inclusión de la Complejidad como Sobre los Aspectos Computacionales de la Medición de la Complejidad 5.5.1 Indicador Ecológico ............................................................................................ 113 5.5.2 Ecológica ............................................................................................................. 113 5.5.3 Comparación con otras Medidas de Complejidad. ............................... 115 La Búsqueda de Soluciones analíticas para la Emergencia ................... 116 5.5.4 Síntesis y Comentario Final. ............................................................................ 116 5.6 6. CONCLUSIONES Y TRABAJO FUTURO ............................................................... 119 Logros Generales ............................................................................................. 119 Trabajo Futuro ................................................................................................ 121 6.1 6.2 7. Anexo: Caracterización de los Lagos Estudiados .............................................. 123 Lago Artico (Ar) .............................................................................................. 123 Lago Templado del Norte de Tierras Altas (North Higland Lake-NH). ........... 124 Lago Templado del Norte de Tierras Bajas (North Lowland Lake-NL)............ 125 Lago Tropical (T)............................................................................................. 126 7.1 7.2 7.3 7.4 Bibliografía ...................................................................................................................... 127 XII Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Contenido Lista de figuras XIII Pág.                Figura 2-1 Relación entre los Espacios Funcionales y Estructurales en SDC (Fernández et al., 2012a) ........................................................................................................................................... 27 Figura 2-2 Representación Simplificada de un SDC -Ag sobre la Base de la Teoría de Modelado y Simulación de Zeigler et al. (2000). Las designaciones A1, A2, A3 se refieren a Agentes. ....... 30 Figura 3-1 Información de Shannon a Diferentes Probabilidades para una Cadena Binaria (Gershenson and Fernández, 2012a) ............................................................................................ 41 Figura 3-2 Emergencia, Auto-organización y Complejidad vs Probabilidad en una Cadena Binaria (Gershenson and Fernández, 2012a) ............................................................................... 46 Figura 3-3 Zona de Viabilidad Zv y Zona de Funcionamiento Óptimo (ZFO) de un factor Abi o Bi para una red computacional. Los valores del eje x dependen del dominio del factor i. (Fernández et al., 2012b) .............................................................................................................. 50 Figura 3-4 Detalle de la Respuesta Global y Local en la Red SD-Ag. Los valores de los ejes son simplemente demostrativos (Fernández et al., 2012b) ................................................................ 51 Figura 3-5 Comportamiento de la resiliencia (Rsi) y la persistencia (Pei):) en el tiempo de retorno y el tiempo de estancia en el punto medio respectivamente, en constraste con la capacidad homeostática (Fernández et al., 2012b)..................................................................... 53 Figura 3-6 Comportamiento de la resistencia máxima (Re-Maxi) y la vulnerabilidad (Vui) en cuanto al incremento del rango de viabilidad (RXi) (Fernández et al., 2012b). ......................... 54 Figura 4-1 Diagramas de Cajas para los resultados de 1000 RBA, N=100 con variación de la conectividad K. A. Emergencia B. Auto-organización (S). C. Homeostasis y D. Complejidad (Gershenson and Fernández, 2012a). ........................................................................................... 63 Figura 4-2 E, S, C y H promedio para 1000 RBA, N = 100 nodos variando el promedio de conectividad K (Gershenson y Fernández, 2012). ........................................................................ 64 Figura 4-3 Autopoiesis (A) promedio para 50 conjuntos de redes Ne=96, Ni=32. Valores de A<1 en rojo. Valores en azul para A>1. El tamaño de los círculos indican que tan lejos está A de A=1. Los valores numéricos se muestran en la tabla 4-1 (Fernández et al., 2014b). .......................... 66 Figura 4-4 Evolución para la Regla 30 desde condiciones iniciales para grupos de tres vecinos. http://reference.wolfram.com/language/ref/CellularAutomaton.html ..................................... 67 Figura 4-5 Evolución de la regla 30 en 50 iteraciones http://reference.wolfram.com/language/ref/CellularAutomaton.html ..................................... 67 Figura 4-6 Promedios de 50 ECA para la 19 reglas, N=256, b=1. (Gershenson and Fernández, 2012a) ........................................................................................................................................... 74 Figura 4-7 Promedio de 50 ACE para 19 reglas, N=256, b=2 (Gershenson and Fernández, 2012a). .......................................................................................................................................... 74 XIV Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨                Figura 4-8 Promedio de 50 ACE para 19 reglas, N=256, b=4 (Gershenson and Fernández, 2012a). ...........................................................................................................................................75 Figura 4-9 Promedio de 50 ACE para 19 reglas, N=256, b=8 (Gershenson and Fernández, 2012a). ...........................................................................................................................................75 Figura 4-10 Perfiles de E,S,C y H a Escalas 1,2,4,8. . (Gershenson and Fernández, 2012a) ..........76 Figura 4-11 Reglas de autómatas celulares elementales usadas en el modelo de tráfico y su función cómo interruptores de los semáforos (Zubillaga et al., 2014). .......................................78 Figura 4-12 Fronteras no orientables para una grilla de cuatro por cuatro (Zubillaga et al., 2014)..............................................................................................................................................79 Figura 4-13 Esquema de una intersección y parámetros de medición para la definición de reglas en semáforos auto-organizantes (Zubillaga et al., 2014) ............................................................80 Figura 4-14 Resultado de la simulación para fronteras no-orientables para la ola verde y método auto-organizante: promedio de velocidad v y promedio de flujo J para diferentes densidades de ρ. Las curvas óptimas se muestran con línea punteada discontinua (Zubillaga et al., 2014). .......................................................................................................................................81 Figura 4-15 Emergencia (E), Auto-organización (S) y complejidad para los intervalos de cambio de luz en semáforos para fronteras no-orientables (Zubillaga et al., 2014). Las líneas verticales señalan la transición de cada una de las 7 fases del flujo (libre, cuasi-libre, subutilizado, capacidad completa, sobre-utilizado, cuasi-embotellamiento, embotellamiento) .....................82 Figura 4-16 Emergencia (E), Auto-organización (S) y complejidad para los intervalos en intersecciones en semáforos para fronteras no-orientables (Zubillaga et al., 2014). Las líneas verticales señalan la transición de cada una de las 7 fases del flujo (libre, cuasi-libre, subutilizado, capacidad completa, sobre-utilizado, cuasi-embotellamiento, embotellamiento) ......................................................................................................................................................83 Figura 4-17 Emergencia (E), Auto-organización (S) y complejidad para los intervalos en las calles en semáforos para fronteras no-orientables (Zubillaga et al., 2014). Las líneas verticales señalan la transición de cada una de las 7 fases del flujo (libre, cuasi-libre, subutilizado, capacidad completa, sobre-utilizado, cuasi-embotellamiento, embotellamiento) .....................83 Figura 4-18 Análisis de Componentes Principales. Las propiedades en Azul y las conectividades en Rojo. .........................................................................................................................................88 Figura 5-1 Zonación de Un Lago. La figura muestra las zonas Superficial, planctónica, de macrófitas acuáticas, bentónica y el substrato.(Fernández et al., 2014b). .................................93 Figura 5-2 Cajas de bigotes para las variables físico-químicas (Fernández and Gershenson, 2014; Fernández et al., 2014b). En ella se representan cada uno de los valores de cada variable representados por puntos. Con ello se observa la distribución real de los valores. Las abreviaciones se dan en las tablas 5-1 ..........................................................................................97 Figura 5-3 Distribución de los valores normalizados a base 10 de las variables físico-químicas en las clases 0-9. (Fernández and Gershenson, 2014; Fernández et al., 2014b). Variables con valores distribuidos de manera similar entre las clases 0 a 9, tendrán mayor información. Es decir, alta emergencia pues la variable puede tomar cualquier valor. Variables con puntos distribuidos de manera más concentrada en una u otra clase, producirán menos información. Es decir, alta auto-organización pues tales clases tendrán una probabilidad alta.. ..................98 Figura 5-4 Emergencia para las variables físico-químicas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) ..................................................................................99 Contenido XV                    Figura 5-5 Auto-organización de las variables físico-químicas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) ................................................................................. 99 Figura 5-6 Complejidad de las variables físico-químicas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) ................................................................................. 99 Figura 5-7 Homeostasis del componente físico-químico de un lago ártico en un ciclo anual. (Fernández and Gershenson, 2014; Fernández et al., 2014b) ..................................................... 101 Figura 5-8 Complejidad para los 3 componentes desde las variables elegidas en la tabla 5, para dos zonas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) ....... 102 Figura 5-9 Autopoiesis de la Biomasa Planctónica y Bentónica respecto de los nutrientes limitantes (PB/LN) y las variables fisicoquímicas (PB/PC), escogidas en tabla 5 (Fernández and Gershenson, 2014; Fernández et al., 2014b). Para ambos casos fue superior a uno por lo que se les da el color azul ..................................................................................................................... 103 Figura 5-10 Complejidad Comparativa en cuatro ecosistemas acuáticos- Ártico-Ar, Templado tierras altas-NH, Templado tierras bajas-NL y Tropical-T (Fernández et al., 2014d) .............. 104 Figura 5-11 Tendencia de la complejidad en una comunidad simulada de mamíferos para 10 sitios. Se muestran las primeras 800 especies en 1000 iteraciones (Fernández et al., 2013). ... 107 Figura 5-12 Mejor Solución Obtenido para Nutrientes Limitantes ........................................... 108 Figura 5-13 Mejor Solución Obtenido para Biomasa ................................................................. 109 Figura 5-14 Mejor Solución Obtenido para el Componente Físico-químico ............................. 109 Figura 5-15 Prueba de Ajuste del Modelo PG al Componente de Nutrientes limitantes ........... 110 Figura 5-16 Prueba de Ajuste del Modelo PG a la Biomasa ....................................................... 111 Figura 5-17 Prueba de Ajuste del Modelo PG al componente Físico-químico ........................... 111 Figura 5-18 Emergencia, Auto-organización, Complejidad y Homeostasis para un el Componente Físico-químico de un Lago Ártico ........................................................................ 112 Figura 5-19 Emergencia, Auto-organización, Complejidad y Homeostasis para un el Componente Físico-químico de un Lago Tropical .................................................................... 112 Figura 7-1 Hidroclimatología del lago Ártico (Randerson and Bowker, 2008). A la izquierda la temperatura en las 3 zonas de estudio. A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. ................................................................................................ 123 Figura 7-2 Hidroclimatología del lago templado del norte-NH (Randerson and Bowker, 2008). A la izquierda temperatura en las 3 zonas de estudio. A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. ................................................................................. 124 Figura 7-3 Hidroclimatología del lago templado del norte de tierras bajas-NL (Randerson and Bowker, 2008). A la izquierda temperatura en las 3 zonas de estudio. A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. ........................................................ 125 Figura 7-4 Hidroclimatología del lago tropical. A la izquierda temperatura en las 3 zonas de estudio (Randerson and Bowker, 2008). A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. ................................................................................................ 126 XVI Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Contenido XVII Lista de tablas             Pág. Tabla 1-1 Tipos de Sistemas Dinámicos (SD) .............................................................................. 14 Tabla 4-1 Autopoiesis (A) promedio para 50 conjuntos de redes Ne=96, Ni=32. En rojo: redes acopladas con 𝑨 < 𝟏, en azul redes acopladas con 𝑨 > 𝟏 Los resultados son los mismos que se ven en la fig. 4-3 (Fernández et al., 2014b). ................................................................................. 66 Tabla4-2 Reglas Incluidas de la Clase I. Condiciones de Evolución y Representación de 20 Iteraciones. .................................................................................................................................. 69 Tabla 4-3 Reglas Incluidas de la Clase II. Condiciones de Evolución y Representación de 20 Iteraciones ................................................................................................................................... 70 Tabla 4-4 Reglas Incluidas de la Clase III. Condiciones de Evolución y Representación de 20 Iteraciones ................................................................................................................................... 71 Tabla 4-5 Reglas Incluidas de la Clase IV. Condiciones de Evolución y Representación de 20 Iteraciones ................................................................................................................................... 72 Tabla 4-6 Reglas de Actualización de ACE ................................................................................. 77 Tabla 5-1 Variables del Componente Físico-químico-PC ............................................................ 94 Tabla 5-2 Variables del Componente de ...................................................................................... 95 Tabla 5-3 Variables del Componente de Biomasa-Bio ................................................................ 95 Tabla 5-4 Categorías, Rangos y Colores de Clasificación de las Propiedades E,S y C ................ 97 Tabla 5-5 Componentes tenidos en cuenta para el cálculo de la complejidad y posterior autopoiesis, en las zonas planctónica y bentónica del lago ártico ........................................... 102 Contenido XVIII Introducción Desde una perspectiva amplia, se puede estimar que todo lo que nos rodea presenta algún grado de dinamismo, es decir de cambio. Aunque puede resultar ser una cuestión de escala (nivel de abstracción), el cambio se da incluso en entidades cómo las rocas, que lo hacen a escalas de tiempo geológico. El cambio en los diversos sistemas naturales y artificiales, ha llamado la atención de la comunidad científica durante mucho tiempo. A pesar que cada día la renovada y aumentada capacidad computacional nos ha permitido acercarnos a su entendimiento, los aspectos relativos a su complejidad ha limitado la posibilidad de su predicción. Mucho de ello se debe a que en los sistemas cambiantes, los comportamientos globales vienen de la diversidad de interacciones surgidas cómo resultado de la naturaleza de sus elementos. Por ello, la ciencia tradicional y su enfoque reduccionista necesita hoy en día complementarse, ahondar en las interacciones, para acercarse más al entendimiento de un sistema dinámico con rasgos de complejidad. Si bien conocemos en diferentes áreas y disciplinas mucho acerca de las partes, hoy tenemos la posibilidad de conocer y ahondar más en las interacciones. El panorama que surge se hace muy interesante. Uno donde podemos armonizar diversos aspectos del universo, de la vida y la computación, tales como la materia, la energía, la cultura, la tecnología, entre muchos otros. Sobre esta base se puede llegar a descifrar los aspectos ocultos del comportamiento y dinamismo en los problemas actuales en el campo de la biología, ecología, sociología y economía. Cómo fundamento para el entendimiento de la complejidad, tenemos que el dinamismo surge desde las interacciones auto-organizantes de las partes. Las interacciones permiten mantener la integridad del sistema, gracias a la autonomía y la autorregulación de las partes. Estos rasgos, renombrados cómo: emergencia, auto-organización autopoiesis y homeostasis, son los que le proporcionan a los sistemas dinámicos la suficiente robustez y posibilidad de cambio ante las variaciones internas y externas. Es decir, le dan la complejidad necesaria para mantener su adaptación y evolución en un contexto de variedad requerida (Ross Ashby, 1956). En ese contexto, se hace posible una refinación y redefinición de las propiedades y conceptos necesarios para la descripción apropiada de un sistema dinámico compuesto de múltiples entidades y en diferentes niveles de abstracción. Este enfoque resulta promisorio, por la posibilidad de contribuir al entendimiento de algunos de los retos actuales, conocidos 2 Introducción cómo las 3 principales emergencias en la historia de la humanidad, planteados en (Gershenson, 2013a): (i) la transición entre lo no vivo y lo vivo, (ii) el surgimiento de la conciencia y (iii) la naturaleza del espíritu humano (cómo el máximo nivel de conciencia). Introducción 3 Motivación En el estudio de los sistemas dinámicos con múltiple número de agentes, han sido apreciables los esfuerzos por generar nuevas nociones, formalismos y metodologías que nos permitan describir los conceptos de emergencia, auto-organización y complejidad (Aguilar, 2014; Bar- Yam, 2004a; Gershenson and Heylighen, 2003; Heylighen, 2011; Koza, 1995; Shalizi et al., 2004). Sin embargo, ha surgido la necesidad de valorar otros aspectos complementarios, cómo la homeostasis y la autopoiesis. Es decir, existe la posibilidad de valorar un número mayor de propiedades en los SD, con el fin de lograr un mejor entendimiento del resultado global de las interacciones que dan origen a su estructura y dinámica. Sobre la anterior base, la investigación doctoral aquí desarrollada buscó lograr una mejor interpretación de las dinámicas colectivas en los SD, y ahondar en el entendimiento de los aspectos relativos a su complejidad. Para ello, se enfocó en generar métricas precisas y formales que representaran adecuadamente los conceptos de emergencia, auto-organización, complejidad, homeostasis y autopoiesis (𝐸𝐴𝐶𝐻𝐴), para fines de modelado de SD. En este sentido, se buscó generar una contribución al conocimiento de los SD que permitiera clarificar el significado de los conceptos que 𝐸𝐴𝐶𝐻𝐴 describiesen. Para tal fin, se diseñaron aplicaciones a fenómenos computacionales y naturales que nos posibilitaran estudiar y analizar el sentido práctico de las formalizaciones a desarrollar. El alcance de los resultados obtenidos en esta tesis, se orientó a brindar una visión de los SD que fuese más allá de los límites del reduccionismo. Nuestro esfuerzo estuvo dirigido a ayudar a responder a las preguntas esenciales en cuanto a la complejidad en SD, que la comunidad científica se viene planteando. Todo ello, a través de un enfoque interdisciplinar, soportado por una red científica colaborativa internacional. 4 Objetivos Introducción General Desarrollar un enfoque metodológico para sistemas dinámicos complejos, que considere la exploración, generación y modelado de los aspectos de auto-organización, emergencia, complejidad, homeostasis y autopoiesis. Específicos  Definir formalmente un sistema multi-agente como un sistema dinámico  Proponer nuevas nociones, con sus formalismos, para la caracterización de los sistemas multi-agente como sistemas dinámicos.  Desarrollar casos de estudio inspirados, entre otros, en un sistema socio-ecológico de referencia. Introducción 5 Antecedentes En décadas recientes, el estudio de los sistemas complejos ha incrementado el entendimiento de un amplio número y rango de fenómenos. Sus fundamentos matemáticos se hallan en las teorías del caos, de la información y de la computación. Se han tratado de explicar desde el estudio de los fractales, los mapas iterativos y la termodinámica. Igualmente, se han desarrollado modelos y técnicas de modelado de gran utilidad, que se han basado en técnicas analíticas y simulación computacional. Se incluyen allí, la mecánica estadística, la dinámicas estocástica, los autómatas celulares y el modelado multi-agentes, por mencionar algunos. (Bar-Yam, 2003; Mitchell, 2009). Los conceptos usados para clarificar el estudio de los SD, tales como emergencia, adaptabilidad, auto-organización y complejidad, han sido usados en diferentes contextos con diferentes significados. Los tópicos abordan casos como:  El análisis de la medidas de complejidad (Crutchfield and Young, 1989; Edmonds, 1999; Lloyd, 2001), en donde se inspeccionan desde la filosofía del modelado, hasta las formas de medir la complejidad, su perspectiva estadística, y la clasificación temática de las diferentes medidas.  El análisis de sistemas naturales (incluidos ecosistemas) desde la teoría de la computación y la información, para cuantificar la complejidad estructural. En específico, el estudio del fenómeno de adaptación desde simulaciones computacionales de sistemas naturales, da el entendimiento para el diseño y construcción de sistemas complejos (Crutchfield, 2011; Koza, 1995).  El estudio de la irreductibilidad de la complejidad, explicada desde el contenido de información algorítmica, dada cuando en una cadena de bits la información contenida es igual a su tamaño. En estos casos, la cadena de bits es incompresible y no tiene redundancia (Chaitin, 2007).  El análisis de la estructura y función de redes complejas de ámbitos biológicos o sociales, partiendo de variadas técnicas de modelado y sobre la base de indicadores estadísticos para su caracterización (Newman, 2003). 6 Introducción  El estudio de la auto-organización como la generación de la propia estructura en organismos biológicos (Camazine et al., 2001), considerando que la auto-organización y la emergencia pueden coexistir como fenómenos separados dado que cada una enfatiza en diferentes aspectos del sistema (De Wolf and Holvoet, 2005), la auto-organización de sistemas fuera del equilibrio (Nicolis and Prigogine, 1977), y la apreciación de la complejidad como una condición intermedia entre el caos y el orden, relacionándose así con la auto-organización (Heylighen, 2011).  La caracterización de la dinámica auto-organizante de estructuras complejas en la sociedad y la naturaleza, en aspectos relacionados como la evolución de la complejidad, la evolución de la optimización y el estudio de casos como los medios de transporte (Schweitzer, 1997a).  La evaluación de las complejidades urbanas y ecológicas que surgen de dinámicas locales a globales (Bottom-Up), desde estas perspectivas, las ciudades son observadas como organismos vivos (Batty, 2012; Batty, 1971). Al mismo tiempo, en la ecología se valora la contribución del enfoque de los sistemas complejos a las dinámicas de acoplamiento del habitat, a las fluctuaciones poblacionales, y a las interacciones en redes tróficas. Desde este punto, se ve como el estudio de la complejidad en ecología es un campo naciente (Parrott, 2010).  El modelado individual basado en dinámicas colectivas de sistemas físicos de partículas que siguen el movimiento Browiniano (Schweitzer, 1997b).  Los intentos de unificación desde la teoría de la información del concepto de auto- organización, y no sólo eso, sino la relación de ésta última con la teoría del caos (Haken, 1989). Desde la perspectiva de Haken (1989) y Prokopenko (2009a), se aborda la existencia de 3 aspectos de la auto-organización para la vida artificial: (a) la evolución del sistema desde la interacción de sus componentes a estados de mayor organización, (b) la manifestación de esta organización en una coordinación y manifestación global, (c) la expresión de patrones no impuestos por influencias externas.  Las necesidades de un nuevo paradigma para la complejidad desde varias teorías, que satisfagan diversas aplicaciones en distintos sistemas (Morin, 1992).  El análisis de la universalidad de la complejidad y su irreductibilidad computacional, que limita la predictibilidad del sistema, por cuanto existe gran diversidad de interacciones locales que brindan apreciable riqueza en el comportamiento del mismo (Wolfram and Gad-el-Hak, 2003; Wolfram, 1984). Como punto de coincidencia, muchos de los anteriores autores tocan una diversidad de nociones, en especial de complejidad, emergencia y auto-organización. Se refieren de manera apreciable a la ambigüedad, confusión, e incluso abuso, de la terminología, en discursos no científicos (Edmonds, 1999). En el caso de los sistemas multi-agentes, han existido trabajos seminales que han permitido consolidar su teoría. Entre ellos se destacan los aportes teóricos y prácticos de Wooldridge Introducción 7 and Jennings (1995), para el diseño y construcción de agentes inteligentes. En el marco de la teoría de agentes se define que es uno y el uso de formalismos matemáticos para su representación y la de su razonamiento. Se abordan también las arquitecturas y modelado desde la ingeniería del software, y el caso no menos importantes de los lenguajes y sistemas de programación y experimentación de agentes. Por su parte, Ferber and Gutknecht (1998), proponen un meta-modelo de organización artificial entre agentes llamado AALAADIN. Basados en AALAADIN, proponen la plataforma MADKIT para apoyar el modelado con agentes heterogéneos. Ante el auge tomado por los sistemas multi-agentes, se ha analizado su vinculación con terminos como situabilidad, flexibilidad y autonomía, para clarificar sus nociones y sus aspectos metodológicos de representación ( Jennings et al. 1998; Teran 2001). Ferber et al. (2004) han avanzado en el modelado con la propuesta de principios generales para sistemas multi-agentes de organización centralizada (OCMAS en Inglés). Trabajos más actuales se han esforzado en involucrar las propiedades de complejidad, auto- organización y emergencia en los sistemas multi-agentes. Muchos de ellos han buscado el desarrollo de arquitecturas, plataformas de modelado y modelos específicos, para que sea viable representar, reproducir y medir la auto-organización y la emergencia en sistemas multi-agentes. (Aguilar et al., 2007; Aguilar, José, Besembel, I., Cerrada, M., Hidrobo, F., Narciso, 2008; Perozo, Niriaska, Aguilar, José, Terán, O, Molina, 2012a, 2012b). 8 Introducción Organización de la tesis La tesis tiene una organización preferencialmente secuencial, de manera que los elementos de un capitulo precedente brinda elementos para el entendimiento de los demás. En la siguiente sección se dan los elementos principales de la teoría de los SD, y las nociones básicas de auto-organización, emergencia, complejidad, homeostasis y autopoiesis (𝐴𝐸𝐶𝐻𝐴). Los elementos de los SD incluyen a la complejidad como una mirada complementaria para su estudio, que va más allá del reduccionismo tradicional; por su parte, las nociones 𝐴𝐸𝐶𝐻𝐴 sientan la base conceptual de la formalización. En el capítulo 2 se discute como puede llegarse a la especificación de un sistema dinámico como una red computacional de agentes. Para ello se propone integrar los elementos de 𝐴𝐸𝐶𝐻𝐴, y se finaliza con varios formalismos que también integran la teoría de SD, de multi-agentes y de grafos. En el capítulo 3 se presentan los formalismos desarrollados para cada noción de 𝐴𝐸𝐶𝐻𝐴. Allí se muestra la riqueza de los formalismos, que dio origen a diversos indicadores. En ese capítulo las nociones derivadas de la teoría de la información para 𝐴𝐸𝐶𝐻𝐴, se constituyen como el eje central de la tesis. El capítulo 4 presente aplicaciones en sistemas discretos, como en las redes booleanas aleatorias, y en autómatas celulares elementales; además, contiene una aplicación urbana en un modelo de tráfico que se basa en tres reglas de autómata celular elementales, de gran utilidad para la planeación en ciudades vivas y auto-organizantes. El capítulo 5 discute y analiza el sentido ecológico de la complejidad, a partir de tres aplicaciones a sistemas ecológicos, ecosistemas acuáticos, y el análisis de comunidades de mamíferos. Se destaca que este capítulo presenta importantes consideraciones para incluir a la complejidad como otro indicador ambiental, útil para el de los ecosistemas. El capítulo 6 finaliza la tesis con las conclusiones y trabajos futuros. Introducción 9 1. CAPITULO 1: SISTEMAS DINÁMICOS; AUTO-ORGANIZACIÓN, EMERGENCIA, COMPLEJIDAD, HOMESTASIS Y AUTOPOIESIS. Resumen Este capítulo presenta los aspectos teóricos y conceptuales básicos acerca de los sistemas dinámicos complejos (SD). Desde una perspectiva histórica, se discuten el elemento de su incertidumbre; a partir de allí, se ahonda en la inclusión de la dimensión de la complejidad. Con lo anterior como base, se muestran nuestras consideraciones acerca de las características de complejidad, emergencia, auto- organización, homeostasis y autopoiesis en SD. Este último apartado tiene como fin brindar una visión lo más clara posible de cada una de ellas, con el objeto de contribuir a su desambiguación y sentar las bases para su posterior modelado matemático. 1.1 Teoría de los Sistemas Dinámicos (TDS) En esta tesis observamos la TDS como el área de las matemáticas usadas para describir el comportamiento de sistemas dinámicos complejos (Mitchell, 2009). En consecuencia las generalizaciones y nociones que se presentan en apartados subsiguientes se enfocan en este tipo de descripción. 1.1.1 Visión General La teoría de sistemas dinámicos (TSD), concierne con la descripción y predicción de sistemas que exhiben un comportamiento complejo en un nivel macroscópico, que emerge desde las acciones colectivas de interacciones de sus componentes. La palabra dinámica es referida al cambio, y los sistemas dinámicos (SD), efectivamente, cambian en el tiempo de alguna forma. Existen variados ejemplos de SD, como lo son: el sistema solar, el corazón y el cerebro de las 12 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- criaturas vivas, la población mundial, o el cambio climático. No obstante, ejemplos como las rocas también son un tipo de SD, nada más que sus cambios se dan a escalas de tiempo geológico (Mitchell, 2009). La TSD describe en términos generales las formas en las que el sistema puede cambiar, qué tipo de comportamientos macroscópicos son posibles, y qué tipos de predicciones se pueden hacer acerca de estos comportamientos. Actualmente esta teoría está en boga debido a los fascinantes resultados obtenidos de uno de sus productos intelectuales, el estudio del caos (Crutchfield, 2011). 1.1.2 Aspectos Históricos El origen de la TSD se remonta a tiempos aristotélicos, dado que este filósofo fue autor de varias teorías de movimiento, una de ellas aceptada ampliamente por más de 1.500 años. Esta teoría básicamente definió dos principios (Gyekye, 1974), (i) el movimiento de la tierra difería del movimiento de los cielos y (ii) los objetos terrestres se movían de diferente forma en dependencia de lo que estuvieran hechos. Posteriormente, los aportes de Galileo sobre el movimiento de los planetas desde bases experimentales, contribuyeron a sentar muchos principios que hemos conocido en diversos cursos de física. Después de la muerte de Galileo, en ese mismo año, nació Newton, la persona más importante en la historia de la dinámica, dado que inventó una rama de las matemáticas que describe el movimiento y el cambio (Madhusudan, 2010). Los físicos de la época llamaron al estudio del movimiento Mecánica (Verlinde, 2011), la cual podía explicarse en términos de acciones combinadas de máquinas simples como palancas, poleas, ruedas y ejes. En particular, un trabajo clave en esta área es el de Newton, conocido como Mecánica Clásica. La mecánica clásica se divide en dos ramas (Christianson, 1998): (i) la cinemática o cómo las cosas se mueven, y (ii) la dinámica que se ocupa por explicar el movimiento de las cosas como producto de las leyes de la cinemática. Newton pudo explicar el cambio en el movimiento elíptico de orbitas planetarias en términos de la fuerza, llamada gravedad. Desde la mecánica newtoniana, que produjo una figuración del "mecanismo del universo", Pierre Sión Laplace vio la importancia que tenía esta figuración para la predicción; por lo que mencionó que desde las leyes de Newton, y al tener la velocidad y posición de una partícula, era posible, en principio, predecir todo (Rivas Lado, 2001). 1.1.3 La Incertidumbre de la Previsibilidad Dos descubrimientos en el siglo XX mostraron que esa capacidad predictiva definida por Laplace no era posible. El primero fue el descubrimiento del principio de incertidumbre, en mecánica cuántica, de Werner Heisenber en 1927, que estableció que no se puede hacer medición de la posición y el momentum (masa y velocidad) de una partícula al mismo tiempo. El segundo fue el descubrimiento del Caos; los Sistemas Caóticos son aquellos en que minúsculas incertidumbres en la medición de la posición inicial, resultan en grandes errores en términos de predicción a largo plazo. Esto es conocido como la sensibilidad a las CAPÍTULO 21 13 condiciones iniciales. Comportamientos caóticos han sido observados en desordenes cardiacos, turbulencias de fluidos, circuitos electrónicos, grifos que gotean, entre otros. El matemático Francés Henri Poincaré, fundador de la topología algebraica, demostró la sensibilidad de las condiciones iniciales en el problema de los tres cuerpos, que trataba de predecir las posiciones futuras de varias masas que se atraían arbitrariamente una a la otra bajo las leyes de Newton. Este hecho marcó uno de los momentos importantes en la historia de la TSD, pues puso en duda la presuposición científica de previsibilidad, y demostró los límites de la previsibilidad heredada hasta ese momento y evidenciada en los trabajos de Galileo, Newton y Laplace (Gershenson, 2011). En nuestros días, la conciencia de falta de previsibilidad de muchos de los SD ha llevado a considerar no solo las partes, sujetos de especial interés desde el método científico tradicional, sino también las múltiples interacciones que existen entre ellas. Esta gran cantidad de interacciones son las que crean sistemas entretejidos cuya dinámica es difícil de separar. Son las interacciones la característica más importante de muchos SD, dado que el comportamiento de un elemento depende de otro, y así el de muchos otros, de forma tal que el rasgo que sobresale es la interdependencia entre elementos del sistema. Esto es lo que hace que los SD se tornen difíciles de predecir, que se tornen complejos, y por ende, que se clasifiquen como tal (Bar-Yam, 2003; Gershenson, 2011). 1.1.4 Implicaciones Matemáticas Típicamente, se considera que un sistema está "resuelto" cuando se llega a un conjunto finito de expresiones finitas, que pueden ser usadas para predecir su estado en un tiempo t, dado el estado del sistema en algún tiempo inicial 𝑡0, o anterior 𝑡1. Las soluciones halladas de esta forma, no son generalmente apropiadas o posibles para muchos SD. En este sentido, la contribución central de la TDS a la ciencia moderna es que las soluciones exactas no son necesarias para entender y analizar los procesos no lineales. Así, en lugar de hallar soluciones exactas o descripciones estadísticas ajustadas, el énfasis de la TDS está en describir la geometría y la estructura topológica del conjunto de soluciones. Es decir, la TDS da una visión geométrica de un proceso estructural de elementos, como atractores, cuencas y separatrices; hecho que la distingue del enfoque puramente probabilístico de la mecánica estadística, en la que la estructura geométrica no es considerada. En adición, la TDS también direcciona las preguntas acerca de qué estructuras genéricas se hallan, esto es, qué tipos de comportamientos son típicos a través del espectro de los SD (Mitchell, 2009; Mitchell et al., 1994). 1.2 Noción General de SD A continuación se hace referencia a las descripciones matemáticas que permiten de manera idealizada representar muchos de los fenómenos reales que acontecen a nuestro alrededor y 14 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- que tienen como característica principal su cambio y evolución en el tiempo. Desde esta consideración básica, se han definido dos componentes para un SD (Scheinerman, 1996): (i) un vector de estado que describe su estado actual, y (ii) una función-regla por medio de la cual, dado el estado actual, se puede obtener el estado del sistema en el siguiente instante de tiempo. Desde esta perspectiva, los SD se han definido como: sistemas que cambian o evolucionan su estado (𝑞) en el tiempo (t), en razón a una regla de evolución (𝑅) que lo demarca. Esta regla puede tener la estructura 𝑅: 𝑄 × 𝑇 → 𝑄, donde 𝑄 es el conjunto de estados posibles y 𝑇 el tiempo. Así, 𝑅 conduce a un estado 𝑞 ∈ 𝑄. De acuerdo a como se especifique la regla se pueden dar diferentes tipos de SDs. Desde nuestra perspectiva, un tanto más amplia e inscrita en el marco de la TDS expuesta en el ítem 1.1., un SD puede ser observado como un sistema dependiente del tiempo, tal que las interacciones de sus elementos según sus estados locales, determinan los estados globales del sistema. La tabla 1-1 lista los más conocidos tipos de SD reportados en la literatura, acorde con la variación del tiempo, su linealidad o no, y su naturaleza determinista. Tabla 1-1 Tipos de Sistemas Dinámicos (SD) Tipo de SD Discreto Característica Formalismo/Modelo de Ejemplo El tiempo transcurre en lapsos pequeños, por lo que el cambio se da Ecuación logística * 𝑞𝑡+1 = 𝑟𝑞𝑡(1 − 𝑞𝑡) en instantes separados de tiempo. Ej. Crecimiento poblacional (r), limitado 𝑡 ∈ ℝ Continuo El tiempo transcurre de manera continua, y así el cambio. 𝑡 ∈ ℤ Lineales Efecto proporcional a la causa. Opera el principio de superposición (Si se conocen dos soluciones para un sistema lineal, la suma de ellas es también una solución). 𝑓(𝑥 + 𝑦) = 𝑓(𝑥) + 𝑓(𝑦) No lineales Efecto no proporcional a la causa. por la capacidad de carga (𝐾) del sistema, que corresponde con la cantidad de elementos que puede soportar el mismo. Ecuación diferencial* 𝑑𝑥 𝑑𝑡 = 𝑟𝑞(1 − 𝑥) Ej. Modelo de predador-presa de Lokta-Volterra Ecuación lineal* 𝑞𝑡+1 = 𝑟𝑞𝑡 Ej. Sistema Masa-Resorte (Oscilador Armónico Simple) Opera el principio de no Péndulo doble, movimiento descrito por superposición, es decir: 𝑓(𝑥 + 𝑦) ≠ ecuaciones de Lagrange. Deterministas Regla 𝑅 puede ser funciones 𝑓(𝑥) + 𝑓(𝑦). iterativas. El determinismo en SD permite generar cierta predictibilidad sobre el sistema basado en las relaciones causa-efecto presentes en él, no obstante, dicha predictibilidad tiene sus límites. 𝑅 está dada por 𝑅(𝑞(𝑡), 𝑟) → 𝑞(𝑡 + 𝜏), donde 𝑞 es un estado del conjunto de estados posibles 𝑄, 𝑟 un parámetro de la regla y 𝜏 es un lapso de tiempo. De esta manera, dado un estado 𝑞(0), la evolución estará determinada ∀𝑡. CAPÍTULO 21 15 Estocásticos Abarcan eventos aleatorios; su evolución puede ser estimada de La variable de estado 𝑞(𝑡) es aleatoria, cuyo dominio son los reales, de acuerdo manera probabilística. con una función de probabilidad. Ej. La planificación de una línea de producción. Ej. el tiempo de funcionamiento de una máquina entre avería y avería, su tiempo de reparación, el tiempo que necesita un operador humano para realizar una determinada operación *Donde 𝑡 es tiempo, 𝑞 es la variable de estado y 𝑟 parámetro que afecta la variable 𝑞. 1.3 La Complejidad en Sistemas Dinámicos: Propiedades de Auto-organización, Emergencia, Complejidad, Homeostasis y Autopoiesis. El punto central del estudio de la complejidad en los Sistemas Dinámicos (SD), está en entender las leyes y mecanismos por los cuales un comportamiento coherente global y difícil de entender, emerge desde las actividades colectivas de componentes locales, relativamente simples. Dada la diversidad de sistemas que se incluyen en la categoría de complejos, el descubrimiento de cualquier rasgo común o ley universal se hace valiosa para conformar marcos teóricos generales (Mitchell, 2009). Actualmente, para observar a un SD bajo la óptica de su complejidad, se requiere de mayores conceptos y profundizar en sus aspectos formales. Se requiere, también, de sofisticadas herramientas, tanto analíticas como computacionales. En este sentido, la representación de un SD con un número mayor de propiedades que faciliten su análisis se hace apreciable (Bar- Yam, 2003; Bersini et al., 2010; Luisi, 2007). Propiedades como el caos (sensibilidad a las condiciones iniciales), la no linealidad (no proporcionalidad entre causa y efecto), la criticalidad auto-organizada (orden desde condiciones críticas), la auto-organización en sistemas fuera del equilibrio, han estado entre las más consideradas en este campo (Bak et al., 1988; Langton, 1990; Nicolis and Prigogine, 1989). Ahora bien, tradicionalmente la emergencia y la auto-organización han constituido los fenómenos más utilizados para explicación de la complejidad en SD. Sin embargo, tales propiedades podrían no ser suficientes para entender, entre otros, la adaptabilidad y la evolución espacio-temporal de los SD. Así, se ha observado como pertinente ahondar en otras propiedades referidas al equilibrio y a la autonomía, como lo son las propiedades de homeostasis y de autopoiesis (Fernández et al., 2010a). La homeostasis (equilibrio dinámico, en general) y la autopoiesis (auto-producción, auto-mantenimiento) constituyen propiedades emergentes que dan soporte al proceso de auto-organización. El sustento de ello está en que: (i) la capacidad auto-organizante del SD de adaptarse a ambientes cambiantes, 16 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- acorde con sus necesidades de sobrevivencia y/o mejoramiento de su funcionamiento colectivo al interior del SD, se logra sobre la base del establecimiento de un equilibrio dinámico u homeostasis. Es decir, la auto-organización se respalda en mecanismos de autorregulación (Martius et al., 2007). (ii) Adicionalmente, la capacidad auto-organizante de los SD se apoya en procesos autopoiéticos (Maturana and Varela, 1980) que le confieren la propiedad de autonomía. La autonomía, es inherente a sistemas cerrados que se auto- referencian, se auto-constituyen en una dinámica de continua producción de sí mismos. A través de la homeostasis y la autopoiesis, se puede desarrollar, conservar, producir y sustentar la propia organización en el tiempo (Garciandía, 2005), desde un enfoque de auto- mantenimiento del SD. Desde la anterior perspectiva, es posible definir un SD como un sistema con rasgos de (𝐸𝐴𝐶𝐻𝐴) emergencia, auto-organización, complejidad, homeostasis y autopoiesis (Fernández et al., 2010a). A continuación presentamos una síntesis de las nociones básicas de 𝐸𝐴𝐶𝐻𝐴 que, desde nuestra visión, contienen los elementos para su posterior formalización. 1.3.1 Emergencia La emergencia es uno de los conceptos que más mal entendido y mal usado ha sido en décadas recientes. Las razones para ello son variadas e incluyen: polisemia (múltiples significados), moda o deseo de impresionar (buzzwording), confusión, Platonismo, y aún misticismo. Podemos decir que la controversia de su concepto ha durado décadas. Aun así, el concepto de emergencia puede ser definido desde ciertas miradas (Anderson, 1972) (Aguilar, 2014). Ahora bien, en general las propiedades de un sistema son emergentes, si ellas no están presentes en sus componentes. En otras palabras, propiedades globales que son producidas por las interacciones locales son emergentes. La emergencia se da a una escala dada, y no puede ser descrita sobre la base de las propiedades de una escala inferior (Fernández et al., 2011). Por ejemplo, la temperatura de un gas se puede decir que es emergente (Shalizi, 2001), pues sus moléculas no poseen esta propiedad: la temperatura es una propiedad colectiva. Igualmente, en una pieza de oro emergen propiedades tales como la conductividad, maleabilidad, color, brillo; las cuales no pueden ser descritas a partir de las propiedades de los átomos de oro (Anderson, 1972). De una manera amplia e informal, la emergencia puede ser vista como las diferencias que son observados, en los sistemas, a diferentes escalas (Prokopenko et al., 2009). Se estima que algunas personas podrían percibir dificultades en describir fenómenos a diferentes escalas (Gershenson, 2013b), pero ello puede ser la consecuencia de tratar de encontrar una única "verdadera" descripción de un fenómeno. Dado que los fenómenos no dependen de las descripciones que hacemos de ellos, podemos tener múltiples y diferentes descripciones del mismo fenómeno. Como alternativa, es más informativo manejar diferentes CAPÍTULO 21 17 descripciones a la vez, lo que actualmente es más apropiado y necesario cuando se estudian sistemas complejos. 1.3.2 Auto-organización Desde la mitad del siglo pasado, la auto-organización ha sido vista como un evento de cambio de la propia organización, sin que una entidad externa influya en el proceso (Ross Ashby, 1947a). Su base está en las reglas locales que conducen al sistema a producir una estructura o comportamientos específicos (Georgé et al., 2009). Este proceso viene desde la dinámica interna del sistema, definida por las interacciones entre los elementos. Su resultado es un grado determinado de regularidad, que puede derivar en la expresión de un conjunto de patrones (Ebeling, 1993; Eriksson and Wulf, 1999). La auto- organización es un proceso por el cual se alcanza dinámicamente la función y el comportamiento global del sistema, a partir de las interacciones de sus elementos. Tal función o comportamiento no es impuesta por uno o pocos elementos, ni determinada jerárquicamente. Así, la auto-organización se adquiere autónomamente por la interacción de los elementos que producen, a su vez, retro-alimentaciones que regulan el sistema (Gershenson, 2007). Desde esta perspectiva, se puede ver la auto-organización como un proceso dinámico, adaptativo, que favorece el mantenimiento de la estructura del sistema, en el que se pueden expresar patrones de comportamiento para hacer frente a los cambios ambientales (Perozo, 2011; Perozo, Niriaska, Aguilar, José, Terán, O, Molina, 2012b). En la naturaleza, la auto-organización ha sido usada para describir los procesos básicos por los cuales se forman enjambres, cardúmenes de peces, parvadas de patos o el tráfico vehicular. En general ha sido descrita en muchos sistemas, donde las interacciones locales además de generar un orden, derivan en la manifestación de un comportamiento global (Camazine et al., 2001; Gershenson, 2007). Cabe destacar que la generación de patrones como un fenómeno emergente vía auto- organización, ha constituido un punto de discusión importante. Esto ha llevado a que la auto- organización y la emergencia hayan sido utilizadas de manera intercambiable. Sin embargo, autores como De Wolf and Holvoet (2005) han expresado que se puede dar auto-organización con y sin emergencia; así como la producción de auto-organización y emergencia al mismo tiempo; o de la emergencia sin la auto-organización. En general, casi cualquier sistema dinámico puede ser visto como auto-organizante: sí su dinámica conduce a un atractor, que puede ser visto como u estado más probable. Podemos decidir llamar ese atractor cómo "organizado". Así, los sistemas dinámicos con un atractor (o más) tenderán a él, de tal manera que el sistema incrementará, por sí mismo, su propia organización (Ross Ashby, 1947b). Sobre esta base, y de manera práctica, podríamos describir casi que cualquier sistema como auto-organizante, más que cómo auto-organizado o no 18 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- (Gershenson and Heylighen, 2003). Lo auto-organizante muestra más el rasgo dinámico, activo de la auto-organización. La conveniencia de tener una medida de auto-organización que capture la naturaleza de la dinámica local en una escala determinada, es especialmente relevante para nuestro tiempo (Ay et al., 2012; Prokopenko, 2009b). Esto es especialmente relevante en el naciente campo de la auto-organización guiada u orientada (AOG) (Ay et al., 2012; Prokopenko, 2009b). La AOG puede describir cómo direccionar la dinámica auto-organizante de un sistema hacia una configuración deseada (Gershenson, 2012a). Esta configuración deseada puede no siempre ser el atractor natural de un sistema controlado. El mecanismo para guiar la dinámica, y el diseño de tal mecanismo, se beneficiará desde la medición y caracterización de la dinámica de un sistema en una forma precisa y concisa. 1.3.3 Complejidad Hay docenas de nociones y medidas de complejidad, propuestas en diferentes áreas con diferentes propósitos (Edmonds, 1999; Lloyd, 2001). Etimológicamente, complejidad viene del latín plexus, lo cual significa entrelazado. Algo complejo es algo difícil de separar. Esto significa que los componentes del sistema son interdependientes, de manera que su futuro está parcialmente determinado por sus interacciones (Gershenson, 2013b). Por lo tanto, estudiar los componentes por separado – como se acostumbra desde el enfoque reduccionista- no es suficiente para describir la dinámica de los sistemas dinámicos complejos. Esto, además, dificulta la predictibilidad de los SD. Un ejemplo de complejidad clásico se halla en el conocido "Juego de la Vida" de Jhon Conway (Berlekamp et al., 1982), en el cual, 4 reglas simples generan dinámicas ricas como estructuras estables, móviles y/u oscilatorias, que son difíciles de predecir. Tradicionalmente se ha pensado que la complejidad se refiere a aquella condición que limita al modelista para formular el comportamiento completo del SD, en un lenguaje dado. En realidad, lo complejo se debe a las interacciones en el SD que generan información nueva y relevante-no presente en las condiciones iníciales ni de frontera- y que influyen en el desarrollo del SD. La complejidad, por tanto, puede estar asociada a los aspectos emergentes y auto-organizantes del SD. Dada la dificultad existente en la adecuada interpretación de la noción de complejidad, y más aún en su medición, tener medidas globales de complejidad es una tarea actual, de gran importancia y utilidad. Una medida útil de complejidad debe estar disponible para permitirnos responder preguntas como: ¿Es un desierto más o menos complejos que una paramo? ¿Cuál es la complejidad de diferentes brotes de influenza? ¿Cuáles organismos son más complejos: predadores o presas; parásitos u hospederos; individuos o sociedades? ¿Cuál es la complejidad de situaciones tan CAPÍTULO 21 19 disimiles cómo diferentes como los géneros musicales y los diferentes regímenes de tráfico? ¿Cuál es la complejidad requerida de una compañía para enfrentar la complejidad de un mercado? ¿Cuándo un sistema es más o menos complejo? 1.3.4 Homeostasis Originalmente, el concepto de homeostasis fue desarrollado para describir funciones internas de regulación fisiológica, tales como la temperatura o los niveles de glucosa. Probablemente, la primera persona que reconoció el mantenimiento de un ambiente interno casi constante para la vida fue Bernard (Bernard, 1859). Subsecuentemente, Cannon (Cannon, 1932) acuño el término homeostasis ( del griego hómoios (similar) y estasis (estado, estabilidad). Canon, definió la homeostasis cómo la habilidad de un organismo para mantener estados estacionarios de operación durante cambios internos y externos. Sin embargo, esto no implicaba un estado inmóvil o inactivo. Aunque algunas condiciones pueden variar, las principales propiedades de un organismo pueden ser mantenidas. Más tarde, el cibernético británico William Ross Ashby propuso, en forma alternativa, que la homeostasis implicaba una reacción adaptativa, para mantener "variables esenciales" dentro de un rango (Ross Ashby, 1960a, 1947b). Estas variables son esenciales, porque en ellas se basa gran porcentaje del funcionamiento del sistema, de forma que si fallan, el sistema lo hará igualmente. Dos conceptos relacionados con la homeostasis son los conceptos de ultra-estabilidad y adaptación homeostática (Di Paolo, 2000). En ellos se basó Ashby para explicar la generación de comportamiento y aprendizaje en máquinas y sistemas vivos. La ultra-estabilidad se refiere a la operación normal de un sistema dentro de una "zona de viabilidad" para hacer frente a los cambios ambientales. Esta zona de viabilidad está definida por los rangos entre los valores mínimos y máximos de las variables esenciales del sistema. Sí al operar, un valor de estas variables sobrepasan los límites de esta zona, el sistema cambia, se regula para encontrar nuevos parámetros que hacen estas variables retornen a su zona de viabilidad. Es allí donde nace la expresión equilibrio dinámico, pues el sistema se ajusta constantemente para estar en una zona de funcionamiento óptimo. Un SD tiene una alta capacidad homeostática para mantener su dinámica cerca a cierto estado o estados (atractores). Cuando las perturbaciones ambientales ocurren, el sistema se adapta para enfrentar los cambios dentro de su zona de viabilidad sin que el sistema pierda su integridad (Ross Ashby, 1947a). La homeostasis puede ser vista como un proceso dinámico de auto-regulación y adaptación, por el cual el sistema adapta su comportamiento en el tiempo (Williams, 2006). En este sentido, los procesos de homeostasis, emergencia y auto-organización guardan relación, 20 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- porque corresponden con una respuesta específica a las presiones ambientales, en cuyo caso se puede decir que el sistema muestra adaptación (Prokopenko et al., 2009). Es de resaltar que para a pesar de haberse descrito gran cantidad de procesos donde opera la homeostasis, la profundización en sus aspectos formales para su medición y valoración es un gran y atractivo campo de estudio (Fernández et al., 2012b, 2010a). 1.3.5 Autopoiesis La Autopoiesis (del griego auto-propio- y poiesis-creación, producción-), fue propuesto como un concepto para definir lo vivo y distinguirlo de lo no vivo. Recientemente Maturana (Maturana, 2011), en respuesta a Froese (Froese and Stewart, 2010), aclara que la noción de autopoiesis fue creada para connotar y describir los procesos moleculares que toman lugar en la realización de los seres vivos cómo entidades autónomas. El significado de la palabra fue usada, desde sus inicios, para describir redes cerradas de producción molecular (Maturana and Varela, 1980). La noción de autopoiesis surgió de una serie de preguntas relacionadas con la dinámica interna de los seres vivos, las cuales Maturana comenzó a considerar en los 60s. Entre ellas estaban: ¿Cuál debería ser la constitución de un sistema que veo cómo vivo, cómo resultado de su operación? ¿Qué pasó hace 3.800 billones de años para que podamos decir que la vida comenzó allí? En la autopoiesis, los organismos vivos ocurren cómo entidades dinámicas, moleculares, autónomas y discretas. Estas entidades están en continua realización de su auto-producción. Así, la autopoiesis describe la dinámica interna de un ser vivo en el dominio molecular. Maturana resalta que los seres vivos son sistemas dinámicos en continuo cambio. Las interacciones entre los elementos de un sistema autopoiético, regulan la producción y la regeneración de los componentes del sistema, lo que le da el potencial de desarrollarse, preservarse y producir su propia organización (Varela et al., 1974). Desde la autonomía expresada en la autopoiesis se puede tener que una bacteria puede producir otra bacteria por división celular, mientras que un virus requiere de un hospedero para producir otro virus. La producción de una nueva bacteria, se da por las interacciones entre los elementos de ella. La producción de un nuevo virus, depende de las interacciones de los elementos con un sistema externo. Así, la bacteria es más autopoiético que un virus. En este sentido, la autopoiesis está relacionada estrechamente con la autonomía (Moreno et al., 2008; Ruiz-Mirazo and Moreno, 2004). Por ejemplo, una bacteria puede producir otra bacteria por división celular, mientras que un virus requiere de un hospedero para producir otro virus. La producción de una nueva bacteria, se da por las interacciones entre los elementos de ella. La producción de un nuevo virus, depende de las interacciones de los elementos con un sistema externo. Así, la bacteria es más autopoiética por su mayor autonomía que un virus. En esta comparación es apropiado resaltar que el virus despliega las propiedades de autoreferencia y heteroreferencia al utilizar la maquinaria de las células CAPÍTULO 21 21 eucarióticas para sus fines reproductivos. Esto visiblemente desencadena una red de procesos para la producción de su material constitutivo. Desde la anterior perspectiva, la autopoiesis está más relacionada con la autonomía (Moreno et al., 2008; Ruiz-Mirazo and Moreno, 2004). Consideramos que la autonomía es una condición emergente que surge del proceso autopoiético y lo representa en gran medida. La autonomía, además, está siempre limitada en sistemas abiertos, dado que su estado dependerá de las interacciones con el ambiente. Desde esta perspectiva, se puede observar a la autopoiesis como un proceso con rasgos homeostáticos que contribuye a los procesos auto- organizantes de la estructura del sistema, y que es expresión de la autonomía del sistema. Normalmente, las diferencias en autonomía entre sistemas, cómo en el ejemplo de bacterias y virus, pueden ser claramente identificadas, lo que nos promueve a generar escalas de autonomía al evaluar la respuesta de los organismos frente a los cambios de su medio. El concepto de autopoiesis se ha extendido a otras áreas más allá de la biología (Froese and Stewart, 2010; Froese et al., 2007; Luisi, 2003; Seidl, 2004). Algunos aspectos formales de la autopoiesis fueron abordados por Varela et al. (1974), quienes proponen pruebas operacionales para determinar sistemas autopoiéticos. A pesar de la oposición de Maturana a la formalización de la autopoiesis en términos matemáticos, (Varela, 1979) y (Urrestarazu, 2011) desarrollaron algunos aspectos formales. Recientemente, Gershenson y Fernández (2011) propusieron una medida de autopoiesis sobre la base de la complejidad comparativa del sistema y su ambiente. 1.4 Síntesis Desde una perspectiva histórica, en este capítulo hemos dado una mirada a la teoría de los SD sobre una base más amplia; una que nos permite entender el porqué de su dificultad de predicción. Dificultad que subyace en las interacciones relevantes del sistema, que lo auto- organiza y que nos produce nueva información. Nueva información que emerge también, cuando el sistema afronta los cambios ambientales. Cambios que muestran la respuesta autónoma del sistema y que definen su grado de adaptabilidad. Todo ello ahora queda inmerso en una visión de SD con rasgos de complejidad. La mayoría son sistemas abiertos, donde el control completo es difícilmente alcanzable, debido a la impredictibilidad que generan las interacciones. Ante lo anterior, surge la posibilidad de adaptarnos como solución. Adicionalmente en el capítulo se presentan, las nociones, y sus características, de auto- organización (incremento del orden-regularidad), emergencia (nuevas propiedades a escala global), complejidad (interconexión), homeostasis (auto-regulación, ultra-estabilidad) y autopoiesis (auto-producción, auto-determinación), lo que sientan las bases y la motivación para su posterior formalización. 22 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 2. CAPÍTULO 2: ESPECIFICACIÓN DE SISTEMAS DINÁMICOS COMO REDES COMPLEJAS DE AGENTES Resumen El capítulo 1 consideró la dimensión de la complejidad en SD y la consideración de un número mayor de propiedades para su análisis. Desde esta base, el capítulo 2 parte de la concepción tradicional de SD, para proponer una noción con propiedades que van más allá de la emergencia y la auto-organización. Se incluyen, entonces, la homeostasis y la autopoiesis. La asociación de tales propiedades genera una interdependencia entre los aspectos funcionales y estructurales de los SD que también son abordados. Posteriormente, al considerar las interacciones entre elementos, que pueden ser descritos como agentes, se muestran los formalismos para la representación de los SD como redes computacionales de agentes. Estos formalismos son la base para el estudio de la complejidad en SD, lo que posibilita la fusión de las teorías multi-agentes y de redes. Desde este enfoque hibrido, un sistema multi-agentes es visto como una red computacional con rasgos a determinar, tales como su emergencia y su auto-organización, entre otros. Con este aporte se espera dar un mejor entendimiento de lo que sucede a partir de las interacciones entre los componentes de los SD. 2.1 Nociones de SD: de lo Tradicional a la Inclusión de lo Complejo. A continuación se presentan dos perspectivas para la definición de SD, una tradicional y otra que incluye la complejidad. Esta última está inscrita en el marco de la TDS expuesta en el ítem 1.1., y se proviene de lo expresado en Fernández, Aguilar, et al. (2010). 24 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 2.1.1 Noción Tradicional de SD Como base general se tiene que, los SD cuentan con características básicas como lo son, el cambio o evolución de su estado (q) en el tiempo, en razón a una regla de evolución (𝑅) que lo demarca. Desde una base determinista, se podría estimar que 𝑅 está dada por 𝑅(𝑞(𝑡), 𝑟) → 𝑞(𝑡 + 𝜏), donde q es un estado del conjunto de estados posibles 𝑄, y 𝒓 un parámetro de la regla. Cabe destacar que r tiene uno o más puntos críticos (𝒓𝒄), el cual o los cuales, son entendidos como el valor de r en el que se genera el cambio o transición de estado de 𝑞𝑖 a 𝑞𝑗. Por su parte, 𝑅 puede estar descrita por funciones iterativas, ecuaciones diferenciales o instrucciones iterativas, según sea su naturaleza discreta, continua o algorítmica, respectivamente. Sobre la anterior base, ha sido posible estudiar gran cantidad de SDs existentes en nuestro medio. No obstante, en ciertos SDs con gran cantidad de componentes que presentan alta conectividad, plasticidad en sus interacciones y roles muy fluidos, las reglas de su cambio y evolución pueden ser no bien entendidas por sus estudiosos. Más aún, cuando a partir de estas reglas que evolucionan de simples a complejas y con la auto-organización como método, muchos SD pueden llevar a cabo tareas que tiene un grado de dificultad considerable. Es allí donde el observador puede apreciar la dimensión emergente de su orden y comportamiento global (Aguilar et al., 2007). En las anteriores condiciones, se hace pertinente observar un SD bajo la óptica de su complejidad. Un evento funcionamiento, que requiere de mayores conceptualizaciones y profundizaciones en cuanto a sus aspectos formales, como a continuación se propone. inherente al 2.1.2 Una Definición de SD que Considera la Complejidad o Sistemas Dinámicos Complejos (SDC). La noción base de SD en la que se fundamentará este manuscrito parte de la presentada en (Fernández et al., 2010a). En ella se incluye los rasgos de complejidad, por lo que podemos definir un Sistema Dinámico Complejo (SDC) y expresa lo siguiente: …"Dado un contexto particular de observación1, un SDC corresponde con un fenómeno unitario que presenta un conjunto de componentes homogéneos o heterogéneos 1 Corresponde a los aspectos propios del observador que, influyen en el proceso de modelación y el consecuente modelo. Estos aspectos son: (i) su cultura, entendida como el conjunto de significados a partir de los cuales el sujeto comprende el mundo y (ii) el lingüístico, como aquel lenguaje especializado que se construye sobre la base de la cultura del sujeto. A partir de ellos el observador determi na el sistema unitario a partir de las relaciones de retro-alimentación y de recursividad circular entre la dupla sistema-entorno. Es de anotar que el entorno se refiere, por ejemplo, a otros miembros de la comunidad de conocimientos, modelos de simulación, teorías, o al mundo físico y social. Mayores ampliaciones de estos elementos pueden hallarse en Fernández et al (2010). CAPÍTULO 22 25 interconectados. Estos componentes, pueden observarse a diversas escalas espacio- temporales y generan, a partir de sus interacciones, diferentes tipos de propiedades, organizaciones, patrones y/o comportamientos propios. Todo ello, con el fin de alcanzar el propósito global que el sistema tiene en sí mismo y sobre la base de su complejidad"… Un análisis de mayor profundidad de la anterior noción, lleva a considerar la existencia de características y elementos importantes como:  La interconexión de los elementos define la topología de sus interacciones, que de acuerdo con el caso y eventos de reforzamiento o debilitamiento, conducirán o no al establecimiento de posteriores interrelaciones. Las interacciones, en este sentido, las entendemos como contactos de corta duración y conexión; mientras que las interrelaciones las entendemos como fenómenos de asociación o combinación de elementos de naturaleza más perdurable. La dinámica de las interconexiones es uno de los elementos básicos de la complejidad del SDC, pues es su dinámica la que conduce a cambios en la organización propia no impuesta. Al mismo tiempo, las interconexiones entre las partes están acordes con los cambios ambientales y las propias posibilidades de los elementos para adaptarse a ellas.  Las adaptaciones en el SDC se dan de manera individual y colectiva; en el tiempo y en el espacio, y sobre la base de los rangos de adaptabilidad de las partes. Desde esta visión, el SDC se auto-organiza a partir de la topología de las interacciones e interrelaciones entre los elementos que constituyen una estructura determinada. Los elementos tienen como rasgo, una potencialidad desconocida de establecer relaciones entre sí. Ellas se establecen sobre la base de su adaptabilidad; otra razón más de la complejidad del SDC.  Las escalas espacio-temporales en las que se ocurren los elementos, corresponden a la ventana de observación de espacio y/o tiempo, donde se definen las características de los elementos y los parámetros del entorno que condicionan sus interacciones. En estas escalas es que operan las variables y procesos que explican los patrones observados en determinado nivel. Las escalas espacio-temporales son un rasgo básico de dinamismo y complejidad sistémica, pues requieren del análisis de su cambio, variabilidad y heterogeneidad, para la explicación y posible predicción, o no, de los patrones observados. Ejemplo de este tipo de escala en SDCs son el nivel de observación local de las interacciones, y el nivel superior o global en el que son observables las propiedades emergentes. No obstante, pueden existir otros niveles intermedios en el SDC. Por otra parte, las relaciones entre niveles se da también por mecanismos de retro-alimentación, en el sentido local-global y global-local. De esta última forma, las propiedades emergentes tienen también influencias en las interacciones del nivel local. 26 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴-  La generación de diferentes tipos de organizaciones, patrones y comportamientos propios a partir de las interacciones locales, en los diferentes niveles del SDC a escalas espacio-temporales definidas, determina claramente el carácter emergente del SDC. Desde esta perspectiva, se visualiza que determinado fenómeno considerado como emergente puede ser, además, producto de la auto- organización. Igualmente, los patrones (Pt) definen la unidad global del SDC, pues son producto de cualidades sintetizadoras que representan, en parte, la emergencia del SD.  La propia organización (auto-organización) incluye, a su vez, la capacidad del SDC para regular su comportamiento en términos adaptativos, en busca de un equilibrio dinámico u homeostasis. Un hecho importante, debido a que el flujo de intercambios entre los elementos y el entorno puede llevar a la falta de regularidad en el SDC. La homeostasis abarca, articula e integra estados del SDC que presentan diferentes condiciones complementarias como las dinámicas-estacionarias, las de estabilidad-inestabilidad, las de síntesis-degradación, entre otras. La funcionalidad de la homeostasis es mantener la auto-organización, ante el constante flujo de materia, energía y/o información en el SDC, en límites que posibiliten el mantenimiento de su autonomía, integridad, compatibilidad y demás atributos esenciales (Zona de Viabilidad del SDC). Todo este fenómeno se da en el marco de la adaptabilidad del SDC.  Consecuente con lo anterior, se estima que un mecanismo homeostático de importancia en un SDC complejo es la autopoiesis; un proceso en el que a partir de las interacciones entre los componentes del SDC logra el auto-mantenimiento, en el contexto de la relación síntesis-degradación del SDC. En esta propiedad se refleja parte de la autonomía del SDC. La autopoiesis favorece la auto-organización del SDC, por medio de mecanismos de auto-regulación que mantienen la propia estructura del SDC, en consonancia con un entorno variable. Los elementos develados de emergencia, auto-organización, homeostasis y autopoiesis en la noción propuesta, son un medio por el cual el SDC logra persistir espacial y/o temporalmente situación que no es otra cosa que su propósito esencial. Al mismo tiempo le confieren una complejidad determinada, de acuerdo con el tipo de SDC. Se puede estimar que la persistencia de un SDC en un medio cambiante se da en términos de la conservación, adaptabilidad y evolución de sus elementos, y del SDC como un todo. Estas CAPÍTULO 22 27 últimas propiedades vienen de las interacciones relevantes2 que generan la auto-organización de la estructura, la emergencia de comportamientos, y en definitiva, la complejidad del SDC. Complejidad que le confiere al SDC su robustez (Conservación) ante la variación de su ambiente, pero también la posibilidad de cambiar de forma adaptativa y evolutiva ante condiciones más drásticas. 2.1.3 Interdependencia Funcional de la Auto-organización, Homeostasis y Autopoiesis, y la Generación del Espacio Estructural del SDC. Normalmente, se ha estudiado cómo la estructura de un sistema determina su función. Sin embargo, en muchos sistemas complejos hay una retroalimentación donde la funcionalidad genera ciertas propiedades en el espacio estructural de un SDC. La interdependencia funcional de auto-organización, homeostasis y autopoiesis, se puede observar en la Figura 2- 1. En ella se observa como la auto-organización y la autopoiesis, sobre la base de mecanismos homeostáticos de auto-regulación, son las que desde el espacio funcional generan el SDC en el espacio estructural. Desde este punto de partida se puede estimar que la relación entre el espacio funcional y estructural se hace cíclica, por cuanto la estructuras también determinan la función. Figura 2-1 Relación entre los Espacios Funcionales y Estructurales en SDC (Fernández et al., 2012a) 2 Son relevantes porque producen información adicional que no estaba en las condiciones iniciales o de frontera. 28 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Entre los aspectos de mayor importancia en el espacio estructural, se hallan la estructura y los patrones. La estructura estará referida a los elementos y el tipo de relación entre ellos (𝐸𝑠𝑡𝑟𝑢𝑐𝑡𝑢𝑟𝑎 = {𝑒𝑙𝑒𝑚𝑒𝑛𝑡𝑜𝑠, 𝑟𝑒𝑙𝑎𝑐𝑖ó𝑛}). La estructura puede emerger como resultado de: (i) una acción auto-organizante que define el tipo de relación entre los elementos y (ii) por procesos autopoiéticos, de manera que se procura así misma, en términos de autonomía. Al mismo tiempo, la estructura expresa en patrones, la mayoría de los casos definibles u observables, de manera directa o indirecta (Fernández et al., 2011). Históricamente, la estructura ha tenido un papel preponderante en el estudio y explicación de los SDC, dada su condicional aparentemente "palpable". No obstante, vale la pena realizar algunas precisiones al respecto. La primera es que la estructura, en muchas ocasiones, no aparece de manera explícita, sino que por su carácter emergente, resulta ser una condición a determinar. En este aspecto, suele suceder que lo que el observador percibe del SDC son algunos patrones de orden global. Los patrones se generan a partir de procesos de retroalimentación, entre otros, y sobre la base de la adaptabilidad del SDC. Descifrar la estructura en un SDC requiere, por tanto, determinar y caracterizar los elementos y sus relaciones, a través de indicadores; así como de la aplicación de modelos matemáticos, que pueden dar razón sobre su condición jerárquica, holárquica o heterárquica, por ejemplo. La condición deseable de llegar a estructuras "más estables" en el SDC, para periodos determinados de tiempo, será el producto del alcance de ciertos objetivos globales, como resultado del dinamismo del proceso auto-organizante. La estabilidad de una estructura dependerá, básicamente, del mantenimiento de la misma en una zona de viabilidad (𝑍𝑣), es decir, de la condición homeostática que le proporciona la auto-organización. No obstante, la tendencia auto-organizante de mantener tal estabilidad puede conllevar a que, en caso de la superación de los umbrales máximos y mínimos en una determinada 𝑍𝑣, se dé un cambio importante en la estructura, como respuesta evolutiva del SDC. De allí se generará un nuevo estado inicial para la estructura, que buscará llegar nuevamente a su estado estable, y su desempeño se dará en una nueva 𝑍𝑣.A este respecto vale la pena tener presente la consideración de (Kaufmann, 1993), quien asevera que los cambios de estado dentro de la auto-organización, al no ser graduales y constantes, se dan por la superación de umbrales de complejidad que llevan a nuevas formas estables. Lo expuesto anteriormente toma sentido al considerar que, hipotéticamente en un SDC, pueden existir tantas estructuras como tantas posibles combinaciones de elementos e interacciones puedan haber. En términos prácticos, esta situación puede ser observada como la variación dinámica en una intrincada red de elementos que se relacionan a partir de sus interacciones. En consecuencia, se puede estimar que un SDC puede ser descrito, en términos matemáticos, sobre la base de un grafo o red dinámica, cuyos aspectos formales se explican a continuación (Fernández et al., 2011). CAPÍTULO 22 29 2.2 Los SDC cómo Redes Computacionales de Agentes. Los siguientes apartados se basan en lo expresado en (Fernández et al., 2012a, 2012b, 2011, 2010b). Como se ha venido observando, la consolidación del SDC como un todo se puede establecer a través de las interacciones relevantes entre sus elementos que dan como resultado propiedades cómo la auto-organización, emergencia, homeostasis y autopoiesis. En este aspecto, los SDC pueden ser especificados cómo un grafo o red de interacciones, de manera que se puede representar en una estructura matemática formal que representa a los elementos y sus relaciones. Las redes, particularmente, se han convertido en una herramienta central en el estudio de sistemas complejos (Gershenson and Prokopenko, 2011), donde los nodos son considerados como "agentes". La definición básica de agente que adoptamos corresponde con la de Gershenson (2009), referida a una entidad que está situada en un entorno y que puede afectarse o afectar lo que en él sucede. De acuerdo con su tipo, un agente puede generar un conocimiento apropiado del sistema por la creación de una representación interna de los componentes del sistema externo. Cuando los agentes interactúan dinámicamente para alcanzar una función o comportamiento global, el sistema que los contiene puede ser descrito cómo auto- organizante (Gershenson, 2006). Estas definición general se sintoniza con lo requerido para la construcción de sistemas multi-agentes de acuerdo con lo expresado por Wooldridge and Jennings (1995). Tanto la red en su globalidad, como en sus componentes locales, vistos cómo agentes, pueden ser especificados como SDs tradicionales. Es decir con cambios y/o evolución de estados en el tiempo, a partir de reglas que pueden ser expresadas funcionalmente. Con una red computacional de agentes, se pueden lograr descripciones de SDC más adecuadas, y se tiene la ventaja de poder analizar la información nueva producida por las interacciones entre agentes. Para la especificación de una red computacional de agentes, se requiere integrar diferentes aspectos de la teoría de los SDC, grafos y agentes, como se describirá en el siguiente apartado. 2.2.1 Aportes de la Teoría de SDs La teoría de modelado y simulación (Zeigler et al. 2000), permite considerar los elementos básicos para la especificación de la red computacional de agentes (SDC -Ag), como un sistema 30 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- dinámico tradicional (SD). Para ello se tendrá en cuenta la siguiente estructura (ver Ecuación 2.1): 𝑆𝐷 = (𝑋, 𝑌, Ω, 𝑄, 𝐹, 𝐹, 𝑇) (2.1) Donde: 𝑆𝐷: Sistema Dinámico 𝑋: Conjunto de Entradas. 𝑌: Conjunto de Salidas. Ω: Conjunto total de entradas posibles. 𝑄: Conjunto de Estados. 𝐹: (𝑄 × Ω −> 𝑄): Función de Transición entre estados, que actualiza el estado. 𝐹𝑠: (𝑄 × Ω −> 𝑌): Función de Salida. 𝑇: Posibles valores para el tiempo. En esta estructura, 𝑋, Ω y 𝑌 definen las entradas y salidas del SD, mientras que 𝑄 y 𝐹 representan la estructura interna y el comportamiento. La función de transición de estados 𝐹 debe garantizar que siempre se puede determinar el próximo estado del SD, a partir del estado actual. La representación gráfica de los elementos básicos en una red SDC -Ag, desde la perspectiva de un SD acoplado como lo propone Zeigler et al. (2000), se observa en la figura 2-2. Figura 2-2 Representación Simplificada de un SDC -Ag sobre la Base de la Teoría de Modelado y Simulación de Zeigler et al. (2000). Las designaciones A1, A2, A3 se refieren a Agentes. CAPÍTULO 22 31 2.2.2 Aportes de la Teoría de agentes y de los Sistemas Multi- agentes-SMA La teoría de agentes (Desalles et al., 2008; Ferber and Gutknecht, 1998; Ferber et al., 2004; Jennings and Wooldridge, 1995; Müller and Et Al., 1997; Wooldridge et al., 1996), brinda elementos de interés que han permito establecer una de las principales áreas de la inteligencia artificial distribuida, como lo son los SMA. Desde una base general, un agente puede describirse como una entidad que actúa sobre su ambiente o entorno (Gershenson, 2007). Sobre esta base, y en términos de un SDC, se estima que los Ag son entidades que perciben su ambiente (entradas) y actúan (cambio de estado) en concordancia, deliberando sobre las posibilidades de sus acciones (salidas). Igualmente, se ha considerado que los Ags pueden tener rasgos inspirados en el comportamiento humano como: autonomía, reactividad, iniciativa, habilidad social. Estos rasgos le permiten al Ag compartir el entorno o ambiente, recursos y conocimiento, además de comunicarse y coordinar sus actividades con otros (Aguilar, 2014; Aguilar et al., 2012). Desde la perspectiva de la teoría de Ags los SMA, como sistemas compuestos por diversos Ags, presentan características como la modularidad, redundancia, descentralización, comportamiento emergente, y especialmente, coordinación. Esta última, basada en la compartimentalización de información y resultados, la optimización de acciones y recursos, y el trabajo conjunto independiente de los objetivos individuales. Los mecanismos propios para lograr una coordinación efectiva se sustentan en la sincronización de acciones, vista como la simultaneidad de las mismas; en la planificación, es decir en la subdivisión y distribución de tareas; y en la coordinación reactiva, a partir de la definición de marcas o reglas de comportamiento reactivo general. Bases como éstas, permiten el desarrollo de SMA auto-organizantes emergentes cuyo sustento son mecanismos que consideran las interacciones directas, la estigmergia, el refuerzo, la cooperación y la definición de arquitecturas genéricas (Perozo et al., 2008). Operacionalmente, a cada Ag le puede ser asignado uno o más roles para la consecución de los objetivos del SDC, y ser responsable tanto de coordinar sus interacciones con otros Ags, como de ejecutar las actividades bajo su cargo. En este aspecto, un Ag inteligente es un "tomador de decisiones interactivo", capaz de alcanzar de manera pro-activa sus objetivos, mientras que se adapta a un ambiente dinámico. De allí que los SMA son vistos como sistemas computacionales auto-organizantes con un grado apropiado de autonomía, dada la inexistencia de un control central (Aguilar, 2014). 32 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 2.2.3 Aportes de las Teorías de Grafos y Redes La teoría de redes se remonta al siglo XVIII en la antigua ciudad alemana Königsberg (hoy Kaliningrado, ciudad rusa), donde se planteó un interesante problema matemático, hoy conocido como el de "los puentes de Königsberg". El planteamiento de éste problema consistía en hallar la forma de cruzar los 6 puentes, en un solo sentido, y sin repetir el tránsito por alguno de ellos. El problema fue finalmente resuelto en 1736 por el suizo Leonhard Paul Euler (Kruja et al., 2002), a través de la representación de las rutas posibles como enlaces o ligas, y los puntos de conexión entre ellas como nodos. Este hecho es frecuentemente citado como el nacimiento de la teoría de grafos, actualmente una de las formas más comunes de representación de las redes (Hedrih, 2007). Sí bien el modelado SMA ha constituido una herramienta computacional adecuada para lograr explicaciones o predicciones sobre fenómenos específicos en SDC, existe dificultad en extraer algunos parámetros estadísticos que den razón de su comportamiento global. La teoría de redes, como extensión de la teoría de grafos, cuenta con las posibilidades que no tienen los SMA, tales como la representación del flujo de información y las interacciones topológica entre agentes-Ag. Entre los parámetros topológicos de interés se hallan el grado (Gd en ecuación 2.2) y el coeficiente de agrupación (Ca en ecuación 2), las cuales pueden ser determinados considerando su carácter dinámico (Habiba and Berger-Wolf, n.d.). Gd se entiende como el cambio de vecindad 𝑉(𝑖𝑡) de un nodo 𝑁𝑖 en el tiempo 𝑡. Ca de un nodo 𝑖 es la fracción de sus vecinos, quienes han sido vecinos entre ellos mismos en cualquier tiempo previo 𝑡 (ecuación 2-3). 𝐺𝑑𝑇(𝑖) = ∑ 1<𝑡<𝑇 𝑉(𝑖𝑡−1)−𝑉(𝑖) 𝑉(𝑖𝑡−1)∪𝑉(𝑖𝑡) × 𝑉(𝑖𝑡) (2.2) 𝐶𝑎𝑇(𝑖) = ∑ 𝑖∈𝑉(𝑖𝑇) ⋃ 1<𝑡<𝑇 (2.3) ∩𝑉(𝑖𝑇) 𝑉(𝑖𝑡) 𝑉(𝑖𝑇) La importancia de estos parámetros se halla en que, a partir de ellos, se podrían determinar patrones resultantes, o no, de procesos de auto-organización en la red SDC. Adicionalmente, existe la posibilidad de determinar estados subyacente para la red SDC, que definen su "estado de equilibrio" como resultado de los procesos homeostáticos. Esto se puede lograr a través de la aplicación de algoritmos como el de Kamada-Kawai (Kamada and Kawai, 1989). Dado que desde la teoría de redes, los nodos pueden verse como estructuras excesivamente simplificadas, la hibridación de redes con nodos tipo Ag puede ser una concepción de gran importancia. De esta manera, se hace posible ver un SMA como un conjunto de nodos y aristas interactuantes que conforman una red computacional, es decir una red SDC-Ag. CAPÍTULO 22 33 En términos generales, una red SDC-Ag, se sintoniza con la estructura definida por (Gershenson, 2010) para una red computacional 𝐶(𝑁, 𝐾, 𝑎, 𝑓), que la define como un conjunto de nodos 𝑁, ligados por un conjunto de aristas 𝐾, usadas por un algoritmo a para computar una función 𝑓. A nivel operativo, 𝑁 y 𝐾 tienen variables internas que determinan su estado, y funciones 𝑓 que determinan sus cambios de estado, respectivamente. Se resalta que, esta generalización ha resultado de gran utilidad para comparar las arquitecturas de redes neuronales, colonias de hormiga y partículas colectivas; al tiempo que, puede ser un punto de partida y alternativa interesante para la representación, igualmente, genérica, de un SMA. No obstante, para describir completamente la dinámica de un SMA como red computacional, se requiere ahondar en los detalles de cada componente de la red y sobre su dinámica. Por ejemplo, el flujo de información en una red da como resultado la evolución de los cambios de estado en los nodos (Smith et al., 2008). El comportamiento global de esta dinámica depende de varios factores como el tipo de nodo, su modelo cognitivo, sus funciones de transición interna, el tipo de información que comparte, entre otros. Desde esta perspectiva, se haría posible descifrar la trama de interacciones entre Ags, y profundizar en la forma autónoma como se logran objetivos globales a partir de reglas locales, que podrían evolucionar de simples a complejas (Aguilar et al., 2007). 2.2.4 Aspectos Formales de la Red SDC-Ag. A continuación se muestran los desarrollos logrados hasta el momento, que serán la base para llevar, en el corto y mediano plazo, al plano del modelado los elementos conceptuales hasta ahora expuestos.  La forma general de la red SDC-Ag La red SDC-Ag se define, básicamente, como un grafo (𝑁, 𝐾). Cada nodo 𝒊  𝑁 representa un elemento 𝑁𝑖, y cada arista (𝑖, 𝑗)  𝐾 representa flujo de información entre 𝑁𝑖, 𝑁𝑗. Los vecinos son elementos (𝑁𝑖, 𝑁𝑗) unidos por un enlace de información (𝑖, 𝑗)  𝐾. El grafo (𝑁, 𝐾) determina la estructura del SDC-Ag. La conectividad de la red SDC se puede representar en un nivel básico por una matriz de adyacencia 𝑁 × 𝑁: 𝑀(𝑖, 𝑗) = { 1, 𝑖 → 𝑗 0, 𝑖 ↛ 𝑗 . Se dice que 𝑀 contiene la topología de la red, y puede contener el "peso" del acoplamiento 𝑖, 𝑗, tal que 𝑀𝑖𝑗 ∈ [0,1]. No obstante, en una formalización más general, cada elemento de la matriz 𝑀 puede ser más complejo, conteniendo vectores, etiquetas, y/o funciones. 34 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴-  El sentido dinámico de la red SDC-Ag Se puede estimar que la dinámica de una red SDC-Ag estará dada por 𝑆𝐷 = (𝑄𝑛 𝐹); Donde 𝑄𝑛 es un conjunto no vacío, compuesto en el caso de sistemas discretos de un número finito de estados 𝑞 𝑄𝑛 . 𝑄𝑛 también es llamada espacio de estados, o espacio de todas las configuraciones posibles. El exponente 𝑛 ℕ hace referencia al número de total de elementos en la red SDC. En el caso de sistemas continuos, la dinámica se desenvuelve en un espacio de estados. Por su parte, 𝐹 es una función de transición, o de activación global, tal que 𝐹: 𝑄𝑛 → 𝑄𝑛. Igualmente, 𝐹 corresponde a una familia de funciones, denominadas funciones de activación local de la red (𝑓𝑖), tal que 𝐹 = (𝑓1, 𝑓2, … , 𝑓𝑛). Igualmente, 𝑓𝑖: 𝑞𝑛 → 𝑞𝑛 dado que 𝐹(𝑞) = (𝑓1(𝑞), … , 𝑓𝑛(𝑞)), ∀𝑞 ∈ 𝑄𝑛. 𝐹 puede determinar a un sistema compuesto por funciones puramente locales. Se destaca que cada estado 𝑞𝑄𝑛 hace parte de una configuración global de la red (𝐶), entendida como la descripción completa de la situación en que se encuentra el SDC en un momento dado. Así, 𝐶 estará representada por su estado actual 𝑞𝑡  𝑄, y un evento 𝑒 que genera la transición 𝐹, donde 𝑒 ∈ ∑. ∑ , por su parte, será el conjunto de todos los eventos que generan transición. Entonces, a partir de un evento 𝑒 se dará una 𝐹(𝑞(𝑡), 𝑒), que producirá la configuración global en el tiempo siguiente 𝑡 + 1, (𝑞(𝑡 + 1)). Dependiendo del caso, el evento 𝑒 puede corresponder con el punto crítico (rc) del parámetro r mencionado anteriormente. Consecuente con lo expuesto, para pasar de una configuración 𝑐𝑖 a 𝑐𝑗, cada Ni tendrá su función de transición 𝑓𝑖, tal que 𝐹(𝑞𝑖 → 𝑞𝑖 (𝑡 + 1)) = 𝑓1(𝑞1𝑖), … , 𝑓𝑛(𝑞(𝑡 + 1)) = 𝑞𝑖 … 𝑞𝑖(𝑡 + 1). Para una configuración global 𝑐  𝐶 en el tiempo t de la red (𝑞(𝑡)), 𝐹(𝑞(𝑡, 𝑒)) dará la configuración global siguiente en el tiempo 𝑡 + 1, (𝑐(𝑡 + 1)). Un sistema donde todo estado tiene un sólo sucesor es determinista, mientras que si hay más de un sucesor posible para por lo menos un estado, habrá cierto grado de no determinismo.  El vecindario y el entorno en la red SDC-Ag. En cuanto a 𝑁𝑖, se tiene para cada uno de ellos un 𝑞𝑖 determinado en el tiempo 𝑡, es decir 𝑞𝑖(𝑡). Por poseer 𝑁𝑖 un conjunto de vecinos 𝑉𝑖, la evolución del SDC estará descrito por la regla 𝑞𝑖(𝑡 + 1) = F(𝑞𝑖(𝑡); 𝑉𝑖) , donde 𝑖 = 1, … , 𝑛. Esto es debido a la influencia que pueden tener los nodos vecinos entre sí. De forma complementaria, el estado de la red SDC en un tiempo 𝑡, puede ser representado por un vector de estado, tal que 𝑞⃗(𝑡) = ( 𝑞1(𝑡), … , 𝑞𝑛(𝑡)). Así, la forma vectorial de la red dinámica sería 𝑞⃗(𝑡 + 1) = F⃗⃗(𝑞⃗(𝑡)) + 𝑉𝑞 ⃗⃗⃗⃗(𝑡) + 𝐴𝑏𝑞 ⃗⃗⃗⃗(𝑡)), donde 𝑉𝑞 ⃗⃗⃗⃗(𝑡) es el estado de los vecinos y 𝐴𝑏𝑞 ⃗⃗⃗⃗⃗⃗(𝑡) corresponde al ambiente o entorno. CAPÍTULO 22 35 Se aclara que los componentes de 𝐴𝑏𝑞 ⃗⃗⃗⃗(𝑡), pueden ser exógenos o endógenos a los Ni. Como también, pueden darse de manera conjunta o separada. Al mismo tiempo, pueden tener efecto sobre la totalidad de los 𝑁𝑖 (incidencia global) o sobre 𝑁𝑖 determinados (incidencia parcial). La intensidad relativa de la incidencia, o la probabilidad de interacción, de 𝐴𝑏𝑞 ⃗⃗⃗⃗(𝑡), de manera totalizada, sobre determinado 𝑁𝑖, estaría dada por un parámetro 𝐵 ∈ [0,1] (González-Avella et al., 2007). Se puede asumir que la influencia de 𝐵 sobre los 𝑁𝑖i será uniforme, más no así la respuesta que los 𝑁𝑖 den al mismo. Esto será relativo a la capacidad homeostática de cada 𝑁𝑖 para responder a influencias bajas, intermedias o altas de 𝐴𝑏𝑞 ⃗⃗⃗⃗(𝑡). 2.2.5 La Especificación de Nodo Tipo Ag y SMA Desde la perspectiva de (Zeigler et al. 2000) se puede afirmar que un nodo 𝑁𝑖, cuando representa a un proceso tendrá reglas de comportamiento internas y externas. Las primeras estarán definidas por sus variables de estado, funciones de transición, y funciones de avance en el tiempo; y las segundas estarán dadas por el conjunto de salidas. De la misma manera, se debe caracterizar la naturaleza discreta o continua de los nodos, lo que definirá su forma particular de especificación. Por ejemplo, en el caso de sistemas naturales como los socio-ecológicos, representados como una red de agentes, los procesos como la contaminación o el cambio climático pueden ser representados como nodos. La dinámica de estos nodos puede ser descrita a través de sistemas de ecuaciones diferenciales de primer orden (SEDO), dada su naturaleza continua. Para el caso de procesos, como los de manufactura industrial, como los que se dan en una fábrica de ensamblado, cada operación puede ser representada como un nodo. No obstante, la dinámica será descrita a través de un sistema de eventos discretos SED (DEVS en inglés), o de ecuaciones en diferencia (EED). Nosotros consideraremos a las EED dentro de las SED, ya que las EED están embebidas en los SED según (Zeigler et al. 2000). Ahora bien, en este modelo también puede haber nodos tipos agentes (NAg) que tienen una estructura dinámica dada por la tupla 2.4: 𝑁𝐴𝑔 =< 𝑋, 𝑌, 𝑄𝐴𝑔, 𝑓𝐴𝑔−𝑖𝑛𝑡, 𝑓𝐴𝑔−𝑒𝑥𝑡,𝑓𝑠, 𝑡𝑎, 𝑊, 𝛤,∧, 𝑂𝑝 > (2.4) Donde :  𝑁𝐴𝑔: Nodo tipo Ag  𝑋: conjunto de eventos externos de entrada al Ag. Pueden corresponder a información proveniente desde el Ab (𝐼𝐴𝑏) o desde otros Ag (𝐼𝐴𝑔). En algunos casos, el ambiente puede ser modelado también como un Ag.  𝑄𝐴𝑔: conjunto de estados del Agi : 𝑄𝐴𝑔 ∈ 𝑄𝑛  𝑌: conjunto de eventos de salida desde el Ag 36 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴-     transición función de 𝑓𝐴𝑔−𝑖𝑛𝑡: 𝑄𝐴𝑔 → 𝑄𝐴𝑔 función de transición interna por la que el Ag cambia de estado. Determina el tiempo de vida de un estado 𝑞. 𝑓𝐴𝑔−𝑒𝑥𝑡: 𝑄𝐴𝑔 × 𝑋 → 𝑄𝐴𝑔 externa, donde: 𝑄𝐴𝑔 = {(𝑞, 𝑡𝑞 𝑄𝐴𝑔, 0 𝑡 𝑡𝑎(𝑞) } es el conjunto total de estados, 𝑡 es el tiempo transcurrido desde la última transición. Especifica la respuesta del Ag a los eventos de entrada. 𝑓𝑠 = 𝑄𝐴𝑔 → 𝑌: función de salida. Es una aplicación de 𝑄𝐴𝑔 dentro del conjunto de eventos de salida 𝑌. Se activa cuando el tiempo que se ha permanecido en un estado dado es igual a su tiempo de vida. Consecuentemente 𝑌 es definida solo por estados activos, p.e. ∀𝑞 ⋮ 𝑡𝑎(𝑞) ≠ . 𝑡𝑎: 𝑄𝐴𝑔 → ℜ + 0,  o función de avance en el tiempo, donde ℜ + 0,  es el conjunto de valores positivos reales entre 0 y .  𝐶𝐴𝑔: 𝑓𝐴𝑔−𝑖𝑛𝑡 ∪ 𝑓𝐴𝑔−𝑒𝑥𝑡 comportamiento del Ag  𝑊 es el conjunto de los estados del ambiente que rodea a los Ags, representado por las variables de estado del SDC (𝑊 = {𝑤𝑖, … , 𝑤𝑛}, donde 𝑤: es variable de estado). Estas variables pueden ser útiles en la computación de las funciones de transición.  𝛤 es el conjunto de influencias del Ag para generar nuevos eventos. Corresponden con acciones propias con las que el Ag intenta modificar el curso de los eventos que ocurrirán.  𝛬 descripción del sistema a través de las leyes o reglas de cambio del SDC.  𝑂𝑝 es un conjunto de operadores del Ag para generar influencias. En razón que la red SDC puede contener más de un 𝑁𝐴𝑔, se conformará entonces un SMA. En este aspecto, un SMA se definirá como un sistema informático compuesto por un grupo de Ags que interactúan entre sí. La interacción dará como resultado respuestas que irán más allá de las acciones individuales de cada 𝑁𝐴𝑔. Desde la teoría multi-agentes de Ferber y Müller (1996) y su extensión por Dávila et al (2005), se puede estimar que la dinámica del Ab estará dada por una tupla de la forma < 𝑊, 𝛤, 𝑂𝑝, 𝐴𝑐𝑐𝑖ó𝑛, 𝑅𝑒𝑎𝑐𝑐𝑖ó𝑛 >. Donde Acción, será el intento del Ag por influenciar a la red SDC-Ag tal que: 𝑂𝑝 × 𝑊 × 𝛤 → 𝛤; y Reacción del Ab ante esas influencias será 𝛬 × 𝑊 × 𝛤 → 𝑊. Al mismo tiempo, se tendrá que la evolución del SDC-Ag será una función infinita recursiva tal que 𝐸𝑣𝑜𝑙𝑢𝑐𝑖ó𝑛: 𝑊 × 𝛤 → 𝜏, donde τ será una función que no retorna valor o que retorna valores en un dominio de errores. Cabe resaltar que en la formalización propuesta, 𝛬, 𝑊, 𝛤 y 𝑂𝑝 no se explicitan en cuanto a sus componentes, dado que dependerían del SDC-Ag a modelar. A partir de lo anterior, el SMA será modelado como un conjunto de Ags que interacionan entre ellos y su Ab. La función global del SDC-Ag permitirá relacionar el cambio en las conexiones de los Ags a partir de las funciones locales en los Ags. CAPÍTULO 22 2.3 Síntesis 37 Hemos revisado la descripción y especificación de un sistema dinámico sobre la base de un conjunto mayor de propiedades, más allá de la auto-organización y emergencia. La definición obtenida incluye las propiedades de complejidad requeridos para descifrar cómo un comportamiento global y coherente, puede surgir desde las actividades locales. Con estas propiedades, es posible describir un gran conjunto de sistemas abstraídos del cosmos, de la biosfera, así como sistemas construidos por nosotros. Los elementos conceptuales y formales desarrollados sugieren un enfoque ampliado para el estudio de los SDC, y de sus propiedades emergentes de auto-organización, homeostasis y autopoiesis. Estas propiedades son producto de las interacciones locales que se expresan como una conducta global y sintética observable de todo el sistema. Como respuesta a la tarea de trasladar a aspectos formales las características emergentes mencionadas, se tiene que la teoría de grafos, y su instanciación en la teoría de redes, brindan un soporte adecuado de interpretación. A partir de ellas, los nodos pueden ser observados como agentes que pueden calcular funciones locales, las aristas pueden describir las interacciones, y una función global puede regular las interacciones entre agentes para alcanzar un estado global. El estado global, consecuentemente, se alcanzaría con la auto- organización como método, la cual se soportará en mecanismos de auto-regulación y auto- mantenimiento. De igual manera, la emergencia del sistema tendrá como base la cuantificación de la auto- organización y los patrones, situación que unifica muchas de las visiones hasta ahora analizadas en el campo de la complejidad. En este sentido, el enfoque y las formalizaciones aquí propuestas, enriquecen las aproximaciones hasta ahora desarrolladas para el estudio de los sistemas emergentes como redes computacionales de agentes. Trabajos futuros plantean la necesidad de ejemplificar los desarrollos aquí logrados en diversos casos de estudio, de manera a clarificar su utilidad. Al mismo tiempo, dada la complejidad del modelado, será apropiado desarrollar una herramienta computacional que facilite su uso en dichos casos de estudios. 3. CAPITULO 3: FORMALISMOS MATEMÁTICOS PARA LA MEDICIÓN DE LA EMERGENCIA, AUTO- ORGANIZACIÓN, COMPLEJIDAD, HOMEOSTASIS Y AUTOPOIESIS EN SISTEMAS DINÁMICOS. Resumen Habiendo definido un SDC con rasgos de complejidad (capitulo 1) y habiéndolo especificado como una red computacional de agentes (capitulo 2), surge la necesidad de formalizar las definiciones relativas a la emergencia, auto-organización, complejidad, homeostasis y autopoiesis. El objetivo del establecimiento de estos formalismos, es permitir la valoración de estas propiedades. El fundamento para todos los formalismos fue la teoría de la información, la cual permitió generar definiciones claras, sencillas y consistentes, que pueden relacionar los procesos relativos a la complejidad en SDC. Estos formalismos son de utilidad para descripciones, a diferentes escalas, en el mundo físico y en diversos campos como la biología, la ecología, la sociología, entre muchos otros, como se verá en los capítulos 4 y 5. 3.1 Teoría de la Información cómo Fundamento Creada por Claude Shannon (Shannon, 1948), en el contexto de las telecomunicaciones, la Teoría de la Información analizó la posibilidad de reconstruir los datos trasmitidos, a través de un canal con ruido. En este modelo, la información es representada por una cadena 𝑋 = 𝑥0, 𝑥1 … 𝑥𝑛 donde cada 𝑥𝑖 es un símbolo desde un conjunto finito de símbolos 𝐴, llamado alfabeto. Cada símbolo en el alfabeto 𝐴, tiene una probabilidad de ocurrencia 𝑃(𝑥). En 40 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- consecuencia, símbolos que sean muy comunes tendrán alta 𝑃(𝑥), mientras que símbolos con baja frecuencia de aparición tendrán baja 𝑃(𝑥). Shannon estuvo interesado en hallar una función para medir que tanta información era producida por un proceso. Citándolo textualmente Shannon (1948)3, argumentó: …Suponga que tenemos un conjunto de posibles eventos cuyas probabilidades de ocurrencia son 𝑝1, 𝑝2,…,𝑝𝑛 . Estas probabilidades son conocidas y es todo lo que nosotros conocemos acerca del evento que podría ocurrir. ¿Podemos encontrar una medida de qué tanta "oportunidad" está involucrada en la selección del evento, o cuanta incertidumbre tenemos de la respuesta o salida? Si existe tal medida, se puede decir que sería razonable que (𝑝1, 𝑝2,…,𝑝𝑛) cumpliera con las siguientes propiedades:  Continuidad: 𝐼 debería ser continua en cada 𝑝𝑖  Monotonía: Sí todas las 𝑝𝑖 son iguales, 𝑝𝑖 = 1 𝑛⁄ , entonces 𝐼 debe ser una función de incremento monotónico de 𝑛. Con 𝑛 eventos de igual probabilidad hay más opción, o incertidumbre, que cuando hay eventos de mayor probabilidad.  Recursión: Sí una opción se divide en dos, y se consideran opciones sucesivas, la 𝐼 original debe ser la suma ponderada de los valores individuales de 𝐼… Con estos pocos axiomas, Shannon demostró que la única función 𝐼 que los satisface, tiene la forma: 𝑛 𝐼 = −𝐾 ∑ 𝑝𝑖 log 𝑝𝑖 𝑖=1 (3.1) Donde 𝐾 es una constante positiva La ecuación 3.1 se representa de manera gráfica en la figura 3-1. De allí podemos extractar las siguientes consideraciones para cadenas binarias: (i) si la probabilidad de recibir unos es máxima 𝑝(1) = 1, y la probabilidad de recibir ceros es mínima 𝑝(0) = 0 , la información es mínima. Esto sucede porque conocemos de antemano que el futuro valor de 𝑥 será 1. La información (I) será 0 debido a que futuros valores de 𝑥 no adicionan nada nuevo, es decir, los valores son conocidos de manera anticipada. Ahora, si no conocemos acerca del futuro valor de 𝑥, como con la probabilidad para cara o sello de una moneda 𝑝(0) = 𝑝(1) = 0.5, entonces la información será máxima (𝐼 = 1), debido a que una futura observación nos traerá información relevante que será independiente de los valores previos. Desde esta condición, la información también ha sido vista como una medida de incertidumbre (Gershenson, 2013a). Sí hay absoluta certeza que el futuro de x será cero (𝑃(0) = 1) o uno (𝑃(1) = 1), 3 Se reemplaza H de Shannon por I de información Capítulo 3 41 entonces la información recibida será cero. Si no existe certidumbre debido a la distribución de probabilidad ((𝑃(0) = 𝑃(1) = 0.5)), la información recibida será máxima (Fig. 3-1). Figura 3-1 Información de Shannon a Diferentes Probabilidades para una Cadena Binaria (Gershenson and Fernández, 2012a) Por ejemplo, si tenemos una cadena '0001000100010001. ..', podemos estimar que 𝑃(0) = 0.75 y P(1) = 0.25, entonces 𝐼 = −(0.75 ∗ log 0.75 + 0.25 ∗ 𝑙og 0.25 ). Sí usamos 𝐾 = 1 y logaritmo base 2, entonces 𝐼 ≈ 0.811. Sí transformáramos a escala porcentual, diríamos que el próximo valor que venga nos traerá un 81.1% de información, pues presumimos que tener un cero tendría mayor probabilidad. No obstante, puede llegar un uno. Shannon usó 𝐻 para describir información (nosotros usamos 𝐼), debido a que él cuando desarrolló la teoría estaba pensando en el teorema de Boltzmann. El llamó a la ecuación 3.1, la entropía de un conjunto de probabilidades 𝑝1, 𝑝2, … , 𝑝𝑛. En palabras modernas, 𝐼 es una función de una variable aleatoria 𝑋. La unidad de información es el bit (binary digit). Un bit representa la información ganada cuando una variable binaria aleatoria se vuelve conocida. Sin embargo, desde la ecuación 3.1 es una suma de probabilidades, por lo que resulta ser una medida adimensional. Más detalles sobre la teoría de la información, pueden ser encontrados en (Ash, 2000, 1965). En años recientes, la teoría de la información (𝑇𝐼) ha sido relacionada con la complejidad, auto-organización y emergencia (Prokopenko, 2009b). De ello se destaca que la teoría de la información presenta, en el contexto de la ciencia de los sistemas complejos-𝐶𝑆𝐶, razones importantes cómo:  Un considerable cuerpo de trabajo en 𝐶𝑆𝐶 ha sido desarrollado dentro de la 𝑇𝐼, cómo pioneros tenemos al instituto Santa Fé, el cual ha direccionado la investigación de muchos de ellos.  La 𝑇𝐼 brinda definiciones que pueden ser formuladas matemáticamente 42 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴-  La 𝑇𝐼 brinda herramientas computacionales disponibles al instante; un número de medidas puede ser actualmente computadas. 3.2 Medidas Las medidas presentadas en este capítulo corresponden con las publicadas recientemente en Gershenson and Fernández(2012), y posteriores más refinadas y basadas en axiomas explicitas en Fernández et al. (2014b). El beneficio de usar axiomas está en que permite centrar la discusión en las presunciones o propiedades que estas medidas pueden tener, más que en la medida en sí misma. Adicionalmente, se incluyen medidas adicionales desde los trabajos de Fernández et al. (2012a, 2011, 2010a), que ahondan en las características de la homeostasis y la autopoiesis. 3.2.1 Emergencia Hemos mencionado que la emergencia se refiere a las propiedades de un fenómeno que están presentes en una escala y no en otra. Las escalas pueden ser temporales, espaciales o numéricas. Si describimos un fenómeno en términos de información, para tener nueva información, la información vieja debe haber sido transformada. Esta transformación puede ser dinámica, estática, activa, o estimérgica (Gershenson, 2012a). Por ejemplo, nueva información se produce cuando un sistema dinámico cambia su comportamiento, pero también cuando una descripción del sistema cambia. Concerniente con el primer caso, se han propuesto enfoques de medición de la diferencia pasada y futura (p.e. Shalizi & Crutchfield (2001)). Podemos llamar esta emergencia, emergencia dinámica. Concerniente con el segundo caso, los enfoques de medición se dan entre las escalas que han sido usadas (Holzer and de Meer, 2011; Shalizi, 2001), podemos llamar esta emergencia, emergencia de escala. Aun cuando existan diferencias entre las emergencias dinámicas y de escalas, ambas pueden ser vistas cómo la producción de información nueva. En el primer caso, la dinámica produce nueva información. En el segundo caso, el cambio de descripción produce nueva información. Entonces, una medida de emergencia basada en información 𝐸 incluiría tanto la emergencia dinámica cómo la de escala. Si retomamos, Shannon propuso una cantidad que medía qué tanta información era producida por un proceso. En consecuencia, podemos decir que la emergencia es la misma información de Shannon 𝐼 (Ecuación 3.2.). En lo sucesivo, consideraremos a la emergencia de un proceso 𝐸, cómo la información 𝐼, y usaremos el logaritmo en base 2. 𝐸 = 𝐼 (3.2) Capítulo 3 43 La idea intuitiva de emergencia cumple con las tres nociones básicas (axiomas) que Shannon usó para derivar la 𝐼 (Shannon 𝐻). Para el axioma de continuidad, se espera que una medida no tenga grandes saltos cuando se presentan pequeños cambios. El segundo axioma es más difícil de mostrar. Este establece que si consideramos una función auxiliar 𝑖, la cual es función de 𝐼, cuando existan 𝑛 eventos con la misma probabilidad 1/𝑛, entonces la función 𝑖 tendrá un incremento monotónico. Si tenemos la misma configuración para la emergencia, entonces podríamos pensar que el proceso tendría igual variabilidad en cualquiera de los 𝑛 estados disponibles. Si algo pasa y ahora el proceso puede estar en 𝑛 + 𝑘 estados, igualmente probables, podemos decir que el proceso ha tenido emergencia, dado que ahora necesitamos más información para conocer en cual estado del proceso está. Para el tercer axioma, necesitamos encontrar una forma para resolver cómo se puede dividir el proceso. Recordemos que la tercera propiedad requerida para la 𝐼 de Shannon, es que si una opción puede ser dividida en dos diferentes, la 𝐼 original debe ser promedio de las otras dos 𝐼. En un proceso, podemos pensar en las opciones cómo una fracción del proceso que actualmente observamos. Para este propósito, podemos hacer una partición del dominio. En nuestro caso, podemos tomar dos subconjuntos cuya intersección es el conjunto nulo y cuya unión es el conjunto original completo. Después de ello, computamos la función 𝐼 para cada uno. Dado que observamos dos diferentes partes de un proceso y de cada observación tenemos el promedio de información requerida para describir el proceso (parcial), entonces toma sentido tomar el promedio de ambos cuando se observa el proceso total. Al estar 𝐸 basada en 𝐼, tenemos que es una medida probabilística. 𝐸 = 1 significa que cuando cualquier variable binaria se vuelve conocida, un bit de información emerge. Sí 𝐸 = 0, entonces no emergerá nueva información, incluso variables aleatorias se vuelven "conocidas" (se conocen de manera anticipada, de antemano). Enfatizamos que la emergencia puede tener lugar en un nivel de observación de un fenómeno o en un nivel de descripción del fenómeno. Uno u otro, pueden producir nueva información.  Múltiples Escalas y Normalización. Para medir la emergencia, auto-organización, complejidad y homeostasis a múltiples escalas numéricas, las cadenas binarias pueden ser convertidas a diferentes bases. Si tomamos dos bits, tendremos cadenas de base 4. Si tomamos tres bits tendremos cadenas de base 8. Tomando 𝑏 bits tendremos cadenas de base 2𝑏. Cuando Shannon definió la ecuación 3.1, él incluyó una constante positiva 𝐾. Este hecho es importante debido a que podemos cambiar el valor de 𝐾 para normalizar la medida en el intervalo [0,1]. El valor de 𝐾 puede depender de la longitud del alfabeto finito 𝐴 que usemos. En el caso particular de las redes Booleanas, que tienen un alfabeto 𝐴 = {0, 1}, su longitud es 𝐴 = 2. Entonces, el valor de 𝐾 = 1 normalizará la medida en el intervalo [0,1]. Debido a la relevancia de la notación binaria en ciencias computacionales y otras aplicaciones, en 44 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- este trabajo se usará a menudo el alfabeto booleano. Sin embargo, podemos computar la entropía para alfabetos de diferentes longitudes. Para ello consideramos la ecuación 3.3 𝐾 = 1 ⁄ 𝑙𝑜𝑔2𝑏 (3.3) Donde 𝑏 es la longitud del alfabeto usado. En esta forma 𝐸 y las medidas derivadas de ella son normalizados, teniendo un máximo de 1 y un mínimo de 0. Por ejemplo, consideremos la cadena en base 4 '0133013301330133. ..'. Podemos estimar que 𝑃 (0) = 𝑃 (1) = 0.25, 𝑃 (2) = 0, 𝑦 𝑃 (3) = 0.5. Siguiendo la ecuación (3.1), tenemos que 𝐼 = −𝐾((0.25 × 𝑙𝑜𝑔 0.25 ) + ( 0.25 × 𝑙𝑜𝑔 0.25) + ( 0.5 × 𝑙𝑜𝑔 0.5)), dado que 𝑏 = 4, entonces 𝐾 = 1 𝑙𝑜𝑔24 , y obtenemos una 𝐼 = 0.75 normalizada. ⁄ 3.2.2 Auto-organización La auto-organización ha sido asociada, correlacionada, con el incremento de orden. Por ejemplo, con la reducción de entropía (Gershenson and Heylighen, 2003). Si la emergencia implica un incremento de información, la cual es análoga a la entropía y al desorden, la auto- organización estará anti-correlacionada con la emergencia. Una medida de auto-organización 𝑆 (por self-organization en inglés), será una función 𝑆: ∑ → ℝ D donde ∑ = 𝐴ℕ, con las siguientes propiedades: a. El rango de 𝑆 está en el intervalo de los reales [0,1] b. 𝑆(𝑋) = 1 sí y sólo sí 𝑋 es determinística. P.e. conocemos de antemano el valor del proceso. c. 𝑆(𝑋) = 0 sí y sólo sí 𝑋 tiene una distribución uniforme. P.e. cualquier estado de un proceso tiene la misma probabilidad. d. 𝑆(𝑋) tiene una correlación negativa con la emergencia E Proponemos cómo medida: 𝑆 = 1 − 𝐼 = 1 – 𝐸 (3.4) Es sencillo comprobar que esta función cumple con los axiomas declarados. Sin embargo, no es la única formulación posible. No obstante, es la única función (lineal) ajustada que satisface los axiomas. Por simplicidad, proponemos usar esta expresión cómo medida de la auto-organización. 𝑆 = 1 significa que existe orden máximo. Es decir, no se produce nueva información (𝐼 = 𝐸 = 0). En el otro extremo, cuando 𝑆 = 0, no existe orden en absoluto. Es decir, cuando ninguna variable aleatoria se vuelve conocida se produce/emerge información. Capítulo 3 45 Cuando 𝑆 = 1, el orden es máximo, la dinámica no produce nueva información, entonces el futuro, desde el pasado, es completamente conocido. De otra parte, cuando 𝑆 = 0 el orden es mínimo, la información pasada no nos dice nada acerca de la información futura. Nótese que la ecuación 3.4 no hace distinción entre si el orden es producido por el (mismo) sistema o por su ambiente. Entonces, 𝑆 podría tener un gran valor en sistemas con alta organización, independientemente sí es un producto de las interacciones locales o impuesta externamente. Esta distinción puede hacerse fácilmente por la descripción en detalle del comportamiento de los sistemas, pero es difícil e innecesario de capturar con una medida de probabilidad general cómo 𝑆. Cómo analogía, podemos medir la temperatura de una substancia, pero la temperatura no nos diferencia (y no es necesaria para diferenciar) entre las substancias que están calientes o frías desde el exterior, y substancias cuya temperatura es principalmente dependiente de sus reacciones químicas internas. Resaltamos que con esta medida de auto-organización los patrones y/o dinámicas estáticas pueden ser valorados adecuadamente por su alta S y baja entropía. Por ejemplo, los sistemas vivos reducen su entropía termodinámica para mantenerse, lo cual ha sido descrito con el concepto de auto-organización (Kauffman, 1993). 3.2.3 Complejidad Sobre la base de Lopez-Ruiz et al. (1995), podemos definir complejidad 𝐶, cómo el balance entre el cambio (caos) y estabilidad (orden). Dado que hemos asociado tales aspectos a las medidas de emergencia y auto-organización, la complejidad es una función 𝐶: ∑ → ℝ, que tiene las siguientes propiedades:  Su rango son los reales en el intervalo [0,1]  𝐶 = 1 sí y sólo sí 𝑆 = 𝐸  𝐶 = 0 sí y sólo sí 𝑆 = 0 y 𝐸 = 0 Es natural considerar que el producto de 𝑆 y 𝐸 satisface los dos últimos requerimientos. Por lo tanto, proponemos: 𝐶 = 4 × 𝐸 × 𝑆 (3.5) Donde 4 es adicionada cómo una constante de normalización para cumplir con el primer axioma, tal que 𝐶 pertenecerá al intervalo [0,1]. 𝐶 puede ser representada en términos de 𝐼 cómo: 𝐶 = 4 × 𝐼 × (1 − 𝐼 ) (3.6) 46 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- La Figura 3-2 muestra las medidas propuestas para diferentes valores de 𝑃(𝑥). Allí es notorio que 𝐸 es máxima cuando 𝑃(𝑥) = 0.5 y mínima cuando 𝑃 (𝑥) = 0 or 𝑃 (𝑥) = 1. Lo opuesto es para 𝑆, esto es mínima cuando 𝑃 (𝑥) = 0.5 y máxima cuando 𝑃 (𝑥) = 0 o P (x) = 1. 𝐶 es máxima cuando 𝐸 = 𝑆 = 0.5, lo cual ocurre cuando 𝑃 (𝑥) ≈ 0.11 or 𝑃 (𝑥) ≈ 0.89. La información de Shannon puede ser vista como un balance de ceros y unos (máxima cuando 𝑃 (0) = 𝑃 (1) = 0.5), en tanto que 𝐶 es vista como un balance de 𝐸 y 𝑆 (máxima cuando 𝐸 = 𝑆 = 0.5). Figura 3-2 Emergencia, Auto-organización y Complejidad vs Probabilidad en una Cadena Binaria (Gershenson and Fernández, 2012a) Cabe destacar que históricamente, sí bien Shannon definió la información cómo entropía (Equivalente a nuestra 𝐸), Wiener y von Bertalanfy definieron información como el opuesto, cómo negentropía (Equivalente a nuestra 𝑆). En este sentido, nuestra medida de complejidad 𝐶, reconcilia estos dos puntos de vista al ser producto del balance entre el orden y el caos. Desde esta perspectiva, las ventajas de ésta medida son importantes, pues pueden caracterizar sistemas dinámicos con máxima complejidad en regiones complejas, cómo las redes booleanas y sistemas vivos que requieren un balance de estabilidad (𝑆) y adaptabilidad (𝐸). Casos que ilustran este tipo de situaciones se observarán en el capítulo de aplicaciones. 3.2.4 Homeostasis  Homeostasis del Sistema Total (Homeostasis de los Estados del Sistema en el Tiempo) Las anteriores medidas (𝐸, 𝑆 y 𝐶) estudian cómo variables individuales cambian en el tiempo. Para calcular las medidas para un sistema total, podemos graficar el histograma, o simplemente promediar las medidas para todas las variables del sistema. Para la homeostasis Capítulo 3 47 𝐻𝑚, estamos interesados en cómo todas las variables del sistema cambian, o no, en el tiempo. Se resalta que 𝐸, 𝑆 y 𝐶 se calculan desde las series de tiempo de las variables halladas en las columnas, mientras 𝐻𝑚 se enfoca en los estados del sistema que cambia en el tiempo, y cuya información se halla en las filas. Sí el sistema ha tenido alta homeostasis, se espera que sus estados no cambien mucho en el tiempo. La homeostasis es una función 𝐻: ∑ × ∑ → ℝ que tiene las siguientes propiedades:  Su rango es el intervalo real [0,1]  𝐻 = 1 , Sí y sólo sí, para los estados 𝑞 y 𝑞′, 𝑞 = 𝑞′, esto es: no hay cambio en el tiempo.  𝐻 = 0 , Sí y sólo sí, ∀𝑖 , 𝑞𝑖 ≠ 𝑞′𝑖, esto es: todas las variables en el sistema cambiaron Una función útil para la comparación de cadenas de igual longitud es la distancia de Hamming, considerada también en la teoría de la información (Hamming, 1950). La distancia de Hamming 𝑑 (ecuación 3.7) mide el porcentaje de símbolos diferentes en dos cadenas 𝑋 y 𝑋′. Para cadenas binarias, puede ser calculada con la función 𝑋𝑂𝑅 (⊕). Su normalización lleva al intervalo [0,1]: ∑ 𝑑(𝑋, 𝑋′) = 𝑥𝑖 ⊕ 𝑥′𝑖 𝑖 ∈{0,…,𝑋} 𝑋 (3.7) 𝑑 mide la fracción de símbolos diferentes entre 𝑋 y 𝑋′. Para casos booleanos 𝑑 = 0 ⇔ 𝑋 = ¬𝑋′, mientras 𝑋 y 𝑋′ están no correlacionadas ⇔ 𝑑 ≈ 0.5. Nosotros podemos usar el inverso de 𝑑 para definir ℎ: ℎ(𝑋𝑡, 𝑋𝑡+1) = 1 − 𝑑(𝑋𝑡, 𝑋𝑡+1) (3.8) La ecuación 3.8 de homeostasis, claramente cumple con las propiedades deseadas para la homeostasis entre dos estados. Para medir la homeostasis de un sistema en el tiempo, podemos generalizarlo como se muestra en la ecuación 3.9: 𝐻 = 1 𝑚 − 1 𝑚−1 ∑ ℎ(𝑋𝑡, 𝑋𝑡+1 𝑡=0 ) (3.9) Donde 𝑚 es el número total de sucesos de tiempo que serán evaluados. 𝐻 será el promedio de las diferentes ℎ desde 𝑡 = 0 a 𝑡 = 𝑚 − 1. Cómo las medidas previas basadas en 𝐼, 𝐻 es una medida adimensional. Cuando 𝐻 es medida a grandes escalas, captura dinámicas periódicas. Por ejemplo, sea nuestro sistema con 𝑛 = 2 variables y un ciclo de periodo 2: 11 → 00 → 11. 𝐻 para base 2 será mínima, dado que en todos los pasos de tiempo todas las variables cambian, los unos se 48 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- cambian a ceros o los ceros a unos. Sin embargo, si la medida de 𝐻 se hace en base 4 tomamos dos eventos de tiempo binarios. Así, en base 4 el atractor será 22 → 22 y la 𝐻 = 1. Lo mismo aplica para bases grandes. Ejemplos de éste funcionamiento se demostrará en el capítulo de aplicaciones.  Homeostasis para Una Variable del Sistema (𝐻𝑚𝑉) Cómo se mencionó en el capítulo 1 numeral 1.3.4 (definición de homeostasis), se ha establecido que las variables esenciales de un sistema se mantienen en una zona de viabilidad. Esta zona, está definida por los límites inferiores y superiores (rango) en los que se mueve la variable (Ross Ashby, 1960b). Con esto en mente, desarrollamos una medida que valorara la dispersión de los estados de una variable (desviación), calculada a partir de los valores de probabilidad de tales estados. Se considera que a mayor dispersión- desviación entre estados, el sistema mostrará menor regularidad y menor será la homeostasis. Para complementar la medida de 𝐻𝑚𝑣 tiene en cuenta la equitabilidad entre estados (evenness, en inglés; relativo a uniformidad, regularidad). Más equitabilidad es más homeostasis, dado que las trayectorias del sistema serán más regulares (mayor auto- organización). En el anterior contexto, definimos la homeostasis de una variable 𝐻𝑚𝑉 como una función de la dispersión entre las probabilidades de sus estados, representada en su desviación estándar (𝜎𝑃𝑖) y el inverso de la equitabilidad (Buzas and Gibson's evenness). El cálculo tiene en cuenta la transformación de la variable según la escala elegida. La ecuación 3.10 muestra la relación planteada. 𝐻𝑚𝑉 = 𝜎𝑃𝑖 𝐸𝑣𝑒𝑛𝑛𝑒𝑠𝑠 = 𝜎𝑃𝑖 𝑒𝐼/𝐶𝑠 Donde: 𝐼: Información de Shannon: 𝐼 = − ∑ 𝑃𝑖 ∗ 𝐿𝑜𝑔𝑃𝑖 𝐶𝑠: Número de casos a partir de los que se calculan las probabilidades. Para efectos de normalización en el rango [0,1] se obtiene de la siguiente manera: 𝐻𝑚𝑉𝑁 = 1 − 1 𝐻𝑚 + 1 (3.10) (3.11)  Mecanismos Homeostáticos en Redes Computacionales de Agentes A través de diferentes mecanismos homeostáticos, un sistema multi-agentes puede hacer frente a los cambios, influencias y/o perturbaciones de su ambiente interno o endógeno, y/o Capítulo 3 49 externo o exógeno 𝐴𝑏. Estos mecanismos de auto-regulación promueven la estabilidad y flexibilidad del sistema, en su paso por los múltiples estados de su ciclo adaptativo, sin que se dé su destrucción. La base del mecanismo homeostático está en el desenvolvimiento de la red en una zona de viabilidad (𝑍𝑣), definida por los rangos de viabilidad (RX), de nodos y aristas, a los factores Ab. Como consecuencia de la búsqueda de equilibrio ante la influencia de Ab, la homeostasis puede generar cambios semipermanentes en la red, como respuesta a los cambios semipermanentes de Ab. En una red computacional de agentes, representados por nodos, cómo se explicó en el capítulo 2, los procesos de auto-regulación se pueden dar de manera complementaria, tanto de forma activa como pasiva. Para el primer caso se parte que la función global F es una función de transición, tal que 𝐹: 𝑄 → 𝑄, donde 𝑄 es un conjunto no vacío compuesto de un número finito de estados 𝑞𝑖 𝑄. Al mismo tiempo, F corresponde a una familia de funciones, denominadas funciones de activación local de la red (𝑓𝑖), tal que 𝐹 = (𝑓1, 𝑓2, … , 𝑓𝑛). Por lo tanto, la auto- regulación activa se puede dar, acorde a la forma en que se den las diferentes combinaciones de los 𝑓𝑖. Así, al tener a "𝑜" como operador, tal que 𝐹 = 𝑓1𝑜 𝑓2𝑜, … , 𝑓𝑛, será evidente que 𝐹 = 𝑓1𝑜 𝑓2 ≠ 𝐹 = 𝑓2𝑜 𝑓1. Lo anterior permite observar que según se combinen las funciones parciales, se generará un cambio en la función global. Esta condición será un mecanismo básico que le dará a la red mayor adaptabilidad ante los cambios del entorno Ab. La auto-regulación pasiva, por su parte, estará definida por situaciones como los rangos de viabilidad RX de la red a la influencia de un factor ambiental Ab, como a continuación se aprecia: – Ante la influencia Ab, la red se desenvuelve en una zona de viabilidad 𝑍𝑣. – Tal 𝑍𝑣 está en correspondencia con los valores máximos (𝑋𝑚𝑎𝑥) y mínimos (𝑋𝑚𝑖𝑛) de un factor i de 𝐴𝑏. Este hecho define la respuesta (tolerancia) de la red al factor i de 𝐴𝑏, acorde con sus rangos de viabilidad (𝑅𝑋𝑖). En consecuencia, para cada factor i de 𝐴𝑏, un nodo j de la red tendrá un rango de viabilidad al factor i, resultante de la diferencia de los valores de extremos, tal que 𝑅𝑋𝑖 𝑗 = 𝑋𝑚𝑎𝑥 − 𝑋𝑚𝑖𝑛 . – Para el caso del rango de viabilidad en las aristas, que considerará igualmente los 𝑗. valores máximos y mínimos ante el factor i de Ab, su notación estará dada por 𝑅𝑋′𝑖 – La respuesta global 𝑅𝑋𝑖 se dará por el promedio de 𝑅𝑋𝑖 𝑗. 𝑗 y 𝑅𝑋′𝑖 – Se estima que la respuesta de la red o tolerancia (T) a un factor i Ab, puede coincidir, con una función gaussiana con un valor promedio (𝜇𝑖) de su rango 𝑅𝑋𝑖 (Fig. 3-3). La escogencia de la función normal, se debe a que ella coincide con la respuesta de muchos sistemas complejos, como los sistemas socio-ecológicos. No obstante, en caso de respuestas diferentes se pueden usar otros tipos de distribuciones. 50 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- – El valor de tolerancia para un determinado valor específico 𝑋𝑖 ∈ 𝑅𝑋𝑖 del factor 𝐴𝑏𝑖, deberá considerar como parámetros para su cálculo la desviación estándar 𝜎𝑖, y el promedio μi. La función también cuenta con las constantes 𝜋 y 𝑒, y su formulación se expresa en la ecuación 3.12. – 𝑇(𝐴𝑖) = 1 𝜎𝑖√2𝜋 −(𝑋𝑖−μi)2 2 𝑒 2𝜎𝑖 (3.12) Figura 3-3 Zona de Viabilidad Zv y Zona de Funcionamiento Óptimo (ZFO) de un factor Abi o Bi para una red computacional. Los valores del eje x dependen del dominio del factor i. (Fernández et al., 2012b) – Se destaca que dentro de la zona de viabilidad 𝑍𝑣, definida por la función de tolerancia T(Ab), existe una zona de funcionalidad óptima 𝑍𝐹𝑂, que se corresponde con una zona cercana al promedio, de cada factor 𝜇𝑖 del factor i de Ab. Estadísticamente, esta área se encuentra entre 𝜇𝑖 ∓ 𝜎𝑖. – En general, los niveles más altos de tolerancia serán aquellos en los cuales el nivel de satisfacción de un nodo o agente es máximo. Ello sucede cuando el agente alcanza un objetivo particular, y al mismo tiempo se minimiza la fricción4 con los demás agentes. Los mecanismos de auto-regulación como el descrito, son los que soportan el proceso auto-organizativo que alcanza niveles más altos de tolerancia y satisfacción, y mínimos de fricción. Lo anterior, coincide lo que Gershenson (2010) ha definido como perfil sigma. 4 interacciones negativas que no favorecen el alcance del objetivo global del sistema. Capítulo 3 51 – En caso que los nodos y/o aristas tengan tolerancias individuales heterogéneas, la tolerancia global de la red será producto del acople de las tolerancias individuales. Esta condición se muestra en la función de tolerancia global desde las funciones locales de los nodos y aristas, representada en la figura 3-4. – Figura 3-4 Detalle de la Respuesta Global y Local en la Red SD-Ag. Los valores de los ejes son simplemente demostrativos (Fernández et al., 2012b) – Consecuente con lo anterior, se puede considerar que un factor i de 𝐴𝑏 puede tener efecto sobre la totalidad de los nodos Ni (incidencia global) o sobre determinados Ni (incidencia parcial). No obstante, si se asumiera en algún caso que la influencia de 𝐴𝑏𝑖 sobre los Ni es uniforme, no se podrá asumir que su respuesta sea igual. La respuesta diferencial de los Ni, será relativa a la capacidad de auto-regulación de cada uno de ellos para responder a las influencias de 𝐴𝑏𝑖. – Si el funcionamiento de la red, nodos y/o aristas individuales se diese por fuera de la zona de viabilidad-𝑍𝑣, ellos operarán en zonas definidas como críticas (𝑍𝑐). Estar en 𝑍𝑐, puede generar su fallo extremo o degradación. En términos de la autopoiesis, esto es una limitación para el auto-mantenimiento pues se da la afectación de la estructura y función en la red. No obstante, al hallarse dentro de 𝑍𝑐, la red deberá compensar este efecto tanto en un sentido homeostático (de auto-regulación) como autopoiético (de auto-mantenimiento), con el fin de seguir operando. El mecanismo básico de compensación para ello se presenta en la siguiente sección.  Mecanismo Homeostático General de Compensación ante Factores del Ambiente Ab La siguiente es la secuencia de compensación en la red. – La compensación será parte del mecanismo homeostático, y consistirá en la escogencia de una acción 𝑦 ∈ 𝑌 apropiada de respuesta, ante una influencia 𝐴𝑏𝑖. Para ello, se 52 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- contará con la ejecución de una función objetivo (𝑓𝑜𝑏) que restablecerá el sistema al llevarlo nuevamente a la zona de viabilidad 𝑍𝑣, tal que 𝑦(𝑖, 𝑓𝑜𝑏,𝑋𝑚𝑎𝑥, 𝑋𝑚𝑖𝑛). La ejecución de 𝑓𝑜𝑏 tendrá como parámetros: (i) que la acción relacionada controle efectivamente el desbalance, y (ii) que no se cause desequilibrio en los demás componentes de la red. Por lo tanto, 𝑓𝑜𝑏 causará el mínimo trauma en la red y compensará efectivamente la influencia. – La anterior visión coincide, nuevamente, con la maximización de la satisfacción y la reducción de la fricción en el sistema, como lo propone Gershenson(2010). Se puede estimar que 𝑓𝑜𝑏 podría ser parte del algoritmo que modifica la red computacional. – La formalización del mecanismo homeostático de compensación es extensible a la autopoiesis. Para tal fin, existirá una función 𝑓𝑜𝑏 involucrada en la producción (síntesis- degradación) de los elementos de la red. Esto, debido a que la autopoiesis puede ser observada como un tipo particular de homeostasis.  Otros indicadores del proceso homeostático Hasta ahora la estabilidad y flexibilidad en la red computacional de agentes se han soportado en la función de tolerancia 𝑇(𝐴𝑏𝑖), que involucra al rango (𝑅𝑋), y en los mecanismos de auto- regulación pasiva y activa. Sin embargo, es posible aportar otros elementos que enriquecen la caracterización de esta misma estabilidad, e incluso determinar la "fragilidad" de la red. Como base se tienen los conceptos expresados por (Pimm, 1991) acerca de la resiliencia, persistencia y resistencia, en sistemas ecológicos. A partir de estas definiciones, se establecen nuevos formalismos para su representación matemática que pueden hacer posible la integración de todos ellos, en una medida extendida de homeostasis. – Resiliencia (𝑹𝒔𝒊): tasa (velocidad) a la que se retorna al punto medio 𝜇𝑖en la zona de viabilidad Zv. Esto puede ser definido como 𝑅𝑠𝑖 : = 𝐷𝑖/𝑡, donde 𝐷𝑖 o distancia es el valor absoluto de la diferencia entre el valor 𝑋𝑖 del factor 𝑖 de Ab y su promedio 𝜇𝑖, tal que 𝐷𝑖 = 𝑋𝑖 − 𝜇𝑖 . Entre tanto, 𝑡 es el tiempo que demora en retornar al valor 𝜇𝑖. 𝑅𝑠𝑖 puede ser extendido a la tasa de retorno a un punto determinado de la zona de funcionalidad óptima-𝑍𝐹𝑂. Mayores valores de 𝑅𝑠𝑖 indican mayor capacidad homeostática. – Persistencia (𝑷𝒆𝒊): es la medida del tiempo que se pasa en el punto medio 𝜇𝑖, o en ZFO. 𝑃𝑒𝑖 = 𝑡𝑟, donde 𝑡𝑟 es el tiempo de residencia en esos valores anteriores. Valores grandes de 𝑃𝑒𝑖 indican más capacidad homeostática. Los dos indicadores anteriores (𝑅𝑠𝑖, 𝑃𝑒𝑖) tienen un comportamiento opuesto en el tiempo (Fig. 3-5), dado que a mayores tiempos de retorno la resiliencia se hace menor, y en consecuencia, la capacidad homeostática. Entre tanto, al aumentar el tiempo de Capítulo 3 53 mantenimiento en el punto medio, la persistencia es mayor, y así mismo, la capacidad homeostática de la red. – Figura 3-5 Comportamiento de la resiliencia (𝑹𝒔𝒊) y la persistencia (𝑷𝒆𝒊):) en el tiempo de retorno y el tiempo de estancia en el punto medio respectivamente, en constraste con la capacidad homeostática (Fernández et al., 2012b). – Resistencia Máxima (𝑹𝒆 − 𝑴𝒂𝒙𝒊): es la medida del grado de desplazamiento máximo con referencia al 𝜇𝑖, tal que el grado de desplazamiento máximo se corresponde con el valor absoluto de 𝑅𝑋𝑖 . Es decir, 𝑅𝑒 − 𝑀𝑎𝑥𝑖 = 𝑅𝑋𝑖/𝜇𝑖. Mayores valores de 𝑅𝑒 − 𝑀𝑎𝑥𝑖 indican mayor capacidad homeostática. Cabe notar que, de la misma forma que se obtiene resistencia máxima se puede obtener otras variantes como la Resistencia 𝑍𝐹𝑂, o la resistencia instantánea, acorde con los valores de 𝑋𝑖 que se tengan en cuenta. – Vulnerabilidad (𝑽𝒖𝒊): la susceptibilidad de la red ante el factor i de Ab estará definida por una función inversa con el rango de viabilidad 𝑅𝑋𝑖, tal que 𝑉𝑢𝑖=1/(1+RXi), donde 𝑉𝑢𝑖 ∈ [0,1]. Mayor 𝑉𝑢𝑖 indica menor capacidad homeostática. El comportamiento de la resistencia máxima y la vulnerabilidad a medida que se incrementa el rango, se puede observar en la figura 3-6. En ella es notorio que la resistencia máxima se incrementa con el rango, y así mismo se transfiere mayor capacidad homeostática a la red. En tanto, la vulnerabilidad se hace menor a medida que el rango se incrementa, y por ende, la capacidad homeostática aumenta. Dado que el proceso de auto-regulación es un proceso dinámico, los anteriores indicadores deben ser observados en el tiempo para determinar su cambio. A manera de síntesis, se puede decir que la capacidad de auto-regulación u homeostasis de la red multi-agentes (𝐻𝑆𝐷−𝐴𝑔) a un factor 𝑖 dado de Ab, estará dado por la caracterización de los indicadores. Así, la 𝐻𝑆𝐷−𝐴𝑔 de la red será una función 𝑔 de la tolerancia, resiliencia, 54 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- persistencia, resistencia máxima, y vulnerabilidad, tal que 𝐻𝑆𝐷−𝐴𝑔 = 𝑔(𝑇(𝐴𝑏𝑖), 𝑅𝑠𝑖, 𝑃𝑒𝑖, 𝑅𝑒 − 𝑀𝑎𝑥𝑖, 𝑉𝑢𝑖). Al normalizar por cualquier procedimiento 𝑅𝑠𝑖, 𝑃𝑒𝑖, 𝑅𝑒 − 𝑀𝑎𝑥𝑖𝑖 a una escala [0,1], la función 𝑔 de 𝐻𝑆𝐷−𝐴𝑔 podrá calcular el promedio de los mismos. Figura 3-6 Comportamiento de la resistencia máxima (𝑹𝒆 − 𝑴𝒂𝒙𝒊) y la vulnerabilidad (𝑽𝒖𝒊) en cuanto al incremento del rango de viabilidad (𝑹𝑿𝒊) (Fernández et al., 2012b). 3.2.5 Autopoiesis  Autopoiesis cómo expresión de la Autonomía del Sistema Los siguientes aspectos formales de la autopoiesis se basan en la consideración que los sistemas adaptativos requieren alta complejidad o 𝐶 para enfrentar los cambios ambientales y mantener su integridad (Kaufmann, 1993; Langton, 1990). Desde esta perspectiva, fue posible definir una expresión de autopoiesis desde la autonomía del sistema. A continuación presentamos las consideraciones axiomáticas para ello. Sea 𝑋 las trayectorias de las variables de un sistema y 𝑌 las trayectorias de las variables del ambiente del sistema. Una medida de autopoiesis 𝐴: ∑ × ∑ → ℝ tiene las siguientes propiedades: – 𝐴 ≥ 0 – 𝐴 deberá reflejar la independencia de 𝑋 y 𝑌, lo que implica: o 𝐴 > 𝐴′ ⇔ 𝑋 produce más de su propia información que 𝑋′ para un 𝑌 dado. o 𝐴 > 𝐴′ ⇔ 𝑋 produce más de su propia información en 𝑌 que en 𝑌′ o 𝐴 = 𝐴′ ⇔ 𝑋 produce mucho más de su propia información que 𝑋′ para un 𝑌 dado. o 𝐴 = 𝐴′ ⇔ 𝑋 produce mucha más de su propia información en 𝑌 que en 𝑌′ o 𝐴 = 0 si toda la información en 𝑋 es producida por 𝑌 . Siguiendo la clasificación de tipos de transformación de información propuesta por Gershenson (2012b), las transformaciones dinámicas y estáticas son internas (un sistema Capítulo 3 55 produce más de su propia información), mientras que las transformaciones activas y estimérgicas son externas (información producida por otro sistema). Desde la alta complejidad requerida por un sistema para afrontar el ambiente, tenemos que sí el sistema 𝑋 tiene alta 𝐸, entonces podría pasar que no esté apto para producir los mismos patrones para diferentes 𝑌. Por otra parte, si se tiene alta 𝑆, 𝑋 podría no estar apta o ser capaz para adaptarse a los cambios en 𝑌. Por lo tanto, proponemos la autopoiesis como una relación de las complejidades del sistema 𝐶(𝑋) y su ambiente 𝐶( 𝑌 ), acorde con la ecuación 3-13 𝐴 = 𝐶(𝑋) 𝐶( 𝑌 ) (3.13) Sí 𝐶(𝑋) = 0, entonces 𝑋 es estática (𝐸(𝑋) = 0) o seudoaleatoria (𝑆(𝑋) = 0). Esto implica que cualquier patrón (complejidad) que pueda ser observado en cualquier 𝑋, debe venir de 𝑌. Este caso da una mínima 𝐴. Por otra parte, sí 𝐶( 𝑌 ) = 0, esto implica que cualquier patrón en 𝑋, sí lo hay, debe venir de sí mismo. Este caso da una 𝐴 = ∞. Un caso particular ocurre sí 𝐶( 𝑋 ) = 0 y 𝐶( 𝑌 ) = 0, 𝐴 resulta indefinada. Pero cómo podemos decir algo sobre la autopoiesis si comparamos dos sistemas los cuales son sín variación (𝑆 = 1) o seudoaleatorios (𝐸 = 1), este debe ser un caso indefinido. El resto de las propiedades son evidentemente cumplidas por la ecuación 3.13. Se aclara, que esto no es ciertamente la única función que cumple con los axiomas deseados. La exploración de alternativas requiere más estudios. Dado que 𝐴 representa la proporción de las probabilidades, esta es una medida adimensional. 𝐴 ∈ [0, ∞), sin embargo ella puede ser mapeada a [0,1] usando una función cómo 𝑓(𝐴) = 1+𝐴 No obstante, nosotros no normalizamos 𝐴 debido a que es útil distinguir cuando 𝐴 > 1 y 𝐴 < 1. El primer caso corresponde cuando el sistema es más autónomo y adaptable que su ambiente, y el segundo cuando el sistema, por su menor complejidad, tiene menores posibilidades de adaptación que su medio, hecho que le confiere menor autonomía. Esto se podrá ver claramente en el capítulo de aplicaciones. 𝐴 . 56 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴-  Procesos Autopoiéticos de Auto-Mantenimiento en Redes Computacionales de Agentes Desde el punto de vista del auto-mantenimiento, 𝐴 es la responsable de los procesos de auto- regulación de la síntesis5 y degradación6 de los agentes y sus interacciones (Fernández et al., 2010a). 𝐴 se soporta en mecanismos homeostáticos específicos, enfocados en la constitución de los elementos de la red computacional, y promueve la "producción del sistema" y la "preservación de la estructura" (Iba, 2011). Por medio de 𝐴 se establece la capacidad de la red para desarrollar, mantener, producir y restablecer su unidad e identidad en un nivel dado. Particularmente, este proceso está basado en la hetero y autoreferenciación7, hecho que involucra un grado determinado de cognición. La cognición, según Gershenson (2010), se refiere al conocimiento que el sistema tiene de cómo actuar en su ambiente. Se puede estimar que en una red de agentes, 𝐴 se corresponde con una función f de auto- mantenimiento (𝑓𝑎) que hace parte de la combinación de las funciones de activación local 𝑓𝑖 que regula el "desgaste" natural de la red. El desgaste, en el sentido homeostático, es visto como una variante de un factor i de Ab. En consecuencia, el mecanismo de ejecución de la autopoiesis corresponde con el mecanismo general de auto-regulación, el cual puede estar inmerso en el algoritmo 𝑎 de la red computacional. El auto-mantenimiento, referido a 𝐴, puede ser medible a través del grado de "auto- mantenimiento" de la red en la constitución de nodos N y aristas K. La medida del auto- mantenimiento autopoiético puede ser expresada a través de un balance de masas (Holzbecher, 2007). El balance considera la dinámica de la constitución de los N, a partir de la diferencia de sus tasas de síntesis y de degradación. Para un N de la red, la dinámica de su auto-mantenimiento será igual a su síntesis (Si), o producción local, menos su degradación local (𝐷𝑔), de tal manera que: 𝑑𝑁 𝑑𝑡⁄ = 𝑑𝑆𝑖 𝑑𝑡⁄ − 𝑑𝐷𝑔 ⁄ 𝑑𝑡 (3.14) La síntesis de los nodos N estará dada por 𝑆𝑖 = 𝛾𝑁, donde γ será el coeficiente o la razón promedio de síntesis de N. La degradación o perdida de N, será proporcional a 𝐷 = −𝜆𝑁𝑑, donde 𝜆 corresponde a la constante de degradación promedio de N. El exponente 𝑑, es el orden de la degradación (Holzbecher, 2007). Así, la ecuación 3.14 toma la forma: 5 La síntesis autopoiético se refiere al proceso de constitución y mantenimiento tanto de los N, como de la estructura de que depende la constitución del elemento. Esto se logra a partir del establecimiento de una red de procesos de producción (Luhmann 1995). 6 La degradación, por su parte, se refiere a la perdida de integridad estructural o unidad de los N, y en consecuencia de la estructura que los soporta y de las condiciones para la reproducción (Niklas Luhmann, 1995). 7 Referida a relación, semejanza o dependencia que como producto de la cognición el elemento hace sobre el exterior (hetero-referencia) y sobre sí mismo (auto- referencia). Capítulo 3 57 𝑑𝑁 𝑑𝑡⁄ = 𝛾𝑁 − 𝜆𝑁𝑑 (3.15) Cuando 𝑑 es de orden 1, la diferencia de los promedios de las tasas de Síntesis y Degradación en una red es la siguiente: 𝑑𝑁 𝑑𝑡 = 𝛾𝑁 − 𝜆𝑁1 = 𝑁(𝛾 − 𝜆) = 𝑁𝑟𝑝 (3.16) Ahora podemos definir la tasa de reparabilidad de la red (𝑟𝑝). Acorde con los valores que rp puede tomar, se tiene que sí 𝑟𝑝 > 0, la capacidad auto-regenerativa se incrementará y habrá auto-crecimiento de la red; sí 𝑟𝑝 = 0, la capacidad de regeneración se mantiene estable. Sí 𝑟𝑝 < 0 la capacidad decrecerá y el sistema evidenciará un proceso de desgaste o degradación. El anterior razonamiento puede ser instanciado en la síntesis-degradación de las aristas, de forma que para una arista 𝐾(𝑖, 𝑗), con sus correspondientes tasas de síntesis (𝛾′) y degradación (𝜆′), la ecuación correspondiente será: 𝑑𝐾 𝑑𝑡 = 𝛾𝐾 − 𝜆′𝐾1 = 𝐾(𝛾′ − 𝜆′) = 𝐾𝑟𝑝′ (3.17) El promedio de 𝑟𝑝 y 𝑟𝑝' dará la tasa de reparabilidad de la red-𝑅. Cabe anotar que si bien todos los sistemas autopoiéticos son homeostáticos, todos los sistemas homeostáticos pueden no ser autopoiéticos. Razón por la cual, para los que sí lo son, la formulación de homeostasis incluirá las tasas de reparabilidad 𝑟𝑝 y 𝑟𝑝′. 3.3 Síntesis A partir de la proposición de diversos axiomas, en este capítulo se han propuesto medidas para la emergencia, auto-organización, complejidad, homeostasis y autopoiesis, basadas principalmente, en la teoría de la información. Estas medidas son de tipo simple y general, y básicamente representan de forma adecuada lo que miden. El beneficio potencial de estas medidas y aplicaciones son múltiples. Aún si mejores medidas son halladas, las nuestras alcanzan el objetivo de caracterizar de forma general la complejidad, emergencia, auto- organización, homeostasis y autopoiesis. Adicionalmente, se profundiza en aspectos importantes para la homeostasis y la autopoiesis. Para la primera, el resultado es el poder medir, no sólo cómo el sistema cambia o no su equilibrio en el tiempo, sino cómo varia a través de sus cambios de estado. Se complementa la homeostasis, con la descripción de sus procesos activos y pasivos, fundamentales para el entendimiento como mecanismo de auto-regulación, y se brinda un conjunto de indicadores interesantes que la describen detalladamente. Para la autopoiesis, desde la perspectiva de 58 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- redes computacionales de agentes, se profundiza en su significado como proceso de auto- mantenimiento, y se brinda un formalismo. Cabe destacar que a nivel formal, son carentes las expresiones de Homeostasis, de manera que los formalismos para ella desarrollados, son de especial interés. En este mismo sentido se puede opinar de la autopoiesis, desde la perspectiva de la autonomía del sistema respecto de su ambiente. La importancia de desarrollar una medida de complejidad, está en que ella debe estar disponible para permitirnos responder preguntas como: ¿Es un desierto más o menos complejos que una paramo? ¿Cuál es la complejidad de diferentes brotes de influenza? ¿Cuáles organismos son más complejos: predadores o presas; parásitos u hospederos; individuos o sociedades? ¿Cuál es la complejidad de situaciones tan disimiles cómo diferentes como los géneros musicales y los diferentes regímenes de tráfico? ¿Cuál es la complejidad requerida de una compañía para enfrentar la complejidad de un mercado? ¿Cuándo un sistema es más o menos complejo? Dada la dificultad existente en la noción de complejidad, y más aún en su medición, el desarrollo de medidas generales y simples de ella, es una tarea de gran importancia y utilidad. Como se dijo antes, la conveniencia de tener una medida de auto-organización que capture la naturaleza de la dinámica local en una escala global, es especialmente relevante en el campo de la auto-organización guiada u orientada (AOG) (Ay et al., 2012; Prokopenko, 2009b). Para el caso de la autopoiesis, desde las tasas de síntesis y reparabilidad, futuros trabajos pueden probar su utilidad al aplicarla a modelos dinámicos de autopoiesis. Para ello trabajos como el de Bourgine & Stewart (2004) pueden dar bases para la especificación de las tasas. Capítulo 3 59 4. Capítulo 4: APLICACIONES EN SISTEMAS DISCRETOS Y SISTEMAS URBANOS Resumen Definidos los aspectos teóricos, conceptuales y formales de la complejidad en SDC con múltiples agentes, en este capítulo desarrollamos experimentos y aplicaciones computacionales, para mostrar el significado y la utilidad de nuestras medidas de complejidad, auto-organización, emergencia, homeostasis y autopoiesis. El primer casos de estudio corresponde a redes booleanas y autómatas celulares elementales, los cuales han sido extensamente utilizados en la literatura científica como modelos clásicos para representar dinámicas ordenadas, críticas y caóticas, al igual que para detectar cambios de fase. El segundo caso corresponde a la caracterización de dos tipos de modelos de tráfico urbano, uno con coordinación tradicional y otro con un método auto-organizante y fronteras no orientables, que demuestra la utilidad de nuestras medidas en la identificación y descripción de diferentes fases dinámicas del flujo y de la velocidad vehicular. Los resultados muestran que los formalismos desarrollados, efectivamente, miden lo que representan, y describen regímenes desde ordenados hasta caóticos, pasando por los regímenes complejos. Dado que todo puede ser descrito y medido en términos de información, con nuestras medidas se puede caracterizar las diferentes fases de un SDC. 4.1 Sistemas Discretos 4.1.1 Las Redes Booleanas Aleatorias-RBA Las redes booleanas aleatorias (RBA) son modelos computacionales abstractos, originalmente propuestos para el estudio de redes genéticas de regulación (Kauffman, 1969; Kaufmann, 1993). No obstante, han sido de utilidad para abordar el conocimiento de diversos casos, como en sistemas auto-organizantes, en los que se estudia cómo cambios en los nodos y las conexiones afectan la dinámica global de la red. Igualmente, las RBA han sido de 62 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- utilidad para estudiar la auto-organización guiada, la cual consiste en que, dependiendo de la configuración y régimen del sistema dinámico, se pueden realizar diferentes intervenciones para llevarlo a un punto de mayor o menor auto-organización (Aldana- González et al., 2003; Gershenson, 2012a, 2004a). Básicamente, una RBA está formada por 𝑁 nodos unidos por 𝐾 conexiones. Cada nodo tiene un estado booleano, como por ejemplo cero o uno, vivo o muerto, perdió o ganó, etc. El estado futuro de cada nodo está determinado por los estados actuales de los demás nodos, con los que tiene conexión. Su estructura se determina por la red de los nodos interactuantes. Su funcionamiento se basa en tablas de búsqueda que especifican cómo es la actualización del estado de cada nodo, en dependencia con el estado de sus entradas. La conectividad y la actualización de los estados de los nodos se generan aleatoriamente al iniciar la red, pero permanece fija la conectividad en cuanto inicia la dinámica. Las RBA clásicas se actualizan sincrónicamente (Gershenson, 2004b), lo que define una dinámica determinística. Cómo su espacio de estados es finito (2𝑁 estados), tarde o temprano, uno o más estados se repetirán. Estos estados pueden ser definidos como atractores. Los atractores alcanzados, pueden ser puntuales (un estado único), o cíclicos de diferentes longitudes. En dependencia con sus propiedades estructurales o funcionales, se han definido para las RBA tres tipos de regímenes, que han sido estudiados de forma extensa (Gershenson, 2004a). (i) Ordenado: muchos nodos son estáticos, lo que da robustez a la RBA a perturbaciones. (ii) Caótico: muchos nodos son cambiantes, lo que genera RBA frágiles a perturbaciones. (iii) Crítico: algunos nodos cambian, lo que le da a la RBA un potencial de adaptabilidad por el balance de los regímenes caóticos con ordenados. Dinámicas críticas están relacionadas con alta complejidad. Se puede decir que el régimen crítico es el balance entre la robustez del régimen ordenado y la variabilidad del régimen caótico. Desde los 90s, se ha argumentado que la computación y la vida requieren de este balance para computar y adaptarse (Kaufmann, 1993; Langton, 1990). 4.1.2 Midiendo la Complejidad en Redes Booleanas con Diferentes Regímenes y Grados de Conectividad. A continuación, se aplican las medidas propuesta en el capítulo 3 al caso de estudio de las redes booleanas con diferentes regímenes. Los resultados se obtuvieron desde el promedio de 1000 RBA, después de 1000 iteraciones, que iniciaron en estado aleatorio y que dieron como resultado la información de entrada. Seguidamente, las propiedades de Emergencia (𝐸), Auto-organización (𝑆), Complejidad (𝐶) y Homeostasis (𝐻) fueron calculadas para los datos generados en 1000 iteraciones adicionales. Los paquetes BoolNet (Müssel et al., 2010) y entropy (Hausser and Strimmer, 2012), del proyecto R (www.r-project.org), fueron usados como apoyo. Capítulo 4 63 Las series de tiempo de nodos simples fueron evaluadas, y con ello se obtuvo el promedio de los resultados para la red. Para la medida de homeostasis H, d fue medida entre los estados 𝑡 y 𝑡 − 1, para la serie de tiempo total. 4.1.3 Resultados para Redes Booleanas La figuras 4-1 y 4-2 muestran los resultados de las RBA con 100 nodos y conectividad K variable entre 0 − 5. Se observa que para bajos K, existe alta S y H (𝑆, 𝐻 ≅ 1), y baja E y C (𝐸, 𝐶 ≅ 0). Esto refleja el régimen ordenado de RBA, que tienen alta robustez y pocos cambios. Se puede decir que en esta fase existe poca o ninguna información emergente y que hay alto grado de auto-organización y homeostasis, debido a que los estados de los nodos tienen alta regularidad, de manera que el estado siguiente no trae información nueva, es decir, hay una alta probabilidad que el estado más común se repita. Figura 4-1 Diagramas de Cajas para los resultados de 1000 RBA, N=100 con variación de la conectividad K. A. Emergencia B. Auto-organización (S). C. Homeostasis y D. Complejidad (Gershenson and Fernández, 2012a). Para altos K, existe alta 𝐸, baja 𝑆 y 𝐶, y una homeostasis de 0.5. Esto refleja los rasgos del régimen caótico (alta variabilidad y aleatoreidad) de las RBA, donde existe alta fragilidad en la red y muchos cambios. Casi cada bit (un nuevo estado para muchos nodos) trae 64 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- información emergente nueva. En este sentido, los constantes cambios de estado en la red implican baja organización (𝑆 = 0; 𝐸 = 1) o regularidad, y baja complejidad o capacidad de adaptación. Para conectividades medias (2 ≤ 𝐾 ≤ 3) existe un balance entre 𝐸 y 𝑆 (cercanas a 0.5), lo que conduce a alta 𝐶. Esto corresponde con el régimen crítico de RBN, el cual ha sido asociado con complejidad y la posibilidad de vida (Kauffman, 2000). Figura 4-2 𝑬, 𝑺, 𝑪 y 𝑯 promedio para 𝟏𝟎𝟎𝟎 RBA, 𝑵 = 𝟏𝟎𝟎 nodos variando el promedio de conectividad 𝑲 (Gershenson y Fernández, 2012). En síntesis se puede decir que la emergencia aumenta con la conectividad, mientras que la auto-organización disminuye con ella, y la complejidad se hace máxima a niveles intermedios de 𝐾. El aumento de la conectividad hace que surja mayor probabilidad que el estado de un nodo, por la afección de otros, cambie en una mayor proporción. Los experimentos de RBA también fueron llevados a cabo para múltiples escalas de observación. Si bien existen las escalas espaciales y temporales, también existen las escalas numéricas. El cambio de escala se logra a través de un cambio de la base, cómo se explica en el capítulo de formulaciones matemáticas, en el ítem de "Múltiples Escalas y Normalización". Las escalas en las que se evaluaron las medidas fueron 2, 4, 6, 8, 16,32 y 64. El análisis a múltiples escalas se dio debido a que la información puede cambiar drásticamente con la escala observada. Por ejemplo, una cadena 10101010 tiene 𝐼1 = 1, pero a una escala mayor, por ejemplo base 4, la cadena anterior se vuelve 2222 y su información es 𝐼2 = 0. Sobre la anterior base, se calcularon 𝐸, 𝑆, 𝐶 y 𝐻 a múltiples escalas. Los resultados obtenidos no mostraron cambios importantes entre escalas, excepto por una pequeña elevación del valor máximo de 𝐶, el cual que se incrementó moderadamente con la escala. Para 𝐸, se halló Capítulo 4 65 un pequeño decremento a altas escalas, que puede ser atribuido al pequeño tamaño de las cadenas evaluadas (1000 por nodo). Los valores de 𝐻 decrecieron con la escala. Para el cálculo de autopoiesis (𝐴), que relaciona la complejidad del sistema con el medio ambiente, de manera que se pueda observar su grado de autonomía del primero sobre el segundo, se modeló a través de dos redes acopladas que representaron el sistema y su ambiente. La primera RBA, fue una red interna que simuló el sistema, con 𝑁𝑖 nodos y 𝐾𝑖 conexiones promedio. La segunda, fue una red "externa", con 𝑁𝑒 nodos y 𝐾𝑒 conexiones promedio, que simuló el ambiente. Se considera una RBA acoplada como 𝑁𝑐 = 𝑁𝑖 + 𝑁𝑒 nodos y 𝐾𝑖 conexiones. En cada iteración, la RBA externa evolucionó independientemente. Sin embargo, su estado fue copiado a los 𝑁𝑒 representados en la RBA acoplada, los cuales ahora evolucionan parcialmente dependiendo de la RBA externa. Así, los 𝑁𝑖 nodos en la RBA acoplada representaron la red interna que puede ser afectada por la dinámica de la RBA externa, pero no viceversa. Se simularon 50 redes acopladas con 96 nodos en la externa y 32 en la interna. Como base para el cálculo de la autopoiesis se obtuvo la complejidad 𝐶 de cada nodo, que luego fue promediada para cada red. De allí se obtuvo una complejidad interna 𝐶𝑖 y una externa 𝐶𝑒. La autopoiesis A fue obtenida desde la relación 𝐴 = 𝐶𝑖 ⁄ . 𝐶𝑒 Los valores resultantes de A se reportan en la tabla 4.1 y se ilustran en la figura 4.3. En ellas, la autopoiesis A cambia con la conectividad. Se espera, entonces, tener 𝐴 ≈ 1 cuando 𝐾𝑖 ≈ 𝐾𝑒. Cuando la complejidad externa 𝐶𝑒 es alta (𝐾𝑒 = 2 o 𝐾𝑒 = 3), entonces el ambiente (RBAe) domina los patrones de la red acoplada, conduciendo a 𝐴 < 1. Esto es debido a que el sistema al tener menor complejidad que su ambiente no puede adaptarse a los cambios del mismo por su menor autonomía. Es por ello que los patrones que muestra el ambiente se imponen a los patrones que muestra el sistema. En caso contrario, cuando 𝐶𝑖 > 𝐶𝑒, o sea 𝐶𝑒 es más baja (𝐾𝑒 < 2 o 𝐾𝑒 < 3), los patrones producidos por el sistema no son afectados, de mayor manera, por su ambiente; entonces 𝐴 > 1, siempre que 𝐾𝑖 < 𝐾𝑒 (de otra forma, el sistema es más caótico que su ambiente, y la complejidad de los patrones vienen de fuera). A no trata de medir cuanta información emerge en el sistema o en el ambiente, sino como los patrones que son interna o externamente producidos se afectan entre sí. Alta 𝐸 significa que no hay patrones, pues existe un cambio constante. Alta 𝑆 implica que el patrón es estático. Alta 𝐶 refleja patrones complejos (balance entre la inexistencia de patrones y patrones estáticos). En este trabajo, hemos estado interesados en una medición de A a partir de la razón de la complejidad de los patrones que se producen por un sistema, comparada con la complejidad de los patrones producidos por su ambiente. A en este aspecto, es una expresión de la autonomía del sistema respecto de su ambiente, sobre la base de su 66 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- complejidad que le profiere adaptabilidad. Un sistema más autónomo no será influido por los patrones del ambiente. Tabla 4-1 Autopoiesis (A) promedio para 50 conjuntos de redes Ne=96, Ni=32. En rojo: redes acopladas con 𝑨 < 𝟏, en azul redes acopladas con 𝑨 > 𝟏 Los resultados son los mismos que se ven en la fig. 4-3 (Fernández et al., 2014b). 𝐾𝑖 (𝑁𝑖 = 32) 𝐶𝑖 𝐴 = ⁄ 𝐶𝑒 1 𝑺 = 𝟏; 𝑪 = 𝟎; R.ordenado 2 3 𝑺 = 𝑬; 𝑪 = 𝟏; R. crítico 4 5 𝑬 = 𝟏; 𝑪 = 𝟎; R. caótico 1 𝑺 = 𝟏; 𝑪 = 𝟎; R.ordenado 2 𝐾𝑒 (𝑁𝑒 = 96) 3 𝑺 = 𝑬; 𝑪 = 𝟏; R. crítico 4 5 𝑬 = 𝟏; 𝑪 = 𝟎; R. caótico 0.4464025 0.5151070 0.7526248 1.6460345 3.4081967 1.6043330 0.9586809 1.1379227 2.0669794 3.2473729 2.4965328 0.9999926 0.9355231 1.3604272 2.6283798 2.1476247 0.7249803 0.6151742 0.8055051 1.38900630 1.8969094 0.4760027 0.3871875 0.4755580 0.8648389 En la figura 4-3 el tamaño de los círculos es relativo a su valor negativo (en rojo) o positivo (en azul). Figura 4-3 Autopoiesis (A) promedio para 50 conjuntos de redes Ne=96, Ni=32. Valores de A<1 en rojo. Valores en azul para A>1. El tamaño de los círculos indican que tan lejos está A de A=1. Los valores numéricos se muestran en la tabla 4-1 (Fernández et al., 2014b). Capítulo 4 67 4.2 Autómatas Celulares Elementales (ACE) Los autómatas celulares elementales son modelos matemáticos para representar sistemas dinámicos que evolucionan en pasos discretos. Los ACE han sido estudiados de manera igualmente extensa (Wolfram and Gad-el-Hak, 2003; Wolfram, 1984; Wuensche and Lesser, 1992). Ellos pueden ser vistos como casos particulares de RBA (Carlos Gershenson, 2002; Wuensche, 1998), donde los nodos tienen la misma función (regla) y la estructura es regular. Esto es, cada nodo tiene 𝐾 = 3 entradas: ellos mismos y sus dos vecinos más cercanos. Dado que hay 8 posibles estados binarios para las 3 celdas de vecinos en una célula dada, existen 256 "reglas" posibles. Es decir, las posibles combinaciones de funciones booleanas para tres entradas 22^3=28=256. Cada regla puede ser indexada por un número binario de 8 dígitos (Wolfram and Gad-el-Hak, 2003; Wolfram, 1984; Zenil, 2009). Por ejemplo, la regla 30 en notación binaria es como sigue: 302 = 00011110. Su regla de evolución se ilustra en la figura 4-4. En ella se puede observar los valores de los conjuntos de tres vecinos en la parte superior que resultan en la condición inferior de la celda central en la siguiente generación. Figura 4-4 Evolución para la Regla 30 desde condiciones iniciales para grupos de tres vecinos. http://reference.wolfram.com/language/ref/CellularAutomaton.html La evolución de la regla 30 para dos pasos, a partir del acostumbrado estado inicial que parte de una única celda central en estado 1, será numéricamente así: Inicial: 0,0,0,1,0,0,0   Paso 1: 0,0,1,1,1,0,0  Paso 2: 0,1,1,0,0,1,0 Pasados 50 pasos o iteraciones, su forma gráfica es como se ve en la siguiente figura. Figura 4-5 Evolución de la regla 30 en 50 iteraciones http://reference.wolfram.com/language/ref/CellularAutomaton.html 68 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Entre las reglas existentes hay 88 clases que son inequivalentes (Li and Packard, 1990; Wuensche and Lesser, 1992). Han existido varias clasificaciones para la dinámica de los ACE, la más popular ha sido realizada por Stephen Wolfram, quien distinguió el comportamiento evolutivo cualitativo a través del establecimiento de 4 clases (Wolfram, 1984). Para Wólfram, las clases I y II evolucionan hasta aproximarse a dinámicas ordenadas de configuraciones estables, homogéneas, simples y/o periódicas, que terminan con celdas de un mismo valor. Las reglas de la clase I tienden a atractores puntuales desde todos los estados iniciales (p.e. reglas 0, 248 y 32), mientras que las reglas de la clase II tienden a atractores cíclicos (p.e. reglas 34,1 y 4 definidas cómo periódicas). Las reglas clase III evolucionan a estados caóticos (p.e. reglas 90, 30 y 165). La evolución de las reglas clase IV generan dinámicas críticas (p.e. 54, 110 y 193); algunas de ellas han sido definidas como capaces de realizar computación universal (Cook, 2004). 4.2.1 Experimentos con ACE. Los experimentos de ACE evaluaron 50 instancias de algunas reglas de las cuatro clases de Wolfram. Se consideraron 256 celdas, 212 pasos iniciales y 212 pasos adicionales fueron usados para generar los resultados. Finalmente, 18 reglas de ACE, escogidas como representativas de cada una de las clases, fueron evaluadas. Una relación de ellas se da en las tablas 4-2 a 4-5. Capítulo 4 69 Tabla4-2 Reglas Incluidas de la Clase I. Condiciones de Evolución y Representación de 20 Iteraciones. Regla Cero ACE Clase I Evolución Regla 8 Regla 32 Regla 40 Regla 128 http://reference.wolfram.com/language/ref/CellularAutomaton.html 70 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Tabla 4-3 Reglas Incluidas de la Clase II. Condiciones de Evolución y Representación de 20 Iteraciones ACE Clase II Evolución Regla 1 Regla 2 Regla 3 Regla 4 Regla 5 http://reference.wolfram.com/language/ref/CellularAutomaton.html Capítulo 4 71 Tabla 4-4 Reglas Incluidas de la Clase III. Condiciones de Evolución y Representación de 20 Iteraciones ACE Clase III Evolución Regla 41 Regla 54 Regla 106 Regla 110 http://reference.wolfram.com/language/ref/CellularAutomaton.html 72 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Tabla 4-5 Reglas Incluidas de la Clase IV. Condiciones de Evolución y Representación de 20 Iteraciones ACE Clase IV Evolución Regla 18 Regla 22 Regla 45 Regla 161 http://reference.wolfram.com/language/ref/CellularAutomaton.html Las simulaciones fueron programadas en R con el uso de los paquetes BoolNet (Müssel et al., 2010), CellularAutomaton (Hugues, 2012) y entropy (Hausser and Strimmer, 2012). 4.2.2 Resultados ACE Los resultados para diferentes clases (I1, I2, I4, e I8) se muestran en la figuras 4-6 a 4-9. En ellas se puede observar que las reglas de la clase I tienen 𝐸 = 𝐶 = 0 y 𝑆 = 𝐻 = 1 para todas las escalas. Esto es debido a que en la evolución de los nodos no hay cambio de estado, así que el ACE alcanza un estado estable. El tiempo que toma ello, es el de básicamente una iteración. Capítulo 4 73 Algunas reglas de la clase II se comportaron de manera similar a la clase I, por ejemplo la regla 4. Otras, como la 1 y 5, tuvieron emergencias medianamente altas (𝐸 > 0,5 en 𝑏 = 1). Esto se debío a que la mayoría de las células oscilaron entre cero y uno en cada iteración. Sin embargo, al incrementar la escala (𝑏 ≥ 2), los patrones se tornaron más regulares y se comportaron como las reglas de la clase I (𝐸 = 𝐶 = 0 ; 𝑆 = 𝐻 = 1). Dado que medimos el cambio de la información por nodo, las reglas cómo la 2 y 3 parecen tener alta 𝐶, a pesar de que sus patrones tienden a la regularidad (Gershenson and Fernández, 2012a). Las reglas de la Clase III, tienden a tener alta E. La regla 18 es particular, dado que tiene una proporción mucho mayor de ceros que unos, que se acumulan en triángulos. Su 𝐸 es menor y su 𝑆, 𝐶, y 𝐻 es mayor, comparativamente con las otras de la clase III. Esto es notable aún más a escalas mayores, en razón a que las otras reglas de la clase III tienen un porcentaje más balanceado de ceros y unos. La regla 161 está en el medio, y tiene un porcentaje ligeramente mayor de unos que ceros. Las reglas de la Clase IV se comportan de manera similar a las de la clase III para 𝑏 = 1, con alta 𝐸, baja 𝑆 y 𝐶. La regla 54 es una excepción para 𝐻, donde la alternancia de patrones del entorno (éter) genera una H anti-correlacionada. Es decir que los patrones espaciales en un tiempo determinado comparado con el siguiente, son completamente opuestos. El éter regular de la regla 54, y la regularidad de sus deslizadores (gliders), genera un decremento en E a altas escalas, por razones similares a las de las reglas 2 y 3. Para la regla 41 la 𝐸 decrece y se incrementa su 𝑆 y 𝐶 con la escala, y se mantiene una 𝐻 no correlacionada. Esto revela patrones más regulares en altas escalas, similar a la regla 18. La regla 110 se comporta similar a la regla 41: 𝐸 es baja, 𝑆 y 𝐶 incrementan para escalas altas. La diferencia está en 𝐻. Para la regla 110, 𝐻 es anti-correlacionada para 𝑏 ≥ 4. Esto puede ser nuevamente explicado por lo que pasa en el éter, el cual tiene un periodo de 7 para la regla 110. 74 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Figura 4-6 Promedios de 50 ECA para la 19 reglas, N=256, b=1. (Gershenson and Fernández, 2012a) Figura 4-7 Promedio de 50 ACE para 19 reglas, N=256, b=2 (Gershenson and Fernández, 2012a). Capítulo 4 75 Figura 4-8 Promedio de 50 ACE para 19 reglas, N=256, b=4 (Gershenson and Fernández, 2012a). Figura 4-9 Promedio de 50 ACE para 19 reglas, N=256, b=8 (Gershenson and Fernández, 2012a). El comportamiento de 𝐸, 𝑆, 𝐶 y 𝐻 a múltiples escalas (bits 1,2,4,8) se puede ver en la figura 4-10. La regla 0 no cambia con la escala. La regla 1 tiene alta 𝐸, baja 𝑆 y 𝐻, y media 𝐶 para 𝑏 = 1, pero es similar a la regla 0 para escalas grandes. En la regla 11, la 𝐸 y 𝐻 decrecen y se 76 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- incrementan su 𝑆 y 𝐶 con la escala. La regla 30 tiene alta E, baja 𝑆, 𝐻 y 𝐶, para todas las escalas. Figura 4-10 Perfiles de E,S,C y H a Escalas 1,2,4,8. . (Gershenson and Fernández, 2012a) 4.3 Medición de la Complejidad en Sistemas Urbanos- Semáforos Auto-organizantes 4.3.1 La Movilidad Vehicular en las Ciudades Modernas y los Métodos Auto-Organizantes para su Solución Vivir en un mundo urbanizado ha traído grandes ventajas a sus habitantes, como la solución a muchas necesidades básicas que se deben satisfacer. Se estima que esta satisfacción impulsa a las personas a buscar cada vez más las ciudades, tanto que para el 2050 el 70% de las Capítulo 4 77 personas vivirán en ciudades. En la actualidad, la mitad de las ciudades en el mundo tienen entre 100 y 500 mil personas, y el 10% de la población mundial viven en megalópolis, ciudades con más de 10 millones de habitantes (http://goo.gl/FYoIfP). Ante el gran número de personas que necesitan desplazarse en las ciudades, su planeación ha tenido cómo uno de sus ejes centrales la movilidad vehicular. La movilidad vehicular es un problema complejo en el que el flujo que cambia constantemente, interactúa con transeúntes y la sincronización de los semáforos. Al ser un problema no estacionario, se requiere de un enfoque adaptativo para su solución, como por ejemplo un enfoque auto- organizante propuesto por Gershenson (Gershenson, 2005). Las preguntas que surgen son: ¿Cómo la auto-organización puede ser orientada para lograr flujos eficientes de tráfico? ¿Cuál puede ser su régimen deseado? ¿Cómo puede ser medida en términos de complejidad? En el contexto anterior, se presenta un modelo de auto-organización guiada. Además, se hace una comparación de la complejidad entre los métodos tradicionales de sincronización, comparado con los semáforos auto-organizantes (Gershenson, 2005). Los métodos y resultados presentados son acordes con lo expresado en Zubillaga et al. (2014). 4.3.2 Modelo de Tráfico El modelo de tráfico usado fue desarrollado previamente por (Gershenson and Rosenblueth, 2012), basado en autómatas celulares elementales explicados en la sección 4.2 (Wolfram and Gad-el-Hak, 2003; Wuensche and Lesser, 1992). Este modelo es determinístico, en tiempo y espacio discreto, velocidad uno o cero y aceleración infinita. Su condición sintética permite medir claramente las diferentes fases dinámicas. Se aclara que el modelo es descriptivo, no predictivo. Cada calle es representada por un autómata celular elemental (ACE), acoplado en las intersecciones. Cada ACE contiene un número de celdas que toman valor de cero (vacío) o uno (vehículo). El estado de la celda se actualiza en correspondencia con los estados previos de la celda y de los vecinos más cercanos. Muchas celdas vacías adelante del vehículo permiten su avance. El comportamiento fue modelado con la regla 184, la cual es uno de los modelos más simples de tráficos en autopistas. Sus reglas de actualización según sus dos vecinos se dan en la tabla 4-7. Tabla 4-6 Reglas de Actualización de ACE 𝑡 − 1 000 001 010 011 100 𝑡184 0 0 0 1 1 𝑡252 0 0 1 1 1 𝑡136 0 0 0 1 0 78 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 101 110 111 1 0 1 1 1 1 0 0 1 Los modelos ACE-184 se mueven a la derecha. Otras direcciones son obtenidas cambiando las posiciones de los vecinos con el uso de la misma regla. En las intersecciones se usaron dos reglas más. El no avance de los vehículos antes de una luz roja, se modeló con la regla ACE- 252. La regla ACE-136 se utilizó para modelar los vehículos que cruzaban la intersección después de una luz roja. Las calles con luz verde utilizaron la regla ECA-184. La figura 4-11 muestra detalles de las reglas. Figura 4-11 Reglas de autómatas celulares elementales usadas en el modelo de tráfico y su función cómo interruptores de los semáforos (Zubillaga et al., 2014). El modelo de tráfico es conservativo. Esto es, la densidad de vehículos ρ (proporción de celdas con un valor de 1) es constante en el tiempo. El promedio de velocidad 𝑣 se calcula por el número de celdas que cambian desde cero a uno (movimiento ocurrido entre el tiempo 𝑡 y 𝑡 + 1, dividido por el número total de vehículos (celdas con 1). 𝑣 = 0 sí el vehículo no se mueve, 𝑣 = 1 cuando se mueven. El flujo 𝐽 se define como la velocidad 𝑣 multiplicada por la densidad ρ; 𝐽 = 𝑣 ∗ 𝑝. 𝐽 = 0 cuando no hay flujo. Es decir, no hay vehículos en la simulación (ρ = 0), o los vehículos están parados (𝑣 = 0). Una 𝐽 máxima ocurre cuando todas las intersecciones son cruzadas por vehículos. En el estudio de escenarios 𝐽𝑚𝑎𝑥 = 0.25. Esto es debido a que para una calle única (ACE-184), los vehículos necesitan espacio para avanzar, limitando la 𝐽𝑚𝑎𝑥 = 0.5. Sin embargo, cuando dos calles se intersectan, una de ellas tiene el pare mientras la otra fluye, reduciendo 𝐽𝑚𝑎𝑥 a 0.25. Capítulo 4 79 Teóricamente, la velocidad y flujo óptimos para una proporción dada en un problema de coordinación, deben ser los mismos 𝑣 y 𝐽 para las intersecciones aisladas. Esto es un límite superior. Lo que implica que cada intersección interactúa con sus vecinos si es tan eficiente como cuando está vacía. Esto puede ser visto con curvas de optimización (definidas como el mejor rendimiento para una intersección aislada, para diferentes densidades, menos el rendimiento actual del sistema de tráfico (Gershenson and Rosenblueth, 2012), que visual y analíticamente compara los diferentes resultados con el óptimo teórico. Las fronteras fueron no orientables (como en una banda de Möbius o una botella Klein) por sugerencia de Bedau and Humphreys (2008), lo que mantiene la entrada y la densidad de vehículos. Si bien estos son deterministas, en la práctica se destruyen las correlaciones que se dan en fronteras cíclicas. El resultado es equivalente a una sola calle con 100 intersecciones. Una muestra de la aplicación de las fronteras no orientables se observa en la figura 4-12. Figura 4-12 Fronteras no orientables para una grilla de cuatro por cuatro (Zubillaga et al., 2014) También es posible extender el modelo a una grilla hexagonal para permitir intersecciones más complejas (Gershenson and Rosenblueth, 2012). El problema de coordinación de los semáforos es EXPTIME-Completo (Gershenson, 2005). Esto implica que la optimización de grandes redes de tráfico es inviable. Además es no- estacionario. Aún si se encontrara una solución óptima, sería obsoleta en segundos. Como alternativa se tiene la adaptación, siendo la auto-organización un método útil para construir sistemas adaptativos. La solución común es la "ola verde", donde los semáforos son sincronizados con el mismo periodo y fase se ajusta acorde con la velocidad esperada de vehículos. Esto es bueno para dos direcciones, mientras que las otras quedan detenidas por estar sus semáforos en rojo, es decir en fases anti-correlacionadas. En densidades medias, la ola verde causa largas esperas y 80 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- colas. Igualmente, cuando las densidades cambian, los vehículos no van a la velocidad esperada. Las fases halladas en la ola verde son dos: intermitentes (algunos vehículos detenidos, algunos se mueven) y embotellamiento (todos los vehículos parados). Por su parte el método auto-organizante se basa en 6 reglas, simples y por intersección, con el que se alcanza un óptimo global de coordinación (22,33,38,40). A continuación se esquematiza una intersección y sus parámetros (Fig. 4.13). Figura 4-13 Esquema de una intersección y parámetros de medición para la definición de reglas en semáforos auto-organizantes (Zubillaga et al., 2014) Las reglas son: i. ii. iii. iv. v. vi. En cada tic se adiciona a un contador el número de vehículos que se aproximan o esperan en una luz roja a una distancia 𝑑. Cuando el contador excede el umbral 𝑛, se enciende la luz verde. Cuando la luz verde se enciende el contador se reinicia a cero. La luz se mantiene en verde un tiempo mínimo 𝑢 Si los vehículos son pocos (𝑚 o pocos, pero más que cero) y están a una distancia corta 𝑟, la luz no pasará a verde. Sí no hay ningún vehículo que se aproxime a la luz verde dentro de una distancia 𝑑, y hay por lo menos un vehículo aproximándose, en la otra dirección, a la luz roja dentro de la distancia 𝑑, la luz cambiará a verde para este último. Sí hay un vehículo parado en una calle a una distancia corte 𝑒, ante una luz verde, entonces la luz pasará a roja. Si hay vehículos parados en ambas direcciones, a una distancia corta 𝑒, entonces ambas luces pasan a rojo. Una vez, una de las direcciones esté libre, se restablecerá la luz verde en esa dirección. 4.3.3 Resultados Los resultados de 𝑣 y 𝐽 para fronteras no orientables se muestran en la figura 14, en la que se indica con la línea punteada las transiciones de fase del método auto-organizante, cuando se incrementa la proporción. Estas fases fueron halladas en estudios previos para fronteras cíclicas y se colocan para propósitos ilustrativos. Ellas son: (i) flujo libre (ningún vehículo Capítulo 4 81 detenido 𝑣 = 1), (ii) cuasi-libre (casi ningún vehículo detenido), (iii) subutilizado- intermitente (algunos vehículos detenidos, las intersecciones no alcanzan su flujo máximo), (iv) capacidad completa-intermitente (algunos vehículos detenidos, 𝐽𝑚𝑎𝑥, todas las intersecciones tienen vehículos siempre), (v) sobre-utilizado-intermitente (algunos vehículos parados, las intersecciones tienen que restringir el flujo en ambos sentidos, usando la regla 6), (vi) cuasi-embotellamiento (casi todos los vehículos parados, pero con algunos espacios de "pelotones" o convoys), (vii) embotellamiento (todos los vehículos parados y las intersecciones bloqueadas desde condiciones iniciales; 𝑣 = 0). En fronteras cíclicas, las transiciones de fase se dieron a valores de ρ: 0.15, 0.22, 0.38, 0.63, 0.77 y 0.95. En la figura 4-14 se observa claramente que el método auto-organizante (en azul) se acerca estrechamente con el óptimo teórico y mantiene mayores velocidades y flujos que la ola verde (en verde). Con ello se evidencia la gran ventaja y bondades del método auto-organizante. Entre tanto, para la ola verde la fase de embotellamiento llega rápidamente a bajas densidades. Figura 4-14 Resultado de la simulación para fronteras no-orientables para la ola verde y método auto- organizante: promedio de velocidad 𝒗 y promedio de flujo 𝑱 para diferentes densidades de 𝝆. Las curvas óptimas se muestran con línea punteada discontinua (Zubillaga et al., 2014). Se puede ver que el método auto-organizante está por encima de la curva optima en la fase de capacidad completa-intermitente (𝐽 > 𝐽𝑚𝑎𝑥). Esto es debido a que un vehículo en una calle, puede alcanzar la luz verde para avanzar sin dejar espacio entre él y el último vehículo en la calle que alcanzó la luz roja. Para una clara diferenciación de los métodos en términos de complejidad, las métricas 𝐸, 𝑆, y 𝐶 fueron aplicadas a tres componentes del modelo: (i) los intervalos de cambio de luz en 82 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- los semáforos-ICLS, (ii) los intervalos de autos en la intersección-ICI y (iii) los intervalos de carros en la calle-ICC. Los resultados se presentan en las figuras 4-15 a 4-17 En la figura 4-15 se puede ver que para ICLS el método de la ola verde tiene periodos cíclicos para los semáforos (𝐸 = 𝐶 = 0 y 𝑆 = 1, puntos verdes). Se aclara que nuestra medida de 𝑆 no distingue entre organización interna (auto-organización) o externa, cómo en este caso que el método depende de un control central. Teniendo un extremo de regularidad en los intervalos en los semáforos representado por 𝑆 = 1, se entiende que no puede haber adaptación a los cambios en el flujo de tráfico. Entre tanto, el método auto-organizante se adapta constantemente a los cambios en la demanda (puntos azules), como puede ser visto desde la variación de las medidas para diferentes densidades (𝜌). Los intervalos en el semáforo son más irregulares (muy alta 𝐸, muy baja 𝑆) en las fases de flujo cuasi-libre (𝜌[0 − 0.15]) y cuasi-embotellamiento (𝜌[0.77 − 0.95]), mientras que alcanza regularidad (muy baja 𝐸 y 𝐶, y máxima 𝑆) para la fase de capacidad-completa intermitente, dado que la demanda viene en todas direcciones. 𝐶 es alta para todas las fases, pero en las fases de capacidad- completa intermitente (𝜌[0.22 − 0.38]) y embotellamiento (𝜌[0.95 − 1]) exhibe un comportamiento similar a la ola verde en estas dos fases. Figura 4-15 Emergencia (E), Auto-organización (S) y complejidad para los intervalos de cambio de luz en semáforos para fronteras no-orientables (Zubillaga et al., 2014). Las líneas verticales señalan la transición de cada una de las 7 fases del flujo (libre, cuasi-libre, subutilizado, capacidad completa, sobre-utilizado, cuasi-embotellamiento, embotellamiento) Para ICI (Fig. 4-16), las medidas de 𝐸, 𝑆, y 𝐶 para diferentes densidades 𝜌 se observó que son similares en las fases de capacidad-completa-intermitente y embotellamiento, dado que hay vehículos que constantemente cruzan la intersección o hay vehículos detenidos. 𝐸 y 𝐶 son altas para bajas densidades en ambos métodos, y se hacen más regulares con 𝜌 (incremento de 𝑆), hasta alcanzar un máximo en la fase de capacidad-completa-intermitente para el método auto-organizante y para la fase de embotellamiento en la ola verde. El método auto- organizante nuevamente incrementa 𝐸 y 𝐶 hasta la fase de cuasi-embotellamiento (𝜌 > 0.6; 𝑝 < 0.9), dado que los vehículos cruzando son menos regulares y se forman pelotones espaciados, solo para decrecer de nuevo en la fase de embotellamiento. Capítulo 4 83 Figura 4-16 Emergencia (E), Auto-organización (S) y complejidad para los intervalos en intersecciones en semáforos para fronteras no-orientables (Zubillaga et al., 2014). Las líneas verticales señalan la transición de cada una de las 7 fases del flujo (libre, cuasi-libre, subutilizado, capacidad completa, sobre-utilizado, cuasi-embotellamiento, embotellamiento) En las últimas pruebas (Fig. 4-17), los intervalos de los vehículos en diferentes calles (no intersecciones) se escogieron aleatoriamente en cada simulación. El comportamiento de las medidas fue similar al descrito para las intersecciones. Sin embargo, 𝑆 = 1 se alcanza sólo en la fase de embotellamiento (𝜌 > 0.95). Parece haber un mínimo de 𝐸 ≈ 0.2 para la fase de capacidad-completa-intermitente. A pesar que se forman pelotones grandes, existen espacios grandes entre ellos. 𝐶 es alta para todas las fases, excepto para el embotellamiento para ambos métodos. Figura 4-17 Emergencia (E), Auto-organización (S) y complejidad para los intervalos en las calles en semáforos para fronteras no-orientables (Zubillaga et al., 2014). Las líneas verticales señalan la transición de cada una de las 7 fases del flujo (libre, cuasi-libre, subutilizado, capacidad completa, sobre-utilizado, cuasi-embotellamiento, embotellamiento) Se pudo constatar que el flujo en tráfico urbano es muy complejo. No obstante, se pueden identificar diferencias entre algunas fases del método auto-organizante. En otras palabras, la fase dinámica puede tener un impacto directo sobre 𝐸, 𝑆 y 𝐶, lo cual puede ser potencialmente utilizado para identificar las fases dinámicas en sistemas más reales de tráfico donde las transiciones no son tan fáciles de identificar. 84 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 4.4 Discusión La siguiente discusión abordará los tópicos de la utilidad general de las medidas a partir de lo observado en sistemas discretos. La relación entre complejidad e información. Los resultados de las múltiples escalas y la relación con otras medidas. Algunos tópicos serán abordados de nuevo en las siguientes aplicaciones, con el fin de ahondar en su significado a partir de los resultados obtenidos en cada una de ellas. 4.4.1 La Aplicación de Las Medidas Aunque pueda ser deducible, nuestras medidas no se enfocan en qué elemento interactúa con cual, cómo y cuándo. En contraste, estamos más interesados en las propiedades globales y tendencias de los estados de los elementos. En este sentido, queremos capturar dinámicas globales a través de medidas que la sintetizan en forma adecuada. Igualmente, estamos interesados en su cambio con la escala. Si bien en al emplear E,S,C, A y H se debe considerar la perdida de información subyacente en el fenómeno que se describe, también se debe considerar el tipo de información en el que estemos interesados, ya que la relevancia es también parcialmente dependiente del observador (Gershenson, 2002). Para el caso de semáforos auto-organizantes, se recalca que dada la generalidad y simplicidad de S, ella no puede distinguir entre un sistema auto-organizado (p.e. por un controlador externo) y un sistema auto-organizante, el cual se adapta constantemente a las demandas de cambio. Esto puede ser comparado con el caso de la temperatura, la cual no indica sí el calor es generado externa o internamente en el sistema, pero sería muy útil saberlo. 𝑆, cómo medida, tiene que ser interpretada adecuadamente en el contexto específico para explotar su utilidad. Una vez definido el contexto, el complemento de las medidas de 𝐸 y 𝐶, son de gran utilidad para la caracterización de fases dinámicas en sistemas con y sin control central. Podemos afirmar que actualmente no hay un acuerdo en la comunidad científica sobre como los términos de complejidad, emergencia y auto-organización deben ser definidos y usados, dada la evidente contradicción entre autores. En este aspecto, nuestras medidas muestran desde los resultados obtenidos su potencial y esclarecen su utilidad, por lo que vale la pena emplearlos. No obstante, acatamos las restricciones que puedan tener y que el revisor manifiesta. En este sentido, no nos apartamos que las sugerencias mostradas por el revisor sirvan para motivar el desarrollo de una mejor terminología; situación que también es una contribución. 4.4.2 La Complejidad cómo Balance Cómo ha sido propuesto por varios autores y confirmado en nuestros experimentos en sistemas discretos, la complejidad puede ser vista como balance entre el orden y el desorden (Gershenson and Fernández, 2012; Kaufmann, 1993; Langton, 1990; Lopez-Ruiz et al., 1995). Capítulo 4 85 Desde esta perspectiva, 𝐶 como un balance entre 𝐸 y 𝑆, cumple con esta condición, que ha sido definida también para la llamada criticalidad. Podría sonar cómo contradictorio o contra-intuitivo definir la emergencia cómo el opuesto de auto-organización, dado que ambas están presentes en muchos fenómenos complejos. Sin embargo, cuando se toma un extremo de los dos (𝐸 o 𝑆), la expresión del otro es muy baja. Es decir, si se tiene emergencia muy alta, se tendrá auto-organización muy baja. Es precisamente cuando 𝐸 y 𝑆 están en balance, que la complejidad se maximiza, pero esto no significa que ambos hayan sido máximos. Un punto importante para resaltar es que existen visiones opuestas de información. Shannon definió la información como entropía (Equivalente a nuestra 𝐸). Wiener (1948) y von Bertalanfy (1968), definieron información como lo contrario, negentropía (Equivalente a nuestra 𝑆). En este sentido, nuestra medida de complejidad C reconcilia estas dos visiones opuestas, como balance entre la regularidad u orden (𝑆) y el cambio o caos (𝐸), donde la complejidad es máxima. Los sistemas dinámicos cómo autómatas celulares y redes booleanas, tienen máxima 𝐶 en la región donde su dinámica es considerada cómo más compleja. 4.4.3 Múltiples Escalas y Perfiles En general, para redes booleanas aleatorias-RBA los resultados para 𝐸, 𝑆, 𝐶 y 𝐻 cambian poco con la escala, precisamente por ser generadas de manera aleatoria. Los autómatas celulares elementales-ACE son mucho más regulares en cuanto a estructura y función (expresada en la dinámica). Por ejemplo, ACE con dinámicas ordenadas y caóticas (Clase I y III, respectivamente), no cambian mucho con la escala. En el medio (Clase II), existen reglas de interesantes patrones que demuestran que la escala tiene importancia. Para RBA y ACE con dinámicas caóticas existe una ligera reducción de la emergencia con la escala. Este es un efecto de tamaño finito, debido a que cadenas cortas tienen una emergencia más reducida que las largas. Es interesante, dado que muchos estudios en sistemas dinámicos, especialmente analíticos, están hechos asumiendo sistemas infinitos. Sin embargo, muchos sistemas reales son mejor modelados como finitos, y aquí hemos visto que la longitud de la cadena -no sólo la escala- puede jugar un papel relevante en determinar la emergencia y la complejidad de los sistemas. Una de las implicaciones del efecto finito, es que a grandes escalas se requiere menor información para describirlo. Extrapolando las "escalas mayores" ( 𝑏 → ∞) implica no información (𝐼 → 0): sí todo esta contenido, cuando hay información no necesaria para describirlo (Gershenson, 2012b). Los perfiles de 𝐸, 𝑆, 𝐶 y 𝐻 a múltiples escalas (figura 4-10), dan una mejor percepción que las escalas particulares o por separado (figura 4-6 a 4-9). Esto se pudo ver claramente en las reglas de la clase II, donde la E es alta a una escala, pero mínima a escalas altas. El cambio de las medidas a través de las escalas, es una herramienta de gran utilidad en el estudio de 86 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- sistema dinámicos (Weeks, 2010; Wolpert and Macready, 1999), cómo se ha formalizado por la teoría de Krohn-Rhodes (Egri-Nagy and Nehaniv, 2008; Rhodes, 2009). Por ejemplo, un sistema con alta (pero no máxima) 𝐸, a través de las escalas sería más emergente que un sistema con E muy alta en una escala, pero con 𝐸 muy baja en otra escala. Con soporte en los resultados visuales de ACE, se puede observar como que existe una condición inicial de una única celda "viva" con emergencia muy baja (E). Este es el caso de la regla cero de la clase I. No importaría con cuanta información se alimente al autómata, esta se perdería en una sola iteración. La regla 1 de la clase II produce gran cantidad de información, y emerge un patrón estriado (Tabla 4-3). No obstante, para base 2, este patrón desaparece. La emergencia, entonces, es mínima para escalas base 2 (Figura 4-7). La regla 110 de la clase IV produce estructuras y patrones complejos, con deslizadores que interactúan de manera regular en su éter. Se puede decir, que estos patrones son más emergentes, dado el incremento de información que se produce. La Regla 30 de la clase III produce aún más patrones, dado que su comportamiento seudo-aleatorio produce máxima información. La emergencia muy alta será la que produce mayor información partiendo de una menor a la inicial. La auto-organización (𝑆) definida como reducción de la información de Shannon, es mínima cuando un patrón ordenado es convertido en un patrón desordenado. Los sistemas caóticos, cómo la regla 30, tienen mínima 𝑆. La máxima auto-organización ocurre cuando un patrón altamente desordenado se convierte en altamente ordenado. Sistemas que tienden a estados estacionarios, I de ACE, exhiben máxima auto- organización(Gershenson and Fernández, 2012a; Wolfram, 1984). tales como los de la clase Casos como el perfil de complejidad de Bary-Yam mide la complejidad cómo la cantidad de información requerida para describir un sistema en diferentes escala (Bar-Yam, 2004b). Desde nuestra perspectiva, esta puede ser mejor llamado "perfil de información" o "perfil de emergencia". Inspirados en el perfil de complejidad (Información) de Bar-Yam, el perfil sigma fue propuesto para medir la organización a múltiples escalas (Carlos. Gershenson, 2011). Siguiendo el enfoque de Lopez-Ruiz et al. (1995) que fue usado en nuestros desarrollos, estos dos perfiles pueden ser combinados para desarrollar un nuevo perfil de complejidad. Esto podría ilustrarse por las propiedades de complejidad a diferentes escalas. 4.4.4 La Complejidad su Diferencia con el Caos Algunos enfoques relacionan la complejidad con la alta entropía (para nosotros, alta emergencia), p.e. el contenido de información Bar-Yam 2004; Delahaye and Zenil 2007. Sin embargo, es importante destacar que la complejidad no debe ser confundida con Caos (Gershenson 2013). El Caos es alta entropía (Información de Shanon) (Prokopenko et al., 2009). Una alta entropía (alta emergencia E) implica mucho cambio, lo que genera que Capítulo 4 87 patrones complejos sean destruidos. De otra parte, baja entropía (alta auto-organización S) previene que patrones complejos emerjan. En los experimentos de ACE puede ser visto que las reglas caen generalmente en dos categorías: las que computan algo (𝐸 > 0) o las que no (𝐸 = 0). Esto está relacionado con el principio de equivalencia computacional de Wolfram (Wolfram and Gad-el-Hak, 2003; Zenil, 2009), el cual conjetura que los sistemas dinámicos son capaces (sistemas complejos) o no (sistemas simples) de hacer computación universal8. Desde los ACE, sólo la regla 110 ha sido capaz de producir computación universal (Cook, 2004), pero otras reglas también posiblemente sean capaces. No solo las de la clase IV, también reglas de la clase III; sin embargo, la variabilidad les hace difícil guardar información. Aun así, hay técnicas para lograr dinámicas complejas de sistemas caóticos, por ejemplo, con memoria (Martínez et al., 2012) o con señales externas regulares (Luque and Solé, 1998, 1997) (Control del Caos). La clase II podría ser considerada muy regular para computar universalmente, pero adicionando ruido o una señal compleja podría ser posible que ella llevara a cabo computación universal. La clase I es muy estática, ella podría requerir más de la computación que viniese de fuera del sistema. En general, por medio de diferentes técnicas se puede direcciona un sistema dinámico a la alta complejidad. 4.4.5 Homeostasis En muchos casos, la homeostasis (𝐻) ha sido relacionada con auto-organización. Alta S indica baja variabilidad, lo cual es característico de una alta 𝐻, esto fue visible en RBA con K menores a 2. Por otra parte, baja S, en muchos casos, está acompañada por una 𝐻 no correlacionada 𝐻 ≈ 1 2𝑏⁄ . Esto también es un rasgo de sistemas con dinámicas caóticas (Alta E). Sin embargo, algunos ACE con baja S tienen una 𝐻 correlacionada 𝐻 > 1 2𝑏⁄ o anti-correlacionada 𝐻 < 1 2𝑏⁄ . Esto es característico de estructuras complejas que interactúan en un éter regular. Al respecto, diferentes reglas de ACE pueden tener alta 𝐶 en una escala particular debido a 𝑆 ≈ 0.5. No obstante, reglas con una H que se desvía de 1 2𝑏⁄ muestran un rasgo de estructura compleja sobre un éter. El éter facilita la computación, de manera que la universalidad puede ser explorada más fácilmente. La desviación de 𝐻 ≈ 1 2𝑏⁄ podría ser usada para explorar el inmenso espacio posible de los sistemas dinámicos. Por otra parte, al implementar un análisis de componentes principales, encontramos que para las RBA la desviación estándar de 𝐻 está inversamente correlacionada con 𝐶 (𝐸 × 𝐴𝑜 en 8 Capacidad de realizar cualquier proceso computable que se le pida, sin necesidad de cambiar su estructura, ni su función de transición local. 88 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- fig. 4-18). RBAs en un régimen ordenado tienen consistentemente alta 𝐻. Entre tanto, RBAs en régimen crítico tienen una 𝐻 variable. Figura 4-18 Análisis de Componentes Principales. Las propiedades en Azul y las conectividades en Rojo. 4.4.6 Autopoiesis y el Requisito de Variedad de Ashby. Nuestro enfoque de autopoiesis cómo la razón entre las complejidades del sistema y el ambiente, nos lleva a considerar la ley de la Variedad Requerida de Ashby (Ross Ashby, 1960a). En ella se establece que un controlador activo requiere mucha variedad (número de estados), para que el sistema controlado sea estable. Por ejemplo, si un sistema puede estar entre cuatro estados, su controlador debe discriminar entre los cuatro estados para regular la dinámica del sistema. Desde esta perspectiva, la medida propuesta de autopoiesis está relacionada con el requisito de variedad requerida. Cuando un sistema tiene 𝐴 > 1, tiene una complejidad (variedad) mayor que su ambiente, lo que refleja una mayor autonomía. Es decir, la complejidad del sistema es mayor que la del ambiente, en consecuencia, un controlador exitoso para este sistema debería tener 𝐴 > 1 (a múltiples escalas; (Gershenson, 2011)). Para que un controlador fuese más eficiente debería tener 𝐴 → 1, lo que le daría más variedad que el Capítulo 4 89 sistema. Lo ideal es que el sistema tenga la capacidad de acoplar su complejidad a la del entorno, o sistema que se está tratando de controlar. Si el entorno cambia mucho, el sistema también. En cuanto a los sistemas vivos desde la perspectiva de la variedad requerida, la complejidad y la autopoiesis existen elementos de importancia. Primero los seres vivos requieren de cumplir con la ley de variedad requerida. Segundo, se ha observado que la variedad puede ser vista como un sinónimo de complejidad (Bar-Yam 2004). Tercero, los organismos pueden ser descritos como sistemas de control. Estos elementos nos llevan a afirmar que la ley de variedad requerida también se encuentra en la vida, dado que los organismos deben hacer frente a la complejidad o variedad de su ambiente a diferentes escalas. Sobre las anteriores bases, la relación de la ley de variedad requerida se relaciona con nuestra medida de autopoiesis. Aún más, todo ello ha dado pie para que Gershenson (2014), haya armonizado y generalizado estos elementos en lo que puede ser denominado la ley de complejidad requerida: un activo y eficiente controlador requerirá por lo menos la misma complejidad que la complejidad de lo controlado. Es decir, un controlador para un sistema complejo requiere ser por lo menos tan complejo como el sistema que intenta controlar. Según Gershenson (2014), en la práctica se requiere de un balance entre predictibilidad y adaptabilidad del controlador para enfrentar la emergencia y auto-organización de lo controlado. 4.5 Síntesis Hemos ilustrado a partir de experimentos computacionales, en sistemas dinámicos discretos (RBA y ACE), y en sistemas de tráfico vehicular, la utilidad de la valoración de la regularidad, cambio, adaptabilidad y equilibrio dinámico, a partir de las medidas propuestas de auto- organización, emergencia, complejidad y homeostasis. La complejidad representa un balance entre auto-organización y emergencia, desde la relación establecida entre orden y caos (Kaufmann, 1993; Langton, 1990). Esta relación, es otra expresión de diversos balances hallados en la naturaleza, como por ejemplo el balance de especies obtenido por selección natural. Desde este conocimiento, el balance entre cambio y orden puede ser alcanzado por diseñadores que buscan explotar los beneficios de la complejidad. La homeostasis brindó información sobre la estabilidad de los estados en el tiempo y su transición. La autopoiesis brindó información acerca de la autonomía del sistema respecto de su ambiente, en términos de complejidad-adaptabilidad. Dado que todo puede ser descrito en términos de información, se hace posible la evaluación de la complejidad, emergencia, auto-organización, homeostasis y autopoiesis, en todo lo que nos rodea. En este sentido, existen diferentes tópicos de investigación en los cuales las 90 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- presentes ideas pueden ser extendidas. Se propone la continuidad de su evaluación a través de la aplicación en diferentes sistemas, como por ejemplo: a las máquinas de Turing, a las máquinas 𝜖-psilon, a los sistemas tecnológicos, ecológicos y sociales, entre otros. Desde una mirada compuesta se puede decir que combinaciones de medidas pueden ser usadas para un mejor entendimiento y combinación de sistemas dinámicos. Observando solo 𝐶 o 𝐻, no es posible distinguir la complejidad desde reglas caóticas, pero su combinación revela que el éter es requerido para la interacción de estructuras complejas. Existe también la pregunta de cómo considerar y medir la información usada por las medidas. Por ejemplo, en ECA diferentes reglas son obtenidas si una cadena de bits es considerada verticalmente (cómo lo hicimos), horizontalmente o diagonalmente. 5. Capítulo 5: APLICACIONES EN SISTEMAS ECOLÓGICOS Resumen El estudio de la complejidad en Sistemas Ecológicos es un tópico de gran interés actual. Sin embargo, las interpretaciones que se hacen de ella han demostrado que se requiere de clarificaciones debido a la novedad del tema. En este capítulo, estudiaremos la complejidad ecológica de lagos y mamíferos terrestres. En estos dos casos de estudio, se analizaron las métricas de emergencia, auto-organización, complejidad, homeostasis y autopoiesis, desarrolladas en este trabajo. Adicionalmente, se desarrolla una solución analítica para la emergencia ecológica por medio de la programación genética, y se prueba su confiabilidad en lagos. Nuestros resultados muestran que la dinámica y complejidad ecológica puede ser efectivamente caracterizada en términos de información. Este enfoque, constituye una manera complementaria de entender la dinámica ecológica. 5.1 Visiones Tradicionales y Complejidad Ecosistémica Existen diferentes vías para describir el estado de un ecosistema y su dinámica. En ecología, lo natural es describir el sistema a través de sus atributos de abundancia, diversidad, frecuencia y/o riqueza de especies. Adicionalmente, los estudios ecológicos buscan establecer la estructura jerárquica a través de análisis de clasificación, que parten de medidas de similitud, y dan como resultados arboles de afinidad de especies y de sitios o dendrogramas (Legendre and Legendre, 1998). Igualmente, se tienen opciones de realizar análisis de ordenación de correspondencia canónica en los que las especies, los sitios y los parámetros fisicoquímicos del entorno, son correlacionados y asociados. El resultado de la ordenación son grupos de variables asociadas de manera significativa (Borcard et al., 2011). En el anterior contexto y desde los resultados obtenidos, se sugiere que la medición de la complejidad ecológica (o biológica en general), puede complementar la descripción de 92 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- ecosistemas y la explicación de la dinámica de especies. Sobre esta base, es posible ahondar en la descripción sintética la dinámica de los procesos ecológicos en términos de regularidad, cambio y adaptabilidad. Las anteriores consideraciones, toman mayor importancia si se considera que el campo de la complejidad ecológica está en construcción y el interés por su estudio es creciente. Se destaca que ya se han dado esfuerzos que tratan de relacionar la complejidad con los descriptores ecológicos comunes, o al menos hacer una distinción entre unos y otros (Boschetti, 2010, 2008; Parrott, 2010, 2005; Proulx and Parrott, 2008). Desde un principio, la complejidad biológica se trató de describir desde la caracterización de los componentes estructurales y sus interacciones funcionales (Pond, 2006). No obstante, se observó la dificultad de describir el comportamiento global de organismos y sistemas. Desde inicio de siglo, se estimó que en la asociación y modularidad de componentes podría darse algunas explicaciones al fenómeno de la complejidad. Desde allí, se estimó que la auto- organización juega un papel importante en la evolución de la complejidad biológica (Camazine et al., 2001; Kauffman, 2000). Enfoques desde la teoría de la información, han relacionado la complejidad de la vida como la cantidad de información que tiene o guarda un organismo o sistema en su genómica acerca de su ambiente (Adami et al., 2000). Se podría pensar que un sistema entre más información contenga, más complejo será. En consecuencia, se pensó que un organismo con alta información será muy auto-organizado, cómo lo propone el enfoque de la negentropía de (von Bertalanfy, 1968). Esta visión ha sido más coincidente con el pensamiento biológico, y se ha contrapuesto a la visión física de la entropía como producción de información nueva, como Shannon lo propuso. Por otra parte, desde la teoría de la información se han desarrollado formalismos alternativos que han relacionado la complejidad con el grado de aleatoriedad de los estados de los sistemas ecológicos (Anand et al., 2010; Ulanowicz, 2011, 2004). Sin embargo, esta visión resultaría ser no la más adecuada, pues sería difícil demostrar que la complejidad es igual a la emergencia. Más aún, la complejidad no se podría igualar al caos, cómo ya se discutió en 4.4.4. En el anterior contexto, y basados en que nuestra medida de complejidad puede reconciliar las dos visiones opuestas (entropía equivalente a nuestra 𝐸, y negentropía, equivalente a nuestra 𝑆), este capítulo mide la complejidad de variables y sistemas ecológicos. 5.1.1 Fundamentos de Limnología Los ecosistemas acuáticos continentales superficiales son estudiados por la Limnología (λίμνη, limne, "lago" y λόγος, logos, "conocer"). Entre ellos se hallan dos tipos: los lénticos, como los lagos, lagunas, estuarios, esteros, pantanos embalses; y los lóticos, como quebradas, riachuelos y ríos. Capítulo 5 93 Al ser sistemas abiertos, los lagos interactúan con su cuenca de drenaje y la atmosfera. A raíz de ello, un lago integran las relaciones funcionales de crecimiento, adaptación, ciclado de nutrientes, productividad biológica y composición de especies. En su estudio, se describen y evalúan cómo los ambientes químicos, físicos y biológicos regulan estas interacciones (Wetzel, 2001). Los lagos presentan una zonación particular, que se presenta en la figura 5-1. Relacionada con la porción terrestre, en el litoral, se halla la zona de macrófitas (macrophytes zone). Esta zona está compuesta por plantas acuáticas que pueden estar enraizadas, sumergidas o flotantes, y constituye una frontera entre los sistemas terrestres y la zona planctónica (planktonic zone). Por su función fotosintética, las macrófitas son de gran importancia en la producción primaria. La zona planctónica corresponde con las aguas más allá de la zona de macrófitas, donde los organismos que fotosintetizan (Fitoplancton) flotan pasivamente, y se dejan llevar por corrientes o gradientes horizontales y/o verticales. Entre tanto, los organismos que los pastorean (zooplancton) los siguen en sus migraciones. Figura 5-1 Zonación de Un Lago. La figura muestra las zonas Superficial, planctónica, de macrófitas acuáticas, bentónica y el substrato.(Fernández et al., 2014b). La zona bentónica (Benthic zone) tiene que ver con el fondo litoral y el fondo profundo del lago. Está relacionada con el sedimento (substratum) y la capa subsuperficial (ubicada a 20 centímetros arriba del substrato). Es bentónico todo organismo que esté en las inmediaciones del fondo, bien sea para escarbarlo, penetrarlo, o posarse en él. La zona de mezcla (Mixing zone), es una zona de intercambio entre las aguas de las zonas planctónica y bentónica. En las zonas descritas arriba, uno o más componentes o subsistemas pueden ser valorados para determinar la dinámica ecológica. Para nuestro primer caso de estudio ecológico, se consideraron tres componentes. 94 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- (i) El físico-químico (PC) referido a la composición física y química del agua. El PC es afectado por varias condiciones y procesos de naturaleza geológica, por la disolución y dispersión, por el ciclo hidrológico, por la generación de solutos y sólidos (de la fotosíntesis, por ejemplo) y por la sedimentación. En este componente, se destacan procesos cómo el equilibrio del pH, que afecta, entre otros, el intercambio de elementos entre un organismo y su ambiente. También, la regulación de la temperatura, que se relaciona con el calor específico del agua. (ii) El componente de nutrientes limitantes (LN), que si bien está relacionado con el componente físico-químico (PC), es básico para la fotosíntesis. LN está asociado con los ciclos bio-geoquímicos del nitrógeno, carbono y fósforo, que permiten la adsorción de gases en el agua o la disolución de los nutrientes. (iii) El componente de biomasa (Bio), considera el peso por unidad de volumen de organismos como productores primarios (fotosintéticos), consumidores primarios (que pastorean los primarios), consumidores secundarios (que predan los consumidores primarios), y los descomponedores. Las variables estudiadas de cada componente se relacionan con la zonificación de un lago en las siguientes tablas. Tabla 5-1 Variables del Componente Físico-químico-PC Variable (Español) Acrónimo Luz Superficial Luz Planctónica Luz Bentónica Temperatura Superficial Temperatura Planctónica Temperatura Bentónica Flujo de Entrada (Afluente) y Salida (Efluente) Tiempo de Retención Evaporación Mezclado entre zonas Conductividad del afluente Conductividad planctónica Conductividad bentónica Oxígeno superficial Oxígeno planctónico Oxígeno bentónico Oxígeno de sedimento pH del afluente pH planctónico pH bentónico SL PL BL ST PT BT IO RT Ev ZM ICd PCd BCd SO2 PO2 BO2 SdO2 IpH PpH BpH Unidades MJ/m2/día MJ/m2/día MJ/m2/día °C °C °C m3/sec days m3/día %/día µS/cm µS/cm µS/cm mg/litro mg/litro mg/litro mg/litro Unidades de pH Unidades de pH Unidades de pH Capítulo 5 Tabla 5-2 Variables del Componente de Nutrientes Limitantes-LN (Todas en mg/litro) 95 Variable (Español) Silicatos Afluente Silicatos Planctónico Silicatos Bentónico Nitratos Afluente Nitratos Planctónico Nitratos Bentónico Fosfatos Afluente Fosfatos Planctónico Fosfatos Bentónico Dióxido de Carbono Afluente Dióxido de Carbono Planctónico Dióxido de Carbono Bentónico Detrito Planctónico Detrito Bentónico Acrónimo IS PS BS IN PN BN IP PP BP ICD PCD BCD Pde Bde (Todas en mg/m3) Tabla 5-3 Variables del Componente de Biomasa-Bio Variable (Español) Acrónimo Diatomeas Planctónicas Cianobacterias Planctónicas Algas verdes Planctónicas Planctónicas Diatomeas Bentónicas Cianobacterias Bentónicas Algas verdes Bentónicas Macrófitas de superficie Macrófitas sumergidas Zooplancton Herbívoro Zooplancton Carnívoro Bentónicos Herbívoro Bentónicos Carnívoro Peces Planctonicos Peces Bentónicos Peces Piscivoros PD PCy PGA PCh BD BCy BGA SurM SubM HZ CZ BH BD PlF BF PiF 96 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 5.1.2 Tipos de Lagos Estudiados La escogencia de los cuatro lagos estudiados se dio en razón a su posición en el gradiente latitudinal ártico-trópico. Ellos correspondieron a un lago ártico (Ar), un lago de alta montaña en zona templada (NH), un lago de tierras bajas en zona templada (NL), y un lago tropical (T). El anexo A muestra la descripción ecológica detallada de cada uno de los lagos. El criterio de escogencia de los lagos se dio por su disposición latitudinal, que genera un gradiente de Ar a T, donde las condiciones de luz, temperatura e hidrología tienen amplia variación. 5.1.3 Simulaciones y Cálculo de las Medidas Los datos de los lagos Ar, NH, NL y T fueron obtenidos usando The Aquatic Ecosystem Simulator-AES (Randerson and Bowker, 2008). Las variables consideradas para cada componente fueron definidas en las tablas 5-1 a 5-3. El modelo para la obtención de valores diarios de las variables de los tres componentes, fue determinista, de manera que no hubo variación entre sus diferentes ejecuciones o corridas. Para todas las variables y lagos fueron obtenidos valores diarios. Los datos resultantes de las simulaciones fueron normalizados a base 10, para ello se utilizó la siguiente ecuación para todos los puntos x de todas las variables X. 𝑓(𝑥) = ⌊𝑏. 𝑥 − min 𝑋 max 𝑋 − min 𝑋 ⌋ (5.1) Donde 𝑥 es la función piso de 𝑥. El software de apoyo para este fin fue R: Comin, desarrollado por Villate et al. (2013). Se destaca que en la comparación entre lagos, el rango máximo y mínimo para todas las variables fue tomado como la mayor y menor cantidad de todos los valores de todos los lagos. En casos individuales de comparación, se tomó solamente el máximo y mínimo de cada variable en el respectivo lago. Una vez las variables transformadas en un alfabeto finito (para 𝑏 = 10), se calculó la emergencia, auto-organización, complejidad, homeostasis y autopoiesis (Fernández et al., 2014b; Gershenson and Fernández, 2012a). Para esto se definió una clasificación para las medidas de acuerdo con las categorías, rangos y colores de la tabla 4. Estos rasgos fueron inspirados en los existentes para índices de calidad y contaminación del agua (Fernández et al., 2005; Ramírez et al., 2003). Capítulo 5 97 Tabla 5-4 Categorías, Rangos y Colores de Clasificación de las Propiedades E,S y C (Fernández et al., 2014b) Categoría Rango Color Muy Alta Alta Media Baja [0.8 − 1] Azul [0.6 − 0.8] Verde [0.4 − 0.6] Amarillo [0.2 − 0.4] Naranja Muy Baja [0.0 − 0.2] Rojo 5.1.4 Resultados  Complejidad en un Ecosistema Ártico (Ar) Ar tiene una marcada dinámica de la temperatura e hidrología que ejerce gran influencia en los componentes estudiados (Fig. 6-1 Anexo A). Este comportamiento se resume en la figura 5-2, que muestra la distribución de cada una de las variables del componente físico-químico en su escala original. Allí, es notoria la amplitud de las variables como temperatura y luz, los altos valores y dispersión del tiempo de retención y la evaporación, los cortos rangos de la conductividad y pH. Por razones como estas, que dificultan la visualización, es que se indican procedimientos de estandarización y normalización. Figura 5-2 Cajas de bigotes para las variables físico-químicas (Fernández and Gershenson, 2014; Fernández et al., 2014b). En ella se representan cada uno de los valores de cada variable representados por puntos. Con ello se observa la distribución real de los valores. Las abreviaciones se dan en las tablas 5-1 En anterior figura (Fig. 5-2) se puede observar las variables transformadas a base 10 y la distribución de cada uno de sus 365 valores en las clases 0 a 9. La aglomeración o distribución de los puntos en una u otra categoría, da indicios del comportamiento regular o cambiante de la variable. Basados en esta distribución, algunos rasgos del comportamiento emergente o auto-organizado de cada variable, pueden ser descritos y comparados. Variables con 98 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- distribuciones más homogéneas producirán más información, debido a que puede ubicarse en cualquier clase de la escala. Esto generará altos valores de emergencia. En cambio, variables con distribuciones más heterogéneas serán más regulares en sus estados, y por lo tanto, producirán valores de auto-organización más altos. La complejidad, sin embargo, requiere de una mirada más detallada, y no puede ser fácilmente deducida. Para ello, se requerirá de su cálculo y clasificación en las categorías propuestas en la tabla 5-3. Figura 5-3 Distribución de los valores normalizados a base 10 de las variables físico-químicas en las clases 𝟎 − 𝟗. (Fernández and Gershenson, 2014; Fernández et al., 2014b). Variables con valores distribuidos de manera similar entre las clases 𝟎 a 𝟗, tendrán mayor información. Es decir, alta emergencia pues la variable puede tomar cualquier valor. Variables con puntos distribuidos de manera más concentrada en una u otra clase, producirán menos información. Es decir, alta auto-organización pues tales clases tendrán una probabilidad alta.. Las probabilidades obtenidas para cada clase en cada variable, será la base para calcular la cantidad de información, y por ende, la emergencia, que será el primer cálculo en la aplicación de las medidas desarrolladas, explicadas en el capítulo 3. El resultado de la aplicación de la métrica propuesta se observa en las figuras 5-4 a 5-6, donde se muestran los valores de emergencia 𝐸 (emergence), auto-organización 𝑆 (self- organization) y complejidad 𝐶 (complexity) para el sistema físico-químico (PC). Variables con muy alta complejidad 𝐶 ∈ [0.8 − 1] (según categorías de la tabla 5-4) reflejan un balance entre cambio/caos (emergencia) y regularidad/orden (auto-organización). Este es el caso del pH planctónico y bentónico (BpH;PpH), flujo de entrada y salida (IO) y tiempo de retención hidráulica (RT). Variables con altas emergencias (𝐸 > 0.92), como conductividad (ICd) y zona de mezcla (ZM), cambian constantemente en el tiempo, una condición necesaria que exhibe el caos. Para las demás variables, los valores de auto-organización se clasifican como bajos y muy bajos (𝑆 < 0.32), lo que refleja baja regularidad. Se hace interesante observar que el lago ártico no tiene variables con alta auto-organización o baja emergencia. Esto nos conduce a sugerir que un sistema con periodos climáticos muy marcados, el sistema tenderá Capítulo 5 99 a su cambio periódico, debido a las fuertes transiciones de temperatura e hidrología, y mantendrá cierta estabilidad en los periodos intermedios. Por ende, la complejidad fisicoquímica, y la biológica se verá limitada. Figura 5-4 Emergencia para las variables físico-químicas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) Figura 5-5 Auto-organización de las variables físico-químicas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) Figura 5-6 Complejidad de las variables físico-químicas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) 100 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Dado que 𝐸, 𝑆, 𝐶 ∈ [0,1], se puede hacer una clasificación de acuerdo a las categorías presentadas en la tabla 4 y definidas por (Fernández et al., 2014b), como sigue.  Variables con muy alta complejidad: 𝐶 ∈ [0.8 − 1] (En Color Azul). Existen dos grupos de variables con un balance entre S y E: (i) pH del flujo de entrada, planctónico y bentónico (BpH, PpH), flujo de entrada y salida (IO), y tiempo de retención (RT). (ii) El oxígeno en el sedimento y el oxígeno Bentónico. Es interesante ver cómo estos dos grupos se correlacionan de manera inversa. Es decir, el incremento del régimen hidrológico durante el verano genera, por la temperatura, estratificación, y la consecuente depleción del oxígeno disuelto (SO2, BO2). En este sentido, se puede observar cómo el comportamiento complejo de variables que inciden en otras, generan respuestas igualmente complejas por asociación. influencia de  Alta Complejidad: 𝐶 ∈ [0.6 − 0.8] (En Color Verde). Este grupo incluye 11 de las 21 variables que tienen alta E y baja S. Estas 11 variables que muestran un comportamiento preferentemente caótico y baja regularidad, están altamente influenciadas por la radicación solar. Ellas son Oxígeno superficial y planctónico (PO2, SO2), temperatura y luz planctónica y bentónica (ST, PT, BT, PL, BL), y Evaporación (Ev). Por su parte, las conductividades en la zona planctónica y bentónica (PCd, BCd) están influenciadas por el régimen hidrológico que se desprende de la temperatura.  Muy Baja Complejidad: 𝐶 ∈ [0 − 0.2] (En Color Rojo). Variables claramente caóticas o cambiantes caen en esta categoría, como se desprende de su muy alta 𝐸 y muy baja 𝑆. Incluye la conductividad en el flujo de entrada (IO) y el porcentaje de mezcla entre las zonas bentónicas y planctónicas (ZM). Las conductividades se aumentan debido al incremento del flujo de entrada, que genera un arrastre de iones y cationes en temporadas de deshielo. Posteriormente, en temporadas de congelación de la capa superior se ve reducida. El flujo también influencia de manera directa la zona de mezcla. Lo anterior hace que se asocien flujo, conductividades y mezclas, como las variables que mayor cambio tienen producto de la estacionalidad. Desde la clasificación realizada, es notorio que en el ártico, dadas las fuertes condiciones climáticas que marcan los periodos de invierno y verano, el cambio físico-químico prevalece sobre la regularidad. Es decir, existe alta emergencia y baja auto-organización en su ambiente físico y químico. Sin embargo, lo pequeños valores de auto-organización permiten estos espacios altamente emergentes, que son la base de su alta complejidad. Alta complejidad físico-química, marcará la pauta para la adaptabilidad y la evolución biológica. Capítulo 5  Homeostasis (H) 101 La homeostasis fue calculada comparando los valores diarios de todas las variables según la ecuación 3.8. Como resultado se obtuvo la variación diaria, como se ve en la figura 5-7. La escala temporal es importante porque a través de la homeostasis podemos considerar si comparar estados cada minuto, días, semanas, mes o año. El lago ártico tuvo un valor promedio de 0.9574 con una desviación de 0.0649. La homeostasis más baja fue de 0.6 y la más alta de 1. Si bien, se podría apreciar que la homeostasis general es alta, el ciclo anual mostró 4 periodos diferentes, cómo se observa en la figura 5-7, que se soportan en la descripción hecha de los lagos en el Anexo A. Estos periodos muestran valores dispersos de homeostasis como resultados de las transiciones entre invierno y verano, y nuevamente invierno. Para el periodo de invierno (primeros y últimos meses del año), se da una homeostasis alta debido a los valores regulares de todas las variables en el lago. El periodo abarca del día 212 hasta el 365, y del 1 al 87, y se caracteriza por la baja luz, temperatura, máximos periodos de retención debido al cubrimiento de la superficie con hielo, bajo flujo y mezcla, bajas conductividades, bajos pH y altos oxigenos presentes (ver anexo A). Figura 5-7 Homeostasis del componente físico-químico de un lago ártico en un ciclo anual. (Fernández and Gershenson, 2014; Fernández et al., 2014b) El segundo periodo comienza en el día 83, y va hasta el 154. Se caracteriza por el incremento del pH en la zona béntica, el incremento de la mezcla, el afluente y el efluente. Las fluctuaciones son fuertes debido al incremento de luz y temperatura por el advenimiento del verano. La homeostasis, por tanto, fluctúa y alcanza su mínimo de 0,6 en el día 116. Al final de este segundo periodo se incrementa la evaporación y el oxígeno decae en el fondo del lago. El tercer periodo va entre los días 155 a 162, y refleja la estabilización de las condiciones de verano. Significa que la evaporación, temperatura, luz, mezcla, conductividad y pH son máximos; mientras el oxígeno alcanza su más bajo nivel. El cuarto periodo (días 163 − 211), en el que la homeostasis fluctúa en altos niveles cerca de 0,9, corresponde al periodo de transición entre el verano y el invierno. 102 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Cómo se puede ver, la utilidad de la homeostasis se da en el seguimiento y estudio de la dinámica estacionaria.  Autopoiesis (A) El análisis de 𝐴 estuvo enfocado a observar la autonomía de la biomasa, con respecto a la fisicoquímica (PC) y nutrientes limitantes, en las zonas planctónica cómo bentónica. Para ello, se establecieron las relaciones más estrechas entre organismos fitoplanctónicos y su entorno funcional directo, representadas por las variables listadas en la tabla 5-5, donde también se muestran los organismos escogidos. Como primera condición para calcular A, se calculó la complejidad de cada grupo de variables, y se clasificaron según las categorías propuestas de muy alta a baja complejidad (Tabla 5-4). Tabla 5-5 Componentes tenidos en cuenta para el cálculo de la complejidad y posterior autopoiesis, en las zonas planctónica y bentónica del lago ártico Zona Planctónica Luz, temperatura, conductividad, pH Componente/Zona Físico-químico Zona Bentónica Luz, temperatura, conductividad, Oxígeno, Oxígeno en Sedimento, pH Silicatos, Nitratos, Dióxido de Carbono Diatomeas, Cianobacterias, Algas verdes Fosfatos, Nutrientes Limitantes Biomasa Silicatos, Nitratos, Fosfatos, Dióxido de Carbono Diatomeas, Cianobacterias, Algas verdes, Clorófitas La complejidad calculada fue promediada para cada componente y zona, como se observa en la figura 5-8. Así, los nutrientes limitantes en ambas zonas tuvieron baja complejidad (𝐶 ∈ [0.2 − 0.4]; color naranja). Las variables fisicoquímicas de la zona planctónica son de alta complejidad (𝐶 ∈ [0.6 − 0.8]; color verde). La biomasa y las variables fisicoquímicas de la zona bentónica están en la categoría de muy alta complejidad ( 𝐶 ∈ [0.8 − 1]; color azul). Figura 5-8 Complejidad para los 3 componentes desde las variables elegidas en la tabla 5, para dos zonas de un lago ártico (Fernández and Gershenson, 2014; Fernández et al., 2014b) Capítulo 5 103 A partir de los valores de 𝐶 para los organismos y su ambiente (ecuación 3.5), se calculó 𝐴 según la ecuación 3.13, para los organismos en relación a su ambiente físico-químico y los nutrientes limitantes. Todos los valores resultantes fueron mayores a 1 (Figura 5-9). Esto significa que las diatomeas, cianobacterias, algas verdes y clorófitas que se hallan en el fitoplancton (biomasa planctónica y bentónica), tienen mayor complejidad que las variables de su ambiente, representado en PC y LN. A raíz de ello, se puede sugerir que algunas variables PC, incluidas LN, tienen un efecto pequeño sobre la dinámica biomasa bentónica y planctónica, pues su grado de complejidad, y por tanto adaptabilidad, es mayor que la de su ambiente. De allí se estima que la biomasa ártica es más autónoma (adaptable) comparada con su ambiente físico y químico. Los altos valores de complejidad de la biomasa, implican que estos organismos pueden adaptarse a cambios de su ambiente debido al mayor balance entre su emergencia y auto-organización. Figura 5-9 Autopoiesis de la Biomasa Planctónica y Bentónica respecto de los nutrientes limitantes (PB/LN) y las variables fisicoquímicas (PB/PC), escogidas en tabla 5 (Fernández and Gershenson, 2014; Fernández et al., 2014b). Para ambos casos fue superior a uno por lo que se les da el color azul 5.2 Complejidad comparativa en el Gradiente Altitudinal Ártico-Trópico (Ar-T). Para el análisis de complejidad comparativa se consideraron, para la transformación a base 10 y la definición de los máximos y mínimos, los valores completos de cada variable para los cuatro lagos. En este sentido, el cálculo queda normalizado para todos los lagos. Una vez obtenidos los valores de complejidad de cada variable, se obtuvo el promedio del componente para un ciclo anual, y su dispersión, cómo se observa en la figura 5-10. 104 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Figura 5-10 Complejidad Comparativa en cuatro ecosistemas acuáticos- Ártico-Ar, Templado tierras altas-NH, Templado tierras bajas-NL y Tropical-T (Fernández et al., 2014d) En general, la complejidad fisicoquímica y de nutrientes limitantes se reduce desde el ártico al trópico, al igual que su variación. En tanto la complejidad biológica aumenta desde Ar a NL. Para ambos casos, podemos considerar al Ar y T cómo los extremos de la complejidad ecosistémica en el gradiente latitudinal. En el trópico se nota que la adaptabilidad de la biomasa, deducida de su complejidad comparativa que se requiere para responder al ambiente, se reduce debido a que en el trópico las condiciones son más estables. En general en el ártico, la dinámica es más caótica y emergente, y en el trópico más regular y auto-organizada. La dinámica del ártico, da como resultado especies de más amplia tolerancia ambiental (Euritípicas) que en el trópico (especies más estenotípicas, o de estrecha tolerancia ambiental). Al haber más auto-organización en el trópico, pero con la suficiente emergencia, se hace posible los procesos de especiación. Es por ello que en el trópico hay más diversidad y riqueza de especies. Esto es, las especies contarán con la suficiente posibilidad de cambio (emergencia, aunque sea poca) para que emerja una nueva especie como solución, si el ambiente cambia de forma considerable. La especie original que no pudo adaptarse, tendrá entonces el aislamiento a un lugar determinado donde pueda tolerar las condiciones ambientales, o el fin como destino dado su alta regularidad u auto-organización. En los dos casos, la especie originaria da paso a una nueva con nuevos rangos para poderse adaptar al nuevo ambiente (Rundell and Price, 2009; Schluter, 1996). Desde esta perspectiva, ser muy auto-organizado disminuye la posibilidad de adaptación biológica (Fernández et al., 2014d). En el recuadro de la figura 5-10, se puede observar que los lagos templados de tierras bajas (NL) parecen representar un punto de cambio para la tendencia de la complejidad. A partir de él, la complejidad físico-química se pasa de alta a muy alta en el trópico. La complejidad Capítulo 5 105 de la biomasa va de alta en NL a muy baja en el trópico, y la complejidad de los nutrientes limitantes disminuye ligeramente, pero se mantiene en la categoría de muy alta. Se podría estimar que NL constituyen un punto crítico, o de mayor complejidad, entre dinámicas emergentes del Ar y auto-organizadas del T. Las variables responsables del cambio de la complejidad en el gradiente fueron, las macrófitas, clorofíceas y fósforo planctónico. En promedio, para los cuatro lagos se puede ver la mayor complejidad de los componentes PC y LN, que Bio muestra que la respuesta de los componentes biológicos tiene menor autonomía dada su menor complejidad. Se puede decir que la complejidad fisicoquímica es mayor que la complejidad de los nutrientes, y ellas dos mayores que la complejidad de la biota (𝐶𝑃𝐶 > 𝐶𝐿𝑁 > 𝐶𝐵𝑖𝑜). No obstante, es apropiado ver que la complejidad promedio de la biomasa, siendo más baja, tiene una mayor dispersión. Con ello, la biomasa tiene posibilidad de responder a su ambiente, acorde con la ley de variedad requerida de Ashby, dentro de su rango de complejidad (0.382 ± 0.22). Adicionalmente, análisis estadísticos de comparación de promedios múltiples (Prueba de Tukey) demostraron que la complejidad promedio de PC y LN entre lagos no tuvo diferencias significativas (𝑝 = 0.85; 𝑝 > 0.05), lo que si sucedió para la biomasa con PC y LN. Esto demuestra que en promedio la complejidad de PC y LN es muy diferente de la Bio que es mucho menor. Este hecho ratificaría que la influencia del ambiente fisicoquímico y de nutrientes sobre la biota tiene una incidencia alta, por lo que los organismos deben hacer, en promedio, un esfuerzo mayor para adaptarse a ello. Lo anterior contrasta con lo sucedido en el lago ártico, donde al observar la complejidad de un grupo de organismos no de toda la biota, como fue el caso de la complejidad comparativa, que fue mayor que la del ambiente PC y LN. Este hecho evidenció que grupos de organismos particulares pueden tener más complejidad local que global, y responder mejor dentro de sus condiciones particulares. Estos resultados sugieren dos formas de ver la complejidad, (i) por grupos de organismos (caso del ártico) y (ii) como un componente completo (caso de la complejidad comparativa), situaciones que pueden verse como complementarias. 5.3 Complejidad del Número Incremental de Especies de Mamíferos Un análisis de gran interés en ecología con fines de conservación de especies, es el relativo a los análisis de ocupancia9 de especies en un territorio (Bailey et al., 2014; Mackenzie and Royle, 2005; Royle, 2006), acorde con su detectabilidad10 (Garrard et al., 2013; Tang et al., 9 Uso y ocupación del territorio por una especie. 10 Probabilidad de detección de la especie, que tiene efecto en los registros de ella en los muestreos por diferentes medios como por ejemplo cámaras trampa o registro de rastros (huellas, heces, restos de consumo de alimentos, etc) 106 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 2006). Así, con el fin de ahondar en los aspectos de la dinámica de la presencia-ausencia de especies animales, se realizaron simulaciones computacionales de ocupancia en una comunidad de mamíferos, con número incremental de especies (Fernández et al., 2013). Para ello se diseñó una formulación jerárquica de dos procesos o ecuaciones estocásticas ligadas, como sigue: 𝑜𝑐~𝐵𝑖𝑛𝑜𝑚𝑖𝑎𝑙(𝑛, 𝑝𝑠𝑖) 𝑦~𝐵𝑖𝑛𝑜𝑚𝑖𝑎𝑙(𝑖 , 𝑜𝑐 ∗ 𝑝) (5.2) (5.3) Donde 𝑜𝑐 es la ocurrencia de especies en un sitio dado, 𝑛 es el número de individuos de especies, 𝑝𝑠𝑖 es la ocupancia. 𝑦 es la detectabilidad de las especies donde: 𝑖 es el número de sitios con cámaras trampa, y 𝑝 es la probabilidad de detección. Para los dos casos, se partió de una distribución de probabilidad binomial. La comunidad simulada tuvo 10 sitios de muestreo con cámara-trampa y un ciclo de 1000 iteraciones variando en 𝑝𝑠𝑖y 𝑝 en cada iteración, como números aleatorios entre 0.1 y 0.9. Una vez generadas las poblaciones multi- específicas vía simulación, se aplicaron las medidas de 𝐸, 𝑆 y 𝐶 para el número incremental de especies en cada uno de los sitios. Los resultados se muestran en la figura 5-11. Desde allí se puede establecer que existen para la emergencia niveles altos para la complejidad, niveles medios para la emergencia y bajos para la auto-organización. Lo anterior muestra mucha más variabilidad que regularidad en las apariciones o ausencias de las especies. Esta variabilidad se mantiene con el incremento de especies. Es decir, no es común encontrar a todos los especies de manera constante (alta probabilidad de 1) o no encontrarlos (alta probabilidad de cero). Más bien, existe una combinación entre mediana y altamente variable de su presencia y ausencia, lo que genera su alta complejidad. Se observa, además, que no por incluir más especies en los 10 sitios se eleva la auto- organización, por suponer que se podría incrementar el número de detecciones. Se hace interesante observar cómo al incrementar el número de especies, la complejidad se mantiene prácticamente estable y tiende a mantener y consolidar valores, en la categoría de alta. La línea que representa la dispersión confirma la poca variación entre sitios, al hacerse cada vez más pequeña con el incremento de 1 a 1000 especies. Resultados como los anteriores son muy estimulantes, dado que muestran una tendencia estable en la dinámica de poblaciones de mamíferos. Capítulo 5 107 Figura 5-11 Tendencia de la complejidad en una comunidad simulada de mamíferos para 10 sitios. Se muestran las primeras 𝟖𝟎𝟎 especies en 1000 iteraciones (Fernández et al., 2013). 5.4 Modelado y Especificación de la Emergencia Acuática a través del uso de programación genética Uno de los mayores esfuerzos en ecología está en entender los aspectos de la emergencia asociada a la complejidad. Al respecto, soluciones analíticas son requeridas. Para llegar a ellas, se hace conveniente la integración del conocimiento y teorías desde las ciencias físicas hasta las sociales. En este apartado, utilizamos la programación genética para develar las relaciones entre la emergencia ecológica y los procesos complementarios de auto- organización, homeostasis y autopoiesis. Los métodos y resultados expuestos se presentan acorde con lo desarrollado en (Fernández et al., 2014c). Respecto de la programación genética (PG) y su aplicación en ecología, se puede estimar que la generalidad, realismo y certeza de los modelos desarrollados para ecología hacen de ella una herramienta apropiada para entender y expresar el comportamiento complejo de los ecosistemas. Basada en la forma en que se da el procesamiento de la información biológica, la PG define un algoritmo evolutivo para encontrar programas de computador que busquen soluciones analíticas más apropiadas a través de procesos de entrecruzamiento y evolución de funciones. Esencialmente, la PG tiene un conjunto de instrucciones y una función de ajuste para medir la bondad del resultado de la tarea llevada a cabo (John R Koza, 1994). La PG tiene ventajas comparativas con otras técnicas de aprendizaje de máquina, debido a que necesita un conocimiento pequeño a priori, que se obtiene de un conjunto de datos del sistema en estudio (Koza and Poli, 2003). La literatura revela los beneficios de la PG dentro de un marco de identificación del sistema (Koza 1994), razón por la cual fue escogida para 108 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- llegar a las soluciones analíticas que requirió la determinación de la emergencia ecológica a partir de los fenómenos de auto-organización, homeostasis y autopoiesis. El modelo desarrollado tuvo los siguientes componentes básicos:  Los individuos. Descritos por los árboles que representaron las ecuaciones matemáticas. Los nodos de las hojas son variables o constantes. Los demás, son funciones matemáticas u operadores.  Funciones y terminales. Las funciones fueron operadores matemáticos básicos como: adición, sustracción, multiplicación y división. Para los terminales (u hojas), se usaron la emergencia, auto-organización, homeostasis y autopoiesis. las variables que representan  Los operadores genéticos. Los principales fueron el cruzamiento y la mutación. El primero fue un operador de punto único para dos individuos. El segundo, tomó parte del árbol (rama) de un individuo y generó uno nuevo.  El ajuste del modelo. Se definió una función de evaluación de ajuste del modelo, entre los valores observados de la emergencia del lago ártico comparado con los valores calculados por la programación genética para el mismo lago. La métrica para el ajuste del modelo, fuero el coeficiente de correlación y análisis de residuos; para ello se usó el paquete CAR (Companion to Applied Regression) de R. Las condiciones experimentales tuvieron un número de generaciones de 3000, con excepción de la biomasa en que 2000 generaciones fueron usadas. El tamaño de la población fue de 50 y la probabilidad de uso del operador de cruzamiento fue de 1.0, la probabilidad de mutación 0.08. Los mejores organismos obtenidos se muestran a continuación en las figuras 5-12 a 5-14. Figura 5-12 Mejor Solución Obtenido para Nutrientes Limitantes Capítulo 5 Figura 5-13 Mejor Solución Obtenido para Biomasa Figura 5-14 Mejor Solución Obtenido para el Componente Físico-químico 109 Estos individuos obtuvieron el mínimo error a partir de la función objetivo entre las 𝐸 observadas y calculadas. Las funciones halladas fueron las siguientes: Para biomasa: 𝑓(𝑆, 𝐴, 𝐻) = √1 − ( 𝐻2𝐴(𝐻 − 𝑆) 𝑆 ) − (𝐴4𝑆𝐻) − √𝑆(𝑆 + 𝐻)) Para nutrientes limitantes 𝑓(𝑆, 𝐴, 𝐻) = √ (𝑆𝐻 + 𝐴 𝑆⁄ ) − (𝑆(𝐴 + 𝑆) − 1 1 − 𝑆 (5.4) (5.5) 110 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- Para el componente físico-químico: 𝑓(𝑆, 𝐴, 𝐻) = √ 𝐴 𝐻⁄ + 𝐴 + 2𝐻 √𝐴 𝐻⁄ √((𝐴 − 𝐻 − 𝑆 − 𝐴 𝐻⁄ )) − 𝑆2 √ (5.6) Donde 𝑓(𝑆, 𝐴, 𝐻), 𝐴, 𝑆, 𝐻 son la emergencia, autopoiesis, auto-organización y homeostasis, respectivamente. La robustez del modelo se evalúa al observar las gráficas 5-15 a 5-17, el ajuste de los datos calculados y predichos en la regresión y la desviación de los residuales a la misma. En ellas, la línea roja indica el ajuste perfecto de 1. Cómo se observa, los puntos que representan las parejas entre calculados y predichos se ubican de manera cercana a la línea de regresión. El análisis se complementa con la graficación de los residuales, representados en la línea verde. Para todos los casos los residuales fueron muy pequeños, lo que se puede notar en la escala del eje 𝑌. Esto denota una pequeña diferencia en la predicción del modelo, que no es considerada como significativa, como se confirma en otros indicadores como el error absoluto promedio del modelo que se muestra más adelante. Figura 5-15 Prueba de Ajuste del Modelo PG al Componente de Nutrientes limitantes Capítulo 5 Figura 5-16 Prueba de Ajuste del Modelo PG a la Biomasa Figura 5-17 Prueba de Ajuste del Modelo PG al componente Físico-químico 111 Se puede observar que en las representaciones gráficas, el adecuado ajuste de las soluciones obtenidas por PG a pesar de la gran diferencia de los lagos. Lo anterior puede ser confirmado por los coeficientes de correlación obtenidos que fueron de 2; 0.9995 y 0.9937 para nutrientes, biomasa y físico-químicos. Los errores medios absolutos para los mismos componentes fueron 0.0004169; 0.005368 𝑦 0.007129. Lo anterior demuestra la gran exactitud del modelo En términos ecológicos, los resultados alcanzados son de gran importancia. En particular, debido a las grandes diferencias que tienen los lagos ártico y tropical. Se podría decir que están en verdaderos extremos ecológicos, el ártico con una alta variabilidad lo que conduce 112 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- a mayores emergencias de sus variables, y el tropical con alta regularidad, lo que incrementa su auto-organización. A manera ilustrativa se da el ejemplo entre el comportamiento del componente físico-químico de los dos lagos, en las figuras 5-18 y 5-19, lo que confirma su diferencia, en especial en cuanto a su homeostasis, la cual muestra mayor equilibrio del sistema tropical. Además, para complementar esta ilustración, en el anexo A se pueden ver la descripción de cada uno de los lagos y sus regímenes hidro-climáticos. Figura 5-18 Emergencia, Auto-organización, Complejidad y Homeostasis para un el Componente Físico- químico de un Lago Ártico Figura 5-19 Emergencia, Auto-organización, Complejidad y Homeostasis para un el Componente Físico- químico de un Lago Tropical Las anteriores diferencias entre los lagos ártico y tropical, demuestran por un lado la importancia de la emergencia como una expresión que mide la nueva información, y por el otro la importancia de las expresiones halladas que permiten medirla desde los procesos auto-organizantes, que se dan sobre la base de procesos homeostáticos y autopoiéticos. Capítulo 5 5.5 Discusión 113 5.5.1 Descripción de los Ecosistemas y la Inclusión de la Complejidad como Indicador Ecológico El uso de la complejidad como indicador ecológico nos permite ver cuánto una entidad ecológica es más compleja que otra, como el caso comparativo entre variables fisicoquímicas, o entre lagos en diferentes latitudes. También podemos generar perfiles de complejidad espacial (latitudinal, latitudinal, entre otros). Podemos identificar entidades que representan cambios de complejidad en los gradientes. Por ejemplo, basados en los valores promedios de 𝐶 para los tres componentes, se pudo observar que en el gradiente altitudinal, NL representa un punto intermedio donde la complejidad latitudinal cambia. Sí bien los resultados muestran dinámicas diferenciales por componentes, se requiere de más estudios para determinarlo con precisión. En este mismo sentido, será apropiado definir si existen variables que exhiben gradientes relacionados con este cambio de complejidad. Por ejemplo para la biomasa, podemos estimar que podría haber un cambio en el transecto Ar-T, en NL, para las cianobacterias en la zona planctónica. El uso de la complejidad como indicador ambiental, también permite ver como en los trópicos la alta 𝑆 y baja 𝐶 pueden conducir a baja 𝐴 en el componente de biomasa. La alta regularidad muestra que los organismos son estables a escalas anuales. En estos ambientes, la biomasa reduce su autonomía y adaptabilidad, motivo por lo que se favorece la especiación, por los cortos rangos que soportan las especies. Esto se asocia con el grado de robustez que tienen los organismos, que al ser alta da baja adaptabilidad, por lo que la generación de nuevas especies se promueve. Todas estas apreciaciones, complementan de manera apropiada lo que se sabe de los trópicos en cuanto a su regularidad, con la ventaja que le dan un sustento numérico para su explicación. En adición a lo anterior, podemos usar la complejidad para ver en qué condiciones los organismos pueden tener rasgos de mayor autonomía al ser más complejos que su ambiente. Esto fue explícito en los resultados para el componente fitoplanctónico de los lagos. La medición de la complejidad del número incremental de especies, vía simulación, es un punto interesante de partida para la deducción de relaciones con aspectos como la riqueza y número de especies, con fines de conservación. 5.5.2 Sobre los Aspectos Computacionales de la Medición de la Complejidad Ecológica Entre los aspectos computacionales de la complejidad ecológica se debe considerar que los sistemas complejos como sujeto de estudio, cuentan con el aporte de biólogos, ecólogos, computistas, matemáticos, físicos, economistas y muchos otros científicos. Como resultado, 114 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- existe una mezcla de enfoques que puede generar confusión en los conceptos. En cada campo, un enfoque distinto es tomado. Los biólogos, por ejemplo, están interesados en la auto-organización y la emergencia, debido a que vienen trabajando con estos conceptos desde Darwin y los trabajos de Mendel. Por otra parte, los científicos de la computación quieren relacionar la medida de complejidad para sistemas complejos con la complejidad que ellos ya han definido. Por ejemplo, con la complejidad de Chaitin, donde la complejidad de un objeto/problema es la longitud del programa mínimo que necesita una máquina universal de Turing para reproducirlo/resolverlo. En este contexto, la teoría de la información es muy útil para obtener una medida de 𝐶, comenzando con la noción de 𝐸 cómo un tipo de entropía, descrita por la medida de información de Shannon, que es equivalente a la entropía de Boltzman. Cómo la temperatura, que caracteriza las propiedades de la energía cinética de las moléculas o fotones, una medida global de la complejidad ecológica puede caracterizar la dinámica de las interacciones espacio-temporales (Fernández and Gershenson, 2014). Por ejemplo, a pesar que cada componente de un sistema ecológico tiene diferentes (micro) estados que son afectados por el estado de los demás elementos (interacciones), su dinámica promedio y los diferentes regímenes dinámicos pueden ser representados en forma sintética en nuestra medida de complejidad. Esto se pudo observar en la aplicación de la complejidad comparativa de lagos. Considerando que 𝐸 puede ser entendida cómo la información producida por un proceso o ecosistema, una 𝐸 = 0 nos permite interferir sobre la predictibilidad de los estados, dado que no se produce nueva información. Es decir, dado que los estados presentan alta regularidad (𝑆 = 1), lo más probable es que el estado que viene sea el mismo que los anteriores. Entre tanto, cuando 𝐸 es máxima (𝐸 = 1) los estados son irregulares o seudo-aleatorios, y por lo tanto, el siguiente estado traerá nueva información. Dado que la regularidad es prácticamente nula (𝑆 = 0), será muy difícil predecir el próximo estado. La integración de los aspectos de 𝑆 y 𝐸 en nuestra medida de 𝐶, tiene ventajas como el observar la dinámica compleja de un ecosistema como un balance entre su regularidad y cambio. En este contexto, podemos definir qué variable, componente o proceso, es más complejo-adaptable que otro. Esta pregunta de gran importancia en ecología, nos puede conducir a generar categorías de complejidad de procesos, poblaciones, comunidades y ecosistemas. También nos puede brindar una medida de la adaptabilidad de nuestra biosfera, y en general, de la complejidad biológica. Igualmente, la medida de 𝐶 puede ser útil cuando se desea relacionar el cambio evolutivo y la sostenibilidad de ecosistemas. Es únicamente cuando 𝐶 es máxima y mayor que el ambiente que la evolución potencial en los sistemas puede ser mejorada, dado que los niveles intermedios de 𝐸 y 𝑆 le dan al sistema el suficiente potencial de cambio, pero con la suficiente robustez para perdurar. Para ello se parte de una condición más o menos estable, que para el momento será el estado más probable del sistema. Capítulo 5 115 Cómo los sistemas vivos requieren de un balance entre adaptabilidad-emergencia y estabilidad-auto-organización (Kaufmann, 1993), 𝐶 es útil para caracterizar sistemas vivos. A partir de ella, se puede relacionar la complejidad del sistema con la del ambiente, para obtener su autopoiesis. En este sentido, la caracterización del comportamiento del componente biótico, ante los cambios o disturbios ambientales, en términos de complejidad, puede dar información acerca de la autonomía de los organismos. Desde esta perspectiva, la complejidad, la homeostasis y la autopoiesis, pueden ser de utilidad en estudios de cambio climático global. En ellos podría confirmarse nuestra hipótesis que sistemas con alta complejidad tienen el suficiente grado de robustez y adaptabilidad para enfrentar los cambios ambientales. Esto lo soportamos en los datos de autopoiesis que se dan en ambientes extremos, cómo el lago ártico. 5.5.3 Comparación con otras Medidas de Complejidad. Desde las aplicaciones realizadas hemos hallado diferencias con algunos autores. Por ejemplo, Shalizi et al.(2004) define auto-organización cómo un incremento interno de la complejidad estadística. Esto puede suceder en sistemas con alta 𝐶 y máxima 𝑆. Entre tanto, para nosotros consideramos una máxima 𝑆 para sistemas con mayor regularidad y de patrones más estáticos. Alineados con la termodinámica y la teoría de la información, creemos que un cristal y un estado congelado, son más organizados que complejos. Sin embargo, un sistema con una 𝑆 mínima, lo más probable es que sea no organizado. En este aspecto creemos que la medida de Shalizi es más apropiada para estudiar la dinámica de sistemas auto-organizados, más que emergentes (Camazine et al., 2001; Gershenson, 2007). Se destaca que un sistema de baja auto-organización podría elevarla bajo la influencia de un factor externo, en lo que podemos considerar como auto-organización guiada (Fernández et al., 2014b). Desde la simplicidad de nuestras definiciones, la auto-organización puede ser vista como el opuesto de la emergencia. La auto-organización es alta para sistemas de dinámicas ordenadas, en tanto la emergencia es alta para sistemas altamente caóticos (en todas las escalas). Qué la auto-organización y emergencia sean opuestos puede ser visto como contra- intuitivo, dado que hay muchos sistemas que presentan alta complejidad al parecer por tener alta emergencia y alta auto-organización. Precisamente, estos sistemas, son los más interesantes para la ciencia. Para nosotros estos sistemas son los que tienen alta 𝐶 , a partir de una combinación de alta auto-organización y baja emergencia, o viceversa. Cabe destacar, que en nuestra propuesta en valores opuestos y altos de S y E encontramos el extremo del orden y del desorden, de allí su carácter opuesto. La información (y emergencia) puede ser vista, en una variable binaria, como un balance de ceros y unos (𝑃(0) = 1 − 𝑃(1) ⇔ 𝑃(0) = 𝑃(1) = 0.5). Así, la complejidad puede ser vista como un balance de emergencia y auto-organización (𝐼 = 0, 𝑆 = 1 − 𝐸; 𝑀𝑎𝑥(𝐶) ⇔ 𝑆 = 𝐸 = 0,5). Ya ha sido discutido en la comunidad científica que la complejidad es máxima cuando 116 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- existe una compensación entre persistencia de información (memoria) y su variabilidad (exploración, computación) (Bar-Yam, 2004a; Langton, 1990). También ha sido argumentado, que existe una tendencia de auto-organización o evolución a estados que incrementan su balance y alta complejidad (a múltiples escalas) (Bak et al., 1988; Gershenson, 2012b; Kauffman, 2000). Lo anterior está en acuerdo con el trabajo de Escalona-Moran et al.(2012). Aun cuando nuestra formulación es diferente, 𝐶 está estrechamente relacionada con la complejidad estadística de Crutchfield and Young (1989), que mide cuanta computación es requerida para producir información. Cadenas ordenadas o aleatorias si bien requieren de computación, cadenas complejas requieren más. Esta complejidad estadística, puede también ser definida como información mínima requerida para predecir el futuro de un proceso (Shalizi et al., 2004). Para nuestra 𝐶, baja 𝐼 implica cadenas que son fácilmente predecibles, mientras altas 𝐼 implican cadenas que no tienen mucho que predecir. Alta 𝐶 ocurre cuando 𝐼 es media, lo cual corresponde con más computación requerida para producir o predecir un patrón determinado. Aun cuando 𝐶 y la complejidad estadística, exhiben comportamientos similares, 𝐶 es mucho más simple de calcular y explicar. 5.5.4 La Búsqueda de Soluciones analíticas para la Emergencia El principal aspecto a resaltar de este tipo de modelado y especificación de la emergencia en sistemas acuáticos, producto del uso de la programación genética, es que se obtuvieron tres expresiones generales para la emergencia basada en sus propiedades asociadas de auto- organización, homeostasis y autopoiesis. Si se considera que la emergencia es información nueva, y que por las interacciones se hace difícil de predecir, debido a los cambios ambientales, esta expresión cobra especial interés. El primer punto de partida para la comprobación de la generalización de las expresiones obtenidas por PG en el lago ártico, fue su comprobación de funcionamiento correcto en la estimación la emergencia en el lago tropical, en los componentes de PC, LN y Bio. Este inicio, sienta las bases para la extensión de la generalización de las expresiones a otros lagos. También nos promueve a seguir en la evaluación de otras propiedades sintéticas, cómo la complejidad ecostémica, además de otros componentes ecosistémicos, de manera que se vaya confirmando su generalización. 5.6 Síntesis y Comentario Final. A partir de las medidas desarrolladas de 𝐸, 𝑆, 𝐶, 𝐻 y 𝐴, basadas en la teoría de la información y compuestas por expresiones matemáticas simples, se pudieron capturar aspectos complementarios a los tradicionales estudiados para la dinámica ecológica. Esto es, indicadores cómo la abundancia, diversidad, riqueza de especies, etc. Capítulo 5 117 En contraste con los anteriores indicadores, y otros análisis ecológicos cómo el de estructura jerárquica, nuestro enfoque permite analizar propiedades sintéticas de los ecosistemas, así como sus tendencias, en términos de la información. Esto puede ser aplicado a cualquier variable en diferentes componentes, al componente como tal, o a uno o más ecosistemas. Para ello se requiere que previamente las variables hayan sido transformadas la escala (base) deseada. Con esto, se logra una forma complementaria a la tradicional de observar los sistemas ecológicos sobre la base de su regularidad, cambio y balance. Las aplicaciones en sistemas acuáticos descritas, han considerado: (i) las dinámicas de ecosistemas por separado en cuanto a sus componentes PC, LN y Bio. (ii) Los análisis comparativos entre ecosistemas (Ar, NH, NL y T). (iii) Estudios de la ocupancia y detectabilidad de especies, un tópico de interés en estudios de conservación. Desde estas últimas, se pueden dar otras extensiones, como por ejemplo, estudios de patrones de movimiento. Los resultados obtenidos muestran como las medidas 𝐸, 𝑆, 𝐶, 𝐻 y 𝐴, pueden ser vistas como indicadores ecológicos útiles a diferentes escalas, que permiten describir cómo los ecosistemas se mantienen cerca de ciertos estados (atractores), pero con la suficiente variabilidad para poderse adaptar de manera autónoma al medio. En este aspecto, la regularidad de los estados puede ser observada en la auto-organización como sucedió para la biomasa en el lago tropical, la variabilidad en la emergencia como en el componente fisicoquímico del ártico, la adaptabilidad en la complejidad como todos los componentes del lago ártico, y la autonomía en la autopoiesis como en el fitoplancton ártico. Como se ha demostrado, es posible describir la dinámica ecológica en términos de información, como también ha sucedido en estudios previos (Fernández and Gershenson, 2014; Fernández et al., 2014b, 2014d, 2013), lo que muestra un campo creciente y una visión complementaria en ecología. Ello sienta las bases para inspeccionar el significado ecológico y aplicación de las medidas propuestas, así como buscar su validación progresiva en otros procesos y sistemas ecológicos. Basados en nuestra expresión de complejidad, puede ser posible dar soporte y explicación al porqué los organismos tendiendo al incremento del orden por el almacenamiento de información, tienen la posibilidad de tener un componente aleatorio que les permite explorar e innovar, para persistir. Lo anterior constituye la base de la adaptabilidad de las especies y ecosistemas (Walker, 2005). Por su parte, los resultados obtenidos desde la programación genética, han sentado las bases para la continuación en la búsqueda de soluciones analíticas, que desde los resultados obtenidos para la emergencia, nos invita a continuar en la inspección de expresiones con las que se puedan caracterizar y especificar las propiedades estudiadas. 118 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- 6. CONCLUSIONES Y TRABAJO FUTURO 6.1 Logros Generales Esta tesis logró el desarrollo de un enfoque metodológico para la especificación y estudios de sistemas dinámicos con un gran número de agentes, desde la perspectiva de su auto- organización, emergencia, complejidad, homeostasis y autopoiesis. Para tal fin, se partió de una definición propia de sistema dinámico que expresa la inclusión de múltiples agentes, interacciones, organización propia, emergencias y complejidad. Desde allí se ahondó en la descripción de un sistema multi-agente como un sistema dinámico, que incorporó la teoría de los sistemas dinámicos con el abordaje de los aspectos de su complejidad. De esta manera, fue posible incluir mayores aspectos para la descripción de un sistema dinámico sobre la base de un número mayor de propiedades, más allá de la auto-organización (cómo orden) y la emergencia (como cambio). Es decir, considerando la homeostasis (como equilibrio dinámico) y autopoiesis (como autonomía). Estas propiedades, por su carácter sintético, permitieron capturar diversos aspectos de la dinámica global desde las interacciones locales. Centrados en las interacciones, se logró la formalización de un sistema dinámico como una red de agentes. Con ello, se pudo integrar la teoría de sistemas dinámicos, la teoría multi- agentes y la teoría de grafos. Estos últimos resultados teóricos sientan las bases para posteriores aplicaciones. Como segundo propósito, la tesis propuso nuevas nociones y formalismos de auto- organización, emergencia, complejidad, homeostasis y autopoiesis. El objeto estuvo en la clarificación de estas medidas y sentar las bases para la posterior formalización. Se quiso desde un principio que las nociones y medidas representasen lo que manifestaban. El alcance de estas medidas contribuye a la desambiguación de los conceptos. No obstante, son un enfoque desde la teoría de la información, en el que se debe tener en cuenta su trasfondo probabilístico. De la emergencia resaltamos su carácter de información nueva, pues corresponde con propiedades que surgen desde las interacciones locales en una escala dada. De la auto-organización como una expresión del grado de orden, sin la participación de un control central, resaltamos su carácter opuesto a la emergencia. De la complejidad, resaltamos ser el resultado del balance entre auto-organización y emergencia, o balance entre orden y caos. De la homeostasis resaltamos su capacidad para determinar si el sistema 120 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- mantuvo a través del tiempo los mismos o diferentes estados, como producto de su equilibrio dinámico. Su valor agregado está en que es útil para identificar los ritmos del sistema a través de la dinámica temporal, como muestra de su funcionamiento en su zona óptima. Finalmente, de la autopoiesis, su capacidad para enfrentar los cambios ambientales. En cuanto a las diversas aplicaciones realizadas, es destacable que los formalismos desarrollados permiten describir muchos fenómenos en términos de información a diversas escalas. Las aplicación en redes booleanas y autómatas celulares elementales, definieron y categorizaron dinámicas ordenadas, críticas y caóticas, de forma apropiada. La aplicación en tráfico, permitió determinar de forma comparativa con el sistema tradicional de la ola verde, que el sistema auto-organizante generó para unos casos un comportamiento regular y una mayor complejidad. Lo anterior se expresó en un número mayor de fases, lo que demostró su adaptabilidad a diferentes flujos y demandas de tráfico. Respecto a las aplicaciones en sistemas ecológicos, se puede estimar que la medición de la complejidad en ella dio interesantes resultados. En términos prácticos, las medidas propuestas contribuyen a darle significado e interpretación a la complejidad ecológica. Como hemos argumentado, en la literatura no se ha definido una única medida de complejidad. La discusión se mantiene en las diferentes interpretaciones dadas a la complejidad ecológica (complejidad funcional y estructural). Como clarificación se puede decir que acorde con Heylighen (2011, 2000), la complejidad estructural es el resultado de la diferenciación espacial y de la selección del ajuste de las interacciones entre los componentes. Por otra parte, la complejidad funcional, está relacionada con la variedad de acciones y procesos para enfrentar las perturbaciones ambientales. Ambos procesos producen una jerarquía de sistemas anidados, subsistemas o meta-sistemas, que tienden al auto-refuerzo. Desde nuestro enfoque de la complejidad como balance entre 𝐸 y 𝑆, está asociada con la caracterización de la complejidad funcional, por cuanto es producto de la dinámica de los estados de los elementos del sistema. Estas dinámicas pueden ser ordenadas (auto-organizadas), caóticas (emergentes) o críticas (complejas). Desde esta condición, los lagos en el gradiente latitudinal pueden se clasificados en uno u otro régimen. Igualmente, la medida de autopoiesis está asociada con la complejidad funcional, dado que refleja la adaptabilidad y autonomía ante los cambios ambientales en el tiempo. Un aspecto interesante que se desprende de esta conclusión, es que surgen nuevas oportunidades de contrastar la complejidad funcional con la complejidad estructural. Esta última puede ser definida desde las diversas topologías de redes de interacción propuestas por la teoría de redes. La base para esto lo constituyen los experimentos realizados con redes booleanas aleatorias, donde las dinámicas ordenadas, caóticas y críticas, estuvieron relacionadas con la conectividad. Sobre las anteriores bases, además de los formalismos propuestos, el uso de la programación genética (ítem 5.5.5) permitió obtener expresiones analíticas que complementan el estudio de los sistemas ecológicos. Esto quedó demostrado por la exactitud del modelo desarrollado, que obtuvo valores de emergencia apropiados para un lago tropical a partir de los datos de CONCLUSIONES 121 un lago ártico, un lago con una dinámica muy diferente. Así, el enfoque de programación genética nos permitió hallar expresiones generales para caracterizar ecosistemas ecológicos, y contribuye al entendimiento de las interacciones entre procesos ecológicos de auto- organización, homeostasis y autopoiesis. La utilidad de este enfoque puede ser evaluado gradualmente en otros ecosistemas acuáticos. Se podría pensar que la metodología desarrollada, en especial las métricas propuestas, tienen limitaciones por su generalidad. Sin embargo, vemos que esto también es una virtud, pues puede sintetizar diversos fenómenos, los cuales pueden ser descritos en términos de información. Desde las aplicaciones y los resultados obtenidos, surge la motivación y la posibilidad de hacer extensiones de la metodología a otros casos, e incluso lograr perfeccionamientos y formalizaciones adicionales. 6.2 Trabajo Futuro Existen varias situaciones que sería interesante explorar, por ejemplo:  Nuestras medidas pueden ser generalizadas para redes computacionales (Gershenson, 2010), con potenciales aplicaciones al estudio de redes complejas de agentes (Fernández et al., 2012b; Newman, 2003).  En consecuencia con lo anterior, se puede llegar a la implementación de una plataforma de modelado basada en agentes que explore el surgimiento de las propiedades de auto-organización, emergencia, complejidad, homeostasis y autopoiesis, desde la representación de una red computacional de agentes.  Es necesario continuar la aplicación de los formalismos propuestos, a grandes series de tiempo de sistemas reales, para ahondar más en su significado.  La información de Shannon ha sido recientemente usada como una medida de la complejidad espacial (Batty et al., 2012), con aplicaciones en áreas urbanas. Será interesante comparar este tipo de trabajos con nuestras medidas de complejidad.  Es necesario ahondar en el ajuste de indicadores homeostáticos y su aplicación, debido a que a que hemos notado su asociación directa con el proceso de auto- organización.  Se debe finalizar un paquete de software libre para el cálculo y graficación ágil de las métricas propuestas. La primera versión llegó hasta la programación de una función para R, que hemos llamado Comin (Complexity and Information) (Villate et al., 2013).  Nuevos casos de estudio pueden ser abordados. Por ejemplo, la complejidad biológica desde la genómica, la proteómica, la metabolómica y la enfermedómica. Desde la base de su medición, se pueden explorar los aspectos de modularidad y complejidad a múltiples escalas de organización. Por otra parte, surgen tópicos especiales como la complejidad del ritmo cardiaco en diferentes cohortes de edad, pues hemos notado, 122 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝐴𝐸𝐶𝐻𝐴- en inspecciones preliminares, que el ritmo cardiaco en los niños es más emergente que en los adultos. Nace también el interés por explorar la complejidad de ciertas patologías cardiacas.  Y finalmente, desde la perspectiva de la complejidad como balance, creemos que podemos abordar tópicos relativos a la cognición, la cual puede tener elementos como el balance entre el aprendizaje y la memoria. El flujo, entendido como aquel estado de plenitud y complacencia en que existe felicidad, también puede ser considerado como un balance entre los retos y las habilidades. Este último hecho favorecería el aprendizaje complejo, que puede ser estimado como un nuevo modelo educativo (Fernández et al., 2014a). 7. Anexo: Caracterización de los Lagos Estudiados 7.1 Lago Ártico (Ar) Los lagos árticos están localizados en el círculo polar ártico. Su temperatura superficial (ST) promedio es de 3°, su máxima de 9°C y su mínima de 0°C. En general, los sistemas Ar se clasifican cómo oligotróficos debido a su baja producción primaria, representada en valores de clorofila de 0.8-2.1 mg/m3. Su zona limnética, o columna de agua, está bien mezclada, lo que significa que no tiene estratificaciones (capas con diferentes temperaturas). Durante el invierno (Octubre a Marzo), la superficie del lago está cubierta de hielo. Durante el verano (Abril a Septiembre), el hielo se derrite y el flujo de agua y evaporación se incrementa. Estos dos regímenes climáticos tienen una fuerte influencia en el comportamiento hidrológico de los lagos, que a su vez influencia la físico-química y la biota. Figura 7-1 Hidroclimatología del lago Ártico (Randerson and Bowker, 2008). A la izquierda la temperatura en las 3 zonas de estudio. A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. Los nutrientes limitantes en las formas de nitratos, silicatos y dióxido de carbono están entre el 90 y 100% de disponibilidad para el fitoplancton en todo el año. El fitoplancton y perifitón están dominados por diatomeas en porcentajes de 38.6 y 45%, respectivamente. En el 124 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨- zooplancton dominan los herbívoros, con un 91.7% de la biomasa. En la zona bentónica, los invertebrados detritívoros alcanzan un 86.8%, y los peces piscívoros el 85.8%. 7.2 Lago Templado del Norte de Tierras Altas (North Higland Lake-NH). NH corresponde a un ecosistema mesotrófico en una latitud templada (T° promedio de 5.3°C). Los niveles de clorofila están entre 2.2-6.2 mg/m3. La superficie está cubierta de hielo en el invierno (final de noviembre a principios de febrero). La cubierta de hielo forma una barrera para el viento, lo que minimiza la pérdida de agua por evaporación, mientras que el fondo del lago permanece sin congelar. La columna de agua no está termo-estratificada, y está permanentemente bien mezclada con niveles del 50% en verano y 90% en invierno. El flujo máximo se da en primavera y otoño (9.6 m3/s). El flujo mínimo en verano (0.6 m3/s). La evaporación se reduce debido a que el agua es más o menos fría y los gradientes de presión de vapor no son grandes (promedio= 9.262 m3/d). El tiempo de retención (RT) es máximo en verano con 100 días. La concentración de oxígeno es superior a 10 mg/lt en las tres capas. El pH ronda las 7 a 7.3 unidades, pero se mueve en rangos de 6.7 a 7.8 unidades desde la superficie al fondo. Figura 7-2 Hidroclimatología del lago templado del norte-NH (Randerson and Bowker, 2008). A la izquierda temperatura en las 3 zonas de estudio. A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. La correlación entre variables es más estacional en NH que en lagos templados de tierras bajas (North Lowland-NH). Esto significa que el periodo de verano está relacionado con altas RT y altos pH. La estación invernal está relacionada con altos niveles de oxígeno en el flujo de entrada y salida. Sin embargo, existe una correlación más estrecha entre el oxígeno bentónico y el oxígeno del sedimento (BO2, SdO2). Anexo 125 Los nutrientes limitantes, como los nitratos y el dióxido de carbono, están en el 95% de disponibilidad para el fitoplancton. Los fosfatos y silicatos muestran variaciones y porcentajes mejores de disponibilidad. Los primeros alrededor del 80%, y los segundos cerca del 85%. La composición de la biomasa es dominada por las diatomeas planctónicas (46.7%) y bentónicas (41%). La composición del zooplancton es dominada por el zooplancton herbívoro (91%), los carnívoros abarcan el restante 8,6%. Entre los invertebrados bentónicos, los detritívoros dominan con el 87.5%. La comunidad de peces es dominada por los peces bénticos en porcentaje de 88.9. 7.3 Lago Templado del Norte de Tierras Bajas (North Lowland Lake-NL). Este lago prototipo de los lagos templados es un lago eutrófico, localizado en un clima templado del hemisferio norte. Su productividad primaria es de 6.3-19.2. Esta bajo la influencia de las estaciones de invierno, primavera, verano y otoño. En verano, las variaciones de flujo de entrada a salida caen a 3.5 de 25.2 m3/s. El tiempo de retención se incrementa en 100 días. La falta de viento y altas temperaturas (24°C), causan termo-estratificación, que se evidencia en los 24°C en la superficie, 20.6°C en la capa plantónica y 17.3°C en el bentos. Por tal motivo, el agua al ser es más densa en el fondo no se mezcla. En invierno, no se cubre de hielo la superficie, y el flujo es mínimo. En primavera y otoño la columna de agua se mezcla en un 100%, y el tiempo de retención hidráulica baja a 14 días. Esto genera un incremento en la conductividad eléctrica. En verano, las depleciones de oxígeno en las tres capas son más drásticas que en el ártico. El oxígeno está directamente correlacionado con el porcentaje de mezcla y los flujos de entrada a salida, y de forma inversa a otros parámetros cómo el pH y la retención hidráulica. Figura 7-3 Hidroclimatología del lago templado del norte de tierras bajas-NL (Randerson and Bowker, 2008). A la izquierda temperatura en las 3 zonas de estudio. A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. 126 Modelado Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨- Los nutrientes limitantes están por encima del 90% de disponibilidad para el fitoplancton en todas las estaciones. De acuerdo con su disponibilidad, la composición de la biomasa del fitoplancton y perifitón es dominado por diatomeas planctónicas (47%) y bentónicas (34.3%). En consecuencia, el zooplancton es 100% herbívoro. La comunidad de peces es dominada por los peces bentónicos (67%). 7.4 Lago Tropical (T) T es un sistema híper-eutrófico (clorofila > 19.2), localizado en un clima tropical húmedo, en el norte del Ecuador. Su temperatura media es de 25°C en la capa superficial. Está sometido a una estación lluviosa y otra seca que marca los flujos máximos y mínimos, respectivamente. La alta radiación conduce a altas temperaturas en el agua y a bajas diferencias entre zonas. A pesar que se puede dar la estratificación debida al intercambio de calor entre capas, esta es menos estable que en los lagos de latitudes altas (NH y NL). Especialmente, debido a que el viento puede tener gran incidencia en la mezcla de la columna de agua. El flujo de nutrientes y la dinámica planctónica se ven afectados por los episodios de calor y mezcla. Se destaca que la producción primaria entre lagos tropicales es de dos veces la de altas latitudes. También se ha reconocido que en T el factor limitante es el nitrógeno, de manera que cuando se agota los demás procesos no tienen lugar. El equilibrio entre especies dentro del fitoplancton y perifiton es alto (33% para diatomeas, algas verdes y cianobacterias). Las poblaciones de zooplancton están dominadas por los herbívoros (90%), el bentos por invertebrados detritívoros (84%) y los peces (87%). Figura 7-4 Hidroclimatología del lago tropical. A la izquierda temperatura en las 3 zonas de estudio (Randerson and Bowker, 2008). A la derecha los flujos hidrológicos de afluente, efluente, mezcla y evaporación. Bibliografía Adami, C., Ofria, C., Collier, T.C., 2000. Evolution of biological complexity. Proc. Natl. Acad. Sci. U. S. A. 97, 4463–4468. Aguilar, J., 2014. Introducción a los Sistemas Emergentes, 1st ed. Talleres Gráficos, Universidad de Los Andes. Aguilar, J., Cerrada, M., Hidrobo, F., 2007. A Methodology to Specify Multiagent Systems, in: Nguyen, N., Grzech, A., Howlett, R. (Eds.), Agent and Multi-Agent Systems: Technologies and Applications. Springer , Berlin / Heidelberg, pp. 92–11. Aguilar, J., Cerrada, M., Hidrobo, F., 2012. Sistemas Multiagentes y sus aplicaciones en Automatización Industrial, 1st ed. Talleres Gráficos, Universidad de Los Andes, Mérida, Venezuela. Aguilar, José, Besembel, I., Cerrada, M., Hidrobo, F., Narciso, F., 2008. Una Metodología para el Modelado de Sistemas de Ingeniería Orientado a Agentes. Rev. Iberoam. Intel. Artif. 12, 39–60. Aldana-González, M., Coppersmith, S., Kadanoff, L., 2003. Boolean Dynamicswith Random couplings, in: Perspectives and Problems in Nonlinear Science. A Celebratory Volume in Honor of Lawrence Sirovich, E. Kaplan, J. E. Marsden, and K. R. Sreenivasan. Springer Appl. Math. Sci. Ser., pp. 23–89. Anand, M., Gonzalez, A., Guichard, F., Kolasa, J., Parrott, L., 2010. Ecological Systems as Complex Systems: Challenges for an Emerging Science. Diversity. 2,3, 395-410 doi:10.3390/d2030395 Anderson, P.W., 1972. More is different. Science. 177, 393–396. Ash, R.B., 1965. Information Theory. Plant Signal. Behav. 6, 1-27. doi:10.1007/s10827-011-0314- 3 Ash, R.B., 2000. Information Theory 4.1. Entropy 97, 1–5. doi:10.3390/s110605695 128 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Ay, N., Der, R., Prokopenko, M., 2012. Guided self-organization: perception–action loops of embodied systems . Theory Biosci. 131, 125–127. Bailey, L.L., MacKenzie, D.I., Nichols, J.D., 2014. Advances and applications of occupancy models. Methods Ecol. Evol. 5,12, 1269–1279 doi:10.1111/2041-210X.12100 Bak, P., Tang, C., Wiesenfeld, K., 1988. Self-organized criticality. Phys. Rev. A, Lecture notes in Physics 38, 364–374. doi:10.1103/PhysRevA.38.364 Bar-Yam, Y., 2003. Dynamics of complex systems, Computers in Physics, Studies in nonlinearity. Westview Press. 12,4,864. doi:10.1063/1.168724 Bar-Yam, Y., 2004a. A mathematical theory of strong emergence using multiscale variety. Complexity 9, 15–24. Bar-Yam, Y., 2004b. Multiscale Variety in Complex Systems. Complexity 9, 37–45. doi:10.1002/cplx.20014 Batty, M., 1971. Modelling Cities as Dynamic Systems. Nature. 231,303, 425-428. doi:10.1038/231425a0 Batty, M., 2012. Building a science of cities. Cities 29,1. Batty, M., Morphet, R., Massuci, P., Stanilov, K., 2012. Entropy, complexity and spatial information ( No. 185). Bedau, M.A., Humphreys, P., 2008. Emergence: Contemporary Readings in Philosophy and Science. MIT Press, Cambridge, MA, USA. Berlekamp, E.R., Conway, J.H., Guy, R.K., 1982. Winning Ways for Your Mathematical Plays, American Mathematical Monthly, Winning Ways for Your Mathematical Plays. Academic Press. 2,5,411. doi:10.2307/2323620 Bernard, C., 1859. Lecons sur les propri´et´es physiologiques et les alterations pathologiques des liquides de l'organisme. Paris. 2,496. Bersini, H., Stano, P., Luisi, P.L., Bedau, M.A., 2010. Philosophical and scientific perspectives on emergence. Synthese 185, 165–169. doi:10.1007/s11229-010-9718-2 Borcard, D., Gillet, F., Legendre, P., 2011. Numerical Ecology with R, Media, Use R! Springer New York. doi:10.1007/978-1-4419-7976-6 Boschetti, F., 2008. Mapping the complexity of ecological models. Ecol. Complex. 5, 37–47. Boschetti, F., 2010. Detecting behaviours in ecological models. Ecol. Complex. 7, 76–85. Bibliografía 129 Bourgine, P., Stewart, J., 2004. Autopoiesis and cognition. Artif. Life 10, 327–345. Camazine, S., Deneubourg, J.-L., Franks, N.R., Sneyd, J., Theraulaz, G., Bonabeau, E., 2001. Self-Organization in Biological Systems, 1st ed. Princeton University Press, Princeton, New Jersey. Cannon, W.B., 1932. The Wisdom of the Body. WW Norton & Co, New York. Chaitin, G., 2007. The halting probability omega: Irreducible complexity in pure mathematics. Milan J. Math. 75, 291–304. Christianson, G.E., 1998. Isaac Newton and the scientific revolution, Oxford portraits in science. Oxford University Press. Cook, M., 2004. Universality in elementary cellular automata. Complex Syst. 15, 1–40. Crutchfield, J., Young, K., 1989. Inferring statistical complexity. Phys. Rev. Lett. 63, 105–108. doi:10.1103/PhysRevLett.63.105 Crutchfield, J.P., 2011. Between order and chaos. Nat. Phys. 8, 17–24. doi:10.1038/nphys2190 De Wolf, T., Holvoet, T., 2005. Emergence Versus Self-Organisation: Different Concepts but Promising When Combined, in: Brueckner, S., Di Marzo Serugendo, G., Karageorgos, A., Nagpal, R. (Eds.), Engineering Self-Organising Systems. Springer, Berlin / Heidelberg, pp. 77–91. Desalles, J.L., Ferber, J., Phan, D., 2008. Emergence in Agent-Based Computational Social Science: Conceptual, Formal and Diagrammatic Analysis, in: Yang, A., Shan, Y. (Eds.), Intelligent Complex Adaptive Systems. IGI Publishing, Hershey, PA, USA, pp. 1–24. Di Paolo, E., 2000. Homeostatic adaptation to inversion of the visual field and other sensorimotor disruptions. From Anim. to Animat. 6 Proc. 6th Int. Conf. Simul. Adapt. Behav. 440–449. Ebeling, W., 1993. Chaos, order, and information in the evolution of strings. Appl. Math. Comput. 56, 131–151. doi:10.1016/0096-3003(93)90119-Y Edmonds, B., 1999. Syntactic Measures of Complexity. Dep. Philisophy PhD, 245. Egri-Nagy, A., Nehaniv, C.L., 2008. Hierarchical coordinate systems for understanding complexity and its evolution with applications to genetic regulatory networks. Artif. Life 14, 299–312. Eriksson, D.M., Wulf, V., 1999. Self-organizing social systems: a challenge to computer supported cooperative work. Cybern. Hum. Knowing 6, 3–4. 130 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Escalona-Moran, M., Paredes, G., Cosenza, M., 2012. Complexity, information transfer and collective behavior in chaotic dynamical networks. Int. J. Appl. Math. Stat. 26, 58–66. Ferber, J., Gutknecht, O., 1998. A meta-model for the analysis and design of organizations in multi-agent systems, Proceedings International Conference on Multi Agent Systems Cat No98EX160. IEEE Comput. Soc. doi:10.1109/ICMAS.1998.699041 Ferber, J., Gutknecht, O., Michel, F., 2004. From Agents to Organizations : an Organizational View of Multi-Agent Systems. AgentOriented Softw. Eng. IV, Lecture Notes in Computer Science 2935, 214–230. doi:10.1007/978-3-540-24620-6_15 Fernández, N., Aguilar, J., Gershenson, C., Terán, O., 2012a. Sistemas dinámicos como redes computacionales de agentes para la evaluacion de sus propiedades emergentes, in: I Simposio Científico Y Tecnológico En Computación SCTC2012. Universidad Central de Venezuela, Caracas. Fernández, N., Aguilar, J., Terán, O., 2010a. Conceptual modeling of emergent processes in dynamics complex systems, in: Rivas-Echavarria, F., Mousallí-Kayat, G. (Eds.), 9th WSEAS International Conference on Computational Intelligence, Man-Machine Systems and Cybernetics (CIMMACS '10). WSEAS Press, Mérida, Venezuela, pp. 75–82. Fernández, N., Aguilar, J., Terán, O., 2012b. Formalization of Emergent Properties in Computing Agent Networks, in: Ana Julia Viamonte (Ed.), 14 WSEAS International Conference on Mathematical Models and Methods in Modern Science. WSEAS, Porto- Portugal, pp. 189–194. Fernández, N., Carvajal, L., Colina, E., 2010b. Sistema Difuso Tipo Mamdani para la Determinación Genérica de la Calidad del Agua. BISTUA 8. Fernández, N., Gershenson, C., 2014. Measuring Complexity in an Aquatic Ecosystem, in: Castillo, F., Cristancho, M., Isaza, G., Pinzón, A., Corchado, J. (Eds.), Advances in Intelligent Systems and Computing. Springer. Springer Berlin / Heidelberg, pp. 83–89. doi:10.1007/978-3-319-01568-2_12 Fernández, N., Gershenson, C., Aguilar, J., Terán, O., 2011. Introducción a los sistemas dinámicos como redes computacionales basados en agentes (SD-Ag)., in: Coloquio de Sistemas Complejos Como Modelos de Computación. Centro de Ciencias de la Complejidad-C3, Instituto de Ciencias Nucleares. Universidad Nacional Autónoma de México, México, D.F. Fernández, N., Huertas, M., Rodriguez, F., Gershenson, C., 2014a. Complex Learning, Leadership and Flow, in: International Congress on Systems and Cybernetics. WOSC, Ibagué, pp. 263–279. Fernández, N., Lizcano, D., Ahumada, J., Hurtado, J., Gershenson, C., 2013. Complexity of Mammal Presence in a Tropical Forest, in: European Conference in Complex Systems. Bibliografía 131 Fernández, N., Maldonado, C., Gershenson, C., 2014b. Information Measures of Complexity, Emergence, Self-organization, Homeostasis, and Autopoiesis, in: Prokopenko, M. (Ed.), Guided Self-Organization: Inception SE - 2, Emergence, Complexity and Computation. Springer Berlin Heidelberg, pp. 19–51. doi:10.1007/978-3-642-53734-9_2 Fernández, N., Marcano-Valero, G., Terán, O., Aguilar, J., Gershenson, C., 2014c. Modeling and Specification of the Aquatic Ecological Emergence using Genetic Programming, in: XL Conferencia Latinoamerican En Informática. Fernández, N., Ramírez, A., Solano, F., 2005. Physico-Chemical Water Quality Indices. BISTUA 2, 19–30. Fernández, N., Villate, C., Terán, O., Aguilar, J., Gershenson, C., 2014d. Comparative Complexity Among Aquatic Ecosystems in a Latitudinal Gradient, in: European Conference in Complex Systems. Froese, T., Stewart, J., 2010. Life After Ashby: Ultrastability and the Autopoietic Foundations of Biological Autonomy. Cybern. Hum. Knowing 17, 7–50. Froese, T., Virgo, N., Izquierdo, E., 2007. Autonomy : a review and a reappraisal Autonomy : a review and a reappraisal. Differences 455–464. Garciandía, J., 2005. Pensar Sistémico: Una Introducción al pensamiento sistémico. Universidad Javeriana, Bogotá. Garrard, G.E., Mccarthy, M.A., Williams, N.S.G., Bekessy, S.A., Wintle, B.A., 2013. A general model of detectability using species traits. Methods Ecol. Evol. 4, 45–52. Georgé, J.-P., Peyruqueou, S., Régis, C., Glize, P., Demazeau, Y., Pavón, J., Corchado, J., Bajo, J., 2009. Experiencing Self-adaptive MAS for Real-Time Decision Support Systems, in: Demazeau, Y., Pavón, J., Corchado, J.M., Bajo, J. (Eds.), 7th International Conference on Practical Applications of Agents and Multi-Agent Systems (PAAMS 2009). Advances in Intelligent and Soft Computing., Advances in Soft Computing. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 302–309. doi:10.1007/978-3-642-00487-2 Gershenson, C., 2002. Classification of Random Boolean Networks, in: Standish, R.K., Bedau, M.A., Abbass, H.A. (Ed.), Artificial Life VIII: Proceedings of the Eight International Conference on Artificial Life. MIT Pres, Cambridge, MA, pp. 1–8. Gershenson, C., 2002. Contextuality: A philosophical paradigm, with applications to philosophy of cognitive science, POCS Essay, COGS. Gershenson, C., 2004a. Introduction to random Boolean networks, in: Bedau, M., Husbands, P., Hutton, T., Kumar, S., Suzuki, H. (Eds.), Workshop and Tutorial 132 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Proceedings, Ninth International Conference on the Simulation and Synthesis of Living Systems (ALife IX). Boston, MA, pp. 160–173. Gershenson, C., 2004b. Updating schemes in random Boolean networks: Do they really matter?, in: Pollack, J., Bedau, M., Husbands, P., Ikegami, T., Watson, R.. (Eds.), Artificial Life IX Proceedings of the Ninth International Conference on the Simulation and Synthesis of Living Systems. MIT Press, pp. 238–243. Gershenson, C., 2005. A General Methodology for Designing Self-Organizing Systems. Gershenson, C., 2007. Design and Control of Self-organizing Systems. CopIt ArXives, Mexico. , México. Gershenson, C., 2010. Computing Networks: A General Framework to Contrast Neural and Swarm Cognitions. Paldyn, J. Behav. Robot. 1, 147–153. Gershenson, C., 2011. Facing Complexity : Prediction vs . Adaptation. Complexity. Gershenson, C., 2011. The sigma profile: A formal tool to study organization and its evolution at multiple scales. Complexity 16, 37–44. Gershenson, C., 2012a. Guiding the self-organization of random Boolean networks. Theory Biosci. 131, 181–191. Gershenson, C., 2012b. The world as evolving information, in: Minai, A., Braha, D., Bar-Yam, Y. (Eds.), Unifying Themes in Complex Systems. Springer Berlin / Heidelberg, Berlin, pp. 100–115. Gershenson, C., 2013a. Information and Computation. Gershenson, C., 2013b. The implications of interactions for science and philosophy. Found. Sci. 18, 781–790. Gershenson, C., 2014. Requisite Variety, Autopoiesis, and Self-organization, in: WOSC (Ed.), 16th Congress of the World Organization of Systems and Cybernetics. WOSC, Ibagué, pp. 45–52. Gershenson, C., Fernández, N., 2012a. Complexity and information: Measuring emergence, self-organization, and homeostasis at multiple scales. Complexity 18, 29–44. Gershenson, C., Fernández, N., 2012b. Complexity and Information: Measuring Emergence, Self-organization, and Homeostasis at Multiple Scales. Arxiv Prepr. arXiv12052026 1– 23. Bibliografía 133 Gershenson, C., Heylighen, F., 2003. When can we call a systems self-organizing?, in: Banzhaf, W., Ziegler, J., Christaller, T., Dittrich, P., Kim, J. (Eds.), Advances in Artificial Life, Lecture Notes in Computer Science. pringer, Berlin / Heidelberg, pp. 606–612. Gershenson, C., Prokopenko, M., 2011. Complex networks. Artif. Life 17, 259–61. doi:10.1162/artl_e_00037 Gershenson, C., Rosenblueth, D.A., 2012. Self-organizing traffic lights at multiple-street intersections. Complexity 17, 23–39. Gyekye, K., 1974. Aristotle and a Modern Notion of Predication. Notre Dame J. Form. Log. Habiba, T., Berger-Wolf, Y., n.d. Theoretic Measures for Idenfying Effective Blockers of Spreading Processes in Dy-namic Networks, in: 6th International Workshop on Mining and Learning with Graphs. Helsinki, Finland. Haken, H., 1989. Information and Self-Organization: A Macroscopic Approach to Complex Systems. Am. J. Phys. doi:10.1119/1.15809 Hamming, R.W., 1950. Error detecting and error correcting codes. Bell Syst. Tech. J. 29, 147– 160. doi:10.1002/j.1538-7305.1950.tb00463.x Hausser, J., Strimmer, K., 2012. Entropy . Hedrih, K., 2007. Leonhard Euler (1707-1783) and Rigid Body Dynamics. Sci. Tech. Rev. LVII, 3–11. Heylighen, F., 2000. Evolutionary Transitions: how do levels of complexity emerge? Complexity 6, 53–57. doi:10.1.1.31.8679 Heylighen, F., 2011. Complexity and self-organization. Encycl. Libr. Inf. Sci. 1–20. doi:10.1063/1.871666 Holzbecher, E., 2007. Environmental Modeling Using MATLAB®. Springer, Berlin/Heildelberg. Holzer, R., de Meer, H., 2011. Methods for approximations of quantitative measures in self- organizing systems, in: Bettstetter, C., Gershenson, C. (Eds.), Self-Organizing Systems, Lecture Notes in Computer Science. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 1–15. doi:10.1007/978-3-642-19167-1 Hugues, J., 2012. R package "CellularAutomaton." Iba, T., 2011. An Autopoietic Systems Theory for Creativity . Procedia - Soc. Behav. Sci. 2, 6610–6625. 134 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Jennings, N.R., Sycara, K., Wooldridge, M., 1998. A Roadmap of Agent Research and Development. Auton. Agents MultiAgent Syst. 1, 7–38. Jennings, N.R., Wooldridge, M.J., 1995. Applying Agent Technology. Appl. Artif. Intell., Appl. Artif. Intell. (USA) 9, 357–369. Kamada, T., Kawai, S., 1989. An algorithm for drawing general undirected graphs. Inf. Process. Lett. doi:10.1016/0020-0190(89)90102-6 Kauffman, S.A., 1969. Metabolic stability and epigenesis in randomly constructed genetic nets. J. Theor. Biol. 22, 437–467. doi:10.1016/0022-5193(69)90015-0 Kauffman, S.A., 2000. Investigations. Oxford University Press, USA. Kaufmann, S., 1993. The origins of order, Oxford University Press. Oxford University Press. Koza, J.R., 1994. Introduction to genetic programming, in: Kinnear, Jr., K.E. (Ed.), Advances in Genetic Programming. MIT Press, pp. 21–42. doi:doi:10.1145/1388969.1389057 Koza, J.R., 1994. Genetic programming as a means for programming computers by natural selection. Stat. Comput. 4, 87–112. Koza, J.R., 1995. Hidden Order: How Adaptation Builds Complexity. Artif. Life. doi:10.1162/artl.1995.2.333 Koza, J.R., Poli, R., 2003. A Genetic Programming Tutorial. Introd. Tutorials Optim. Search Decis. Support. Kruja, E., Marks, J., Blair, A., Waters, R., 2002. A Short Note on the History of Graph Drawing. Graph Draw., Lecture Notes in Computer Science 2265, 602–606. doi:10.1007/3-540-45848-4 Langton, C.G., 1990. Computation at the edge of chaos: Phase transitions and emergent computation. Phys. D Nonlinear Phenom. 42, 12–37. doi:10.1016/0167-2789(90)90064-V Legendre, P., Legendre, L., 1998. Numerical ecology, Numerical Ecology Second English Edition, Developments in evironmental modelling. Elsevier. doi:10.1021/ic050220j Li, W., Packard, N., 1990. The structure of the elementary cellular automata rule space. Complex Syst. 4, 281–297. Lloyd, S., 2001. Measures of complexity: a nonexhaustive list. Control Syst. IEEE 21. doi:10.1109/MCS.2001.939938 Bibliografía 135 Lopez-Ruiz, R., Mancini, H.L., Calbet, X., 1995. A statistical measure of complexity. Phys. Lett. A 209, 321–326. Luisi, P.L., 2003. Autopoiesis: a review and a reappraisal. Naturwissenschaften 90, 49–59. Luisi, P.L., 2007. The Emergence of Life: From Chemical Origins to Synthetic Biology, Angewandte Chemie International Edition. Cambridge University Press. doi:10.1002/anie.200685454 Luque, B., Solé, R.V., 1997. Controlling chaos in Kauffman networks. Eur Lett 37, 597–602. Luque, B., Solé, R.V., 1998. Stable core and chaos control in random Boolean networks. J Phys A Math Gen 31, 1533–1537. Mackenzie, D.I., Royle, J.A., 2005. Designing occupancy studies: General advice and allocating survey effort. J. Appl. Ecol. Madhusudan, H.R., 2010. 1610: The year that improved our 'vision. Resonance 15, 948–953. doi:10.1007/s12045-010-0104-7 Martínez, G.J., Adamatzky, A., Alonso-Sanz, R., 2012. Complex dynamics of elementary cellular automata emerging from chaotic ruleS. Int. J. Bifurc. Chaos 22, 1250023. doi:10.1142/S021812741250023X Martius, G., Herrmann, J., Der, R., 2007. Guided Self-organisation for Autonomous Robot Development , in: Almeida e Costa, F., Rocha, L.M., Costa, E., Harvey, I., Coutinho, A. (Eds.), Advances in Artificial Life, Lecture Notes in Computer Science. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 766–775. doi:10.1007/978-3-540-74913-4 Maturana, H., 2011. Ultrastability...Autopoiesis? Reflective Response to Tom Froese and John Stewart. Cybern. Hum. Knowing 18, 143–152. Maturana, H., Varela, F., 1980. Autopoiesis and Cognition: The realization of living. Reidel Publishing Company, Dordrecht. Mitchell, M., 2009. Complexity: A Guided Tour, Complexity. Oxford University Press, USA. doi:10.1080/09602011.2012.701865 Mitchell, M., Crutchfield, J.P., Hraber, P.T., 1994. Dynamics computation and the edge of chaos: A re-examination. St. FE Inst. Stud. Sci. Complexity proceedings Vol., Santa Fe Institute Proceedings, Volume 19 19, 497–513. doi:10.1016/0167-2789(90)90064-V Moreno, A., Etxeberria, A., Umerez, J., 2008. The autonomy of biological individuals and artificial models. Biosystems. 91, 309–319. 136 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Morin, E., 1992. From the concept of system to the paradigm of complexity. J. Soc. Evol. Syst. doi:10.1016/1061-7361(92)90024-8 Müller, J., Et Al., 1997. Towards agent systems engineering. Data Knowl. Eng. 23, 217–245. doi:10.1016/S0169-023X(97)00013X Müssel, C., Hopfensitz, M., Kestler, H.A., 2010. BoolNet – an R package for generation, reconstruction and analysis of Boolean networks. Bioinformatics 26, 1378–1380. Newman, M.E.J., 2003. The Structure and Function of Complex Networks. SIAM Rev. doi:10.1137/S003614450342480 Nicolis, G., Prigogine, I., 1977. Self-Organization in Non-Equilibrium Systems: From Dissipative Structures to Order Through Fluctuations. Wiley, Chichester. Nicolis, G., Prigogine, I., 1989. Exploring Complexity: An Introduction, New York NY WH Freeman. W.H. Freeman. doi:10.1063/1.2810725 Niklas Luhmann, 1995. Social Systems . Standford University Press. Parrott, L., 2005. Quantifying the complexity of simulated spatiotemporal population dynamics. Ecol. Complex. 2, 175–184. Parrott, L., 2010. Measuring ecological complexity. Ecol. Indic. 10, 1069–1076. Perozo, N., 2011. Modelado multiagente para sistemas emergentes y auto-organizados. Universidad de los Andes. Perozo, N., Aguilar, J., Terán, O., 2008. Proposal for a Multiagent Architecture for Self- Organizing Systems (MASOS), in: PAISI, PACCF and SOCO '08 Proceedings of the IEEE ISI 2008 PAISI, PACCF, and SOCO International Workshops on Intelligence and Security Informatics . Springer-Verlag , Berlin, Heidelberg, pp. 434–4390. Perozo, Niriaska, Aguilar, José, Terán, O, Molina, H., 2012a. A Verification Method for MASOES. IEEE Trans. Syst. Man, Cybern. Part B, IEEE Soc. 99. doi:10.1109/TSMCB.2012.2199106 Perozo, Niriaska, Aguilar, José, Terán, O, Molina, H., 2012b. Un Modelo Afectivo para la Arquitectura Multiagentes para Sistemas Emergentes y Auto-organizados (MASOES). Rev. Técnica Ing. Univ. del Zulia 35, 80–90. Pimm, S.L., 1991. The Balance of Nature? The University of Chicago Press, Chicago. Pond, F., 2006. The Evolution of Biological Complexity. NCSE 26, 22, 27–31. Bibliografía 137 Prokopenko, M., 2009a. Information and Self-Organization: A Macroscopic Approach to Complex Systems, (3rd enlarged ed.). H. Haken. (2006, Springer.) euro79.95 (hardcover), 262 pages. ISBN: 978-3-540-33021-9. Artif. Life 15, 377–383. doi:10.1162/artl.2009.Prokopenko.B4 Prokopenko, M., 2009b. Guided self-organization. HFSP J. 3, 287–289. Prokopenko, M., Boschetti, F., Ryan, A.J., 2009. An information-theoretic primer on complexity, self-organization, and emergence. Complexity 15, 11–28. doi:10.1002/cplx Proulx, R., Parrott, L., 2008. Measures of structural complexity in digital images for monitoring the ecological signature of an old-growth forest ecosystem. Ecol. Indic. 8, 270–284. Ramírez, A., Restrepo, R., Fernández, N., 2003. Evaluación de Impactos Ambientales Causados por Vertimientos Sobre Aguas Continentales. Ambient. y Desarro. 2, 56–80. Randerson, P., Bowker, D., 2008. Aquatic Ecosystem Simulator (AES) - a learning resource for biological science students. Rhodes, J., 2009. Applications of Automata Theory and Algebra via the Mathematical Theory of Complexity to Finite-State Physics, Biology, Philosophy,and Games. World Scientific, Singapore. Rivas Lado, A.L., 2001. Una hipotesis innecesaria. Laplace y el sistema del mundo. Agora 20, 183–214. Ross Ashby, W., 1947a. Principles of the self-organizing dynamic system. J. Gen. Psychol. 37, 125–128. Ross Ashby, W., 1947b. The Nervous System as Physical Machine: with Special Reference to the Origin of Adaptive Behaviour. Mind 56. Ross Ashby, W., 1956. An Introduction to Cybernetics, New York. Ross Ashby, W., 1960a. Design for a Brain: The Origin of Adaptive Behaviour. Chapman & Hall, London. Ross Ashby, W., 1960b. Design for a Brain - The origin of adaptive behavior, Cybernetics. Chapman & Hall. doi:SYST ASHB Royle, J.A., 2006. Site occupancy models with heterogeneous detection probabilities. Biometrics 62, 97–102. 138 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Ruiz-Mirazo, K., Moreno, A., 2004. Basic autonomy as a fundamental step in the synthesis of life. Artif. Life 10, 235–259. doi:10.1162/1064546041255584 Rundell, R.J., Price, T.D., 2009. Adaptive radiation, nonadaptive radiation, ecological speciation and nonecological speciation. Trends Ecol. Evol. 24, 394–399. Scheinerman, E., 1996. Invitation to Dynamical Systems. Prentice Hall , Upper Saddle River, N.J. Schluter, D., 1996. Ecological causes of adaptative radiation. Am. Nat. 148, S40–S64. Schweitzer, F., 1997a. Self-Organization of Complex Structures: From Individual to Collective Dynamics - Some Introductory Remarks, in: Schweitzer, F. (Ed.), Self- Organization of Complex Structures: From Individual to Collective Dynamics. Gordon and Breach, pp. xix–xxiv. Schweitzer, F., 1997b. Active Brownian Particles: Artificial Agents in Physics, in: Schimansky-Geier, L., Pöschel, T. (Eds.), Stochastic Dynamics, Lecture Notes in Physics. Springer-Verlag Berlin and Heidelberg GmbH & Co. KG, pp. 358–371. doi:10.1007/BFb0105623 Seidl, D., 2004. Luhmann ' s theory of autopoietic social systems Ludwig-Maximilians- Universität München. Munich Bus. Res. 2, 1–28. Shalizi, C., Crutchfield, J., 2001. Computational mechanics: Pattern and prediction, structure and simplicity. J. Stat. Phys. 816–879. Shalizi, C., Shalizi, K., Haslinger, R., 2004. Quantifying Self-Organization with Optimal Predictors. Phys. Rev. Lett. 93, 118701. doi:10.1103/PhysRevLett.93.118701 Shalizi, C.R., 2001. Causal Architecture, Complexity and Self-Organization in Time Series and Cellular Automata. Time University, 182. Shannon, C.E., 1948. The mathematical theory of communication. 1963. MD Comput. Comput. Med. Pract., The mathematical theory of communication 14, 306–17. doi:10.1145/584091.584093 Smith, B.S., Wang, J., Egerstedt, M.B., 2008. Persistent formation control of multi-robot networks, in: Proceedings of the IEEE Conference on Decision and Control. pp. 471– 476. Tang, H., Arnold, R.J., Alves, P., Xun, Z., Clemmer, D.E., Novotny, M. V., Reilly, J.P., Radivojac, P., 2006. A computational approach toward label-free protein quantification using predicted peptide detectability, in: Bioinformatics. Bibliografía 139 Teran, O., 2001. Emergent Tendencies in Multi-Agent-based Simulations using Constraint- based Methods to Effect Practical Proofs over Finite Subsets of Simulation Outcomes Table of Contents, Facilities. Manchester Metropolitan University. Ulanowicz, R.E., 2004. Quantitative methods for ecological network analysis. Comput. Biol. Chem. 28, 321–339. Ulanowicz, R.E., 2011. 9.04 - Quantitative Methods for Ecological Network Analysis and Its Application to Coastal Ecosystems, in: Wolanski, E., McLusky, D. (Eds.), Treatise on Estuarine and Coastal Science. Academic Press, pp. 35–57. doi:http://dx.doi.org/10.1016/B978-0-12-374711-2.00904-9 Urrestarazu, H., 2011. Autopoietic systems: A generalized explanatory approach - Part 1. Constr. Found. 6, 307–324. Varela, F., 1979. Principles of Biological Autonomy. Elsevier North Holland., New York, NY. Varela, F.G., Maturana, H.R., Uribe, R., 1974. Autopoiesis: the organization of living systems, its characterization and a model. Curr. Mod. Biol. 5, 187–196. Verlinde, E., 2011. On the origin of gravity and the laws of Newton. J. High Energy Phys. 2011. Villate, C., Gershenson, C., Fernández, N., 2013. Exploring Complexity with a Function for R:Comin, in: European Conference in Complex Systems. Von Bertalanfy, L., 1968. General System Theory: Foundations, Development, Applications. George Brazilier, New York. Walker, L., 2005. Margalef y la sucesión ecológica. Ecosistemas 14, 66–78. Weeks, A., 2010. Neutral Emergence and Coarse Graining Cellular Automata. University of York. Wetzel, R.G., 2001. Limnology: Lake and River Ecosystems, Journal of Phycology. Acedemic Press. doi:10.1046/j.1529-8817.2001.37602.x Wiener, N., 1948. Cybernetics; or, Control and Communication in the Animal and the Machine. Wiley and Sons, New York. Williams, H.T.P., 2006. Homeostatic adaptive networks. Computing. University of Leeds. Wolfram, S.A., 1984. Universality and complexity in cellular automata. Phys. D 10, 1–35. Wolfram, S.A., Gad-el-Hak, M., 2003. A New Kind of Science. Appl. Mech. Rev. 56, 18–20. 140 Modelación Multi-agentes de Sistemas Dinámicos- Un Enfoque 𝑨𝑬𝑪𝑯𝑨 Wolpert, D.H., Macready, W.G., 1999. Self-dissimilarity: An empirically observable complexity measure, in: Bar-Yam, Y. (Ed.), Unifying Themes in Complex Systems: Proceedings of the First International Conference on Complex Systems. Westview Press, Boulder, CO,USA, pp. 626–643. Wooldridge, M., Müller, J., Tambe, M., 1996. Agent theories, architectures, and languages: A bibliography. Intell. Agents II Agent Theor., Lecture Notes in Computer Science 1037, 408–431–431. Wooldridge, M.J., Jennings, N.R., 1995. Intelligent Agents: Theory and Practice. Knowl. Eng. Rev., Knowl. Eng. Rev. (UK) 10, 115–152. Wuensche, A., 1998. Discrete dynamical networks and their attractor basins, in: Standish, R, Henry, B., Watt, S., Marks, R., Stocker, R., Green, D, Keen, S., Bossomaier, T. (Ed.), Complex Systems '98. Sydney, Australia, pp. 3–21. Wuensche, A., Lesser, M.., 1992. The Global Dynamics of Cellular Automata An Atlas of Basin of Attraction Fields of One-Dimensional Cellular AutomataNo Title. Addison- Wesley, Reading, MA. Zeigler. Bernard, Kim, T., Praehofer, H., 2000. Theory of Modeling and Simulation, 2nd ed. ACADEMIC PRESS , San Diego. Zenil, H., 2009. Compression-based investigation of the dynamical properties of cellular automata and other systems 28. Zubillaga, D., Cruz, G., Aguilar, L., Zapotécatl, J., Fernández, N., Aguilar, J., Rosenblueth, D., Gershenson, C., 2014. Measuring the Complexity of Self-Organizing Traffic Lights. Entropy 16, 2384–2407. doi:10.3390/e16052384
1910.13905	1	1910	2019-10-30T14:49:14	Interplay between Topology and Social Learning over Weak Graphs	[ "cs.MA", "cs.SI" ]	We consider a social learning problem, where a network of agents is interested in selecting one among a finite number of hypotheses. We focus on weakly-connected graphs where the network is partitioned into a sending part and a receiving part. The data collected by the agents might be heterogeneous. For example, some sub-networks might intentionally generate data from a fake hypothesis in order to influence other agents. The social learning task is accomplished via a diffusion strategy where each agent: i) updates individually its belief using its private data; ii) computes a new belief by exponentiating a linear combination of the log-beliefs of its neighbors. First, we examine what agents learn over weak graphs (social learning problem). We obtain analytical formulas for the beliefs at the different agents, which reveal how the agents' detection capability and the network topology interact to influence the beliefs. In particular, the formulas allow us to predict when a leader-follower behavior is possible, where some sending agents can control the mind of the receiving agents by forcing them to choose a particular hypothesis. Second, we consider the dual or reverse learning problem that reveals how agents learned: given a stream of beliefs collected at a receiving agent, we would like to discover the global influence that any sending component exerts on this receiving agent (topology learning problem). A remarkable and perhaps unexpected interplay between social and topology learning is observed: given $H$ hypotheses and $S$ sending components, topology learning can be feasible when $H\geq S$. The latter being only a necessary condition, we examine the feasibility of topology learning for two useful classes of problems. The analysis reveals that a critical element to enable faithful topology learning is the diversity in the statistical models of the sending sub-networks.	cs.MA	cs	Interplay between Topology and Social Learning over Weak Graphs Vincenzo Matta, Virginia Bordignon, Augusto Santos, and Ali H. Sayed 1 9 1 0 2 t c O 0 3 ] A M . s c [ 1 v 5 0 9 3 1 . 0 1 9 1 : v i X r a Abstract We consider a distributed social learning problem, where a network of agents is interested in selecting one among a finite number of hypotheses. We focus on weakly-connected graphs where the network is partitioned into a sending part and a receiving part. The data collected by the agents might be heterogeneous, meaning that different sub-networks might be governed by different statistical models. For example, some sub-networks might intentionally generate data from a fake hypothesis in order to influence other agents. This work focuses on a two-step diffusion strategy where each agent: i) updates individually its belief function using its private data; ii) computes a new belief function by exponentiating a linear combination of the logarithmic beliefs of its neighbors. We provide two main contributions. First, we examine what agents learn over weak graphs (social learning problem). We obtain closed-form analytical formulas for the limiting beliefs at the different agents, which allow us to examine in some detail the learning performance of each agent. These formulas reveal how the agents' detection capability and the network topology interact to influence the asymptotic beliefs of the agents. In particular, the formulas allow us to predict if and when a leader-follower behavior is possible, where some sending agents can control the mind of the receiving agents by forcing them to choose a particular hypothesis. Second, we consider the dual or reverse learning problem that reveals how agents learned: given a stream of beliefs collected at a receiving agent, we would like to discover the global influence that any sending component exerts on this receiving agent (topology learning problem). A remarkable and perhaps unexpected interplay between social and topology learning is observed: given H hypotheses and S sending components, topology learning can be feasible when H ≥ S. The latter being only a necessary condition, we then examine the feasibility of topology learning for two useful classes of problems. The conducted analysis reveals that a critical element to enable faithful topology learning is a sufficient degree of diversity in the statistical models of the sending sub-networks. V. Matta is with DIEM, University of Salerno, via Giovanni Paolo II, I-84084, Fisciano (SA), Italy (e-mail: [email protected]). V. Bordignon and A. H. Sayed are with the ´Ecole Polytechnique F´ed´erale de Lausanne EPFL, School of Engineering, CH-1015 Lausanne, Switzerland (e-mails: {virginia.bordignon, ali.sayed}@epfl.ch). A. Santos was with the Adaptive System Laboratory, EPFL, CH-1015 Lausanne, Switzerland (e-mail: [email protected]). This work was supported in part by grant 205121-184999 from the Swiss National Science Foundation (SNSF). An early version with partial results from this paper was presented at ICASSP [1]. October 31, 2019 DRAFT 2 Social learning, topology learning, weakly-connected networks, Bayesian update, diffusion strategy. Index Terms I. INTRODUCTION I n a social learning problem, several agents linked through a network topology form their individual opinions about a phenomenon of interest (learning process) by observing the beliefs of their neighboring agents (social interaction) [2] -- [8]. One relevant paradigm for social learning is that of weakly-connected graphs, which are prevalent in social networks and have been studied in [9] -- [11]. Under this model, there are two categories of sub-networks: the sending sub-networks, which feed information to the receiving sub-networks without getting back information from them. This scenario is common over social networks. For example, a celebrity may have a large number of followers, whose individual opinions are not necessarily followed by the celebrity. Another example is that of media networks, which promote the emergence of opinions by feeding data to users without paying attention to feedback from them. This work addresses two fundamental challenges arising in the study of social learning problems. One challenge is to understand the fundamental mechanism and implications of specific social learning strategies on opinion formation. In particular, over weak graphs, it is critical to understand how the receiving agents are influenced by the sending sub-networks. It is not difficult to envisage that the network topology plays an important role in determining the opinion formation. This motivates the second problem, which can be regarded as a dual learning problem. Given observation of the receiving agents' behavior, we want to establish whether it is possible to learn the strength of connections (weighted topology) from the sending components to the receiving agents. This second question is useful in identifying the main sources for opinion formation over a network. A. Related Work The goal of social learning is to let each individual agent create its own opinion (formally, choose one from among a finite set of hypotheses) through local consultation with its neighbors. One classical distinction for social learning models is Bayesian vs. non-Bayesian models. In the former category, the belief of each agent is obtained by computing posterior distributions through Bayes' rule [2] -- [4], [12] -- [15]. In order to accomplish this task, each agent must have some detailed knowledge of the other agents' likelihoods, and some prior knowledge about the phenomenon of interest. In the latter category, this level of knowledge is not assumed and the agents implement suitable distributed algorithms to interact with their neighbors and to aggregate their beliefs into their own [9] -- [11], [16] -- [22]. The present work considers non-Bayesian learning. There are many useful implementations for non-Bayesian learning. The implementations differ in terms of the rule the agents adopt to update their beliefs. One first distinction concerns the type of distributed strategy. October 31, 2019 DRAFT 3 In particular, we can distinguish between consensus [19], [20] or diffusion [9] -- [11], [21], [22] implementations. A second distinction regards the way the beliefs are combined. In [10], [11], [19], they are combined linearly, while a linear combination of the logarithmic beliefs is used in [21], [22]. This latter form is motivated by the fact that, in many detection problems, the best detection statistic is given by the linear combination of log-likelihoods and not of likelihoods. As a matter of fact, using the log-belief combination can help achieve an improved (i.e., faster) learning rate [22]. Once a particular learning rule has been chosen, the behavior of opinion formation will depend heavily on the type of network where the information propagates. In this respect, the majority of prior works in the literature focus on strongly-connected networks. In these networks, there always exist paths linking any two agents in both directions (which makes them connected) and, in addition, at least one agent has a self-loop and places some partial trusts in its own data (which makes them strongly connected). Under a homogeneous setting where the underlying true hypothesis is the same for the entire network, it has been shown that over strongly-connected networks all agents are able to discover and agree on the true hypothesis. This result is available in [11], [19] for diffusion and consensus rules with linear belief combination, and in [22] for the diffusion rule with linear log-belief combination. There are results available also for the heterogeneous setting, where different agents might have different data models, different likelihoods, and promote different opinions across the network. In particular, the diffusion rule with linear log-belief combination with a doubly-stochastic combination matrix is considered in [21], where it is shown that, over a strongly-connected network, all agents reach a common opinion, by minimizing cooperatively the sum of Kullback-Leibler (KL) divergences across the network. In contrast, the important case of weakly-connected networks has received limited attention and was addressed more recently in [9], [10] by using the linear-belief-combination rule. Several interesting phenomena arise over weak graphs, which are absent from strongly-connected networks. The main difference in this work from [9], [10] is that we now consider the following general setting: i) diffusion-type algorithms with linear combination of log-beliefs (as opposed to the beliefs themselves), ii) heterogeneous data and inference model, and iii) inverse topology learning. B. Main Results This work leads to two main contributions. First, we characterize (Theorem 1) the limiting (as learning time elapses) agents' belief through analytical formulas that depend in a transparent manner on inferential descriptors (Kullback-Leibler divergences) and network descriptors (limiting combination matrix power). Some revealing behaviors can be deduced from these formulas. For example, we will be able to characterize a mind-control effect in terms of the interplay between the detection capacity of each agent and the centrality of the different agents. We will see that useful analogies arise with what has been observed in [10]. For example, some of the effects shown in [10] will now be shown to hold under greater generality since our formulas can be obtained by October 31, 2019 DRAFT 4 relaxing some assumptions used in [10], in particular, the all-truths-are-equal assumption, and the prevailing- signal assumption. However, we will observe also some remarkable distinctions with respect to [10]. For instance, we will observe that the individual belief of each receiving agent will necessarily collapse to (or concentrate at) a single hypothesis (which might be different across the agents). This is in contrast with [10], where the beliefs of the receiving agents could end up assigning some probabilities to more than one hypothesis. The reason for this difference in behavior arises from the difference in the combination rule used in this work in comparison to [10]. The second contribution concerns the topology learning problem. We are interested to learn the topology linking the receiving agents to the sending components. This question is interesting because it would then allow us to identify the main sources of information in a network and how they influence opinion formation. Nevertheless, the inverse topology problem is challenging because we will assume that we can only observe the beliefs evolving at the receiving agents. In particular, we will be able to recover some macroscopic topology information, in terms of the limiting weights that each receiving agent sees from each sending component. We call this a macroscopic information since these weights incorporate: i) the global effect coming from all agents belonging to a sending component, and ii) the effect of intermediate receiving agents linked to the receiving agent under consideration. The relevance in estimating these global weights relies on the fact that the limiting beliefs of the receiving agents depend solely on this aggregate information. We will establish conditions under which topology inference becomes feasible. More specifically, given H hypotheses and S sending components, under the assumption of homogeneous statistical models within each sending component, we will ascertain that a necessary condition to achieve consistent topology learning is (Lemma 1): H ≥ S (1) Once established a necessary condition, we will examine some useful models to see whether topology learning can be in fact achieved. We consider first a structured Gaussian model where: i) the true underlying (Gaussian) distributions are distinct across the sending sub-networks; and ii) the (Gaussian) likelihoods are equal across the sending sub-networks, and contain the true distributions. For this setting, we will show in Theorem 2 that topology learning is feasible only when S = 2. We then recognize that one fundamental element for topology learning is the diversity between the sending sub-networks. Adding this further element, we will establish in Theorem 3 that the problem is feasible for any S provided that (1) holds, and even under more general (e.g., non-Gaussian) models. In summary, we remark that there are two learning problems coexisting in our work: a social learning problem and a topology learning problem. The former is the primary, or direct inferential problem that the agents are deployed for. The latter is the dual, or reverse problem, which is in fact based on observation of the output (the October 31, 2019 DRAFT 5 beliefs) of the primary learning problem. One useful conclusion stemming from our analysis is to reveal some unexpected interplay between these two coexisting learning problems -- see Sec. VIII further ahead. Notation. We use bold font notation for random variables, normal fonts for their realizations. Capital letters are generally used for matrices, whereas small font letters for vectors or scalars. Given a matrix M, the symbol M† denotes its Moore-Penrose pseudoinverse. The L× L identity matrix is denoted by IL. Likewise, the L× 1 vector of all ones is denoted by 1L. The notation "a.s." signifies "almost-surely". A. Data Model and Inference II. BACKGROUND We consider a network of N agents that collect streaming data from the environment. Formally, the random variable ξk,i ∈ Xk denotes the data at agent k ∈ {1, 2, . . . , N} collected at time i ∈ N. The data are assumed to be independent over time, whereas they can be dependent across agents (i.e., over space). We work under a heterogeneous setting. First, the space where the data are defined, Xk, is allowed to vary across agents. For example, the data at different agents can be random vectors of different sizes. In the social learning context, this scenario is not that uncommon since different types of attributes can be observed by different users across the network. Second, the data ξk,i are assumed to be generated according to certain distributions fk(ξ), which are allowed to vary across the agents as well, namely, for k = 1, 2, . . . , N: ξk,i ∼ fk [true distribution] (2) This type of heterogeneity can arise in practice for several reasons, for example, some agents may intentionally inject fake data to let the other agents have a bias towards fake hypotheses. Based on the available data {ξk,i}, the network agents aim to solve an inferential problem that consists in choosing one state of nature from among a finite collection Θ = {1, 2, . . . , H}, with H denoting the number of possible hypotheses. To solve such inferential problem, the agents rely on a family of likelihood functions. Specifically, for k = 1, 2, . . . , N, the likelihood function of agent k is denoted by: (3) We will often write Lk(θ) instead of Lk(ξθ) for simplicity. In our treatment, we assume that the data can be modeled either as continuous or discrete random variables, with the same (continuous or discrete) nature ξ ∈ Xk, θ ∈ Θ [likelihoods] Lk(ξθ), across all agents. Accordingly, both the true distributions and the likelihoods will be either probability density or probability mass functions, depending on the considered type of random variables. We remark that the considered model includes the possibility that the true distribution fk(ξ) is equal to a certain likelihood Lk(ξθ0) (as assumed, e.g., in [10], [22]) and in this case θ0 can be considered as the true underlying hypothesis. More generally (e.g., in [21]) the true distribution fk(ξ) is not equal to any of the likelihoods Lk(ξθ), which might happen when the agents have an approximate knowledge of the statistical October 31, 2019 DRAFT models, and even the likelihood more similar to the true distribution can be a mismatched version thereof. In order to quantify the dissimilarity between the true distribution and a certain likelihood, we will use the Kullback-Leibler (KL) divergence [23]: 6 D[fkLk(θ)] (cid:44) Efk(cid:20)log fk(ξk) Lk(ξkθ)(cid:21) , k = 1, 2, . . . , N. (4) We remark that we have written ξk in place of ξk,i to highlight identical distributions across time, and that the expectation is computed under the actual distribution of ξk, i.e., under fk. In the forthcoming treatment we use the following regularity condition [21], [22]. Assumption 1 (Finiteness of KL Divergences): All the KL divergences in (4) are well-posed, namely, for all k = 1, 2, . . . , N, and all θ ∈ Θ we have that: D[fkLk(θ)] < ∞ (5) (cid:3) Remark 1 (Likelihood Support): Assumption 1 implies that the likelihood Lk(ξk,iθ) can be equal to zero (cid:3) only for an ensemble of realizations ξk,i having zero probability under the distribution fk. B. Social Learning Algorithm Motivated by the diffusion strategy in [10], [11] for opinion formation over social networks, in this article we consider the useful variations studied in [21], [22]. The learning procedure employs a two-step algorithm that can be described as follows. For each admissible hypothesis θ ∈ Θ at time i, each agent k uses its own fresh private observation, ξk,i, to compute the local likelihood Lk(ξk,iθ). Using this likelihood, agent k updates its local belief, µk,i−1(θ), obtaining an intermediate belief ψk,i(θ) through the following update rule: ψk,i(θ) = µk,i−1(θ)Lk(ξk,iθ) )Lk(ξk,iθ µk,i−1(θ (cid:48) (cid:88)θ(cid:48)∈Θ (cid:48) ) (6) Then, agent k aggregates the intermediate beliefs received from its neighbors through the following combination rule (the division by the denominator term in (7) is meant to ensure that µk,i(θ) is a probability measure with its entries adding up to one): µk,i(θ) = exp(cid:40) N(cid:88)(cid:96)=1 exp(cid:40) N(cid:88)(cid:96)=1 (cid:88)θ(cid:48)∈Θ a(cid:96)k log ψ(cid:96),i(θ)(cid:41) )(cid:41) a(cid:96)k log ψ(cid:96),i(θ (cid:48) (7) where a(cid:96)k is the nonnegative combination weight that agent k uses to scale the intermediate log-belief received from agent (cid:96). It is assumed that a(cid:96)k is equal to zero if k does not receive information from (cid:96), which means October 31, 2019 DRAFT that agent k can combine only intermediate beliefs received from its neighbors. When collected into a combi- nation matrix A (with ((cid:96), k)-entry equal to a(cid:96)k), these combination weights are assumed to obey the standard requirements that make A a left-stochastic matrix, namely, we have that: 7 A (cid:62)1N = 1N ⇔ N(cid:88)(cid:96)=1 a(cid:96)k = 1, ∀k = 1, 2, . . . , N. (8) The value of µk,i(θ) provides an estimate for the likelihood by agent k at time i that the true hypothesis value is θ. We remark that, differently from [10], the second step in (7) combines linearly the logarithm of the intermediate beliefs, log ψ(cid:96),i(θ), in the neighborhood of agent k. Exponentiation and normalization are used to construct the final belief. We invoke the following standard initial condition, motivated by the fact that, at time i = 0, the agents have no elements to discard any hypothesis [21], [22]. Assumption 2 (Initial Beliefs): All agents start by assigning strictly positive probability mass to all hypotheses, (cid:3) namely, µk,0(θ) > 0 for all θ ∈ Θ and k = 1, 2, . . . , N. Remark 2 (Strict Positiveness of Beliefs): From (6) we see that, if µk,i−1(θ) > 0, then ψk,i(θ) > 0 since Lk(ξk,iθ) > 0 (but for zero-probability sets) -- see Remark 1. Strict positiveness of µk,i(θ) now follows (cid:3) from (7) since the combination weights are nonnegative and convex. III. WEAK GRAPHS In this section, we consider the case of a weak graph (or weakly-connected network), which is defined as follows [9], [10]. The overall network N = {1, 2, . . . N} is divided into S +R disjoint components -- see Fig. 1. The first S sub-networks form the sending part S, whereas the remaining R sub-networks form the receiving part R: N = S ∪ R, S (cid:44) Ns, R (cid:44) NS+r. (9) S(cid:91)s=1 R(cid:91)r=1 We remark that S and R denote the number of sending and receiving components, respectively. They do not denote the cardinalities of S and R, which are instead given by: S = Ns, R = NS+r, (10) S(cid:88)s=1 R(cid:88)r=1 where the notation Nj denotes the number of agents in sub-network Nj. Communication from a sending component to a receiving component is permitted, whereas communication in the reverse direction is forbidden. Communication between sending agents is possible, but a sending sub-network is identified by the following two conditions: i) each sending sub-network is strongly connected [24]; and ii) agents belonging to different sending sub-networks do not communicate with each other. If some sending sub-networks communicate with each other, then they can be blended into a larger sending sub-network. The R receiving sub-networks are all individually assumed to be connected (but not necessarily strongly connected), with communication among the October 31, 2019 DRAFT 8 Fig. 1. One example of weakly-connected network, with sending sub-networks N1 and N2, and receiving sub-networks N3 and N4. R sub-networks being allowed. In particular, we assume that each receiving sub-network is connected to at least one agent in each sending sub-network. Without loss of generality, we assume that the network nodes across the S + R components are listed in increasing order. According to the above description, the combination matrix corresponding to a weakly- connected network admits the following convenient block decomposition [9], [10]: A = 0 AS ASR AR  where the matrix AS = blockdiag {AN1, AN2, . . . , ANS} contains the combination weights within the sending sub-networks, and has a block-diagonal form since communication between sending sub-networks is forbidden. The matrix block ASR contains the combination weights for the communication that takes place from sending agents to receiving agents. The left-bottom matrix block is zero since there are no direct links from the receiving agents to the sending ones. Finally, the matrix block AR contains the combination weights ruling the communication among the receiving agents. Figure 1 offers a graphical illustration of the weakly-connected paradigm. It was shown in [9] that the limiting combination matrix power has the following structure: (11) (12) (13) DRAFT E EW A∞ (cid:44) lim i→∞ Ai = E = blockdiag(cid:110)p(1)1(cid:62) 0 0 E Ω 0  = 0  , NS(cid:111) N2, . . . , p(S)1(cid:62) N1, p(2)1(cid:62) where October 31, 2019 N2<latexit sha1_base64="apOwlPSaog+sD+hSTDPx7xNUKk4=">AAACFHicbVC7TgJBFL2LL8QXamJjM5GYWJFdLLQk2FgZSOSRwEpmhwEmzM5uZmZNyGZ/w9hq7x/Yqa29jf/gHzgLFAKeZJKTc869c3O8kDOlbfvLyqysrq1vZDdzW9s7u3v5/YOGCiJJaJ0EPJAtDyvKmaB1zTSnrVBS7HucNr3RVeo376lULBC3ehxS18cDwfqMYG2ku46P9ZBgHt8k3VKumy/YRXsCtEycGSmUj2rf7KXyXu3mfzq9gEQ+FZpwrFTbsUPtxlhqRjhNcp1I0RCTER7QtqEC+1S58eTqBJ0apYf6gTRPaDRR/07E2Fdq7HsmmV6pFr1U/M9rR7p/6cZMhJGmgkw/6kcc6QClFaAek5RoPjYEE8nMrYgMscREm6LmNs3acWMaKZMMdZJ25Cw2skwapaJzXizVTFkVmCILx3ACZ+DABZThGqpQBwISHuEJnq0H69V6sz6m0Yw1mzmEOVifv/8SovA=</latexit>N3<latexit sha1_base64="Tg3cNi4Q2IaW/AKN0pMZvmbfrW8=">AAACFHicbVC7TgJBFL3rE/GFmtjYTCQmVmQXCi0JNlYGEnkksJLZYYAJs7ObmVkTstnfMLba+wd2amtv4z/4B84ChYAnmeTknHPv3Bwv5Exp2/6yVlbX1jc2M1vZ7Z3dvf3cwWFDBZEktE4CHsiWhxXlTNC6ZprTVigp9j1Om97oKvWb91QqFohbPQ6p6+OBYH1GsDbSXcfHekgwj2+SbinbzeXtgj0BWibOjOTLx7Vv9lJ5r3ZzP51eQCKfCk04Vqrt2KF2Yyw1I5wm2U6kaIjJCA9o21CBfarceHJ1gs6M0kP9QJonNJqofydi7Cs19j2TTK9Ui14q/ue1I92/dGMmwkhTQaYf9SOOdIDSClCPSUo0HxuCiWTmVkSGWGKiTVFzm2btuDGNlEmGOkk7chYbWSaNYsEpFYo1U1YFpsjACZzCOThwAWW4hirUgYCER3iCZ+vBerXerI9pdMWazRzBHKzPXwDFovE=</latexit>N4<latexit sha1_base64="YJhH6kS2i8SMT3edlXdDQ2OOO04=">AAACFHicbVC7SgNBFL3rM8ZXVLCxGQyCVdiNgpYhNlaSgHlAsobZySQZMju7zMwKYdnfEFvt/QM7tbW38R/8A2eTFCbxwMDhnHPvXI4Xcqa0bX9ZS8srq2vrmY3s5tb2zm5ub7+ugkgSWiMBD2TTw4pyJmhNM81pM5QU+x6nDW94lfqNeyoVC8StHoXU9XFfsB4jWBvpru1jPSCYxzdJ5zzbyeXtgj0GWiTOlORLh9Vv9lJ+r3RyP+1uQCKfCk04Vqrl2KF2Yyw1I5wm2XakaIjJEPdpy1CBfarceHx1gk6M0kW9QJonNBqrfydi7Cs18j2TTK9U814q/ue1It27dGMmwkhTQSYf9SKOdIDSClCXSUo0HxmCiWTmVkQGWGKiTVEzm6btuDGNlEmGOkk7cuYbWST1YsE5KxSrpqwyTJCBIziGU3DgAkpwDRWoAQEJj/AEz9aD9Wq9WR+T6JI1nTmAGVifvwJpovI=</latexit>N1<latexit sha1_base64="nRFT06pUrNWj/Ds2PdtE2OyDKJo=">AAACFHicbVC7TgJBFL2LL8QXamJjM5GYWJFdLLQk2FgZSOSRwEpmhwEmzM5uZmZNyGZ/w9hq7x/Yqa29jf/gHzgLFAKeZJKTc869c3O8kDOlbfvLyqysrq1vZDdzW9s7u3v5/YOGCiJJaJ0EPJAtDyvKmaB1zTSnrVBS7HucNr3RVeo376lULBC3ehxS18cDwfqMYG2ku46P9ZBgHt8kXSfXzRfsoj0BWibOjBTKR7Vv9lJ5r3bzP51eQCKfCk04Vqrt2KF2Yyw1I5wmuU6kaIjJCA9o21CBfarceHJ1gk6N0kP9QJonNJqofydi7Cs19j2TTK9Ui14q/ue1I92/dGMmwkhTQaYf9SOOdIDSClCPSUo0HxuCiWTmVkSGWGKiTVFzm2btuDGNlEmGOkk7chYbWSaNUtE5L5ZqpqwKTJGFYziBM3DgAspwDVWoAwEJj/AEz9aD9Wq9WR/TaMaazRzCHKzPX/1uou8=</latexit>is a block diagonal matrix that stacks the Ns × 1 Perron eigenvectors p(s) associated with the s-th sending sub-network1, and where 9 (15) We denote the entries of Ω by [ω(cid:96)k] and we keep indexing the columns of the S × R matrix Ω with an W = ASR (IR − AR) −1, Ω = EW. index k ∈ {S + 1,S + 2, . . . ,S + R} . (16) Since the limiting matrix power is left-stochastic and has a zero bottom block, the limiting weights ω(cid:96)k obey: i.e., Ω is left-stochastic. From (15) we can also write: ω(cid:96)k = 1, (cid:88)(cid:96)∈S Ω = EASR(IR + AR + A2 R + . . . ), (17) (18) whence we see that ω(cid:96)k embodies the sum of influences over all paths from sending agent (cid:96) to receiving agent k. IV. LIMITING BELIEFS OF RECEIVING AGENTS Let us momentarily consider a single-agent scenario where agent (cid:96) operates alone. A natural way for agent (cid:96) to choose a hypothesis would be to choose the θ that gives the best match between a model L(cid:96)(θ) and the distribution of the observed data, f(cid:96). One measure of the match between f(cid:96) and L(cid:96)(θ) is the KL divergence D[f(cid:96)L(cid:96)(θ)]. The smaller the value of this divergence is, the higher the match between the data and the model. For this reason, a strategy could be that of choosing the θ that minimizes the divergence D[f(cid:96)L(cid:96)(θ)]. In the social learning context, this optimization problem turns into a distributed optimization problem. In particular, under our social learning setting over weak graphs, we will show soon (Theorem 1) that the log- belief diffusion strategy in (6) -- (7) will end up minimizing (without knowing the true distributions) the following average divergence at receiving agent k ∈ R: Dk(θ) (cid:44)(cid:88)(cid:96)∈S ω(cid:96)kD[f(cid:96)L(cid:96)(θ)] (19) which is a weighted combination, through the limiting combination weights {ω(cid:96)k}, of the KL divergences of the sending agents reaching k. The role of average divergence measures like the one in (19) already arose in the case of strongly-connected networks. For example, it was shown in [21], [22] that with the log-belief diffusion 1For s = 1, 2, . . . , S, the Perron eigenvector of the sub-matrix ANs corresponding to the s-th sending sub-network is given by: (cid:62) Ns p(s) = 1, p(s) ANs p(s) = p(s), 1 (cid:96) > 0, (cid:96) = 1, 2, . . . , Ns. (14) DRAFT October 31, 2019 10 strategy in (6) -- (7), each agent ends up minimizing the same weighted combination of divergences. Under classical identifiability conditions, such minimization leads each individual agent to discover the true underlying hypothesis [22] or the best available approximation thereof [21]. In our weak-graph setting, however, the effect of minimizing Dk(θ) (which depends on the particular receiving agent k) will be less obvious. We already see from (19) that the average divergence combines topological attributes, encoded in the limiting combination weights, with inferential attributes, encoded in the local KL divergences. The interplay arising between the network topology and social learning will be critical in determining the choices of the receiving agents. Throughout the work, we will invoke the following classical identifiability assumption that, as we will see in our examples, arises naturally in several models of interest. Assumption 3 (Unique Minimizer): For each k = 1, 2, . . . , N, the function Dk(θ) has a unique minimizer: θ(cid:63) k (cid:44) argmin θ∈Θ Dk(θ) (20) (cid:3) We are now ready to characterize the limiting belief of the receiving agents. The following theorem is an extension to the case of weakly-connected graphs of similar theorems proved in [21], [22] for the case of strongly-connected graphs. Theorem 1 (Belief Collapse at Receiving Agents): Let k ∈ R. Under Assumptions 1-3 we have that: i→∞ µk,i(θ(cid:63) k) lim a.s. = 1 (21) where the symbol a.s. convergence of the belief to zero takes place at an exponential rate as: = denotes that the pertinent limit exists almost-surely. Moreover, for all θ (cid:54)= θ(cid:63) k, the lim i→∞ log µk,i(θ) i a.s. = Dk(θ(cid:63) k) − Dk(θ) (22) Proof: The proof combines the techniques to establish the convergence of the social learning algorithm used, e.g., in [21], [22] for strongly-connected graphs, with the convergence results of the combination matrix over weak graphs used in [10], [11]. The detailed steps are reported in Appendix A. Several insightful conclusions arise from Theorem 1. Remark 3 (Collapse): The limiting belief of each receiving agent is always degenerate, meaning that it (cid:3) collapses to a single hypothesis, when sufficient time for learning is allowed. Remark 4 (Discord): Different agents can in principle be in discord, since they can converge to different hy- potheses. The particular behavior (who chooses what) will depend on a weighted combination of KL divergences. (cid:3) Remark 5 (Mind Control): We see from (19) and (20) that only the local divergences corresponding to the k at sending agents, (cid:96) ∈ S, determine the value of Dk(θ) and, hence, of θ(cid:63) k. Therefore, the limiting hypothesis θ(cid:63) October 31, 2019 DRAFT 11 agent k is determined by the KL divergences pertaining only to the statistical models within the sending sub- networks, and, hence, irrespective of the data sensed at agent k within its receiving sub-network. In a nutshell, we see the emergence of a mind-control effect: i) the final states of the receiving agents are dependent only upon the properties of the detection problems at the sending agents; and ii) different network topologies allow the sending agents to drive the receiving agents to potentially different decisions. The emergence of a mind-control effect over weakly-connected networks was already discovered in [9], [10]. Here, we establish a similar effect albeit one where the receiving agents attain degenerate beliefs. In comparison, in [9], [10], receiving agents end (cid:3) up assigning nonzero probabilities to more than one belief. Remark 6 (Distinctions relative to [10]): Besides these commonalities, there are nevertheless important distinctions between the behavior observed in our setting and what was observed in [10]. First, for the linear- belief-combination algorithm used in [10], the limiting belief of a receiving agent was shown to be a convex combination of the limiting beliefs of the sending agents, with the convex weights coming from the matrix W in (15). This means that if two sending sub-networks have, e.g., limiting beliefs that collapse to different hypotheses, then the limiting belief of a receiving agent can have nonzero values a these two different locations. In comparison, in the log-belief-combination algorithm considered here the limiting beliefs are always concentrated at a single hypothesis. Moreover, in [10], the analysis required some regularity assumptions called all-truths-are-equal and prevailing- signal assumptions. These assumptions are not required in Theorem 1. In a sense, the lack of these assumptions ascertains that some relevant effects, such as mind control, hold under greater generality and more relaxed settings. Finally, it is useful to observe that one fundamental role in our setting is played by the weighting matrix (cid:3) Ω = EW , while in [10] the main role was played by W alone. A. Canonical Examples In order to examine in more detail the implications of Theorem 1, we consider a simple yet insightful example. The sending and receiving components are: S = N1 ∪ N2, R = N3, (23) namely, we have two sending sub-networks, N1 and N2, and one receiving sub-network N3. For what concerns the inferential model, we assume there are three possible hypotheses, θ ∈ {1, 2, 3}. The likelihood functions are the same across all agents. In particular, we assume that, for all ξ ∈ R, and for θ ∈ {1, 2, 3}: exp(cid:26)− (ξ − mθ)2 2 (cid:27) , L(ξθ) = 1√ 2π October 31, 2019 (24) DRAFT where the means corresponding to the different hypotheses are chosen as, for some ∆ > 0: m1 = −∆, m2 = 0, m3 = +∆. 12 (25) We further assume that the true distributions of the sending sub-networks (recall that only the sending sub- networks determine the limiting beliefs of the receiving sub-network) are Gaussian distributions, with expec- tations chosen among the expectations in (25). In particular, we assume that agents belonging to sub-network N1 generate data according to model θ = 1, i.e., with expectation equal to −∆, whereas agents belonging to sub-network N2 generate data according to model θ = 3, i.e., with expectation equal to +∆. Formally we write: f(cid:96)(ξ) = f(cid:96)(ξ) = 1√ 2π 1√ 2π exp(cid:26)− (ξ + ∆)2 exp(cid:26)− (ξ − ∆)2 2 2 (cid:27) , (cid:27) , ∀(cid:96) ∈ N1, ∀(cid:96) ∈ N2. (26) (27) Recalling that the KL divergence between two unit-variance Gaussian distributions of expectations a and b is given by 0.5(a − b)2, under the setting described above we can write, for all k ∈ R: ω(cid:96)kD[f(cid:96)L(cid:96)(θ)] Dk(θ) = (cid:88)(cid:96)∈S = (cid:88)(cid:96)∈N1 (−∆ − mθ)2 ω(cid:96)kD[f(cid:96)L(θ)] + (cid:88)(cid:96)∈N2 2 (cid:88)(cid:96)∈N1 ω(cid:96)k + ω(cid:96)kD[f(cid:96)L(θ)] (∆ − mθ)2 2 (cid:88)(cid:96)∈N2 = which further implies: ω(cid:96)k, (28) (29) Dk(1) = 2∆2 (cid:88)(cid:96)∈N2 Dk(2) = ∆2 2 , Dk(3) = 2∆2 (cid:88)(cid:96)∈N1 ω(cid:96)k, ω(cid:96)k, where, in the intermediate equality, we used (17). As a result, we can compute the limiting hypothesis, for each k ∈ R, as: θ(cid:63) k = argmin(cid:40)4(cid:88)(cid:96)∈N2 ω(cid:96)k, 1, 4(cid:88)(cid:96)∈N1 ω(cid:96)k(cid:41) . (30) From (18), one can argue that(cid:80)(cid:96)∈Ns ω(cid:96)k reflects the sum of influences over all paths connecting all sending agents in sub-network s to receiving agent k. In order to find the minimizer in (30), we start by using (17) in (30), which yields: θ(cid:63) k = argmin(cid:40)1 − (cid:88)(cid:96)∈N1 ω(cid:96)k, 0.25, (cid:88)(cid:96)∈N1 ω(cid:96)k(cid:41) . October 31, 2019 (31) DRAFT In view of Theorem 1, the belief of the k-th receiving agent will converge to θ(cid:63) conditions are simultaneously verified: k = 1 if the following two 13 Taking the most stringent condition in (32) reveals that: ⇔ (cid:88)(cid:96)∈N1 ω(cid:96)k ⇔ (cid:88)(cid:96)∈N1 ω(cid:96)k > 0.75, ω(cid:96)k > 0.5. 1 − (cid:88)(cid:96)∈N1 1 − (cid:88)(cid:96)∈N1 ω(cid:96)k < 0.25 ω(cid:96)k < (cid:88)(cid:96)∈N1 k = 1 ⇔ (cid:88)(cid:96)∈N1 θ(cid:63) ω(cid:96)k > 0.75. (32) (33) In summary, we conclude that agent k follows the opinion promoted by sending sub-network N1 if the influence of sub-network N1 on agent k is "sufficiently large". The situation is reversed if the influence of sub-network N2 is sufficiently large, namely, θ(cid:63) k = 3 ⇔ (cid:88)(cid:96)∈N2 ω(cid:96)k > 0.75, (34) where we recall that hypothesis θ = 3 is promoted by sub-network N2. However, there is another possibility. It occurs when: ω(cid:96)k < 0.75 ω(cid:96)k < 0.75. (35) (cid:88)(cid:96)∈N1 and (cid:88)(cid:96)∈N2 In this case, no clear dominance from one sub-network can be ascertained, and each receiving agent will choose k = 2, i.e., an opinion that does not coincide with any of the opinions promoted by the sending sub-networks. θ(cid:63) From (33) and (34), we see that the dominance of one of the sending sub-networks is determined by the aggregate influence(cid:80)(cid:96)∈N1 ω(cid:96)k, with the complementary aggregate influence being(cid:80)(cid:96)∈N2 ω(cid:96)k = 1−(cid:80)(cid:96)∈N1 ω(cid:96)k. The main way to manipulate these factors consists in varying the sizes of the sending sub-networks or their connections with the receiving agents. In order to illustrate more carefully the possible scenarios, we consider the following simulation framework: • The strongly-connected sending components N1 and N2 are generated as Erdos-R´enyi random graphs with connection probability q, and the entries of the corresponding combination matrix are determined by the averaging rule, namely,2 1/nk, if k (cid:54)= (cid:96) are neighbors or k = (cid:96) 0, otherwise (36) a(cid:96)k = where nk is the number of neighbors of node k (including node k itself). In our experiments we set q = 0.7. 2When drawing the random graph, we have verified that there exists at least one self-loop. October 31, 2019 DRAFT 14 • An agent k is connected to a sending agent through a Bernoulli distribution with parameter πs, which depends on the sending sub-network s. Given the total number dk, of directed edges from sending agents to agent k, we initially set a(cid:96)k = 1/dk. The combination matrix A of the overall network N1 ∪ N2 ∪ N3 is normalized so that it is left-stochastic. It is now possible to examine different scenarios by manipulating the size of the sending sub-networks as well as the send-receive connection probabilities πs. network graph belief evolution Fig. 2. How majorities build a majority: Convergence of beliefs when the size of sending sub-network N1 is dominant. -- Setup 1 or "How majorities build a majority". In Fig. 2, we set π1 = π2 = 0.5, i.e. it is equally probable that a receiving agent connects to any sending agent, irrespective of the sending sub-network. In view of this uniformity, we can expect that the limiting weights ω(cid:96)k are sufficiently uniform across the two sending sub- networks and, hence, that the value of(cid:80)(cid:96)∈N1 ω(cid:96)k is primarily determined by the sub-network size N1. In the example we are going to illustrate, we assume that the number of agents in sub-network N1 is three times larger than the size of sub-network N2. From the lowermost panel in Fig. 2, we observe that receiving agents October 31, 2019 DRAFT 1234567891011121314151617181920050100i0.00.51.0µ17,i(θ)Agent17050100i0.00.51.0µ18,i(θ)Agent18050100i0.00.51.0µ19,i(θ)Agent19050100i0.00.51.0µ20,i(θ)Agent20θ=1θ=2θ=315 network graph belief evolution Fig. 3. How filter bubbles build a majority: Convergence of beliefs when the connectivity from sending sub-network N1 is dominant. (18, 19, 20) converge to θ = 1, i.e., to the opinion promoted by N1. We see also that agent 17 takes a minority position and opts for θ = 2, i.e., it does follow neither the opinion promoted by N1 nor by N2. This shows the following interesting effect. Even if sub-network N1 is bigger, for the specific topology shown in the example (see uppermost panel of Fig. 2), the aggregate weight of agent 17 is (cid:80)(cid:96)∈N1 ω(cid:96) 17 = 0.645. This means that condition (35) is actually verified, which explains why agent 17 opts for θ = 2. In summary, we observed that building a majority of agents in N1 relative to N2 yields a majority of receiving agents opting for the hypothesis promoted by N1. -- Setup 2 or "How filter bubbles build a majority". Under this setup, we assume that both sending components have the same size, however πs is different for each of the two components. We set π1 = 0.9 and π2 = 0.1 in order to motivate agent k to have more connections with sub-network N1 than with N2. This scenario is considered in Fig. 3, where we see that all agents end up agreeing with opinion θ = 1, i.e., with the opinion promoted by the sending component N1. Therefore, closing a receiving agent into the "filter bubble" determined October 31, 2019 DRAFT 123456789101112050100i0.00.51.0µ9,i(θ)Agent9050100i0.00.51.0µ10,i(θ)Agent10050100i0.00.51.0µ11,i(θ)Agent11050100i0.00.51.0µ12,i(θ)Agent12θ=1θ=2θ=316 network graph belief evolution Fig. 4. Truth is somewhere in between: Convergence of beliefs under balanced influences. by the overwhelming flow of data coming from N1 essentially makes these agents blind to the solicitations coming from N2. We notice that, while in the example of Fig. 2 one receiving agent behaves differently from the majority of agents, for the specific parameters used in Fig. 3 all agents opt for the same hypothesis. However, another type of distinction arises in terms of convergence rate. We observe that agent 10 is reluctant to the received solicitations, since for the first half of the observation window the preferred hypothesis is θ2, and the convergence to θ1 is significantly slower than the convergence of the other agents. -- Setup 3 or "Truth is somewhere in between". We now address the balanced case where the sending sub- networks have the same size and similar number of connections to the receiving sub-network (π1 = π2 = 0.5). Under this setting, it is expected that no dominant behavior emerges, and (35) holds. We see in Fig. 4 that the receiving agents' opinions tend to converge with full confidence to hypothesis θ = 2 (mθ = 0), which is an opinion pushed by none of the sending agents. How can we explain this effect? One interpretation is that, in the presence of conflicting suggestions coming from the two sub-networks, the receiving agent opts for a October 31, 2019 DRAFT 123456789101112050100i0.00.51.0µ9,i(θ)Agent9050100i0.00.51.0µ10,i(θ)Agent10050100i0.00.51.0µ11,i(θ)Agent11050100i0.00.51.0µ12,i(θ)Agent12θ=1θ=2θ=317 conservative choice. If sending sub-network N1 says "choose −∆", while sending sub-network N2 says "choose +∆", then the receiving agent prefers to be agnostic and stays in the middle, i.e., it chooses 0. Referring to real-life situations, we can think of one person betting on a soccer match between teams A and B. Assuming that discordant solicitations come from the environment, i.e., the person receives data suggesting to bet on the victory of team A, as well as data suggesting to bet on the victory of team B. If there is no sufficient evidence to let one suggestion prevail, then the most probable choice would be betting on a draw! This "truth-is-somewhere- in-between" effect is a remarkable effect that is peculiar to the weakly-connected setting, and that has been not observed before, e.g., it was not present in [10]. In summary, it is the cumulative influence of a sending group over a receiving agent that determines whether it will follow the group's opinion or not. This situation emulates the social phenomenon of herd behavior: agents choose to ignore their private signal in order to follow the most influencing group of agents. When none of the above dominance situations occurs, the receiving agent can opt for an opinion that is not promoted by any of the sending agents. V. TOPOLOGY LEARNING In the previous section we examined the effect of the network topology on the social learning of the agents. In particular, we discovered how the topology and the states of the sending agents determine the opinion formation by the receiving agents. The way the information is delivered across the network ultimately determines the minimizers in (20), i.e., the value that each receiving agent's belief will converge to. We now examine the reverse problem. Assume we observe the belief evolution of part of the network. We would like to use this information to infer the underlying influences and topology. This is a useful question to consider because understanding the topology can help us understand why a particular agent adopts a certain opinion. The main question we consider now is this: given some measurements collected at the receiving agents, can we estimate their connections to the sending sub-networks? We shall answer this question under the following assumption of homogeneity of likelihoods and true distributions inside the individual sending sub-networks. Assumption 4 (Homogeneity within sending sub-networks): For s = 1, 2, . . . , S, we assume that the distribution and the likelihood functions within the s-th sending sub-network are equal across all agents in that sub-network, namely, for all (cid:96) ∈ Ns: f(cid:96) = f (s), L(cid:96)(θ) = L(s)(θ) (37) (cid:3) DRAFT October 31, 2019 One main consequence of Assumption 4 is that (19) becomes: ω(cid:96)kD[f(cid:96)L(cid:96)(θ)] Dk(θ) = (cid:88)(cid:96)∈S = S(cid:88)s=1(cid:32)D[f (s)L(s)(θ)](cid:88)(cid:96)∈Ns 18 (38) ω(cid:96)k(cid:33) , where Ns denotes the collection of agents in the s-th sending sub-network. Equation (38) has the following relevant implication. Under Assumption 4, the network topology influences the average divergence Dk(θ) only through an aggregate weight: xsk (cid:44) (cid:88)(cid:96)∈Ns ω(cid:96)k = (cid:88)(cid:96)∈Ns w(cid:96)k (39) The latter equality, using w(cid:96)k instead of ω(cid:96)k, comes straightforwardly from (13) and (15). This equality reveals that the aggregate weights depend solely on the matrix W , and not on the matrix E of Perron eigenvectors. In other words, the inner structure of the pertinent sending sub-network s does not influence the aggregate weight xsk. We notice that, while a combination weight a(cid:96)k accounts for a local, small-scale pairwise interaction between agent (cid:96) and agent k, the aggregate weight xsk accounts for macroscopic topology effects, for two reasons. First of all, xsk is determined by the limiting weights ω(cid:96)k, which embody not only direct connection effects between (cid:96) and k, but also effects mediated by multi-hop paths connecting (cid:96) and k. Second, from (39) we see that xsk embodies the global effect coming from all agents belonging to the s-th sending component. In other words, xsk is a measure of the effect from all agents in sending sub-network s on agent k. Since, in view of Theorem 1, the average divergence determines the behavior of the limiting belief, we conclude from (38) that the network topology ultimately influences the particular hypothesis chosen by a receiving agent only through these global weights {xsk}. We assume that the data available for estimating xsk are the shared (intermediate) beliefs, ψk,i(θ). We will say that consistent topology learning is achievable if the xsk can be correctly guessed when sufficient time is given for learning, i.e., we will focus on the limiting data, for all θ (cid:54)= θ(cid:63) k:3 yk(θ) (cid:44) lim i→∞ log ψk,i(θ) i a.s. = Dk(θ(cid:63) k) − Dk(θ). (40) Accordingly, the topology inference problem we are interested in can be formally stated as follows. For any receiving agent k, introduce its global-weight vector: (cid:62) xk (cid:44) [x1k, x2k, . . . , xSk] , and consider the vector stacking the H limiting beliefs yk(θ) (i.e., the data): (cid:62) yk (cid:44) [yk(1), yk(2), . . . , yk(H)] . 3We remark that, in view of (68) in Appendix A, the asymptotic properties of ψk,i(·) are the same as µk,i(·). October 31, 2019 (41) (42) DRAFT 19 Fig. 5. Macroscopic topology inference problem. The object of topology inference is constituted by the global weights xsk from sending sub-network s to receiving agent k. For example, the weight x1k in the figure embodies the influence of all sending agents in N1, from all paths (possibly including intermediate receiving agents) leading to receiving agent k ∈ N3. The main question is whether we can estimate xk consistently from observation of yk. In the sequel we will sometimes refer to this problem as a macroscopic topology inference problem -- see Fig. 5 for an illustration. As compared to other topology inference problems, we are faced here with one critical element of novelty. We have no data coming from the sending agents. This means that correlation between sending and receiving agent pairs cannot be performed. This is in sharp contrast with traditional topology inference problems, where the estimation of connections between pairs of agents is heavily based on comparison (e.g., correlation) between data streams coming from these pairs of agents [25] -- [27]. In contrast, we focus here on the asymmetrical case that, when estimating the weights xsk from sending to receiving agents, no data are available from the sending agents. For this reason, the topology learning problem addressed in this work is significantly different from other traditional topology problems studied in the literature. VI. IS MACROSCOPIC TOPOLOGY LEARNING FEASIBLE? We now examine the feasibility of the topology learning problem illustrated in the previous section. Let us preliminarily introduce a matrix D = [dθs], which collects the H × S divergences between any true distribution in the sending sub-networks and any likelihood, and whose (θ, s)-th entry is: [D]θs = dθs = D[f (s)L(s)(θ)]. (43) Using (41) and (43) in (38), the network divergence of receiving agent k, evaluated at θ, can be written as: S(cid:88)s=1 Through (42) we can rewrite the limiting data in (40) as: k) − D(θ) = yk(θ) = D(θ(cid:63) Dk(θ) = October 31, 2019 dθsxsk. ks − dθs) xsk. (dθ(cid:63) S(cid:88)s=1 (44) (45) DRAFT N2<latexit sha1_base64="apOwlPSaog+sD+hSTDPx7xNUKk4=">AAACFHicbVC7TgJBFL2LL8QXamJjM5GYWJFdLLQk2FgZSOSRwEpmhwEmzM5uZmZNyGZ/w9hq7x/Yqa29jf/gHzgLFAKeZJKTc869c3O8kDOlbfvLyqysrq1vZDdzW9s7u3v5/YOGCiJJaJ0EPJAtDyvKmaB1zTSnrVBS7HucNr3RVeo376lULBC3ehxS18cDwfqMYG2ku46P9ZBgHt8k3VKumy/YRXsCtEycGSmUj2rf7KXyXu3mfzq9gEQ+FZpwrFTbsUPtxlhqRjhNcp1I0RCTER7QtqEC+1S58eTqBJ0apYf6gTRPaDRR/07E2Fdq7HsmmV6pFr1U/M9rR7p/6cZMhJGmgkw/6kcc6QClFaAek5RoPjYEE8nMrYgMscREm6LmNs3acWMaKZMMdZJ25Cw2skwapaJzXizVTFkVmCILx3ACZ+DABZThGqpQBwISHuEJnq0H69V6sz6m0Yw1mzmEOVifv/8SovA=</latexit>N3<latexit sha1_base64="Tg3cNi4Q2IaW/AKN0pMZvmbfrW8=">AAACFHicbVC7TgJBFL3rE/GFmtjYTCQmVmQXCi0JNlYGEnkksJLZYYAJs7ObmVkTstnfMLba+wd2amtv4z/4B84ChYAnmeTknHPv3Bwv5Exp2/6yVlbX1jc2M1vZ7Z3dvf3cwWFDBZEktE4CHsiWhxXlTNC6ZprTVigp9j1Om97oKvWb91QqFohbPQ6p6+OBYH1GsDbSXcfHekgwj2+SbinbzeXtgj0BWibOjOTLx7Vv9lJ5r3ZzP51eQCKfCk04Vqrt2KF2Yyw1I5wm2U6kaIjJCA9o21CBfarceHJ1gs6M0kP9QJonNJqofydi7Cs19j2TTK9Ui14q/ue1I92/dGMmwkhTQaYf9SOOdIDSClCPSUo0HxuCiWTmVkSGWGKiTVFzm2btuDGNlEmGOkk7chYbWSaNYsEpFYo1U1YFpsjACZzCOThwAWW4hirUgYCER3iCZ+vBerXerI9pdMWazRzBHKzPXwDFovE=</latexit>N4<latexit sha1_base64="YJhH6kS2i8SMT3edlXdDQ2OOO04=">AAACFHicbVC7SgNBFL3rM8ZXVLCxGQyCVdiNgpYhNlaSgHlAsobZySQZMju7zMwKYdnfEFvt/QM7tbW38R/8A2eTFCbxwMDhnHPvXI4Xcqa0bX9ZS8srq2vrmY3s5tb2zm5ub7+ugkgSWiMBD2TTw4pyJmhNM81pM5QU+x6nDW94lfqNeyoVC8StHoXU9XFfsB4jWBvpru1jPSCYxzdJ5zzbyeXtgj0GWiTOlORLh9Vv9lJ+r3RyP+1uQCKfCk04Vqrl2KF2Yyw1I5wm2XakaIjJEPdpy1CBfarceHx1gk6M0kW9QJonNBqrfydi7Cs18j2TTK9U814q/ue1It27dGMmwkhTQSYf9SKOdIDSClCXSUo0HxmCiWTmVkQGWGKiTVEzm6btuDGNlEmGOkk7cuYbWST1YsE5KxSrpqwyTJCBIziGU3DgAkpwDRWoAQEJj/AEz9aD9Wq9WR+T6JI1nTmAGVifvwJpovI=</latexit>N1<latexit sha1_base64="nRFT06pUrNWj/Ds2PdtE2OyDKJo=">AAACFHicbVC7TgJBFL2LL8QXamJjM5GYWJFdLLQk2FgZSOSRwEpmhwEmzM5uZmZNyGZ/w9hq7x/Yqa29jf/gHzgLFAKeZJKTc869c3O8kDOlbfvLyqysrq1vZDdzW9s7u3v5/YOGCiJJaJ0EPJAtDyvKmaB1zTSnrVBS7HucNr3RVeo376lULBC3ehxS18cDwfqMYG2ku46P9ZBgHt8kXSfXzRfsoj0BWibOjBTKR7Vv9lJ5r3bzP51eQCKfCk04Vqrt2KF2Yyw1I5wmuU6kaIjJCA9o21CBfarceHJ1gk6N0kP9QJonNJqofydi7Cs19j2TTK9Ui14q/ue1I92/dGMmwkhTQaYf9SOOdIDSClCPSUo0HxuCiWTmVkSGWGKiTVFzm2btuDGNlEmGOkk7chYbWSaNUtE5L5ZqpqwKTJGFYziBM3DgAspwDVWoAwEJj/AEz9aD9Wq9WR/TaMaazRzCHKzPX/1uou8=</latexit>k<latexit sha1_base64="YUeHDbwsAC00t7bOZIwoEHZPjpo=">AAACCHicbVDLSgMxFM34rOOr6tLNYCm4KjN1oRux6MZlC/YB7VAy6Z02NJMJSUYoQ39At/oF/oA7cevGTxB/wy8w03ZRWw8EDuee+8gJBKNKu+63tbK6tr6xmduyt3d29/bzB4cNFSeSQJ3ELJatACtglENdU82gJSTgKGDQDIY3Wb15D1LRmN/pkQA/wn1OQ0qwNlJt2M0X3JI7gbNMvBkpXH3al+Lly6528z+dXkySCLgmDCvV9lyh/RRLTQmDsd1JFAhMhrgPbUM5jkD56eTQsVM0Ss8JY2ke185Ene9IcaTUKAqMM8J6oBZrmfhfrZ3o8MJPKReJBk6mi8KEOTp2sl87PSqBaDYyBBNJza0OGWCJiTbZ2MX5UdlwgpmfQqKMVeixbULyFiNZJo1yyTsrlWtuoXKNpsihY3SCTpGHzlEF3aIqqiOCAD2iJ/RsPViv1pv1PrWuWLOeI/QH1scvYFadjQ==</latexit>`<latexit sha1_base64="2srABkeiBvKIU4unFMHsAqLjNzw=">AAACC3icbVDLSgMxFM3UVx1fVZdugqXgqszUhW7EohuXFewD2qFk0jttaCYzJBmhDP0E3ereT3AnbsVPEH/DLzDTdlFbDwQO5577yPFjzpR2nG8rt7K6tr6R37S3tnd29wr7Bw0VJZJCnUY8ki2fKOBMQF0zzaEVSyChz6HpD6+zevMepGKRuNOjGLyQ9AULGCU6kzrAebdQdMrOBHiZuDNSvPy0L+KXL7vWLfx0ehFNQhCacqJU23Vi7aVEakY5jO1OoiAmdEj60DZUkBCUl05uHeOSUXo4iKR5QuOJOt+RklCpUegbZ0j0QC3WMvG/WjvRwbmXMhEnGgSdLgoSjnWEs4/jHpNANR8ZQqhk5lZMB0QSqk08dml+VDacEu6lkChjjfXYNiG5i5Esk0al7J6WK7dOsXqFpsijI3SMTpCLzlAV3aAaqiOKBugRPaFn68F6td6s96k1Z816DtEfWB+/wUie2Q==</latexit>x1k<latexit sha1_base64="T02HHD5lTmTurdeSVZGgf6p9QMM=">AAACDXicbVDLSsNAFJ34rPFVFdy4CZaCq5LUhS5L3bhswT6gDWUynbRjJ5MwcyOWkG/Qrf6DS3fiSvAb3PgRfoGTtovaemDgcO65jzlexJkC2/4yVlbX1jc2c1vm9s7u3n7+4LCpwlgS2iAhD2Xbw4pyJmgDGHDajiTFgcdpyxtdZfXWHZWKheIGxhF1AzwQzGcEg5aa973EGaW9fMEu2RNYy8SZkULluP7NXqoftV7+p9sPSRxQAYRjpTqOHYGbYAmMcJqa3VjRCJMRHtCOpgIHVLnJ5NrUKmqlb/mh1E+ANVHnOxIcKDUOPO0MMAzVYi0T/6t1YvAv3YSJKAYqyHSRH3MLQiv7utVnkhLgY00wkUzfapEhlpiADsgszo/KhhPM3YTGSlsjSE0dkrMYyTJplkvOealc12lV0RQ5dIJO0Rly0AWqoGtUQw1E0C16RE/o2XgwXo03431qXTFmPUfoD4zPX/LeoB8=</latexit>x2k<latexit sha1_base64="+d0tAHdVx3TSR1+CCnGPr9t7E2Q=">AAACDXicbVDLSsNAFJ34rPFVFdy4CZaCq5LUhS5L3bhswT6gDWUynbRjJ5MwcyOWkG/Qrf6DS3fiSvAb3PgRfoGTtovaemDgcO65jzlexJkC2/4yVlbX1jc2c1vm9s7u3n7+4LCpwlgS2iAhD2Xbw4pyJmgDGHDajiTFgcdpyxtdZfXWHZWKheIGxhF1AzwQzGcEg5aa972kPEp7+YJdsiewlokzI4XKcf2bvVQ/ar38T7cfkjigAgjHSnUcOwI3wRIY4TQ1u7GiESYjPKAdTQUOqHKTybWpVdRK3/JDqZ8Aa6LOdyQ4UGoceNoZYBiqxVom/lfrxOBfugkTUQxUkOkiP+YWhFb2davPJCXAx5pgIpm+1SJDLDEBHZBZnB+VDSeYuwmNlbZGkJo6JGcxkmXSLJec81K5rtOqoily6ASdojPkoAtUQdeohhqIoFv0iJ7Qs/FgvBpvxvvUumLMeo7QHxifv/SEoCA=</latexit>x2`<latexit sha1_base64="4Pw3tJ1U9R15QJ/UDbk1s2xYy+I=">AAACEHicbVDLSsNAFJ34rPFVFdy4CZaCq5LUhS5L3bhswT6kDWUyvWmHTiZhZiKWkJ/QrX6Ce3ciuPIP3PgRfoGTtovaemDgcO65jzlexKhUtv1lrKyurW9s5rbM7Z3dvf38wWFThrEg0CAhC0XbwxIY5dBQVDFoRwJw4DFoeaOrrN66AyFpyG/UOAI3wANOfUqw0tLtfS8pd4GxtJcv2CV7AmuZODNSqBzXv+lL9aPWy/90+yGJA+CKMCxlx7Ej5SZYKEoYpGY3lhBhMsID6GjKcQDSTSYHp1ZRK33LD4V+XFkTdb4jwYGU48DTzgCroVysZeJ/tU6s/Es3oTyKFXAyXeTHzFKhlf3e6lMBRLGxJpgIqm+1yBALTJTOyCzOj8qGE8zcBGKprZFKTR2SsxjJMmmWS855qVzXaVXRFDl0gk7RGXLQBaqga1RDDURQgB7RE3o2HoxX4814n1pXjFnPEfoD4/MXW12hbA==</latexit>x1`<latexit sha1_base64="ef0vtaNweRoBweVkFQa1RbVV7Wg=">AAACEHicbVC7SgNBFJ31GddXVLCxGQwBq7AbCy1DbCwTMA9JljA7uZsMmX0wMyuGZX9CW/0EezsRrPwDGz/CL3A2SRETDwwczj33MceNOJPKsr6MldW19Y3N3Ja5vbO7t58/OGzKMBYUGjTkoWi7RAJnATQUUxzakQDiuxxa7ugqq7fuQEgWBjdqHIHjk0HAPEaJ0tLtfS+xu8B52ssXrJI1AV4m9owUKsf1b/ZS/aj18j/dfkhjHwJFOZGyY1uRchIiFKMcUrMbS4gIHZEBdDQNiA/SSSYHp7iolT72QqFfoPBEne9IiC/l2He10ydqKBdrmfhfrRMr79JJWBDFCgI6XeTFHKsQZ7/HfSaAKj7WhFDB9K2YDokgVOmMzOL8qGw4JdxJIJbaGqnU1CHZi5Esk2a5ZJ+XynWdVhVNkUMn6BSdIRtdoAq6RjXUQBT56BE9oWfjwXg13oz3qXXFmPUcoT8wPn8BWbShaw==</latexit>It is useful to introduce the matrix: (46) where em is an H × 1 vector with all zeros and a one in the m-th position. It is important to note that Bk has its θ(cid:63) k-th row equal to zero. We can now formulate the topology problem in terms of the following constrained system: Bk (cid:44)(cid:16)1H e (cid:62) θ(cid:63) k − IH(cid:17) D, Find (cid:101)xk ∈ RS such that  yk = Bk(cid:101)xk, (cid:80)S s=1(cid:101)xsk = 1, (cid:101)xk > 0, where we remark that the notation (cid:101)xk > 0 signifies that all entries in the solution vector (cid:101)xk must be strictly positive. This positivity constraint is enforced because by assumption, each receiving sub-network is connected to at least one agent from each sending sub-network, which implies that the true vector we are looking for, xk, has all positive entries. The equality constraint in (47) can be readily included in matrix form by introducing 20 (47) (48) (49) the augmented matrix and vector: which allow rewriting (47) as: Bk 1(cid:62) S  , Ck (cid:44) (cid:101)yk (cid:44) Find (cid:101)xk ∈ RS : (cid:101)yk = Ck(cid:101)xk, yk 1 , (cid:101)xk > 0 We are now ready to state formally the concept of feasibility for the topology learning problem. First, we want to solve the problem under the assumption that the matrix of divergences, D, is known, i.e., that sufficient knowledge is available about the underlying statistical models (likelihoods and true distributions). In this respect, we remark that the matrix Bk in (46) depends on θ(cid:63) k, which in turn depends on the unknowns xsk as well through (20). However, from Theorem 1 we know that the beliefs (and also the intermediate beliefs) converge k from the limiting data yk(θ), which is tantamount to assuming k. Therefore, we can safely estimate θ(cid:63) to 1 at θ(cid:63) that the matrix Bk is known. Therefore, achievability of a consistent solution for the topology learning problem translates into the condition that the linear system in (49) should admit a unique solution. We will now prove the following result. Lemma 1 (Necessary Condition for Macroscopic Topology Learning): The topology learning problem de- scribed by the system in (49) admits a unique solution if, and only if: rank(Ck) = S (50) Thus, a necessary condition for topology learning is that the number of hypotheses is at least equal to the number of sending sub-networks, namely, that: October 31, 2019 H ≥ S (51) DRAFT Proof: We remark that we are not concerned with the existence of a solution for the constrained linear +, which by system (49). In fact, this system admits at least a solution, namely, the true weight vector, xk ∈ RS 21 (cid:4) assumption fulfills the equation (cid:101)yk = Ck xk. whose set of solutions is given by [29]: Let us now focus on the unconstrained system (i.e., the system in (49) without the inequality constraints), (cid:101)xk = C † where z ∈ RS is an arbitrary vector, and C k is the Moore-Penrose pseudoinverse of Ck. If rank(Ck) = S, it k Ck)−1C(cid:62) is well known [29] that C k , which implies that the second term on the RHS in (52) is zero, which in turn implies that the unconstrained system has the unique solution: † k = (C(cid:62) (52) † kCk)z, † k(cid:101)yk + (IS − C (cid:101)xk = (C (cid:62) k Ck) −1C (cid:62) k(cid:101)yk = xk (53) The latter equality holds because, if the unconstrained system has a unique solution, this is also the unique solution for the constrained system, i.e., it coincides with xk and satisfies the positivity constraints. Accordingly, we have proved that whenever rank(Ck) = S, the constrained system has the unique solution corresponding to the true vector xk. We now show that when rank(Ck) < S the constrained system has infinite solutions. Since any solution of the unconstrained system takes on the form (52), and since xk is a particular solution, there will exist a certain vector z0 such that the xk can be written as: xk = C † k(cid:101)yk + (IS − C † kCk)z0. (54) (cid:101)xk = C k(cid:101)yk + (IS − C Consider a solution (cid:101)xk in (52) that corresponds to another vector, z = z0 + , where is a perturbation vector: † kCk)(z0 + ) = xk + (IS − C † kCk). (55) † Since by assumption xk > 0, we conclude from (55) that for sufficiently small perturbations it is always possible to obtain a distinct (cid:101)xk > 0, which implies that the constrained system in (49) has infinite solutions. In summary, we conclude that the topology learning problem is feasible if, and only if, rank(Ck) = S. Finally, by observing that the augmented matrix Ck is an (H + 1) × S matrix with an all-zeros row, we have in fact proved the claim of the lemma. Lemma 1 has at least three useful implications. First, it reveals a fundamental interplay between social learning and topology learning: the possibility of estimating xk depends on the comparison between two seemingly unrelated quantities, the number of hypotheses H (an attribute of the social inferential problem) and the number of sending sub-networks S (an attribute of the network topology). Second, the necessary condition in (51) highlights that topology learning over social networks is challenging. For example, if the agents of the social network want to solve a binary detection problem (H = 2), then the October 31, 2019 DRAFT 22 maximum number of sending sub-networks that could allow faithful topology estimation is S = 2. Increasing the complexity of the social learning problem (i.e., increasing H) is beneficial to topology estimation, since it allows to increase also S. Third, we see that having more sending sub-networks makes topology learning more complicated. This is because increasing the number of sending sub-networks increases the number of unknowns (i.e., the dimension of xk), while not adding information since in our setting we are not allowed to probe the sending nodes. Remarkably, when examining jointly the social learning and the topology learning problems, the role of the data and of the unknowns is exchanged. In the social learning problem, more hypotheses means more unknowns and more sending sub-networks means more data; in the topology learning problem, the situation is exactly reversed. A. Structured Gaussian Models In this section we consider the practical case of a Gaussian model, defined as follows. • All agents use the same family of likelihood functions {L(θ)}, for θ = 1, 2, . . . , H. • These likelihoods are unit-variance Gaussian likelihoods with different means {mθ}. • Each true distribution coincides with one of the likelihoods. This implies that the distribution of the s-th sending sub-network, f (s), is a unit-variance Gaussian distribution with mean νs that is chosen among the means {mθ}, namely, for s = 1, 2, . . . , S: νs ∈ {m1, m2, . . . , mH}. (56) • The sending sub-networks have different means. Using (43) and the definition of KL divergence between Gaussian distributions, the matrix D is given by: . (57)  D = 1 2  (m1 − ν1)2 (m2 − ν1)2 ... (mH − ν1)2 (m1 − ν2)2 (m2 − ν2)2 . . . . . . (m1 − νS)2 (m2 − νS)2 ... (mH − ν2)2 . . . (mH − νS)2 From (57) it is readily seen that, if the sending sub-networks share the same true distribution (i.e., if ν1 = ν2 = ··· = νS), then the matrix D has rank 1, and, hence, the topology learning problem is obviously not feasible. As said, we will instead focus on the opposite case where the true expectations are all distinct. For ease of presentation, and without loss of generality we can assume that the sending sub-networks are numbered so that the expectations of the true distributions are: ν1 = m1, ν2 = m2, . . . , νS = mS, (58) DRAFT October 31, 2019 23 . (59) (61) (62) (63) which implies that (57) takes on the form: D = 1 2 0 (m1 − m2)2 (m2 − m1)2 ... 0 . . . . . . (m1 − mS)2 (m2 − mS)2 ... (mH − m1)2 (mH − m2)2 . . . (mH − mS)2   The structure in (59) implies that, for H = S, the matrix D is a Euclidean distance matrix (but for the constant 1/2) [28]. These matrices are constructed as follows. Given points r1, r2, . . . , rL, belonging to Rdim, the (i, j)-th entry of the matrix EDM(r1, r2, . . . , rL) is given by the squared Euclidean distance between points ri and rj. Accordingly, we see from (59) that, for H = S: D = 1 2 EDM(m1, m2, . . . , mH ). (60) For H > S, the matrix D can be described as an extended Euclidean distance matrix, constructed as follows. Let: ES (cid:44) 1 2 EH (cid:44) 1 2 EH−S (cid:44) 1 2 EDM(m1, m2, . . . , mS), EDM(m1, m2, . . . , mH ), EDM(mS+1, mS+2, . . . , mH ), and let F be the (H − S) × S matrix with entries, for θ = S + 1, S + 2, . . . , H and s = 1, 2, . . . , S: [F ]θs = (mθ − ms)2. 1 2 Then, we have the following representation: D = ES F  , EH = F (cid:62) ES F EH−S  . The following theorem, which establishes the feasibility of the topology learning problem for the considered Gaussian model, relies heavily on some fundamental properties of Euclidean distance matrices. Theorem 2 (Macroscopic Topology Learning under Structured Gaussian Models): Let S ≥ 2 and H ≥ S. Assume that all sending sub-networks have the same family of unit-variance Gaussian likelihood functions L(θ) with distinct means {mθ}, for θ = 1, 2, . . . , H. Assume that the true distributions f (s), within the sending sub-networks s = 1, 2, . . . , S, are unit-variance Gaussian with distinct means νs, chosen from the collection {mθ}. Then, under Assumption 3 (so that the matrix Bk in (46) is well defined), for all receiving agents k ∈ R we have that: Proof: The proof is reported in Appendix B. rank(Ck) = 2 October 31, 2019 (64) DRAFT 24 Remark 7 (Topology Learning under Structured Gaussian Models is Challenging): In view of Lemma 1, Eq. (64) has the following implication. Under the considered Gaussian model, topology learning is feasible only when S = 2. We remark also that, when S = 2, condition (51) plays no role, since any meaningful classification problem has at least H = 2. In summary, Theorem 2 reveals that the structure of the Gaussian model makes topology learning very challenging, as this problem is not solvable for networks with more than 2 sending sub-networks. Thus, the theorem reveals that H ≥ S is not a sufficient condition for consistent topology (cid:3) learning. B. Diversity Models We can now examine the effect that diversity in the models of the sending sub-networks can have on topology learning. Since the limiting beliefs are essentially determined by the divergence matrix D, it is meaningful to impose a form of diversity in terms of the divergences between distributions and likelihoods. In other words, differently from the Gaussian case illustrated in the previous section, we now require that the entries of D are not tightly related to each other, namely, we allow them to assume values in RH×S (where we denote by R+ the nonnegative reals) with no strong structure linking them. + One typical model for this type of diversity is that the divergences perceived by the different agents (i.e., across index s), and corresponding to different hypotheses (i.e., across index h), are modeled as absolutely continuous random variables. This randomness is a formal way to embody some degree of variability in how the agents "see" the world. For example, this is a useful model to consider when the agents, due to imperfect knowledge, have likelihoods that are slightly perturbed versions of some nominal model. Examples of this type are illustrated in the next section. In order to avoid confusion, it is important to remark one fundamental property. Under the diversity setting, the matrix D is random4 with entries modeled as absolutely continuous random variables. The full-rank property for this type of matrices is a classical result. However, we observe from (46) that the matrix Bk is obtained from k, which in turn depends statistically upon the D by multiplying a matrix that depends on a random variable θ(cid:63) entries of D. Finally, we know from (48) that Ck is obtained from Bk by adding an all-ones row. Accordingly, to determine the rank of Ck we need to address carefully these intricate dependencies. This is accomplished in the proof of the forthcoming Theorem 3. Theorem 3 (Macroscopic Topology Learning under General Models with Diversity): Let H ≥ S, and assume that the array {dθs}, with θ = 1, 2, . . . , H and s = 1, 2, . . . , S, is made of random variables that are jointly 4Accordingly, we will now use the bold notation for the matrix entries, dθs, as well as for other related quantities. October 31, 2019 DRAFT network graph belief evolution estimated weights 25 Fig. 6. Unperturbed Gaussian model. Left. Network topology. Middle. Belief convergence at the receiving agents. Right. Estimated macroscopic topology. For each of the four panels, the numbers on the right denote the true values {xsk}, with different colors denoting different s, according to the legend. absolutely continuous with respect to the Lebesgue measure on RH×S we have that, with probability 1, Assumption 3 is verified and the matrix Ck is full column rank, namely, . Then, for all receiving agents k ∈ R + P [θ(cid:63) k is unique and rank(Ck) = S] = 1 (65) Proof: The proof is reported in Appendix C. The meaning of Theorem 3 is that configurations of KL divergence that lead to a rank-deficient matrix Ck are rare. In other words, if some diversity exists in the statistical models of the sending components, then the topology inference problem is feasible for almost all configurations. VII. SIMULATION RESULTS We now present some illustrative examples. The first example refers to the Gaussian model presented in Sec. VI-A. The other two examples refer to the setting with diversity presented in Sec. VI-B. a) Gaussian with H = S = 2. We consider the topology shown in the leftmost panel of Fig. 6. The likelihoods and true distributions for the sending sub-networks are unit-variance Gaussian with means ν1 = m1 = 1, ν2 = m2 = 2. The receiving agents5 employ the same likelihoods of the sending agents, and their true distributions are unit-variance Gaussian with mean equal to 1. In Fig. 6 (middle) we show the convergence of the receiving agents' beliefs. Next, we address the topology learning problem. First, for an observation time i, we construct the empirical data (cid:98)yk(θ) = (1/i) log ψk,i(θ), and construct an estimate (cid:98)θ(cid:63) hypothesis where (cid:98)yk(θ) will collapse to 1). We can then construct an estimate for Bk as: (cid:62)(cid:98)θ(cid:63) (cid:98)Bk =(cid:16)1H e k as the value of θ that maximizes (cid:98)yk(θ) (i.e., the − IH(cid:17) D, (66) 5We recall that the models of the receiving agents will be ultimately immaterial as regards their limiting beliefs. k October 31, 2019 DRAFT 1234567891011120100200300i0.00.51.0µ9,i(θ)Agent90100200300i0.00.51.0µ10,i(θ)Agent100100200300i0.00.51.0µ11,i(θ)Agent110100200300i0.00.51.0µ12,i(θ)Agent12θ=1θ=2010002000i0.000.250.500.751.00bxs90.650.35Agent9010002000i0.000.250.500.751.00bxs100.590.41Agent10010002000i0.000.250.500.751.00bxs110.510.49Agent11010002000i0.000.250.500.751.00bxs120.520.48Agent12s=1s=2network graph belief evolution estimated weights 26 Perturbed Gaussian model. Left. Network topology. Middle. Belief convergence at the receiving agents. Right. Estimated Fig. 7. macroscopic topology. For each of the four panels, the numbers on the right denote the true values {xsk}, with different colors denoting different s, according to the legend. from which we obtain (cid:98)Ck by adding an all-ones row, according to (48). At this point, we have verified on the simulated data that, for any receiving agent k ∈ {9, 10, 11, 12}, the matrices (cid:98)Ck are full column rank. Then, we used (53) with empirical matrices replacing the exact ones to estimate the connection-weight vector xk as:6 (67) † (cid:62) k (cid:98)yk 1 . k(cid:98)yk 1 = ((cid:98)C −1(cid:98)C k (cid:98)Ck) (cid:98)xk = (cid:98)C (cid:62) We see from Fig. 6 (right) that this procedure allows us to retrieve the topology coefficients {xsk}, provided that the system evolves for a sufficiently long time. b) Randomly perturbed Gaussian with H = S = 3. The network topology has three sending sub-networks and one receiving sub-network as shown in the leftmost panel of Fig. 7. When S > 2, we know from Theorem 2 that for the structured Gaussian model, diversity in the sending components is not enough to ensure the full column rank of the matrix Ck. In order to increase diversity, we consider a randomly perturbed model for the likelihood functions, where the likelihood of the s-th sending sub-network, evaluated at hypothesis θ, is unit- variance Gaussian with mean θ + θs. The random variables {θs} are equally correlated zero-mean Gaussian with variance equal to 0.02 and Pearson correlation coefficient equal to 0.5. For the receiving sub-network we use the same type of random perturbation of the likelihoods. The true distributions for all sending and receiving agents are unit-variance Gaussian with mean equal to 1. The belief convergence for the receiving agents can be seen in the middle group of panels of Fig. 7. In the rightmost group of panels, we see how the estimates {(cid:98)xsk} of the topology weights converge to the true values {xsk}. In contrast with the structured Gaussian case, topology learning is now feasible for S > 2 and even if the true distributions are equal across all sending components. This change in behavior is due to the diversity in the models of the sending sub-networks, represented by the 6The symbol (cid:98) is used for quantities estimated from the data, to be not confused with the symbol (cid:101) used for the exact quantities appearing in (49). October 31, 2019 DRAFT 123456789101112131415160100200300i0.00.51.0µ13,i(θ)Agent130100200300i0.00.51.0µ14,i(θ)Agent140100200300i0.00.51.0µ15,i(θ)Agent150100200300i0.00.51.0µ16,i(θ)Agent16θ=1θ=2θ=3010002000i0.000.250.500.751.00bxs130.480.200.32Agent13010002000i0.000.250.500.751.00bxs140.370.380.25Agent14010002000i0.000.250.500.751.00bxs150.300.540.16Agent15010002000i0.000.250.500.751.00bxs160.300.420.28Agent16s=1s=2s=3network graph belief evolution estimated weights 27 Fig. 8. Perturbed Beta model. Left. Network topology. Middle. Belief convergence at the receiving agents. Right. Estimated macroscopic topology. For each of the four panels, the numbers on the right denote the true values {xsk}, with different colors denoting different s, according to the legend. different means of the likelihoods. Moreover, we see from the parameters of the random variables {θs} that a relatively small perturbation is already sufficient to enable consistent topology learning. c) Beta with H = S = 3. Finally, we consider a non-Gaussian example. The network topology is the same as in the last example. However, now the likelihood functions follow a Beta distribution with scale parameter equal to 2 and with shape parameters given by θ + 1 + uθs, where {uθs}, for θ ∈ {1, 2, 3} and s ∈ {1, 2, 3}, are independent random variables sampled from a uniform distribution with support [−0.1, 0.1]. The true distributions coincide with the unperturbed likelihoods, i.e., the true distribution of the s-th sending sub-network is a Beta distribution with scale parameter equal to 2 and shape parameter equal to s + 1. For the receiving sub-network we apply the same type of random perturbation of the likelihoods, whereas the true distributions are Beta with scale and shape parameters equal to 2. The belief convergence for the receiving agents can be seen in the middle group of panels of Fig. 8. In the rightmost group of panels, we see the convergence of the topology estimates. VIII. SOCIAL LEARNING VS. TOPOLOGY LEARNING In this work we have considered two learning problems. The first problem is the Social Learning (SL) problem, which is the goal of the agents in the network. These agents aim at forming their opinions after consulting the beliefs of their neighbors through an iterative update-and-combine SL algorithm. The second problem is the Topology Learning (TL) problem, where a receiving agent (or some entity monitoring its behavior) attempts to get knowledge about the connections between that receiving agent and the sending sub-networks. We can refer to the SL problem as the direct learning problem, in the sense that it is the original inferential problem the network is deployed for. Likewise, we can refer to the TL problem as the dual learning problem, since it is an inferential procedure that takes as input data the output of the direct TL problem. The analysis conducted in this work has revealed some interesting interplay between SL and TL problems. Let us make a summary of the main results. We recall that S denotes the number of sending sub-networks, and October 31, 2019 DRAFT 123456789101112131415160100200300i0.00.51.0µ13,i(θ)Agent130100200300i0.00.51.0µ14,i(θ)Agent140100200300i0.00.51.0µ15,i(θ)Agent150100200300i0.00.51.0µ16,i(θ)Agent16θ=1θ=2θ=3010002000i0.000.250.500.751.00bxs130.480.200.32Agent13010002000i0.000.250.500.751.00bxs140.370.380.25Agent14010002000i0.000.250.500.751.00bxs150.300.540.16Agent15010002000i0.000.250.500.751.00bxs160.300.420.28Agent16s=1s=2s=328 H the number of hypotheses. First, we established in Lemma 1 that H ≥ S is a necessary condition to achieve consistent TL. This condition has a remarkable interpretation. In a sense, the number of hypotheses is an index (even if not the only one) of complexity associated to the SL problem since, other conditions being equal, more hypotheses make the SL problem more complicated. Likewise, the number of sending components represents an index of complexity of the TL problem, since, other conditions being equal, estimating more links is more complicated. According to these remarks, the condition H ≥ S implies that the TL problem can be feasible when its complexity is not greater than the complexity of the SL problem. Such an interplay appears to be not obvious at all. As a matter of fact, in the traditional topology inference problems, the connections between agents are inferred from some kind of pairwise measure of their dependency. In our setting, since we cannot measure the output of the sending sub-network, we cannot get direct data quantifying dependency between a receiving and a sending agent. Our TL inference is based instead on the belief functions. The belief function contains some richness of information in that it is evaluated for the H different values of θ. This richness (i.e., H) is critical to enable feasibility of the TL problem. In particular, H ≥ S means that the richness of information in the belief function should be greater than or equal to the number of unknown topology weights to be estimated, S. Having established a necessary condition for consistent TL, we moved on to examine some useful models to see whether and when consistent TL is in fact achievable. First, we have considered a structured Gaussian model where all sending sub-networks use the same family of Gaussian likelihoods, and the sending sub-networks have distinct true distributions, each one coinciding with one of the likelihoods. We have shown in Theorem 2 that the TL problem is feasible only if S = 2, for any H ≥ 2. The limited possibility of achieving consistent TL can be ascribed to the limited diversity existing between the different sub-networks (which all use the same family of likelihoods). This observation motivated the analysis of more general models with a certain degree of diversity, a condition formalized by saying that the KL divergences between true distributions and likelihoods are not structured, i.e., they are nonnegative real numbers with no particular relationship among them. Under this setting we have ascertained that, if H ≥ S, the TL problem becomes feasible for almost all configurations, in a precise mathematical sense as stated in Theorem 3. In summary, two critical features that enable consistent TL are: more hypotheses than sending components and a sufficient degree of diversity. APPENDIX A PROOF OF THEOREM 1 Exploiting (6) we can write, for θ, θ(cid:48) ∈ Θ: ψ(cid:96),i(θ) ψ(cid:96),i(θ(cid:48)) log = log µ(cid:96),i−1(θ) µ(cid:96),i−1(θ(cid:48)) + log L(cid:96)(ξ(cid:96),iθ) L(cid:96)(ξ(cid:96),iθ(cid:48)) . Using (7) we have: October 31, 2019 log µk,i(θ) µk,i(θ(cid:48)) = a(cid:96)k(cid:18)log N(cid:88)(cid:96)=1 µ(cid:96),i−1(θ) µ(cid:96),i−1(θ(cid:48)) + log L(cid:96)(ξ(cid:96),iθ) L(cid:96)(ξ(cid:96),iθ(cid:48))(cid:19) . (68) (69) DRAFT By iterating over i, we can write: 1 i log µk,i(θ) µk,i(θ(cid:48)) = + 1 i 1 i N(cid:88)(cid:96)=1 N(cid:88)(cid:96)=1 [Ai−t+1](cid:96)klog L(cid:96)(ξ(cid:96),tθ) L(cid:96)(ξ(cid:96),tθ(cid:48)) i(cid:88)t=1 [Ai](cid:96)k log µ(cid:96),0(θ) µ(cid:96),0(θ(cid:48)) . 29 (70) Under Assumptions 1 -- 2, thanks to the integrability of the log-ratios between the true distributions and the likelihoods implied by (5), through standard limiting arguments (see, e.g., [21], [22]) it is possible to determine the asymptotic behavior of (70) by: i) replacing the powers of matrix A with their limit A∞ in (12); and ii) applying the strong law of large numbers to conclude that: lim i→∞ 1 i log µk,i(θ) µk,i(θ(cid:48)) a.s. = N(cid:88)(cid:96)=1 = (cid:88)(cid:96)∈S = Dk(θ [A∞](cid:96)kE(cid:20)log ω(cid:96)kD[f(cid:96)L(cid:96)(θ ) − Dk(θ), (cid:48) f(cid:96)(ξ(cid:96),t) L(cid:96)(ξ(cid:96),tθ(cid:48)) (cid:48) )] −(cid:88)(cid:96)∈S − log f(cid:96)(ξ(cid:96),t) L(cid:96)(ξ(cid:96),tθ)(cid:21) ω(cid:96)kD[f(cid:96)L(cid:96)(θ)] (71) where in the second-last equality we used (4), and we performed the replacement [A∞](cid:96)k = ω(cid:96)k, which holds in view of the block representation in (12) since k is a receiving agent and since we adopt the indexing in (16). Now, in light of Assumption 3, we conclude that: lim i→∞ 1 i log µk,i(θ) µk,i(θ(cid:63) k) a.s. = Dk(θ(cid:63) k) − Dk(θ) < 0 (72) for all θ (cid:54)= θ(cid:63) is converging to zero. Since the belief function must sum to 1, the result in (21) holds. k. Since the denominator of µk,i(θ) µk,i(θ(cid:63) k) is bounded by 1, Eq. (72) implies that the numerator µk,i(θ) APPENDIX B PROOF OF THEOREM 2 Preliminarily, it is useful to introduce some auxiliary matrices. We let, for all θ = 1, 2, . . . , H: I(θ) (cid:44) 1H e θ − IH , (cid:62) and S In view of Eqs. (46) and (48), the definitions in (73) and (74) imply: B(θ) (cid:44) I(θ)D, C(θ) = B(θ) 1(cid:62)  . Bk = B(θ(cid:63) k), Ck = C(θ(cid:63) k). October 31, 2019 (73) (74) (75) DRAFT We continue by showing some useful properties of the matrix D under the considered Gaussian model. Let us focus on the representation in (63). It is a known result that the rank of a Euclidean distance matrix with n points in Rdim is at most dim + 2 [28]. Since in our case dim = 1, we, can write: 30 Moreover, for the cases S = 2 and S = 3 we have that: E2 = E3 = and, hence: rank(ES) ≤ 3. 0 (m1 − m2)2 (m2 − m1)2 0 (m2 − m1)2 (m3 − m1)2 0 (m1 − m2)2 0 (m3 − m2)2 1 2  1 2  , (m1 − m3)2 (m2 − m3)2 0 ,  (m1 − m2)2, det(E2) = − 1 4 (m1 − m2)2(m1 − m3)2(m2 − m3)2. 1 4 det(E3) = (76) (77) (78) (79) (80) (81) Therefore, when the points that determine the Euclidean distance matrix are all distinct, both the above matrices are full rank. Thus, when S = 2, we have that rank(ES) = 2. When S > 2, since E3 is full rank, and in view of (76), we have instead rank(ES) = 3. From the representation of D in (63), we then conclude that: Next we state and proof a useful lemma. Lemma 2: Let I(θ) be defined as in (73). Then, for all θ = 1, 2, . . . , H we have that: rank(D) = 2, 3, if S = 2, if S > 2. IH − I† (θ)I(θ) = 11(cid:62) 1 H Proof of Lemma 2: For ease of notation, in the following proof the explicit dependence on θ is suppressed, and we write I in place of I(θ). By definition of the Moore-Penrose inverse, matrix I† satisfies: Then we note that: II†I = I, (I†I) (cid:62) = I†I. I(IH − I†I) = I − II†I = I − I = 0, (82) (83) where in the second equality we used the first identity in (82). Equation (83) implies that the columns of (IH − I†I) belong to the null space of I, denoted by N (I) = {v : Iv = 0}. On the other hand, in view of (73) we can write: October 31, 2019 θ v − v = 1H vθ − v = 0, (cid:62) Iv = 1H e (84) DRAFT with vθ the θ-th element of v. As a result, Eq. (84) will be satisfied only if vh = vθ for all h = 1, . . . , H. Therefore, we obtain: 31 N (I) = {α1H : α ∈ R}, (85) further implying, in light of (83), that, for each h = 1, 2, . . . , H, the h-th column of IH − I†I is of form αh1H for some {αh}. On the other hand, since IH − I†I is symmetric in view of the second identity in (82), we conclude that αh = α for all h, namely, (86) (87) (88) (89) IH − I†I = α1H 1(cid:62) H for some α ∈ R. Finally, since in particular 1H ∈ N (I), we can write: (IH − I†I)1H = 1H − I†I1H = 1H , which, in view of (86), yields: α 1H 1(cid:62) H 1H = αH 1H = 1H ⇒ α = 1 H , and we have in fact proved (81). We are now ready to prove Theorem 2. Proof of Theorem 2: We will now show that rank(C(θ)) = 2 for all θ = 1, 2, . . . , H, which clearly implies the claim of the theorem in view of the second equation in (75). For the case H = S = 2, it is immediately seen that the matrix C(θ) (assuming, e.g., θ = 1) takes on the form: C(1) = 1 2 which reveals that rank(C(θ)) = 2. 0 1  0 d12 0 1 ,  (90) Let us move on to examine the other cases where H ≥ S (excluding H = S = 2). We will examine first the properties of the matrix B(θ) in (74). As done before, the dependence on θ is suppressed for ease of notation, and, in particular, we write B, C, and I in place of B(θ), C(θ), and I(θ), respectively. Applying Sylvester's inequality to the first equation in (74) we can write [29]: rank(B) ≥ rank(D) + rank(I) − H = rank(D) − 1, (91) where in the latter equality we used the fact that rank(I) = H − 1. Therefore, from (80) and (91) we conclude that: rank(B) ≥ 1, rank(B) ≥ 2, if S = 2 if S > 2. (92) (93) DRAFT October 31, 2019 Now we would like to see if equality is satisfied for the cases S = 2 (with H > 2) and S > 2 (with H ≥ S). To this end, we start by noticing that equality in Sylvester's inequality holds if, and only if, there exist 32 matrices X and Y that solve [29]: DX + Y I = IH , which in turn admits a solution if, and only if, [30]: (IH − DD † )(IH − I†I) = 0. Applying Lemma 2, from (95) we get: which means that the equality sign in (92) or (93) holds if, and only if: (IH − DD † ) 1H 1(cid:62) H = 0 1 H †1H = 1H DD (94) (95) (96) (97) In particular, we will now show that (97) does not hold for S = 2, while it holds for S > 2. Let us start with the case S = 2 (and H > 2). We will appeal to the representation of D in (63), which for the case S = 2 can be written as: D = 1 2 0 (m1 − m2)2 . (98)  (m2 − m1)2 (m3 − m1)2 ... (mH − m1)2 0 (m3 − m2)2 ... (mH − m2)2  October 31, 2019 Let us now consider the linear system Dv = 1H. From the first two rows of D, we get the unique solution: v = 2(m1−m2)−212. Considering now the third row, we get the identity (m3−m1)2 +(m3−m2)2 = (m1−m2)2, which is true only if the third point, m3, is equal to one of the previous points. We conclude that there exist no v such that Dv = 1H, which further implies that DD†1H (cid:54)= 1H Therefore, for S = 2 Eq. (92) gives rank(B) > 1, which since B is of dimension H × 2, with H > 2, implies that rank(B) = 2. Let us move on to examine the case S > 2 and H ≥ 2. It is known that, for an L × L Euclidean distance matrix M, one has M M†1L = 1L, implying that 1L belongs to the range space of M [31]. We can apply this result to the matrices ES and EH in (63), since they are proportional to Euclidean distance matrices. In particular, we can say that there exist vectors uS and uH such that ESuS = 1S, EH uH = 1H. In particular, one of the (infinite) solutions is given by u(cid:63) H = uS 0 . (99) DRAFT H = † 1H(cid:124)(cid:123)(cid:122)(cid:125)DuS F (cid:62) ES F EH−S  D(cid:124) (cid:123)(cid:122) (cid:125)D † DD = DD uS = DuS = 1H . 33 (100) (101) (103) (104) (105) (106) Applying now (99) into (63), we can write: 1H = EH u(cid:63) Equation (97) now follows by observing that: uS 0 = ES F  uS = DuS. We have in fact shown that (97) holds true for S > 2, which implies that (93) becomes an equality for S > 2. In summary, we have shown so far that rank(B) = 2 for all H ≥ S (but for the case H = S = 2, which has been examined separately). We will now use this result to prove the claim of the theorem, namely, that rank(C) = 2. Since C is obtained from B by adding an all-ones row, determining the rank of C from that of B amounts to check whether the row vector 1(cid:62) S lies in the row space of B, which is tantamount to ascertaining whether there exists z such that: (102) Since we exclude the case H = S = 2, we have always H ≥ 3. Now, let us consider an EDM E3 defined on 3 distinct points p1, p2, p3. Since in this case E3 is full rank, the system v(cid:62) 3 has the following (unique) solution: z (cid:62)ID = 1(cid:62) S . v (cid:62) 3 =(cid:104) e13+e12−e23 e13e12 e12+e23−e13 e12e23 e13+e23−e12 e13e23 3 E3 = 1(cid:62) (cid:105) , where we denoted by eij = 1/2(pi − pj)2 the (i, j)-th entry of E3. Let us now introduce the vector: (cid:62) H = [v (cid:62) 3 v (cid:62) H−3]. 0 Since, for H ≥ 3, we know that rank(EH ) = 3, we conclude that: (cid:62) H EH = 1(cid:62) H , (cid:62) 3 E3 = 1(cid:62) 3 ⇒ v v which, using the block representation of D in (63), yields: (cid:62) H D = 1(cid:62) S . H, that is, if v(cid:62) In view of (106), one solution z to (102) exists if z(cid:62)I = v(cid:62) v H lies in the row space of I. October 31, 2019 DRAFT On the other hand, from the definition in (73), we see that the matrix I can be represented as: 34 I = , (107)  0 0 ... 0 0 0 ... 0 0 −1 . . . 0 −1 . . . ... 0 ... 0 1 1 0 . . . 0 0 . . . 0 0 0 ... 0 . . . 0 . . . −1 1 0 0 . . . 0 . . . 1 −1 . . . 0 . . . 0 0 ... 0 . . . 1 0··· − 1  where the bold notation highlights the θ-th row and column. According to (107), the row space of I is: which is equivalent to: Row(I) = Row(I) =(cid:110)[α1 α2 . . . αH ] : α αh [α1 α2 . . . αH ] : αθ = −(cid:88)h(cid:54)=θ (cid:62)1H = 0(cid:111) . , (108) (109) Examining (103), from straightforward algebra it can be shown that v(cid:62) that v(cid:62) H ≥ S (excluding the case H = S = 2) that rank(C) = 2. 1H = 0. Using (109), we conclude that v(cid:62) H 13 = 0, which, in light of (104), implies H lies in fact in the row space of I, which finally implies, for 3 APPENDIX C PROOF OF THEOREM 3 We remark that in our setting the divergences are modeled as random variables, which implies that the value k is random as well. We should take this into account when proving the claim of the theorem. First, we of θ(cid:63) observe that: P[θ(cid:63) k is unique and rank(Ck) = S] = = P[θ(cid:63) k is unique and rank(C(θ(cid:63) k)) = S] We now show that, for all θ = 1, 2, . . . , H: = P[θ(cid:63) k = θ, rank(C(θ)) = S]. H(cid:88)θ=1 P[rank(C(θ)) = S] = 1 October 31, 2019 (110) (111) DRAFT It is useful to visualize the matrix C(θ) as follows: 35 . (112) dθ1 − d11 dθ2 − d12 dθ1 − d21 dθ2 − d22 ... ... dθ1 − d(θ−1)1 dθ2 − d(θ−1)2 . . . . . . dθS − d1S dθS − d2S ... . . . dθS − d(θ−1)S 0 0 . . . 0 dθ1 − d(θ+1)1 dθ2 − d(θ+1)2 . . . dθS − d(θ+1)S ... dθ1 − dH1 ... dθ2 − dH2 1 1 ... dθS − dHS 1 . . . . . .   C(θ) = BS−1(θ) bS(θ) 1(cid:62) S−1 1  . (114) The matrix C(θ) has H − 1 random rows (i.e., excluding the all-zeros and all-ones rows). Thus, when H > S there are at least S rows with random entries. These random entries are jointly absolutely continuous since i) so are the entries of D; and ii) the mapping from D to (the random entries of) C(θ) is non-singular.7 This implies that, for H > S: P[rank(C(θ)) = S] = 1, (113) which proves (111) for the case H > S. We switch to the case H = S. Let us denote by BS−1(θ) the sub-matrix of B(θ) obtained by deleting its last column, and with bS(θ) the last column of B(θ). We can write: We notice that BS−1(θ) depends only on the sub-matrix DS−1 that is obtained by deleting from D the last column. It is thus meaningful to introduce the set of matrices: (115) Recalling that BS−1(θ) contains an all-zeros row, we see that, given a matrix DS−1 ∈ E, there exists a unique sequence of weights: E (cid:44) {DS−1 : rank(BS−1(θ)) = S − 1} . w1, w2, . . . , wθ−1, wθ+1, . . . , wS, (116) to obtain the row vector 1(cid:62) S−1 as a weighted linear combination of the rows of BS−1(θ). Accordingly, given a matrix DS−1 ∈ E, the rank of C(θ) will be equal to S if the last row in C(θ) cannot be obtained as a linear 7For example, property ii) can be grasped by noting that, conditioned on dθ1, . . . , dθS, the random entries in (112) are jointly absolutely continuous. October 31, 2019 DRAFT combination of the rows of B(θ). In view of (114), this corresponds to check whether the linear combination of the elements in bS with the same weights is equal to 1, namely, if: 36 Consider now a matrix DS−1 ∈ E We have that: wh(dθS − dhS) = 1. (cid:88)h(cid:54)=θ wh(dθS − dhS) = 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) DS−1 = 0, P(cid:88)h(cid:54)=θ (117) (118) (119) (120) (121) since (also conditioned on DS−1) the random variables {dhS}, with h = 1, 2, . . . , H, are jointly absolutely continuous. We then conclude that: P[rank(C(θ)) = SDS−1] = 1, which implies (111) since, in view of the joint absolute continuity of the entries in D, we have that: P[rank(BS−1(θ)) = S − 1] = 1 ⇒ P[DS−1 ∈ E] = 1. If we now apply (111) in (110), we conclude that: P[θ(cid:63) k is unique and rank(Ck) = S] = The proof of the theorem will be now complete if we show that the probability of having a unique θ(cid:63) k is equal = P[θ(cid:63) k = θ] = P[θ(cid:63) k is unique ]. H(cid:88)θ=1 to 1. To this aim, by using (20) and (44), we see that: θ(cid:63) k = argmin θ∈Θ S(cid:88)s=1 xskdθs. (122) Let us consider the summations in (122) corresponding to different values of θ. Since the random variables {dθs} are jointly absolutely continuous (and since xk is not an all-zeros vector), the probability that two or (cid:4) more summations are equal is zero, which finally implies that θ(cid:63) k is unique. REFERENCES [1] V. Matta, A. Santos, and A. H. Sayed, "Exponential collapse of social beliefs over weakly-connected heterogeneous networks," in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Brighton, UK, May 2019, pp. 5267 -- 5271. [2] C. Chamley, Rational Herds: Economic Models of Social Learning. Cambridge, UK: Cambridge Univ. Press, 2004. [3] D. Acemoglu and A. Ozdaglar, "Opinion dynamics and learning in social networks," Dyn. Games Appl., vol. 1, no. 1, pp. 3 -- 49, 2011. [4] V. Krishnamurthy and H. V. Poor, "Social learning and Bayesian games in multiagent signal processing: How do local and global decision makers interact?" IEEE Signal Process. Mag., vol. 30, no. 3, pp. 43 -- 57, May 2013. [5] A. Jadbabaie, P. Molavi, and A. Tahbaz-Salehi, "Information heterogeneity and the speed of learning in social networks," Columbia Business School Research Paper, pp. 13 -- 28, May 2013. October 31, 2019 DRAFT 37 [6] C. Chamley, A. Scaglione, and L. Li, "Models for the diffusion of beliefs in social networks: An overview," IEEE Signal Process. Mag., vol. 30, no. 3, pp. 16 -- 29, May 2013. [7] E. Yildiz, A. Ozdaglar, D. Acemoglu, A. Saberi, and A. Scaglione, "Binary opinion dynamics with stubborn agents," ACM Trans. Econ. Comput., vol. 1, no. 4, pp. 19:1 -- 19:30, Dec. 2013. [8] A. Nedi´c and A. Olshevsky, "Distributed optimization over time-varying directed graphs," IEEE Trans. Autom. Control, vol. 60, no. 3, pp. 601 -- 615, Mar. 2015. [9] B. Ying and A. H. Sayed, "Information exchange and learning dynamics over weakly connected adaptive networks," IEEE Trans. Inf. Theory, vol. 62, no. 3, pp. 1396 -- 1414, Mar. 2016. [10] H. Salami, B. Ying, and A. H. Sayed, "Social learning over weakly connected graphs," IEEE Trans. Signal Inf. Process. Netw., vol. 3, no. 2, pp. 222 -- 238, Jun. 2017. [11] X. Zhao and A. H. Sayed, "Learning over social networks via diffusion adaptation," in Proc. Asilomar Conference on Signals, Systems and Computers, Nov. 2012, pp. 709 -- 713. [12] L. Smith and P. Sorensen, "Pathological outcomes of observational learning," Econometrica, vol. 68, no. 2, pp. 371 -- 398, 2000. [13] D. Acemoglu, M. Dahleh, A. Ozdaglar, and A. Tahbaz-Salehi, "Observational learning in an uncertain world," in Proc. IEEE Conf. Decision Control (CDC), Dec. 2010, pp. 6645 -- 6650. [14] V. Krishnamurthy, O. N. Gharehshiran, and M. Hamdi, "Interactive sensing and decision making in social networks," Found. Trends Signal Process., vol. 7, no. 1 -- 2, pp. 1 -- 196, Apr. 2014. [15] D. Acemoglu, M. Dahleh, I. Lobel, and A. Ozdaglar, "Bayesian learning in social networks," Rev. Econ. Studies, vol. 78, no. 4, pp. 1201 -- 1236, Oct. 2011. [16] M. H. DeGroot, "Reaching a consensus," J. Amer. Statist. Assoc., vol. 69, no. 345, pp. 118 -- 121, 1974. [17] L. G. Epstein, J. Noor, and A. Sandroni, "Non-Bayesian learning," BE J. Theor. Econ., vol. 10, no. 1, pp. 1 -- 20, 2010. [18] D. Acemoglu, A. Ozdaglar, and A. ParandehGheibi, "Spread of (mis)information in social networks," Games and Economic Behavior, vol. 70, no. 2, pp. 194 -- 227, Nov. 2010. [19] A. Jadbabaie, P. Molavi, A. Sandroni, and A. Tahbaz-Salehi, "Non-Bayesian social learning," Games and Economic Behavior, vol. 76, no. 1, pp. 210 -- 225, Sep. 2012. [20] P. Molavi, A. Jadbabaie, K. R. Rad, and A. Tahbaz-Salehi, "Reaching consensus with increasing information," IEEE J. Sel. Topics Signal Process., vol. 7, no. 2, pp. 358 -- 369, Apr. 2013. [21] A. Nedi´c, A. Olshevsky, and C. A. Uribe, "Fast convergence rates for distributed non-Bayesian learning," IEEE Trans. Autom. Control, vol. 62, no. 11, pp. 5538 -- 5553, Nov. 2017. [22] A. Lalitha, T. Javidi, and A. D. Sarwate, "Social learning and distributed hypothesis testing," IEEE Trans. Inf. Theory, vol. 64, pp. 6161 -- 6179, Sep. 2018. [23] T. Cover and J. Thomas, Elements of Information Theory. [24] A. H. Sayed, "Adaptation, Learning, and Optimization over Networks," Found. Trends Mach. Learn., vol. 7, no. 4-5, pp. 311 -- 801, John Wiley & Sons, NY, 1991. 2014. [25] V. Matta and A. H. Sayed, "Consistent tomography under partial observations over adaptive networks," IEEE Trans. Inf. Theory, vol. 65, no. 1, pp. 622 -- 646, Jan. 2019. [26] A. Santos, V. Matta, and A. H. Sayed, "Local tomography of large networks under the low-observability regime," IEEE Trans. Inf. Theory, Oct. 2019, doi: 10.1109/TIT.2019.2945033. [27] G. Mateos, S. Segarra, A. Marques, and A. Ribeiro, "Connecting the dots: Identifying network structure via graph signal processing," IEEE Signal Process. Mag., vol. 36, no. 3, pp. 16 -- 43, May 2019. [28] I. Dokmanic, R. Parhizkar, J. Ranieri, and M. Vetterli, "Euclidean distance matrices: Essential theory, algorithms, and applications," IEEE Signal Process. Mag., vol. 32, no. 6, pp. 12 -- 30, Nov. 2015. [29] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University Press, 2012. October 31, 2019 DRAFT [30] J. Baksalary and R. Kala, "The matrix equation AX − Y B = C," Linear Algebra and its Applications, vol. 25, pp. 41 -- 43, Jun. 1979. [31] J. C. Gower, "Properties of Euclidean and non-Euclidean distance matrices," Linear Algebra and its Applications, vol. 67, pp. 81 -- 97, Jun. 1985. 38 October 31, 2019 DRAFT
1708.02361	1	1708	2017-08-08T03:07:41	Verification & Validation of Agent Based Simulations using the VOMAS (Virtual Overlay Multi-agent System) approach	[ "cs.MA", "cs.AI", "cs.SE", "nlin.AO", "nlin.CG" ]	Agent Based Models are very popular in a number of different areas. For example, they have been used in a range of domains ranging from modeling of tumor growth, immune systems, molecules to models of social networks, crowds and computer and mobile self-organizing networks. One reason for their success is their intuitiveness and similarity to human cognition. However, with this power of abstraction, in spite of being easily applicable to such a wide number of domains, it is hard to validate agent-based models. In addition, building valid and credible simulations is not just a challenging task but also a crucial exercise to ensure that what we are modeling is, at some level of abstraction, a model of our conceptual system; the system that we have in mind. In this paper, we address this important area of validation of agent based models by presenting a novel technique which has broad applicability and can be applied to all kinds of agent-based models. We present a framework, where a virtual overlay multi-agent system can be used to validate simulation models. In addition, since agent-based models have been typically growing, in parallel, in multiple domains, to cater for all of these, we present a new single validation technique applicable to all agent based models. Our technique, which allows for the validation of agent based simulations uses VOMAS: a Virtual Overlay Multi-agent System. This overlay multi-agent system can comprise various types of agents, which form an overlay on top of the agent based simulation model that needs to be validated. Other than being able to watch and log, each of these agents contains clearly defined constraints, which, if violated, can be logged in real time. To demonstrate its effectiveness, we show its broad applicability in a wide variety of simulation models ranging from social sciences to computer networks in spatial and non-spatial conceptual models.	cs.MA	cs	1 Verification &Validation of Agent Based Simulations using the VOMAS (Virtual Overlay Multi-agent System) approach Muaz A. Niazi, Amir Hussain and Mario Kolberg Abstract-Agent Based Models are very popular in a number of different areas. For example, they have been used in a range of domains ranging from modeling of tumor growth, immune systems, molecules to models of social networks, crowds and computer and mobile self-organizing networks. One reason for their success is their intuitiveness and similarity to human cognition. However, with this power of abstraction, in spite of being easily applicable to such a wide number of domains, it is hard to validate agent-based models. In addition, building valid and credible simulations is not just a challenging task but also a crucial exercise to ensure that what we are modeling is, at some level of abstraction, a model of our conceptual system; the system that we have in mind. In this paper, we address this important area of validation of agent based models by presenting a novel technique which has broad applicability and can be applied to all kinds of agent-based models. We present a framework, where a virtual overlay multi-agent system can be used to validate simulation models. In addition, since agent-based models have been typically growing, in parallel, in multiple domains, to cater for all of these, we present a new single validation technique applicable to all agent based models. Our technique, which allows for the validation of agent based simulations uses VOMAS: a Virtual Overlay Multi-agent System. This overlay multi-agent system can comprise various types of agents, which form an overlay on top of the agent based simulation model that needs to be validated. Other than being able to watch and log, each of these agents contains clearly defined constraints, which, if violated, can be logged in real time. To demonstrate its effectiveness, we show its broad applicability in a wide variety of simulation models ranging from social sciences to computer networks in spatial and non-spatial conceptual models. Index Terms-Agent-based Modeling and Simulation, Multiagent System, Verification, Validation, Agent Oriented Software Engineering I. INTRODUCTION V ALIDATION of any simulation model is a crucial task[1, 2]. Simulations, however well-designed, are always only an approximation of the system and if it was so easy to build the actual system, the simulation approach would never have been used [3]. Of all the simulation models, agent-based modeling and simulation paradigm has recently gained a lot of popularity by being applied to a very wide range of domains such as [4-9]. Validation of models typically requires experts to look at data or animation as errors and un-wanted artifacts can appear in the development of agent-based models [10]. However, because of the complex nature of agent-based models comprising of multiple interacting entities and the strong dynamics and frequent emergence patterns in the system, it can be hard to validate agent-based models in the same way as traditional simulation models. In the case of agent-based simulations, it is even easier to fall into the trap of tweaking the variables, especially since occasionally, the inputs can tend to be quite numerous [11]. Because of the complex nature of agent based models and resulting emergence as shown in [12-14], coupled with an enormous variation possibility of the variables, the results of the simulation study can vary considerably by changing the range or even the step size of just one or two variables. Thus, it is vitally important to be able to validate the agent-based simulation. The problem however, comes from the grounds up since validation is not to be an after-thought; it needs to be initiated alongside at the start of the simulation study. Now, validation of agent based models can be quite a challenging task [15, 16]. One problem lies in the fact that validation typically requires SME (Subject Matter Experts) to analyze [3] the simulation data or animation for comparison with another system or model. However, because of appearance of complex phenomenon such as emergence of behavior, where one plus one is not necessarily two as it depends more on the two "ones" and the behavior of the addition operation as is the norm in complex systems as compared to complicated systems [17]. Thus it can be very difficult to be sure if the behavior that we are observing is truly representative of the actual system[18]. Also, it is important to note here that even models, which cannot be validated might have merit and use such as bookkeeping devices or as an aid in selling ideas or as a training aid or even as part of an automatic management system. In the social sciences literature and ACE (Agents in Computation Economics), empirical validation of agent-based models has been described in [19]. Alternate approaches to empirical validation are discussed in [20]. Replication of agent-based models has been considered very important by some authors and has been discussed in [21]. An approach of validation based on philosophical truth theories in simulations has been discussed in [22]. Another approach called "companion modeling" is an iterative participatory approach where multidisciplinary researchers and stakeholders work together continuously throughout a four-stage cycle: field study and data analysis; role-playing games; agent-based 2 and and design intensive model computational experiments social simulation has also been used for validation and calibration [24]. implementation; Agent-based [23]. In the past, although agent-based simulation has been shown to be useful in the validation of multi-agent systems[25, 26], multi-agent systems have not been used to validate agent- based models. On the other hand, simulations have been used in conjunction with software engineering for a long time[27]. Our work can be considered as pertaining to the last two stages of i.e. Agent-Based Model Design/Implementation as well as Intensive Computational Experiments. Specifically, in this paper, we present the following innovations: "Companion Modeling" • We show how to develop a VOMAS (Virtual Overlay Multi-Agent System), which can be used for the validation of agent based simulation models. • We thus further develop social science based validation techniques that can be applicable to both social science as well as other relevant domains. • We present an object-oriented software engineering based methodology for validation of agent-base models, which provides for both logging as well as animation based validation approaches in addition to test-case/invariant based approaches. The rest of the paper is structured as following: First we give an overview of the terms "Verification", "Validation" and "Credibility" as discussed in the literature. We also discuss how these terms have been considered traditionally in simulation models. Next, we give an overview of performing Validation using VOMAS. We show the design of VO (Virtual Overlay) and Logger agents. Next, we show an example of developing a VOMAS for an existing model from Agent- Based Modeling literature, and demonstrate its usefulness, and ease in validation. Finally we conclude the paper. II. VERIFICATION, VALIDATION AND CREDIBILITY Researchers transform real-world systems to models by applying abstraction. This transformation requires propagating concepts from the real world to useful computational models. These, in turn, are used to develop simulations. Simulation models, in essence end up giving back results which can be useful for the real world. As such, the more effective the abstraction mechanism, the better would be the expected real world benefits. A. Peculiarities of Agent-Based Models In case of agent-based models, the simulation comprises of one or more agents. These agents can work independently or else interact with each other. These computational entities, which are typically, simplifications of real-world counter-parts, need to have some meaningful semantics which can include anywhere from simple behaviors as well as variables for storing different to complex such as states, items, representations such as artificial neural networks, artificial immune systems, cognitive models etc. B. Definitions of the terms: Validation is the process by which we can determine if the model is a representation of the system.[3]. This is always performed while keeping the specific abstraction by the designer in mind. Verification is basically the debugging of the system where we ensure that the model that we build is working correctly. Credibility is achieved when the decision- makers and other key project personnel accept the model as well as its results as "correct". C. Correlation with VOMAS? VOMAS approach has been designed to cater for all kind of agent-based models. As such, it has capability to monitor spatial as well as non-spatial concepts in agent-based models. Fig. 1 VOMAS relation with an Agent Based Model D. Verification & Validation of agent-based models One sure way to establish the validity of agent-based model is to have Subject Matter Experts, who give the specification as well as examine the results and logs of simulation runs. VOMAS approach allows experts to be involved in the design of the agent-based model as well as the custom-built VOMAS from scratch. By involving SMEs from the start of the project, which are essentially equivalent to clients in the software engineering domain, VOMAS approach allows the simulation study to be a stronger candidate for success. III. VALIDATION USING VOMAS A. Validation in agent based simulations To understand VOMAS, let us examine figure 1. The Virtual Overlay Multi-agent System is created for each simulation model separately by a discussion between the simulation 3 specialist as well as the SMEs (Subject Matter Expert). When the actual simulation is executed, the VOMAS agents perform monitoring as well as logging tasks and can even validate constraints given by the system designer at design time. B. A Taxonomy of Agent-Based Validation techniques using VOMAS Now, let us examine how agent-based models are structured. Since agent based models have one or more agents, what these agents really mean in the real-world is entirely up to the designer of the simulation. These elements can be spatial in nature, where distance between agents in the simulation is important or else non-spatial, where there is no concept of distance in the simulation as shown in Fig. 2. In case of spatial models, it is also entirely possible that the exact distance may not be important, but the links between agents could be important. An example of this is HIV based models, where interaction between agents can be shown as links. A detailed description of each of these follows. Fig. 2 A Taxonomy of Agent Based Validation techniques 1. Visual Validation: Visual validation is a face validation technique based on an animation based validation technique where the SME can examine the animation to see if the behavior appears to be similar to that expected in the actual domain. 3. Spatial Validation: In spatial validation, the placement of agents in the simulation is important. This includes the placement of some of the VOMAS agents, which interact with the actual agent based simulation. 4. Non-Spatial Validation: In non-spatial validation, the actual distance is not important. These could be used to validate for aggregate data and constraints/invariants etc. 5. Networked or Link-Based Validation In spatial validation, it is possible that the actual placement is less important than the links between them. In case of social simulation, the example could be links to show social network friendships. In case of computer science based networks, these could represent e.g. Connectivity of Peer-to-Peer overlay networks. 6. Proximity Based Validation In this case, the actual proximity of agents to each other and especially to VOMAS agents is important. An example of this is pred-prey models where VOMAS agents can verify certain characteristics of agents passing by them at a certain time. 7. Log based validation: In log based validation, the SME can specify what things to be watched and logged so that they can be examined after the fact and see how e.g. the 2. Validation using VOMAS: In case of VOMAS, we can validate both spatially as well as non-spatially. populations evolved over time, or else how wireless sensor networks lost their power over time etc. 4 8. Constraint-based validation or Invariant Based Validation: It is entirely possible that the SME says that there are certain constraints, which should never be violated in a certain simulation experiment. If these were ever to be violated, then the simulation system should notify the user via some console or else log the event as a Fig. 3 Use case model of simulation model design and V&V special case. E.g. Wolves must never all die in a wolf-sheep predation. If all of the wolves die, then the simulation needs to be stopped etc. as further data collection exercise might not be useful. C. Analysis of VOMAS detailed verification (debugging) is checked by the simulation specialist but in case of any ambiguity, the SME can be referred. 2) Validate the Model This validation is done in three ways a. Validation using animations: This validation is face validation by the SME by means of analyzing the animations. b. Validation using Logs In this case, logs are generated based on watches specified by the SME. These logs show after the fact, the entire scenarios like black boxes from airplanes. c. Validation using Invariants These can be cases where the SME wants The analysis of VOMAS has been conducted based on a scenario-modeling approach. In figure 3, we see the use cases, some of which are described below. The rest should be self- explanatory and we are not listing them for shortage of space: 1) Verify the Model The SME verifies the model by means of execution of the simulations by the Simulation Specialist. The immediate either feedback even while running large scale parameter sweeps. So, if the invariants or constraints are ever violated, the user can be notified. Or at least, this is definitely logged in the simulation log. 5 3) Design and Develop Models b) Virtual Console Fig. 4 Class diagram of agents in a VOMAS This use case is to be conducted by the simulation specialist in conjunction with the SME. D. Design of VOMAS 1) Motivation One of the most popular approaches in Validation is the three step approach given in [28] . The approach has the following steps: Virtual Console agent is an agent, which can be used to dynamically display various messages at run-time. c) Invariant Invariant is any condition, which the designer of the VOMAS and the agent-based simulation, feels that must not be violated during the execution of the simulation. If the Invariant is violated, the violation is logged. Build a model that has high face validity. d) Logger Agent The logging capability is provided by the Logger Agent. e) Watch If the designer of the system wants some value to be observed, it can be made a watch. f) Watch Log Entry Each watch can also be logged as a logged entry. g) Invariant Violation Invariant violations can be logged at run-time to the Console Virtual agent or else the log as a log entry. h) Log Entry The base class of all log entries. i) Sim Agent This is an agent which is part of the agent based simulation model. a) b) c) Validate model assumptions Compare the model input-out transformations to corresponding input-output transformations for the real system. VOMAS has been designed to cater for both face validity as well as model assumptions and io-transformations. Model assumptions are ensured by the use of invariants. Face validation is ensured by means of various techniques based on spatial and non-spatial validation and animation-based validation. IO-transformations are ensured by means of essential logging components. Thus, in other words VOMAS provides the complete validation package. 2) Description of Class Diagram In figure 4, we see the class diagram of the VOMAS agents and how they interact with the agents in the simulation. The description of each of these agents is given below: a) VO Manager VO manager agent is the key agent handling the interaction of all of the other agents. 6 j) VO Agent These are agents which can be located spatially or non- spatially to monitor the entire simulation. IV. CASE STUDY Here, we present application of a VOMAS to an agent-based simulation mode of the "Simulation of the research process". Recently an agent-based simulation model of researchers attempting to present research in International publication venues was presented in [5]. We demonstrate how to develop and use the associated VOMAS on this model. A. The Publishing Researchers' model In the publishing researcher model, the abstraction is that researchers are modeled as agents in the simulation. The higher the publications of an agent, the higher the agent goes. Thus space in this simulation model is essentially used to show the capability of the researcher. A screenshot of the simulation model is shown in fig 5. For more details, the interested reader is advised to consult the original article. The model has been developed using NetLogo [29]. So, let us formally define some of the entities involved: SME: An Expert Researcher with experience of publishing in various venues. Objective of Simulation Study: To examine how the policies of researchers in selection of publication venues impacts an overall organization. Example Invariant: Basis: In a particular simulation experiment, enough time of simulation run should be given to ensure that journal preferring researchers publish at least ten times during the simulation. Invariant: If simulation stops before each journal preferring researcher is able to publish at least ten times, note an invariant violation in the console and/or the log. Example watches: Measure the total number of researchers with the best policy. Measure the number of researchers above a certain threshold. Measure the number of overall publications. Fig. 5 Screenshot of the researchers' model [5] showing researchers according to their publication count. (Lime = Conference preferring, Red = Journal Preferring, Cyan = No Preference) V. CONCLUSION AND FUTURE WORK In this paper, we have presented a novel framework for the validation of agent based simulation models. We have given a description of how VOMAS agents can be constructed for validation. As a case study, we have shown its application on an existing published model. In the future, we shall apply VOMAS on various types of simulation models and demonstrate how it can be effective in validation. Some of the models we intend to explore VOMAS application on, include pred-prey models, tumor growth models, Peer-to-Peer unstructured overlay network models. REFERENCES [1] [2] [3] [4] [5] [6] O. Balci, "Verification, validation, and accreditation," in Proceedings of the 30th conference on Winter simulation Washington, D.C., United States: IEEE Computer Society Press, 1998. J. Banks, J. S. C. II, B. L. Nelson, and D. M. Nicol, Discrete-Event System Simulation, Fourth ed.: Peason Education, 2005. A. M. Law, "How to build valid and credible simulation models," in Simulation Conference, 2008. WSC 2008. Winter, 2008, pp. 39- 47. M. Niazi and A. Hussain, " Agent based Tools for Modeling and Simulation of Self-Organization in Peer-to-Peer, Ad-Hoc and other Complex Networks," IEEE Communications Magazine, vol. 47, No. 3, pp. 163 - 173., March 2009. M. Niazi, A. Hussain, A. R. Baig, and S. Bhatti, "Simulation of the research process," in Winter Simulation Conference, Miami, FL, 2008, pp. 1326-1334. M. Niazi, "Self-organized customized content delivery architecture for ambient assisted environments," in Proceedings of the third international workshop on Use of P2P, grid and agents for the development of content networks, Boston, MA, USA, 2008, pp. 45-54. 7 T. H. Naylor and J. M. Finger, "Verification of Computer Simulation Models," Management Science, vol. 2, pp. B92-B101, 1967. Evanston, IL, 1999: Center for U. Wilensky, "NetLogo," Connected Learning Comp.-Based Modeling, Northwestern University, 1999. [28] [29] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] in Proceedings of N. Gilbert and K. G. Troitzsch, Simulation for the social Scientist, Second ed.: McGraw Hill Education, 2005. C. M. Macal and M. J. North, "Agent-based modeling and the 39th simulation: desktop ABMS," conference on Winter simulation: 40 years! The best is yet to come Washington D.C.: IEEE Press, 2007. A. Siddiqa, M. Niazi, F. Mustafa, H. Bokhari, A. Hussain, Noreen Akram, S. Shaheen, F. Ahmed, and S. Iqbal, " A New Hybrid Agent-Based Modeling & Simulation Decision Support System for Breast Cancer Data Analysis " in ICICT Karachi, Pakistan: IEEE Press, 2009. J. M. Galán, L. R. Izquierdo, S. S. Izquierdo, J. I. Santos, R. d. Olmo, A. López-Paredes, and B. Edmonds, "Errors and Artefacts in Agent-Based Modelling," Journal of Artificial Societies and Social Simulation, vol. 12, no. 11, 2009. T. W. Lucas, S. M. Sanchez, F. Martinez, L. R. Sickinger, and J. W. Roginski, "Defense and homeland security applications of multi-agent simulations," in Proceedings of the 39th conference on Winter simulation: 40 years! The best is yet to come Washington D.C.: IEEE Press, 2007. A. Ilachinski, "Exploring self-organized emergence in an agent- based synthetic warfare lab," Kybernetes, vol. 32, pp. 38 - 76, 2003. M. Chli and P. De Wilde, "The Emergence of Knowledge Exchange: An Agent-Based Model of a Software Market," Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on, vol. 38, pp. 1056-1067, 2008. M. Cartier, "An Agent-Based Model of Innovation Emergence in Organizations: Renault and Ford Through the Lens of Evolutionism," Comput. Math. Organ. Theory, vol. 10, pp. 147- 153, 2004. C. Bianchi, P. Cirillo, M. Gallegati, and P. A. Vagliasindi, "Validating and Calibrating Agent-Based Models: A Case Study," Comput. Econ., vol. 30, pp. 245-264, 2007. C. Bianchi, P. Cirillo, M. Gallegati, and P. Vagliasindi, "Validating and Calibrating Agent-Based Models: A Case Study," Computational Economics. J. H. Miller and S. E. Page, Complex Adaptive Systems: An Introduction to Computational Models of Social Life: Princeton University Press, 2007. J. S. Hodges and J. A. Dewar, "Is It You or Your Model Talking? A Framework for Model Validation," RAND Corporation, Santa Monica, California. G. Fagiolo, C. Birchenhall, and P. Windrum, "Empirical Validation in Agent-based Models: Introduction to the Special Issue " Computational Economics, vol. 30, pp. 189-194, October, 2007 2007. S. Moss, "Alternative Approaches to the Empirical Validation of Agent-Based Models," Journal of Artificial Societies and Social Simulation, vol. 11, no. 15, 2008. U. Wilensky and W. Rand, "Making Models Match: Replicating an Agent-Based Model," Journal of Artificial Societies and Social Simulation, vol. 10, p. 2, 10/31 2007. A. Schmid, "What is the Truth of Simulation?," Journal of Artificial Societies and Social Simulation, vol. 8, p. 5, 10/31 2005. O. Barreteau and e. al., "Our Companion Modelling Approach " Journal of Artificial Societies and Social Simulation, vol. 6, 2003. M. Makowsky, "An Agent-Based Model of Mortality Shocks, Intergenerational Effects, and Urban Crime," Journal of Artificial Societies and Social Simulation, vol. 9, No. 2, 2006. M. Cossentino, G. Fortino, A. Garro, S. Mascillaro, and W. Russo, "PASSIM; a simulation-based process for the development of multi-agent systems," Int. J. Agent-Oriented Softw. Eng., vol. 2, pp. 132-170, 2008. G. Fortino, A. Garro, and W. Russo, "From Modeling to Simulation of Multi-Agent Systems: An Integrated Approach and a Case Study," in Multiagent System Technologies, LNAI 3187: Springer-Verlag, 2004, pp. 213-227 R. S. Pressman, Software Engineering, A Practitioner's Approach, Sixth ed.: McGraw Hill, 2005.
1009.6050	1	1009	2010-09-30T07:18:20	Comments on "Consensus and Cooperation in Networked Multi-Agent Systems"	[ "cs.MA", "cs.NI", "math.OC" ]	This note corrects a pretty serious mistake and some inaccuracies in "Consensus and cooperation in networked multi-agent systems" by R. Olfati-Saber, J.A. Fax, and R.M. Murray, published in Vol. 95 of the Proceedings of the IEEE (2007, No. 1, P. 215-233). It also mentions several stronger results applicable to the class of problems under consideration and addresses the issue of priority whose interpretation in the above-mentioned paper is not exact.	cs.MA	cs	Comments on “Consensus and Cooperation in Networked Multi-Agent Systems” 1 = ,...,1 n i - tx )( & i (1) PAVEL CHEBOTAREV,2 Member IEEE Key words: consensus algorithms, cooperative control, flocking, graph Laplacian, networked multi-agent systems The objective of this note is to give several comments regarding the paper [1] published in the Proceedings of the IEEE and to mention some closely related results published in 2000 and 2001. I will focus on the graph theoretic results underlying the analysis of consensus in multiagent systems. As stated in the Introduction of [1], “ Graph Laplacians and their spectral properties […] are important graph-related matrices that play a crucial role in convergence analysis of consensus and alignment algorithms.” In particular, the stability properties of the distributed consensus algorithms ( ) = (cid:229) txta tx ,)( )( )( i ij j iNj for networked multi-agent systems are completely determined by the location of the Laplacian eigenvalues of the network. The convergence analysis of such systems is based on the following lemma [1, p. 221]: Lemma 2: (spectral localization) Let G be a strongly connected digraph on n nodes. Then rank( L) = n − 1 and all nontrivial eigenvalues of L have positive real parts. Furthermore, suppose G has c ‡ 1 strongly connected components, then rank(L) = n − c. Here, L is the Laplacian matrix of G, i.e., L = D – A, where A is the adjacency matrix of G, and D is the diagonal matrix of vertex out-degrees. I would like to make four comments regarding this lemma. 1. The last statement of the lemma is wrong. Indeed, recall that the strongly connected components (SCC’s) of a digraph G are its maximal strongly connected subgraphs. Let, for example, G be a converging tree (in-arborescence), i.e., a directed tree having a node r (a root) such that all nodes can be linked to r via directed paths (for r itself it is a path of length 0). Then G has c = n strongly connected components, so Lemma 2 yields that rank( L) = n − c = 0. But in fact, rank(L) = n − 1. To make the last statement of Lemma 2 valid, one should additionally require that all the SCC’s of G are disjoint. 2. In [1], the proof of the rank property (the first statement of Lemma 2) is attributed to [3]. Let me note L) for digraphs was solved earlier in [2]. More specifically, by that the general problem of finding rank( Proposition 11 of [2], for any digraph G, rank(L) = n – d, where d is the so-called in-forest dimension of G, i.e., the minimum possible number of converging trees in a spanning converging forest of G. Furthermore, it was shown (Proposition 6) that the in-forest dimension of G is equal to the number of its sink SCC’s (the SCC’s having no edges directed outwards) and that the in-forest dimension of a strongly connected digraph is one (Proposition 7)3. A corrected version of the above Lemma 2 immediately follows as a special case. 3. Remark 1 given after Lemma 2 says: “Lemma 2 holds under a weaker condition of existence of a directed spanning tree for G.” Here, by Lemma 2 the authors presumably mean the conclusion that rank(L) = n − 1 and by a directed tree they mean a converging tree. Next, they note that such a weaker condition has appeared in several papers published in 2003 and 2005. Let us observe that the existence of a spanning converging tree for G is tantamount to d = 1, so this statement follows from Proposition 11 of [2]. 1 Manuscript received August 14, 2008; revised January 22, 2010. Date of current version June 18, 2010. Digital Object Identifier: http://dx.doi.org/10.1109/JPROC.2010.2049911. 2 The author is with the Institute of Control Sciences, Russian Academy of Sciences, 65 Profsoyuznaya Street, Moscow 117997, Russia (e-mail: [email protected]; [email protected]). 3 These results have also been presented in [4]. = + i = ,...,1 n , - + )1 kx ( i kx )( i (2) (3) 4. For the study of alignment algorithms for arbitrary digraphs, it is important to observe that the statement of Lemma 2 that “all nontrivial eigenvalues of L have positive real parts” holds true for any digraphs [5, Proposition 9], and not only for strongly connected digraphs or digraphs with spanning converging trees. In Section II.C of [1], a discrete-time counterpart of the consensus algorithm (1) is considered n ( ) (cid:229) kxa kx ,)( )( e i ij j j 1 = 0>e is the step size. In the matrix form, (2) is represented as follows: where kx kxP )1 ( ( ), =+ is referred to in [1] as the Perron matrix with parameter e of G. IP L where e-= IP L The matrices were studied in [2] and [5]; in particular, (i) of Lemma 3 in [1] actually coincides e-= with Proposition 12 of [2]. Finally, let me mention a few additional results [2, 5] that are applicable to the analysis of consensus algorithms (1) and (3) and flocking algorithms. In the general case where the primitivity of a stochastic matrix P, P2, P3,… may diverge, the P is not guaranteed and the sequence long-run transition matrix m (cid:229) = k 1 P m P ¥P always exists and, by the Markov chain tree theorem [6, 7], it lim is considered. ¥ = - m ¥fi k 1 coincides with the normalized matrix J of maximal in-forests of G. J is the eigenprojector of L; by ( J = d, where d is the in-forest dimension of G. The columns of J span the Proposition 11 of [2], rank ) kernel (null space) of L; as a result, they determine the main properties of the trajectories of (1) and the flocking trajectories [8] in the general case. The elements of J were characterized in graph theoretic terms in '2 and 3 of [2]; a finite algebraic method for calculating J was proposed in [5] (see also [4]). Theorems Thus, [2, 5, 4] published before the recent avalanche of papers on distributed consensus algorithms ([2] and [5] were sent to J.A. Fax in 2001 and a reference to [4] was sent to R. Olfati-Saber apropos of Lemma 2 in 2003, both on their requests) contained the basic graph theoretic results needed for the analysis of these algorithms. A number of related theorems were proved in [9] and [10]. Some of these results were surveyed in [11]. REFERENCES [1] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proc. IEEE. vol. 95, pp. 215–233, Jan. 2007. [2] R. P. Agaev and P. Yu. Chebotarev, “The matrix of maximum out forests of a digraph and its applications,” Automation and Remote Control, vol. 61, pp. 1424–1450, Sep. 2000. [3] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Trans. Autom. Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004. [4] P. Chebotarev and R. Agaev, “Forest matrices around the Laplacian matrix,” Linear Algebra and Its Applications, vol. 356, pp. 253–274, 2002. [5] R. P. Agaev and P. Yu. Chebotarev, “Spanning forests of a digraph and their applications,” Automation and Remote Control, vol. 62, pp. 443–466, Mar. 2001. [6] A. D. Wentzell and M. I. Freidlin, “On small random perturbations of dynamical systems,” Russian Mathematical Surveys, vol. 25, no. 1, pp. 1–55, 1970. [7] T. Leighton and R. L. Rivest, “The Markov chain tree theorem,” Computer Science Technical Report MIT/LCS/TM–249, Laboratory of Computer Science, MIT, Cambridge, MA, 1983. [8] J. J. P. Veerman, G. Lafferriere, J. S. Caughman, and A . Williams, “Flocks and formations,” J. Statistical Physics, vol. 121, no. 5–6, pp. 901–936, 2005. [9] R. Agaev and P. Chebotarev, “On the spectra of nonsymmetric Laplacian matrices,” Linear Algebra and Its Applications, vol. 399, pp. 157–168, 2005. [10] R. Agaev and P. Chebotarev, “Which digraphs with ring structure are essentially cyclic?” Advances in Applied Mathematics, vol. 45, pp. 232–251, 2010. [11] P. Yu. Chebotarev and R. P. Agaev, “Coordination in multiagent systems and Laplacian spectra of digraphs,” Automation and Remote Control, vol. 70, pp. 469–483, Mar. 2009.
1802.01194	2	1802	2018-04-26T20:02:04	Anatomy of Leadership in Collective Behaviour	[ "cs.MA", "physics.soc-ph" ]	Understanding the mechanics behind the coordinated movement of mobile animal groups (collective motion) provides key insights into their biology and ecology, while also yielding algorithms for bio-inspired technologies and autonomous systems. It is becoming increasingly clear that many mobile animal groups are composed of heterogeneous individuals with differential levels and types of influence over group behaviors. The ability to infer this differential influence, or leadership, is critical to understanding group functioning in these collective animal systems. Due to the broad interpretation of leadership, many different measures and mathematical tools are used to describe and infer "leadership", e.g., position, causality, influence, information flow. But a key question remains: which, if any, of these concepts actually describes leadership? We argue that instead of asserting a single definition or notion of leadership, the complex interaction rules and dynamics typical of a group implies that leadership itself is not merely a binary classification (leader or follower), but rather, a complex combination of many different components. In this paper we develop an anatomy of leadership, identify several principle components and provide a general mathematical framework for discussing leadership. With the intricacies of this taxonomy in mind we present a set of leadership-oriented toy models that should be used as a proving ground for leadership inference methods going forward. We believe this multifaceted approach to leadership will enable a broader understanding of leadership and its inference from data in mobile animal groups and beyond.	cs.MA	cs	Anatomy of Leadership in Collective Behaviour Joshua Garland,1, a) Andrew M. Berdahl,1, 2 Jie Sun,3, 4, 5, 6 and Erik M. Bollt3, 6, 7 1)Santa Fe Institute, Santa Fe, NM 87501 2)School of Aquatic & Fishery Sciences, University of Washington, Seattle, WA 98195 3)Department of Mathematics, Clarkson University, Potsdam, NY 13699 4)Department of Physics, Clarkson University, Potsdam, NY 13699 5)Department of Computer Science, Clarkson University, Potsdam, NY 13699 6)Clarkson Center for Complex Systems Science, Clarkson University, Potsdam, NY 13699 7)Department of Electrical and Computer Engineering, Clarkson University, Potsdam, NY 13699 (Dated: October 17, 2018) 8 1 0 2 r p A 6 2 ] A M . s c [ 2 v 4 9 1 1 0 . 2 0 8 1 : v i X r a Understanding the mechanics behind the coordinated movement of mobile animal groups (collective motion) provides key insights into their biology and ecology, while also yielding algorithms for bio-inspired technologies and autonomous systems. It is becoming increasingly clear that many mobile animal groups are composed of heterogeneous individuals with differential levels and types of influence over group behaviors. The ability to infer this differential influence, or leadership, is critical to understanding group functioning in these collective animal systems. Due to the broad interpretation of leadership, many different measures and mathematical tools are used to describe and infer "leadership", e.g., position, causality, influence, information flow. But a key question remains: which, if any, of these concepts actually describes leadership? We argue that instead of asserting a single definition or notion of leadership, the complex interaction rules and dynamics typical of a group implies that leadership itself is not merely a binary classification (leader or follower), but rather, a complex combination of many different components. In this paper we develop an anatomy of leadership, identify several principle components and provide a general mathematical framework for discussing leadership. With the intricacies of this taxonomy in mind we present a set of leadership-oriented toy models that should be used as a proving ground for leadership inference methods going forward. We believe this multifaceted approach to leadership will enable a broader understanding of leadership and its inference from data in mobile animal groups and beyond. PACS numbers: 89.70.Cf,89.70.-a,87.10.Vg,02.50.Tt When observing the collective motion of animal groups (e.g., schooling, herding, or flocking), an immediate question is, what is the leadership structure? Who (if anyone) is in charge and who is following, and does such structure stay the same or change over time? Recent techno- logical advances in image processing and animal- mounted sensors make it possible to record the simultaneous movement trajectories of every an- imal in a group. Such abundance of data makes the present a promising time to progress in un- derstanding leadership structure in mobile animal groups. Despite the availability of data and the central importance of understanding leadership in collective motion, there is surprisingly little explicit mathematical description or even a con- sistent and well-defined approach to this subject. Here, as a first step toward addressing this de- ficiency, we construct a framework for inferring leadership in collective motion. We review var- ious sources and characteristics of leadership to provide an anatomy and a language for describ- ing the multifaceted aspects of leadership across a)Electronic mail: [email protected] a variety of animal societies. We then present a suite of leadership-focused toy models, which can be used as a proving ground for any proposed leadership inference method, before being naively applied to (empirical) data. Together, this lays the groundwork for a principled exploration of a perennial question: how is control of a collec- tive system distributed? Such understanding will not only contribute to the ecology and conserva- tion of group-traveling species, but will also aid in the design of control algorithms for emerging distributed technologies. I. OVERVIEW Mobile animals groups (e.g., flocks, herds, schools, swarms) are ubiquitous in nature. In such collective sys- tems, the interactions between individuals may be as im- portant as characteristics of the individuals themselves1. Insight into these interactions and their impact on the group dynamics is of fundamental importance for our un- derstanding of both the ecology of these systems2 as well as design and control principles underlying general com- plex systems3. A key challenge in the study of collective animal behav- ior is understanding how groups of organisms make deci- sions as a whole4, for example about where5 or when6,7 to go. Group decision-making processes range from despotic to shared8, although even in systems with shared or dis- tributed decision making there are likely inter-individual differences (e.g., sex, rank, personality, size, nutritional state, informational state) that produce asymmetry in influence. Models suggest that such heterogeneity is po- tentially important to group-level dynamics9,10, but in- ferring differential influence and leadership from empiri- cal data, though often attempted, is an open challenge. As we elaborate in some detail in this paper, a key step toward tackling this challenge lies in the recognition that the notion of leadership is not merely a simple, unidi- mensional concept. Instead, a rich palette of different types and forms of leadership often coexists, even for the very same system. Thus, we argue that a precursor step to the "correct" inference of leadership is the clarifica- tion of what (type of) leadership is sought of. Without such, any inferred leadership can potentially be deemed inappropriate. The need to distinguish between the definition and in- ference of leadership is standing out as a central problem partly because of the acceleration of technical progress that enabled collection of "big" data. For example, new technologies to collect the simultaneous trajectories of all members of a mobile animal group11, along with in- creases in computing power, make the near future a fruit- ful time to meet this challenge. Will having large amount of real-world data alone be sufficient to address ques- tions about leadership, or do we (still) need conceptual advances? As recently reviewed by Strandburg-Peshkin et al.12, most efforts to infer leadership have used po- sition within a group13 -- 16 (e.g., leaders are assumed to be at the front), initiator-follower dynamics17,18 or time- delayed directional correlations19 -- 23. Information theo- retic measures provide additional, potentially more pow- erful and less subjective, tools to infer leadership and influence24 -- 26. However, a central viewpoint of this pa- per is that any measurement of leadership needs to start by clarifying the particular type or form of leadership one is after. Without such clarification, the "leadership" re- sulted from the application of any inference method can be subject to misinterpretation, and perhaps more se- riously, lead to fundamentally flawed conclusions about the interaction mechanisms of an animal system. To illustrate the many facets of leadership and thus the need to distinguish between its definition and measure- ment, consider, for example, the case of migrating cari- bou. Older, more experienced individuals are thought to guide the migration-scale movements27, however, preg- nant or nursing females might have increased nutritional requirements28 and thus guide movements along that path towards habitat with better forage opportunities29. Therefore, who is leading depends on the time- and length-scale of the movements considered. Additionally, for some populations fall migration coincides with the 2 rut, so mating behaviors drive social interactions: a dom- inant male may attempt to herd females or drive other males away. Such a male is certainly influential, but perhaps should not always be considered a leader, at least in the context of the migration. Finally, whether or not an individual is a leader might depend on who (or which group) one is considering as a potential follower. A nursing (and thus infertile) female might be ignored by the libidinous male, but will be closely followed by her calf30. Because there are many scales and types of in- fluence/leadership, we argue that one should begin such explorations with a clear question and select analytical methods to match. The central goal of this paper is to develop a formal language and multifaceted framework for defining and (potentially) inferring the many aspects of leadership. In addition, we aim to provide a set of leadership-oriented toy models to serve as a proving ground for leadership inference methods. Thus our work here offers a practical language and set of tools for researchers hoping to match questions about leadership with the appropriate meth- ods while avoiding potential pitfalls. We hope that the combination of mathematical rigor, biological intuition, together with several real and synthetic examples will make our framework accessible and interesting to both biologists and applied mathematicians. II. GENERAL MATHEMATICAL FRAMEWORK To capture various forms of leadership, consider dy- namics of individuals (with potential interactions among them via a network) together with dynamics of the group determined by the individuals, modeled by the general form of ODEs: (cid:40) xi = f (Si1(t)x1(t), . . . , Sin(t)xn(t); µi(t); ξi(t)), y(t) = h(x1(t), . . . , xn(t)). (1) In this general model class, xi(t) represents the state of the i-th individual at time t (i = 1, . . . , n), S = [Sij(t)]n×n is the (time-dependent) adjacency matrix (also known as the sociality matrix) of a network en- coding the structure of interactions, where Sij (cid:54)= 0 if it is possible for j to (directly) impact the state of i. Further- more, µi(t) denotes the parameter (vector) associated with i, and ξi(t) is noise. Here a "parameter" can be anything that describes the heterogeneity of the individ- uals in the group. For example, in the Viscek model31, the parameter µi can represent the preferred direction an individual takes, or it can also be used to represent the speed of an individual that might differ from one to an- other, or both by associating a parameter vector to each individual. The function f models how the dynamics of each individual depends on their own state and pa- rameter(s), the state of others in the network, and noise. Finally, the state of the group, y(t), is determined by the state of the individuals through the function h; for (cid:80)n i=1 xi(t) de- example, taking h(x1(t), . . . , xn(t)) = 1 n fines the group state as the average of the individuals states. A separate and complementary perspective is to model/represent the individual and group dynamics as a multivariate stochastic process, focusing on stationary variables Xi(t) and Y (t). From this perspective, the re- lationship between the group variable and the variables are encoded in the conditional distribution function p(y(t)x1(t−), x2(t−), . . . , xn(t−)), (2) where t− = (t − τ, t) denotes time history of the system, taking into account a time lag of τ ∈ (0,∞). We point out that there is intimate connection between a dynamical system [such as one defined by Eq. (1)] and a stochastic process, generally through an underlying (er- godic) measure32, where the uncertainty associated with the state of the variables is generally related to the dis- tribution of initial conditions and noise in addition to the coupled dynamics. For a deterministic system, the randomness initiates exclusively from (experimental) im- perfection of choosing and determining the initial condi- tion, and the evolution of uncertainty can be treated as a stochastic process. Thus entropy methods are naturally associated even with otherwise deterministic dynamical systems Eq. (1) in terms of the associated stochastic pro- cess. From the stochastic representation (2) of the dynam- ics, we can define an individual's (observed) influence on the group using various forms of conditional mutual information. For example, the (unconditioned) mutual information (MI) I(xi(t−); y(t)) (3) measures the apparent influence of i on the group, aggre- gated over both direct and indirect factors. On the other hand, after factoring out indirect factors, the "net" influ- ence of i on the group can be measured by the conditional mutual information (CMI) I(xi(t−); y(t)x¯i(t−)), (4) where ¯i = {1, . . . , i − 1, i + 1, . . . , n}. As suggested re- cently by James et al.33, Eq. 4 may not capture influence entirely, therefore care should be taken when quantifying net influence in this way. Note that Eq. (1) itself does not uniquely determine the distribution in Eq. (2), due to the possibly different states/trajectories the system can follow depending on initial conditions, parameters, and other factors; unique ergodicity and fixed parameters are possible assumptions if we wish to discuss uniqueness. Equation (1) can be interpreted as modeling the possible interactions among the individuals, although these interactions may or not not be realized in a particular setting depending on the states the system operates in; on the other hand, the PDF in Eq. (2) encodes (intrinsic) dependence between the group variable and those of the individual variables 3 without necessarily matching the structural information in Eq. (1), even if such dependence comes from dynamics of Eq. (1). Next, we distinguish between intrinsic states of the sys- tem versus observed states, as a key aspect in mathemat- ical interpretation of any process, including group roles of leadership, is the concept of measurement of observ- ables, from the underlying process. In fact, the concept of leadership and information flow can be dramatically ob- scured depending on the details of the observables (ex- trinsic variables) relative to the underlying system (in- trinsic variables). We use xi(t) to represented the ob- served state regarding xi(t), and similarly, y(t) for the observed state regarding y(t). We represent the observa- tions over a finite time window, producing observational data { xi(t); y(t)}t∈I. (5) Proper characterization and interpretation of leader- ship requires the (subjective) identification of a "refer- ence frame", namely, choosing the (observable) variables, groups, as well as time and spacial scales. That is, we argue that defining such a frame needs to include making at least the following three choices: 1. Variables (e.g., position, velocity, acceleration, di- rection of motion or some combination of these). Depending on the choice of variables, different types of leadership can be defined and (potentially) identified. 2. Temporal resolution and time lag. What is the tem- poral resolution of the actions of interest (e.g., sec- onds, days, or years)? Additionally, there is an is- sue of time lag. How far into the future is an action thought to have potential impact? If the time lag is larger than the time-scale of the typical response to an individual's action, then each individual will appear to have a similar random influence on the others. On the other hand, too small of a time lag might prevent detection of the (time-delayed) dy- namics of the group in response to an individual's actions. 3. Definition of a group and what it represents. For example, a group can contain everyone within a spatial domain, or can be a certain class of individ- uals based on age, gender, etc. III. PRINCIPLE COMPONENTS OF LEADERSHIP In broad terms, we define leadership as an individual having asymmetric potential to impact the trajectory of agents in the group. As we explore below, the source of this asymmetrical impact or influence may be due to group structure, individual information or emerge from social interaction rules alone. Further, the distribution and time and length scales of the resulting leadership may vary considerably. In this section we construct a series of informative classifications which we will refer to as the components of leadership. We further divide these com- ponents into sources and characteristics of leadership. A. Sources of leadership Structural Leadership. Structural leadership encom- passes a wide range of leadership which fundamentally re- lies on the structure of the animal society. This structure could be an explicit dominance hierarchy, or more sub- tly due to unequal social influence due to semi-persistent traits (e.g., age, gender, reproductive status). Depend- ing on the particular taxa, the driving mechanism for such asymmetric interactions differ and deriving such a mechanism is not the purpose of this manuscript. For simplicity, we assume all of this rich societal structure has been pre-encoded in the sociality matrix defined in In particular, Sij (cid:54)= 0 if and only if j has the Eq. 1. capacity to lead i directly. Where "capacity to lead" is defined by the particular society. To formalize this component of leadership, let G be the directed graph associated with the sociality matrix S, where there exists an edge from j to i if Sij (cid:54)= 0. For each node (cid:96) ∈ G, denote the reachability set of node (cid:96) as F(cid:96). In particular, node k is a member of F(cid:96) if there exists a directed path from (cid:96) to k in G. If F(cid:96) (cid:54)= ∅ then (cid:96) is defined to have capacity for structural leadership. We define the set of individuals with non-zero capacity to exhibit structural leadership (have a nonempty reacha- bility set on the sociality matrix) as L. Of course, the degree to which an individual is a structural leader ex- ists on a continuum. Quantifying the strength of such leadership is a highly non-trivial and potentially system- specific task (e.g.,34 -- 36). However, to first order, indi- viduals with many individuals downstream of them and fewer individuals upstream of them in the sociality ma- trix will tend play a stronger leadership role, or at least have the potential to do so. In the same example, In our caribou example from the Introduction, we might expect to find strong hierarchical relationships be- tween males during the rut. With these hierarchies en- coded in the sociality matrix then the dominant males would be labeled as strong structural leaders and the weaker males would be members of various reachabil- ity sets. if a nursing offspring closely followed their mother, then the mother would ex- hibit structural leadership over her calf. Finally, note that while the mother is a structural leader to the calf, she may be influenced by a dominant male; making this mother a structural leader and a follower simultaneously and making the male an indirect structural leader of the calf. Therefore a binary classification of 'leader vs. fol- lower' is generally not appropriate. To further illustrate this point, consider the canonical 4 Figure 1. Hierarchical Leadership in Pigeon Flocks. Here each directed edge represents the capacity for an individual to lead as defined by Nagy et al.19. For example, L has the ability to lead J, and J has the ability to lead no one. example of hierarchical dynamics in pigeon flocks from Nagy et al.19 depicted in Figure 1. In this example, nodes C and J have no structural leadership capacity as they have empty reachability sets. All other nodes however have the capacity to lead at least one other individual and thus all have some degree of structural leadership capacity. Notice, that with the exception of node A, each of the remaining individuals both lead and follow, i.e., they have non-empty reachability sets and are also members of others reachability set. The strength of their structural leadership would roughly mirror their vertical position in Figure 1. Structural leadership is simply the capacity for a mem- ber of an animal society to lead other members of that society as dictated by the societies rules. In this sense structural leadership should really be seen more as a nec- essary but not sufficient condition for leadership to occur within a mobile animal group. However, in reality this component of leadership is quite important because it encodes the potentially important heterogeneity in inter- actions between specific pairs of individuals and more generally any hierarchies in the group. Informed Leadership. Informed leadership arises when a subset of the group are differentially informed and motivated to act on that information, e.g., a subset of the group senses a resource37,38, or has information about a migration route5,39. Such leaders may be anonymous9, or may indicate that they have information, for example by changing speed40 or signaling41. In the case of our migrating caribou, both the experi- enced individuals leading the long-scale migration move- ment, and the individuals responding to local food and predation cues provide complementary examples of in- formed leadership. Informed leadership generally arises from some under- lying intent or motivation e.g., hunger or fear. For this reason, while the concept of informed leadership is in- tuitively sensible, from a mathematical standpoint it is both difficult to define and perhaps impossible to accu- rately infer without additional knowledge of the system. Target-Driven Leadership. Target-driven leadership is a specific subset of informed leadership. A target- driven leader is an informed leader ("informed by tar- get") that uses a series of deliberate control inputs such as calls, explicit motions, etc. to guide a group toward a particular target state or set of target states. How- ever, not all informed leadership is target-driven. For example, when an individual from a group of animals detects a predator, that individual becomes "informed" and tries to move away, and such abrupt change of motion may cause the rest of the group to follow. In this case, the first-reacting individual exhibits informed-leadership, but its sole "target", if any, is to move away from the predator instead of trying to lead the entire group away from the predator. To be more precise, we characterize a target-driven leader as an individual that not only influences the group, but deliberately controls the group toward some target state. In addition, the removal of such an individual should result in the group not going towards the tar- get state. Mathematically, we define this component as follows. Given that A is a set of target states, then in- dividual i is a target-driven leader (with respect to A) if the net influence of i on the group (see Eq. 4) is nonzero and y(t) → A as t → ∞. (6) That is, the individual directly influences the group as a whole and that influence results in the group progressing toward the target states. An example of a target-driven leader is a sheep dog. These dogs runs behind a group of sheep and through an intentional series of signals such as barking, eye contact and body posture the dog deliberately controls the sheep herd toward a given target state such as a barn or field. Emergent Leadership. Asymmetrical influence, and thus leadership, may arise from social interactions rules alone, in the absence of social structure or differential in- formation; we term this emergent leadership. This would be the case if animals used anisotropic social interaction rules. For example when individuals are more influenced by other individuals that are in front of them, then indi- viduals in more frontal positions of the group are more influential, even if they have no additional information, motivation or status. Such emergent leadership has re- cently been shown to be the case in our migratory caribou example30. if Alternately, individuals are more influenced by faster-moving group-mates42, then those faster-moving individuals will have more influence. If those individuals are moving more quickly in response to information, or 5 to signal dominance, then this would be informational or structural leadership, respectively, but if the increased speed is purely a function of the group dynamics, this would be an example of emergent leadership. B. Characteristics of Leadership Distribution of Leadership. In animal groups deci- sions range from full distributed among all group mem- bers ('democratic') to dominated by a single or a few in- dividuals ('despotic')8,12. It can be informative to quan- tify the number of individuals involved in a leadership role within the group. Similar to12 we refer to this as the distribution of leadership which we define on a continuum that lies between centralized and distributed leadership. At the scale of the entire herd, we might expect our migrating caribou to fall somewhere on this spectrum, bookended by primate societies with an alpha individual on one end and leaderless fission-fusion fish schools on the other. If we consider the mother-calf pairs as subgroups, we would expect the mother to be a centralized leader. However, in a larger group containing many such pairs, we would expect distributed leadership shared between the mothers. The pigeon example in Figure 1 illustrates that many systems fall somewhere between these two ex- tremes. In this example nearly all of the individuals have some influence, yet it has a clear hierarchy so it is not fully decentralized; it therefore lies somewhere between centralized and distributed. Temporal Scale of Leadership. A leader may not be actively influencing the motion of other agents at all times and it is thus useful to quantify and understand the time scales for which a leader qualifies as a leader under any of the components of leadership. Here, we con- sider two notions of time scales -- consistency and gran- ularity. For the following discussion consider dynamics of individuals, represented by discrete-time observations {xi(t)}T t=0. Consistency of leadership is simply defined as the pro- portion of the observation window for which a leader qualifies as a leader. More specifically, we classify leaders as persistent over the observation window if it is identified as a leader for the entire time window. Conversely, we classify a leader as ephemeral if it only qualifies as a leader for some small time window [t, t + τ ], with τ (cid:28) T . A similar temporal leadership scale is presented in12 which ranges from variable to consistent but attempts to cap- ture the same notion. The granularity of leadership concerns the resolution of time steps for which an individual acts as a leader. For example, a leader for daily activities might be different from one that is for seasonal activities. We can check for granularity by altering the time step we examine the dynamics under. In particular, quantify leadership using only the observations {xi(kt)}T /k t=0 (k > 1) for a large range of k. If a leader only acts on a coarse basis then 6 a particular source of leadership. In particular, let G be a graph where there exists a directed edge from node j to node i if j has the capacity to lead i, where capacity to lead may be structural, emergent or informed leadership. Then the reach of agent i is the reachability set of i on G. Consider Figure 2, where the graph represents the po- tential for structural leadership. In this example, individ- ual G has a reachability set of {D, B, H, L, I, C, J} and thus those 7 agents are within the reach of structural leader G. Reach naturally lies on a continuum between local and global. If an agent exemplifies some form of leadership over all individuals this would be global reach; if an individual only leads some small subset of the group then this leader is considered local. In Figure 2 Agent A has global reach and agent I has local reach. In the case of our migrating caribou, the experienced migrants leading the entire herd on its broad migration path, would have global reach, while the mother leading her calf on a finer-scale would have local reach. Observability of Leadership. When we observe an animal society we do so imperfectly, mainly in two ways. First, any observed quantity is subject to noise and mea- surement errors. Secondly, and perhaps more impor- tantly, there may be elements of the society which go unobserved. Such hidden variables and states may in turn act in our interpretation of leadership. In fact, the strongest leadership might not be detectable if the data are not appropriate. Across various taxa, leaders may use vocal cues43,44, gestures45 or movements that are too fine to be picked up by GPS (e.g. pre-flight flapping46) to initiate or control movement. If the resulting movement is synchronized, leadership inference based on trajecto- ries will fail. Worse, if in the resulting movement, the least dominate individuals respond first to the cues, it could appear as though those individuals are leading. In the case of our migrating caribou, lead animals may stand up to signal departure, or motivate others to start moving. This would not be captured by GPS tags and so would be hidden to inference methods based on tra- jectories alone. Quantification of hidden leadership in practice is quite difficult by definition. Namely, if you have detected lead- ership it was observed. Doing this in theory however is quite trivial. As defined in Section II we define the full system dynamics via x, y, S (and/or some mix of these). When the system is being observed, the observed variables, denoted x, y, and S, can differ from the true ones. We term an individual's leadership role 'hidden' if it exhibits leadership defined in terms of the intrinsic variables (x, y, S) but does not appear to do so given the observed variables ( x, y, S); a leader that is not hidden is then called an observable leader. Figure 2. Reach of Agent G. Each node with a red circle is within the reachability set and thus the reach of agent G. they may not register as a leader for small k but may then register as a leader for some larger k. In contrast a fine-scaled leader may register for many k. In our migrating caribou example, the experienced in- dividuals leading the broad migration path exhibit lead- ership that is persistent, but perhaps has coarse granu- larity. In contrast, the leadership of those animals re- sponding to resources or predation threats along the way is ephemeral and has fine granularity. Time scales present several challenges when attempt- If the ing to infer leadership roles from a time series. granularity or observation window length do not match the natural time-scales of leadership then leadership events may be completely missed or misclassified. For example, consider a structural leader (cid:96) with the property that I(xi(t−); y(t)x¯i(t−)) = 0, i.e., a structural leader that does not directly influence the group -- although it has the potential to. Regardless of the inference method such a potential leader will always be misclassified. Sim- ilarly consider an informational leader that only leads when they are within some radius to a known resource. Say that this event only occurred for a very short time window [t, t+τ ], with τ (cid:28) T . If you only consider leaders that lead for the entire observation window, most aggre- gate measures will wash out such an ephemeral leadership event. For these reasons, carefully considering both con- sistency by studying sub samples of the data set as well as granularity by down sampling the data and retesting one will be able to obtain a much clearer picture of the leaders that are present in a mobile animal group. Reach of Leadership. The reach of a leader quantifies the members of the group that the leader has potential influence over, di- rectly and indirectly through subsequent interactions. Formally, we define the reach of a leader as the members of that leaders reachability set on a graph associated with C. Real World Animal Behavior and the Anatomy of Leadership IV. A MODEL SANDBOX FOR VALIDATING LEADERSHIP-INFERENCE METHODS 7 Here we discuss real world animal interactions, and we do so in a manner to emphasize the terminology of our anatomy of leadership taxonomy. We expect to find structural leadership in relatively stable animal groups, often having complex social hierar- chies, such as cetaceans, wolves, wild dogs, elephants and primates15,47 -- 50. The canonical example is the so-called 'alpha' individual in a primate society, who has some level of control over an entire group over a long period of time (assuming he society is stable)51. In our taxonomy, this dominant individual would be a persistent, centralized, structural leader with a large reach. It is important to note that in such societies, structural leadership may be well correlated with informational leadership. For exam- ple, a matriarch elephant may have better information about rarely visited water holes, as well as greater per- capita influence to lead her group to them. We expect informational leadership to dominate in an- imal groups composed of unrelated individuals and un- stable membership (i.e., fission-fusion dynamics), such as fish schools and bird flocks. A single arbitrary member of a fish school may perceive a respond to a threat, caus- ing those around it to also startle, or the entire group to make an evasive maneuver52. This is an example of centralized, ephemeral, informational leadership with a limited or global reach, depending how much of the group responded. Similarly, some fraction of the same school might have information about where or when a food re- source might occur and lead the entire school to that time-space location37,38,53. In our terms those fish are distributed, ephemeral, informational leaders with global reach. Informational leadership is also common for movement at long length scales. In flocks of pigeons, better informed individuals act as leaders during homing flights54. (How- ever, it should be noted that pigeons also exhibit a struc- tural hierarchy too19.) During migratory movements, older, more experienced, birds guide groups on efficient migration routes5,39. In both of these examples the in- formed birds are centralized, persistent, target-driven, in- formational leaders with global reach. In migrating white storks some individuals actively seek thermals updrafts, which are necessary for them to get efficient lift to complete the migration, while others tend to copy, by moving towards individuals who are al- ready in thermals55. This is a specific example of a gen- eral phenomenon, emergent sensing5, in which a group spans an environmental gradient and individuals in the 'preferred' end of the gradient alter their behaviour (pur- posefully or not) in a way that cause the entire group to climb the gradient41,56. In general such leadership would be distributed and ephemeral (although could be persis- tent if, like in the storks, the same individuals always find the thermals) informational leadership with global reach. Ultimately one would like to develop methods to in- fer and classify leadership from empirical data. This is of course a long-standing and non-trivial challenge, and a pragmatic approach is to first test inference methods on simulated data where the leadership type and dis- tribution is known because it is programmed in explic- itly. For mobile animal groups an obvious starting point is to modify classic flocking/schooling/herding models (e.g.,31,57,58) to include known leadership structures. In this section we first describe a canonical model of col- lective motion -- the so-called zonal model9,58. Following that, we modify the model to incorporate the various leadership sources and characteristics described in this paper. A. Basic Collective Motion Model Following Couzin et al.9,58, for each agent, numbered i = 1, .., N , and each time t, a position vector ci(t), a direction vector vi(t) and a speed si are maintained. At each time step agent i computes a desired direction di(t) based on neighbors in three different zones, depicted in Figure 3. Figure 3. Schematic of Zonal Flocking Model. The black triangle is the focal individual. The red ring demarks the zone of repulsion R. The blue circle is the orientation zone O, the focal individual attempts to align with the agents in this zone (blue triangles in the figure). The outer ring is the attraction zone A and the focal individual attempts to get closer to these agents (the green triangles in the picture). The resulting desired direction is then the sum of the green and blue vectors. The first zone to consider is called the repulsion zone and is denoted by R. This zone ensures that 'personal space' is maintained for each agent. If any other agent is in the repulsion zone R, for focal individual j, then the desired direction in the next time step is defined by di(t + ∆t) = −(cid:88) j(cid:54)=i j∈R cj(t) − ci(t) cj(t) − ci(t) . (7) This desired direction ensures that a collision will not occur at time t + ∆t. However, if for the focal individual R = ∅ then the focal individual attempts to get closer to agents in their attraction zone A and orient with agents in their orientation zone O. This is accomplished by choosing a desired direction at time t+∆t in the following way: di(t + ∆t) = α (cid:88) cj(t) − ci(t) cj(t) − ci(t) j(cid:54)=i j∈A + (1 − α) (cid:88) j(cid:54)=i j∈O (8) vj(t) vj(t) . Where α is a parameter that controls the relative strength of attraction and alignment. For example, a flock of geese -- dominated by alignment -- would have a relatively low α, while a swarm of insects -- dominated by attraction -- would have a relatively high α. The desired direction vector, d, is normalized to a unit vector d(t + ∆t) = di(t+∆t) di(t+∆t) . Next, to represent uncer- tainty stemming from limitations of sensory and cognitive abilities, this unit vector is transformed into d(cid:48)(cid:48)(t + ∆t) by rotating it by a small angle drawn from a circular- wrapped Gaussian distribution centered at zero. Finally, it is assumed that individuals can turn at a maximum rate of θ radians per unit time. Therefore, if the differ- ence between an individual's current direction, vi(t), and its desired direction for the next time step, d(cid:48)(cid:48) i (t + ∆t), is less than θ∆t then the desired direction is achieved and vi(t + ∆t) = d(cid:48)(cid:48) i (t + ∆t). Otherwise, that individual's direction vi(t + ∆t) is the result of rotating vi(t) by θ∆t radians towards their desired direction d(cid:48)(cid:48) i (t + ∆t). After the heading is assigned, the position at t + ∆t can be computed by ci(t + ∆t) = ci(t) + v(t + ∆t)si∆t, (9) where si is the speed of individual i. B. Explicitly Adding Sources of Leadership (emergent leadership). 8 Structural Leadership To add structural leadership we introduce a sociality matrix S = [Sij]N×N . Sij (cid:54)= 0 if agent i can be influenced by agent j. More generally, Sij is a continuous value that gives the relative influence of individual j on individual i. To take this into account the desired-direction computation is modified to weight the influence of each neighbor relative to Sij (rather than an equal weighting of everyone in A and O, i.e., cj(t) − ci(t) cj(t) − ci(t) di(t + ∆t) = α (cid:88) Sij j(cid:54)=i j∈A (1 − α) (cid:88) j(cid:54)=i j∈O + (10) Sij vj(t) vj(t) . Adding this sociality matrix to the base model allows for structural leadership to be explicitly built in. This is an advantage as you can then see if post-facto if the structural leadership placed in the model can be extracted by a candidate inference method. Informed Leadership To simulate informed leadership, a subset of the agents are given knowledge of a pre- ferred direction g (more generally each informed agent is given their own not necessarily equal preferred direction gi)9. This preferred direction may be part of a migration route or the direction of a prey or known resource. Non- informed group members, have no knowledge of g and may or may not know which individuals are informed. Following Couzin et al.9, to integrate this into the model the informed individuals balance between the social in- teractions and their preferred direction with a weighting term ω. In particular, informed individuals have a de- sired dircetion d(cid:48), given by d(cid:48) i(t + ∆t) = di(t + ∆t) + ωgi di(t + ∆t) + ωgi . (11) If ω = 0 the preferred direction is completely ignored and only social interactions are followed. As ω increases toward 1 the influence of the preferred direction is balanced with influence of the social interactions. With ω > 1 the preferred direction is favored over social interactions. While this base model captures a wide variety of flocking, swarming and schooling behavior it does not account for leadership explicitly. In order to explicitly test leadership inference methods, it is helpful to make a few simple modifications to this base model: (1) add a sociality matrix59 (structural leadership) (2) add "informed" individuals to the group9 (informed leadership) and (3) make interaction rules isotropic58 Emergent Leadership One way to make a test case for inferring emergent leadership, is to make interactions spatially asymmetric. In particular, one can simply add 'blind zones'58 to the model described in Eqs. 7-9. In this case the zones A and O are missing wedges behind them and individuals in those wedges are ignored. If these blind zones are large enough, individuals are more influenced by individuals in front of them30. C. Testing Characterizations of Leadership Distribution of Leadership Using the framework presented in the previous sections one can explore a variety of distributions of leadership, ranging from centralize to distributed12. For structural leadership, the spectrum could range from a sociality matrix with a hub structure (centralized) to a one with a random dense connectivity, or even fully connected (decentralized). For informed leadership, the fraction of the group having a non-zero value of ω would roughly span the spectrum of the distribution of leaderhip. We note that the distributions of structrural and informed leadership are potentially orthoginal. For example a group could have highly centralized structural leadership in tandem completely distributed informational leader- ship, or vise versa. Temporal Scale of Leadership Temporal consistency and granularity of leadership can be built into this leadership model by making the model parameters associated with leadership time dependent, e.g, [Sij(t)], ω(t) and g(t). For example, one could remove or change the preferred direction at regular time intervals by defining time-varying ω(t) and then see if an inference algorithm could detect this change. Reach of Leadership By setting specific examples of the sociality matrix one can experiment with a variety of leadership reach scenarios and test the ability of various inference measures to recover them. Observability of Leadership There is a vast set of variations that could be made to the framework presented here to encode potential for leadership to to be driven by non-trajectory based cues or signals43 -- 46,60. One obvious example (which is also ubiquitous in nature) is auditory signalling, which could provide long-range interactions41. D. A Potential Pitfall: Influence vs. leadership Consider a mobile-animal group where each member is governed by Eq. 9 and the direction is decided by Eqs. 7 & 10. Furthermore, define Si(i+1) = 1 and 0 otherwise for i ∈ {1, . . . , N − 1} and let α ∈ [0, 1]. This describes a simple chain topology, where each individual has the capacity for structural leadership over at most one other agent. In particular, each agent orients and attracts to (follows) at most one other agent in the group. However, it is important to note that every agent avoids collisions with all other agents (the sociality metrix applies to Eq. 10, but not Eq. 7). In this example the incidental social interactions, such as those caused by repulsion, cause a real prob- lem for the majority of influence/causal inference algo- 9 rithms. For example, if one blindly applied optimal causation entropy26 or transfer entropy61 to infer who leads whom then these algorithms would conclude an all-to-all leadership graph. By construction however we know this is incorrect and that the underlying influ- ence graph is a simple chain. The issue here is that these measures26,61, and causal inference from informa- tion in general, are not explicitly measuring leadership but reductions in uncertainty about a particular variable. In this example, the minor local repulsion interactions cause enough "information flow" over time to trigger these algorithms/measures. However, as discussed in Ap- pendix A conflating influence, information flow, causality and leadership is a non-trivial challenge, which is nicely highlighted by the present example. V. AFTERWARD Traditional approaches to leadership inference have fo- cused on a single defining characteristic, e.g., position within a group, social hierarchy, information flow or influ- ence. We believe that, in general, none of these concepts alone fully captures leadership. In this manuscript we have begun to show that a multifaceted approach where multiple axes of leadership are analyzed provides a more complete classification of the leadership structure. This formalism should serve to link questions about empirical systems with the appropriate analytical tools to address those questions. While this taxonomy we provided is surely not complete we hope that this effort will serve as a starting point for formalizing a multifaceted approach to leadership inference. Multiple technological advances in sensors, computer vision have led to the availability of more high-resolution collective motion data than ever before11. As such the near future is an opportune time to make meaningful advances in leadership inference. Causal inference and information theory show a lot of promise in this arena but as we have shown throughout this manuscript lead- ership is a highly intricate and multifaceted subject and neither causal inference nor information theory may be up for the task alone. We hope that as new inference algorithms come to be the formal language and toy mod- els developed here will serve as a proving ground. We believe that being able to carefully classify the compo- nents of leadership being inferred will be invaluable for practitioners and theorists as they begin to tackle all the high-resolution data as it becomes available. VI. ACKNOWLEDGMENTS JG and AMB were supported by Omidyar Fellow- ships from the Santa Fe Institute. AMB was also sup- ported a grant from the John Templeton Foundation. EB and JS were in part by the Army Research Office grant W911NF-16-1-0081, and JS by the Simons Foundation grant 318812, and EB by the Office of Naval Research grant N00014-15-1-2093. The opinions expressed in this publication are those of the authors and do not necessar- ily reflect the views of the John Templeton Foundation. REFERENCES 1T. Vicsek and A. Zafeiris, "Collective motion," Physics Reports 517, 71 -- 140 (2012). 2P. A. Westley, A. M. Berdahl, C. J. Torney, and D. Biro, "Col- lective movement in ecology: from emerging technologies to con- servation and management," (2018). 3P. W. Anderson, "More is different," Science 177, 393 -- 396 (1972). 4L. Conradt and T. J. Roper, "Consensus decision making in an- imals," Trends in ecology & evolution 20, 449 -- 456 (2005). 5A. M. Berdahl, A. B. Kao, A. Flack, P. A. Westley, E. A. Codling, I. D. Couzin, A. I. Dell, and D. Biro, "Collective animal naviga- tion and migratory culture: from theoretical models to empirical evidence," Phil. Trans. R. Soc. B 373, 20170009 (2018). 6B. Helm, T. Piersma, and H. Van der Jeugd, "Sociable schedules: interplay between avian seasonal and social behaviour," Animal Behaviour 72, 245 -- 262 (2006). 7A. Berdahl, P. A. Westley, and T. P. Quinn, "Social interactions shape the timing of spawning migrations in an anadromous fish," Animal Behaviour 126, 221 -- 229 (2017). 8L. Conradt and T. J. Roper, "Group decision-making in ani- mals," Nature 421, 155 (2003). 9I. D. Couzin, J. Krause, N. R. Franks, and S. A. Levin, "Effective leadership and decision-making in animal groups on the move," Nature 433, 513 -- 516 (2005). 10M. d. M. Delgado, M. Miranda, S. J. Alvarez, E. Gurarie, W. F. Fagan, V. Penteriani, A. di Virgilio, and J. M. Morales, "The importance of individual variation in the dynamics of animal col- lective movements," Phil. Trans. R. Soc. B 373, 20170008 (2018). 11L. F. Hughey, A. M. Hein, A. Strandburg-Peshkin, and F. H. Jensen, "Challenges and solutions for studying collective animal behaviour in the wild," Phil. Trans. R. Soc. B 373, 20170005 (2018). 12A. Strandburg-Peshkin, D. Papageorgiou, M. C. Crofoot, and D. R. Farine, "Inferring influence and leadership in moving ani- mal groups," Phil. Trans. R. Soc. B 373, 20170006 (2018). 13J. E. Smith, S. Gavrilets, M. B. Mulder, P. L. Hooper, C. El Mouden, D. Nettle, C. Hauert, K. Hill, S. Perry, A. E. Pusey, M. van Vugt, and E. A. Smith, "Leadership in mam- malian societies: Emergence, distribution, power, and payoff," Trends in Ecology & Evolution 31, 54 -- 66 (2016). 14J. S. Lewis, D. Wartzok, and M. R. Heithaus, "Highly dynamic fission -- fusion species can exhibit leadership when traveling," Be- havioral Ecology and Sociobiology 65, 1061 -- 1069 (2011). 15L. J. Brent, D. W. Franks, E. A. Foster, K. C. Balcomb, M. A. Cant, and D. P. Croft, "Ecological knowledge, leadership, and the evolution of menopause in killer whales," Current Biology 25, 746 -- 750 (2015). 16D. M. Jacoby, Y. P. Papastamatiou, and R. Freeman, "Inferring animal social networks and leadership: applications for passive monitoring arrays," Journal of The Royal Society Interface 13, 20160676 (2016). 17A. Strandburg-Peshkin, D. R. Farine, I. D. Couzin, and M. C. Crofoot, "Shared decision-making drives collective movement in wild baboons," Science 348, 1358 -- 1361 (2015). 18J.-B. Leca, N. Gunst, B. Thierry, and O. Petit, "Distributed leadership in semifree-ranging white-faced capuchin monkeys," Animal Behaviour 66, 1045 -- 1052 (2003). 19M. Nagy, Z. Akos, D. Biro, and T. Vicsek, "Hierarchical group dynamics in pigeon flocks." Nature 464, 890 -- 893 (2010). 20Z. ´Akos, R. Beck, M. Nagy, T. Vicsek, and E. Kubinyi, "Lead- ership and path characteristics during walks are linked to domi- 10 nance order and individual traits in dogs," PLoS computational biology 10, e1003446 (2014). 21L. Jiang, L. Giuggioli, A. Perna, R. Escobedo, V. Lecheval, C. Sire, Z. Han, and G. Theraulaz, "Identifying influential neigh- bors in animal flocking." PLoS Computational Biology 13 (2017). 22I. Watts, M. Nagy, R. I. Holbrook, D. Biro, and T. B. de Per- era, "Validating two-dimensional leadership models on three- dimensionally structured fish schools," Royal Society Open Sci- ence 4, 160804 (2017). 23H. Ye, E. R. Deyle, L. J. Gilarranz, and G. Sugihara, "Distin- guishing time-delayed causal interactions using convergent cross mapping," Scientific reports 5 (2015). 24W. M. Lord, J. Sun, N. T. Ouellette, and E. M. Bollt, "Inference of causal information flow in collective animal behavior," IEEE Transactions on Molecular, Biological and Multi-Scale Commu- nications 2, 107 -- 116 (2016). 25S. Butail, V. Mwaffo, and M. Porfiri, "Model-free information- theoretic approach to infer leadership in pairs of zebrafish," Phys- ical Review E 93, 042411 (2016). 26J. Sun, D. Taylor, and E. M. Bollt, "Causal network inference by optimal causation entropy," SIAM Journal on Applied Dy- namical Systems 14, 73 -- 106 (2015). 27S. Eyegetok, Tuktu, N. Project, and N. Hakongak, Thunder on the tundra: Inuit qaujimajatuqangit of the Bathurst caribou (Generation Printing, Vancouver, 2001). 28P. S. Barboza and K. L. Parker, "Allocating protein to reproduc- tion in arctic reindeer and caribou," Physiological and Biochem- ical Zoology 81, 835 -- 855 (2008). 29L. Conradt, J. Krause, I. D. Couzin, and T. J. Roper, "Lead- ing according to need in self-organizing groups," The American Naturalist 173, 304 -- 312 (2009). 30C. J. Torney, M. Lamont, L. Debell, R. J. Angohiatok, L.-M. Leclerc, and A. M. Berdahl, "Inferring the rules of social interac- tion in migrating caribou," Phil. Trans. R. Soc. B 373, 20170385 (2018). 31T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen, and O. Shochet, "Novel type of phase transition in a system of self-driven parti- cles," Physical review letters 75, 1226 (1995). 32E. M. Bollt and N. Santitissadeekorn, Applied and Computa- tional Measurable Dynamics (SIAM, 2013). 33R. G. James, N. Barnett, and J. P. Crutchfield, "Information flows? a critique of transfer entropies," Physical review letters 116, 238701 (2016). 34J. C. Flack and D. C. Krakauer, "Encoding power in communica- tion networks," The American Naturalist 168, E87 -- E102 (2006). 35E. R. Brush, D. C. Krakauer, and J. C. Flack, "A family of algo- rithms for computing consensus about node state from network data," PLoS computational biology 9, e1003109 (2013). 36C. De Bacco, D. B. Larremore, and C. Moore, "A phys- ical model for efficient ranking in networks," arXiv preprint arXiv:1709.09002 (2017). 37A. Strandburg-Peshkin, C. R. Twomey, N. W. Bode, A. B. Kao, Y. Katz, C. C. Ioannou, S. B. Rosenthal, C. J. Torney, H. S. Wu, S. A. Levin, and I. Couzin, "Visual sensory networks and effective information transfer in animal groups," Current Biology 23, R709 -- R711 (2013). 38S. G. Reebs, "Can a minority of informed leaders determine the foraging movements of a fish shoal?" Animal behaviour 59, 403 -- 409 (2000). 39T. Mueller, R. B. OHara, S. J. Converse, R. P. Urbanek, and W. F. Fagan, "Social learning of migratory performance," Science 341, 999 -- 1002 (2013). 40K. M. Schultz, K. M. Passino, and T. D. Seeley, "The mechanism of flight guidance in honeybee swarms: subtle guides or streaker bees?" Journal of Experimental Biology 211, 3287 -- 3295 (2008). 41C. J. Torney, A. Berdahl, and I. D. Couzin, "Signalling and the evolution of cooperative foraging in dynamic environments," PLoS Computational Biology 7, e1002194 (2011). 42B. Pettit, Z. ´Akos, T. Vicsek, and D. Biro, "Speed deter- mines leadership and leadership determines learning during pi- 11 Chaos: An Interdisciplinary Journal of Nonlinear Science 25, 043106 (2015). 66C. W. Granger, "Testing for causality: a personal viewpoint," Journal of Economic Dynamics and control 2, 329 -- 352 (1980). 67J. Sun, C. Cafaro, and E. M. Bollt, "Identifying the coupling structure in complex systems through the optimal causation en- tropy principle," Entropy 16, 3416 -- 3433 (2014). 68B. Russell, "Mysticism and logic," New Statesman (1913). 69B. Russell, "Human knowledge: Its scope and its limits," New York: Simon &Schuster (1948). 70J. Bigelow and R. Pargetter, "Metaphysics of causation," Erken- ntnis 33, 89 -- 119 (1990). 71J. Pearl and T. S. Verma, "A theory of inferred causation," in Studies in Logic and the Foundations of Mathematics, Vol. 134 (Elsevier, 1995) pp. 789 -- 811. 72J. Pearl, "Bayesian networks," (2011). 73J. Pearl, "The structural theory of causation," in Causality in the Sciences (Clarendon Press, Oxford, 2011) pp. 697 -- 727. 74B. Skyrms and W. L. Harper, Causation, Chance and Credence: Proceedings of the Irvine Conference on Probability and Causa- tion, Vol. 1 (Springer Science & Business Media, 2012). 75D. A. Freedman, "Linear statistical models for causation: A critical review," Encyclopedia of statistics in behavioral science (2005). 76L. Barnett, A. B. Barrett, and A. K. Seth, "Granger causal- ity and transfer entropy are equivalent for gaussian variables," Physical Review Letters 103, 238701 (2009). geon flocking," Current Biology 25, 3132 -- 3137 (2015). 43K. J. Stewart and A. H. Harcourt, "Gorillas' vocalizations during rest periods: signals of impending departure?" Behaviour 130, 29 -- 40 (1994). 44D. Fossey, "Vocalizations of the mountain gorilla (gorilla gorilla beringei)," Animal Behaviour 20, 36 -- 53 (1972). 45J. M. Smith and D. Harper, Animal signals (Oxford University Press, 2003). 46J. M. Black, "Preflight signalling in swans: a mechanism for group cohesion and flock formation," Ethology 79, 143 -- 157 (1988). 47A. J. King, C. M. Douglas, E. Huchard, N. J. Isaac, and G. Cowlishaw, "Dominance and affiliation mediate despotism in a social primate," Current Biology 18, 1833 -- 1838 (2008). 48K. Payne, Sources of social complexity in the three elephant species. (Harvard University Press, 2003). 49R. O. Peterson, A. K. Jacobs, T. D. Drummer, L. D. Mech, and D. W. Smith, "Leadership behavior in relation to dominance and reproductive status in gray wolves, canis lupus," Canadian Journal of Zoology 80, 1405 -- 1412 (2002). 50D. Lusseau and L. Conradt, "The emergence of unshared con- sensus decisions in bottlenose dolphins," Behavioral Ecology and Sociobiology 63, 1067 -- 1077 (2009). 51J. C. Flack, M. Girvan, F. B. De Waal, and D. C. Krakauer, "Policing stabilizes construction of social niches in primates," Nature 439, 426 (2006). 52S. B. Rosenthal, C. R. Twomey, A. T. Hartnett, H. S. Wu, and I. D. Couzin, "Revealing the hidden networks of interaction in mobile animal groups allows prediction of complex behavioral contagion," Proceedings of the National Academy of Sciences 112, 4690 -- 4695 (2015). 53K. N. Laland and K. Williams, "Shoaling generates social learn- ing of foraging information in guppies," Animal Behaviour 53, 1161 -- 1169 (1997). 54A. Flack, B. Pettit, R. Freeman, T. Guilford, and D. Biro, "What are leaders made of? the role of individual experience in de- termining leader -- follower relations in homing pigeons," Animal Behaviour 83, 703 -- 709 (2012). 55M. Nagy, I. D. Couzin, W. Fiedler, M. Wikelski, and A. Flack, "Synchronization, coordination and collective sensing during thermalling flight of freely migrating white storks," Phil. Trans. R. Soc. B 373, 20170011 (2018). 56A. Berdahl, C. J. Torney, C. C. Ioannou, J. J. Faria, and I. D. Couzin, "Emergent sensing of complex environments by mobile animal groups," Science 339, 574 -- 576 (2013). 57C. W. Reynolds, "Flocks, herds and schools: A distributed behavioral model," in ACM SIGGRAPH computer graphics, Vol. 21 (ACM, 1987) pp. 25 -- 34. 58I. D. Couzin, J. Krause, R. James, G. D. Ruxton, and N. R. Franks, "Collective memory and spatial sorting in animal groups," Journal of theoretical biology 218, 1 -- 11 (2002). 59N. W. Bode, A. J. Wood, and D. W. Franks, "The impact of social networks on animal collective motion," Animal Behaviour 82, 29 -- 38 (2011). 60C. R. Brown, M. B. Brown, and M. L. Shaffer, "Food-sharing signals among socially foraging cliff swallows," Animal Behaviour 42, 551 -- 564 (1991). 61T. Schreiber, "Measuring information transfer," Physical review letters 85, 461 -- 4 (2000). 62D. Darmon, E. Omodei, and J. Garland, "Followers are not enough: A multifaceted approach to community detection in on- line social networks," PloS one 10, e0134860 (2015). 63E. M. Bollt, "Synchronization as a process of sharing and trans- ferring information," International Journal of Bifurcation and Chaos 22, 1250261 (2012). 64J. Sun and E. M. Bollt, "Causation entropy identifies indirect influences, dominance of neighbors and anticipatory couplings," Physica D: Nonlinear Phenomena 267, 49 -- 57 (2014). 65C. Cafaro, W. M. Lord, J. Sun, and E. M. Bollt, "Causation entropy from symbolic representations of dynamical systems," Appendix A: Information Flow, Causality, Influence and Leadership Information theory provides sophisticated measures for rigorously quantifying concepts like "the reduction in un- certainty about the present state of X given past states of Y ." As such, these measures are often associated with concepts like information flow, causality, influence and even leadership -- and often all of these terms are used interchangeably. These measures are often viewed as less subjective inference methods because almost no assump- tions need to be made about the structure of the system being observed. As a result, information theory has be- come a popular tool for inferring leadership from time series19 -- 26,62. However, while influence, information flow and causality are all closely related to the notion of lead- ership, these concepts are inherently different and there- fore are not readily interchangeable. Furthermore, recent work has begun to show that these information measures fail to even capture information flow33 let alone leader- ship. The following appendix discuses information flow, causality and influence and provides motivation for why we do not believe any of these alone fully quantifies lead- ership. Information flow and entropy, as we have argued in previous mathematical works24,32,63, is a fundamental concept in coupled (dynamical) systems, and the associ- ated stochastic processes. Information theory, as formu- lated upon Shannon entropy and its variants, basically describes the average "surprise" one should attribute to observing a specific value or state of a random vari- able. More formally, such quantification of surprise or (un)predictability is referred to as "entropy" and can be defined rigorously as a function of the underlying proba- bility distributions. When the time evolution of multiple variables are considered, the state of a variable often de- pends on the history of a set of related variables, and such inter-variable dependencies can be viewed as "infor- mation flow". Explicit characterization of information flow in coupled systems can be done by quantifying how informative (again as a notion of surprise) one should be in measured observations conditioned on given previ- ous observations, giving rise to commonly used measures such as transfer entropy61 and causation entropy26,64,65. In other words, information flow describes the reduction in uncertainty regarding forecasts for predictions asso- ciated with conditioning on the past in various combi- nations. Thus whether by Granger causality66, trans- fer entropy61, causation entropy26,64,65,67, or some other method, the idea is to ask if there is a reduction in uncer- tainty with knowledge of the past of a perhaps coupled variable. Clearly, this question is universally relevant from a wide range of scientific fields of science or math- ematics. However, part of the theme of this paper is that these information flow concepts themselves are not sufficient or equivalent as leadership. Causation is a related but not identical con- 12 to statistical71 -- 75, from philosophical68 -- 70, cept as information flow. The notion of causal- ity has many interpretations, depending on the con- text, to dynamical61,64,66,76. Here we will avoid the philosophical direction entirely, but note that some of these do coincide with the others. Statistical perspectives are sometimes relevant to a stochastic process, especially from the in- fluential work of Pearl71 -- 73, associated with a calculus for understanding interventions, but not always relevant to our context. We are more so interested in understanding interpretations of causal influence, of a free running sys- tem, that is, a system that is passively observed rather than actively probed. As such, this relates more closely, almost synonymously to the concepts of information flow in a stochastic process, but not quite identically. We take the same perspective as Granger in his line of rea- soning that eventually lead to the 2003 award of the No- bel Prize in Economics; Granger's fundamental principles were that 1) cause happens before effect, and 2) a cause necessarily contains unique information concerning fu- ture states of its effect66. In details the so-called Granger causality is a specific computation that assumes a linear stochastic process, and as such, it was shown76 to be en- tirely equivalent to transfer entropy computed by other means (in information theoretic by the Kullback-Liebler divergence appropriately conditioned) in the special case of a linear stochastic process with Gaussian noise. So said, while the underlying principles of Granger are the same, the details of computation may differ. Influence can now be described within this formal- ized framework as related to, but somewhat distinct from leadership, depending on if we are relating interactions between agents in terms of information theory, reduc- tion of uncertainty, or some other underlying principle, including the potential goal of controlling the system. Consider that some agents in a group may be leaders, with various ways to interpret this phrase to be stated subsequently below. A measure of leadership may be as- sociated with information flow for example, or as a proxy for causal influences that leaders may change states, be- fore other agents, a concept which will follow analogously to cause that comes before effect. An influential mem- ber of a group is not necessarily a leader, although in some sense influence is a kind of leadership de facto in the sense that influence is comparable to the possibility to cause others to change their behavior (dynamics). So said then what is the difference between influence, causation, and leadership, from the perspective of infor- mation flow? In some interpretations then, influence or causation over others and leadership are almost synony- mous but with important distinctions. When leadership is viewed through the lens of reduction of uncertainty (thus measurable by causation inference and information flow), then causation and influence becomes a synonym for leadership. Therefore, if a leadership action is ac- tive and observable, then causation and information flow are relevant concepts that enable one to define and em- pirically score the leadership. However, there are other notions of leadership that are clearly beyond the scope of information flow. Herein, by using a taxonomy of lead- ership, we expand beyond the typical causation and in- formation flow concepts24,25,37 to allow for those features which may be missed through the narrow interpretation of entropy, including structure, degree to which agents are informed, distribution, time and space scales, and target-drive are some of the other aspects that we will discuss here. 13
1608.01430	2	1608	2018-02-12T12:42:16	Demand Control Management in Micro-grids: The Impact of Different Policies and Communication Network Topologies	[ "cs.MA" ]	This work studies how the communication network between proactive consumers affects the power utilization and fairness in a simplified direct-current micro-grid model, composed by three coupled layers: physical (an electric circuit that represents a micro-grid), communication (a peer-to-peer network within the micro-grid) and regulatory (individual decision strategies). Our results show that, for optimal power utilization and fairness, a global knowledge about the system is needed, demonstrating the importance of a micro-grid aggregator to inform about the power consumption for different time periods.	cs.MA	cs	Demand Control Management in Micro-grids: The Impact of Different Policies and Communication Network Topologies Florian Kühnlenz, Pedro H. J. Nardelli, Hirley Alves 1 8 1 0 2 b e F 2 1 ] A M . s c [ 2 v 0 3 4 1 0 . 8 0 6 1 : v i X r a Abstract -- This work studies how the communication network between proactive consumers affects the power utilization and fairness in a simplified direct-current micro-grid model, com- posed by three coupled layers: physical (an electric circuit that represents a micro-grid), communication (a peer-to-peer net- work within the micro-grid) and regulatory (individual decision strategies). Our results show that, for optimal power utilization and fairness, a global knowledge about the system is needed, demonstrating the importance of a micro-grid aggregator to inform about the power consumption for different time periods. Index Terms -- Communication networks, Agents-based sys- tems, Network topology, Smart grids, Decision making. I. INTRODUCTION The growth of communication networks within the electric power grid allows for different new applications to its end- users [1]. This, on its turn, increases the complexity involved in the management of such a system. In this context, there is a need to develop models that aim at capturing new features that might emerge [2], as for example synchronization of individual behaviors due to signals and incentive structures. Game theory illustrates one approach that can capture such features since it is able to model strategic interactions of individual agents. It is worth mentioning that this theory was firstly introduced to model distributed economic relations, but is now widespread within other research communities (e.g. [3], [4] and references therein). The present work focuses on electricity micro-grids (specific region that can be self-sustained in relation to the large power grid, e.g. isolated rural areas with local energy generation) and is related to the numerous studies done on the fields of economics [5], power systems [6], [7] and communication sys- tems [8]. Here we follow the ideas presented in [9], where the authors pointed out some limitations of the standard economic models for energy systems while indicating the importance of complexity science approaches. Like some articles as [6], [7], [10], [11], we incorporate here aspects from different fields to include how peer interactions or collective individual actions may affect the macro-level outcomes as in gossip- based algorithms or electricity markets. The authors are with the Centre for Wireless Communications (CWC) at University of Oulu, Finland. P. H. J. Nardelli is also with Laboratory of Control Engineering and Digital Systems, Lappeerntanta University of Tech- nology, Finland. This work is partly funded by Finnish Academy (n. 271150) and CNPq/Brazil (n.490235/2012-3) as part of the joint project SUSTAIN, by SRC/Aka BCDC Energy (n.292854) and by the European Commission through the P2P-SmarTest project (n.646469). Contact: [email protected] To do so, we use an agent-based model, recently proposed by the authors in [12], to model agents' interactions within a system constituted of three layers (physical, communication and regulatory, to be explained later) and then obtain a better understanding of both the inter- and intra-layer dynamics. We therefore expect to provide a different perspective about demand-side management policies [13] -- [16] that goes beyond individual utility maximization and purely competitive strate- gies (usual assumptions of many economic researches [9]). In specific terms, we extend our previous model of a fully distributed system operation by introducing an entity that provides a centralized signal. This was motivated by the demand control problem in direct current (DC) micro- grids. We consider three different scenarios related to how information is shared among consumers: no signaling (similar to [12]), global state signaling and pricing scheme. The first strategy is fully distributed, while the other two require an entity with global state information about the whole system; this entity is the micro-grid aggregator. Our main contribution is to assess these different demand-side management policies (without or with micro-grid aggregator) in both micro and macro levels looking at the agent decision dynamics and considering different communication network topologies. It is worth mentioning the present paper is the first attempt to study micro-grids using the model introduced in [12]. Our model is described as follows. The physical layer (a DC micro-grid) is a circuit composed by a power source and resistors in parallel. Individual agents (the proactive con- sumers, the "prosumers") can add, remove or keep the resistors they have. Agents' decisions aim at maximizing their own delivered power, which is a non-linear function dependent on the others' behavior, and they are based on (i) their internal state, (ii) their global state perception, (iii) the information received from their neighbors in the communication network, and (iv) a randomized selfishness related to their willingness of answering to a demand-side control request. Different peer- to-peer communication network topologies and randomized communication errors in the deployment of micro-grids are important aspects to be included as in [17], [18]. But differ- ently from other papers, we analyze here their systemic effects; for detailed results of communication network implementation in power grids using, for instance, spatial and/or temporal spectrum sharing, refer to [19] -- [23]. Looking at the proposed model, by individually modifying the peer-to-peer communication network topology and the demand-side management policies keeping fixed the other TABLE I NOTATIONS Notation t P Z A " t1, 2, ..., Nu Aztiu i P A Ni Ă A nrts P tN, N ` 1, ...u airts P N` rirts " nrts ´ airts Pirts ą 0 λirts P R λmin P R Sirts P t´1, 0,`1u si P r0, 1s perr P r0, 1s cavg P r0, 1s R RV Meaning discrete time set of agents; N is the system size set A without agent i agent i neighborhood set of agent i active number of resistors at t active resistors of agent i active resistors excluding agent i agent i consumed power [units of power] gain in power of agent i system-wide predefined minimum gain agent i's state: `1 (defect), 0 (ignore), or ´1 (cooperate) selfishness gene of agent i error probability in a communication link average value of cooperating agents resistor value added as a load by agents source resistor parameters including the link error probability, we expect to show how these factors affect the power utilization and fairness in the system for different number of agents. Note that our study assumes that each one of those three layers (physical, communication and regulatory) is equally important to deter- mine the system dynamics and the individuals' (re)actions. In this case, the system can neither be reduced to one or two of these layers nor the dynamics of the whole system cannot be assessed based on individual (typical) agents. Our previous results in [12] can be seen as the benchmark scenario where no demand-side management is considered; there we showed: (i) different communication network topolo- gies (ring and Watt-Strogatz-Graph [24]) lead to different levels of power utilization and fairness at the physical layer and (ii) a certain level of error induces more cooperative behavior at the regulatory layer. Now, if demand-side control or pricing schemes are considered, which is the focus of the present work, the system behaves in a more predictable manner and is less dependent on its size. In this case, the global knowledge about the system state enables much higher utilization and fairness, which indicates the need of an entity to guide the peer relations. It is important to say that, although there is a wide understanding that global signaling leads to more stable outcomes, this is not always the case; many times global signaling induces destabilizing collective behavior (e.g. [25] and references therein). The rest of this paper is divided as follows. Section II describes the multi-layer model employed here. Section III presents the numerical results used to analyze the proposed scenario. In Section IV, we discuss the lessons learned from our model, indicating potential future works. The most impor- tant symbols are presented in Table I. II. MULTI-LAYER MODEL A. Background The power grid has been thought as a one-way network, where the generators produce the electricity that needs to travel long distances in high-voltage transmission lines to arrive at the distribution network, which delivers the electricity to 2 final consumers [2], [12]. However, in order to integrate ever increasing wind and solar power, this network has to change in many ways. First, solar and wind energy need more physical space than classical power plants, since they are not just transforming energy from one high density form to another, but rather collecting low density energy while distilling it in to a high density form. Second, they will produce energy in a much less predictable and/or consistent way. This volatility in supply means that the consumption and distribution need to react to these changes, either though storing energy or shifting consumption in time. In this case, automation is needed to balance the system; therefore, the so-called smart grids are planned to heavily use informa- tion and communication technologies (ICTs) to collect and take autonomous actions. Communication technologies should then enable the grid to use electricity from diverse sources whenever they are available, delivering to wherever they are needed. This would improve the efficiency of both production and distribution since the smart grid is designed to provide the necessary capabilities to enhance the reliability of energy infrastructures, and decrease the impact of supply disruptions, as well as create new services and applications. To achieve the needed degree of automation to integrate volatile energy sources, the power grid needs to be coupled with the communication network. This creates a codepen- dency, where the power grid will only work when the com- munication system is working, which in turn will only work if the power system is working. Furthermore, the volatility of the power sources demands a new interaction paradigm between producers and consumers. Consumers need to react to changes in the supply, by shifting their demand, whenever possible, to synchronize with production. This means that the smart grid will also be much more coupled to a regulatory layer, which will influence consumer behavior and decisions. In summary, more than any specific technological challenge existing in the generation, transmission and distribution, a proper smart grid model should include the following different layers of analysis. ‚ Physical that electricity grids include intermittent sources of energy (e.g. solar panels or wind turbines), micro-grids and mobile batteries/loads (e.g electric ve- hicles). The energy interchange occurs in this layer, transported in form of electricity. ‚ Information networks to acquire, process and dissem- inate information respecting the specific application re- quirements. The communication between agents happens in this layer, which can be seen as the agent reality (i.e. the agent only knows about itself and whatever is communicated to it). ‚ Regulatory layer or markets where consumers and operators can interchange electricity, evolving differ- ent strategies under different exchange regulations. This layer therefore include the socio-economical context and decision-making processes, which directly affects the physical layer. Instead of analyzing each one of them separately, consider- ing the others as given (the most usual approach), this research will model these three layers as constitutive parts of the smart 3 Fig. 1. (a) Electrical circuit representing the physical layer of the system. The circuit is composed by a power source V and RV, and resistors of R in parallel, generating a current I. These resistors are related to N agents that can add, remove or keep the resistors under their control in the circuit. The minimum number of resistors an agent can have is one and there is no maximum. We also consider N as the size of the system. (b) The agents are connected in a communication network so that a given agent has access to the information related to the previous action of their first-order neighbours. In the ring topology ř illustrated here, every agent is connected with two other agents. In this case agent i is connected with agents i ´ 1 and i ` 1 with i " 1, 2, ..., N. In the ring topology, agents 1 and N are neighbors. (c) At time step t, the normalized power delivered Pallrts{Ptyp to the agents with rising number of resistors nrts in the system, where Ptyp " V 2 iPA Pirts with Pirts given by equation (1). (d) The decision tree representing the decision making RV process of each agent for each round. and Pallrts " grid system. By doing so, we will be able to provide a deeper understanding of the smart grid spatio-temporal multi-layer dynamics and the emergent systemic features that arise from them. It is important to say that this proposal is built upon the assumption that effective cross-layer management strategies targeting the realization of the ambitious goals planned for the smart grids can be only realized if the way that the system elements interact within one layer and across different layers is properly characterized. This approach is philosophically inspired by the Rus- sian/Soviet tradition of reflexive systems and control [26]. Such a perspective in their current version has two funda- mental aspects captured by our model: the importance of the information layer as the filter of the physical reality, and the reflexive effects of agents and contexts that they take their decisions. In other words, what individual agents perceive is as important as (or even more important than) what is actually happening in the physical layer. Their individual decisions at the regulatory layer, in turn, depend on a combination from their own internal state and their own perceptions of +−VRVIR1...Ra1...Agent1R1...RaNAgentN(a)Electriccircuitii+1i−1(b)Communicationnetwork0100200300400500n[t]/N0.000.050.100.150.200.250.30Pall[t]/Ptyp(c)Systempowerλ>λminrepeatpreviousstepany(Si[Ni]==−1)orrand()>siifai[t]>1:ai[t+1]=ai[t]−1rand()<siai[t+1]=ai[t]+1ai[t+1]=ai[t]yesnoyesnoyesno(d)Decisiontreethe system state. The individual decisions, which are built upon different perspectives, will determine the physical layer dynamics, which shall affect the future decisions of individual agents (i.e. reflexive behavior). In practical terms, the present work is a multi-layer model of DC micro-grids, as part of smart grids, that considers also peer-to-peer exchange of information within a network with different topology; this model will be briefly described in the following. Our discrete-time, agent-based model assumes these three layers as constitutive parts of the system composed by an electric circuit as the micro-grid physical infrastructure, a communication network where agents exchange local informa- tion and a set of regulations that define the agents' behavior, as exemplified in Fig. 1. It is worth saying that this model -- as any other model -- is a clear simplification of the real grid. Our goal here is to put light on the systemic multi-layer dynamics, which is most often neglected. B. Agents' decision process As stated before the model assumes discrete time steps, denoted by t. Therefore, the interactions between agents might be viewed as a round-based game [27] based in the state Sirts, defined at Table I and discussed later on. At each step in time t, every agent aims to maximize its own consumed power. To achieve this goal, the agent has three options: ‚ Add a resistor R (defecting); ‚ Remove a resistor R (cooperating); or ‚ Do nothing (ignoring). ř We assume that the value of R is fixed and the same for all agents. Our preliminary investigation indicated that relaxing this assumption within a reasonable range of values has no significant effects on the system dynamics; then, we decided to use the homogeneous case for simplicity. The decision is based on the gain from the previous strategy Sirt´ 1s in order to decide its new state Sirts in the following manner. If the gain λirt ´ 1s is greater than or equal to a system-wide predefined minimum λmin, the agent continues to its (successful) strategy at time t, i.e. Sirts " Sirt ´ 1s. If λirts ă λmin, then agent i compares its strategy with its neighborhood Ni, which is related to the communication network and will be defined later in this section. jPNi In the case that a majority of the neighborhood is cooperat- Sjrt´ 1s ă 0, this will appear to agent i as an ing, e.g. indication that the system is in a condition of overusing. The agent will then also change to a cooperative state, leading to Sirts " ´1. Otherwise, the agent decision will be related to its individual selfishness as follows: a random number between 0 and 1 is drawn to be compared to its own selfishness gene si in order to decide whether it will start cooperating or not. Each agent is assigned with selfishness gene at the beginning of the simulation, which is also a random number between 0 and 1. In the case of not cooperating, another random number is drawn and once again compared to the selfishness gene si, but now to decide if stays inactive (i.e. Sirts " 0) or adds another load in the circuit (i.e. Sirts " `1). The agent decision procedure is shown in Fig. 1 (d). 4 C. Communication network The communication network enables agent i to know the state Sjrt ´ 1s of the agents j P Ni which are in his neighborhood. We assume that agent j always transmits its actual state Sjrts to agent i. However, the communication links can also experience errors. An error event means the received message by agent i contains a different information than agent j has sent. So that if SjÑirt ´ 1s " Sjrt ´ 1s is the state information send from j to i and SjÑirt´ 1s be the " ı information received by i. We assume that error events are independent and identically distributed (i.i.d.) such that the probability of the event: SjÑirt ´ 1s ‰ SjÑirt ´ 1s " perr for all t P Z, i P A Pr and j P Ni, Prrs refers to the probability that a given event occurs. The network is a bidirectional graph so that an error at i Ñ j does not imply an error at j Ñ i, and vice-versa. If an error happens, the received information SjÑirt ´ 1s is assumed to be i.i.d. between the other possible states. In this study we compare two different communication networks [24]: ‚ Ring: Each agent connects to exactly two other nodes to form a ring. It is a simple deterministic topology. The agents have then the same degree. ‚ Watt-Strogatz-Graph: Random graph that has small- world properties, namely short average path lengths and high clustering. In this case, the agents have, in average, similar degrees. We did not consider here Barabasi-Albert graphs since the scenario under investigation presents a similar macro-level behavior to Watts-Strogatz, as shown in [12]. D. Physical system In the physical systems as presented in Fig. 1, there exists an optimal number of resistors that leads to the maximum power delivered from the source to the agents. If the delivered power is below the maximum on the right, then there will be a gain by removing a resistor until the system has reached such point. Conversely, if it is below on the left, then there will be a gain by adding a resistor. One may ask the following question: Is it possible for the agents to reach the optimal point with limited knowledge about the system? In the presence of a central controller this would be a fairly easy problem. First, we need to find the number of resistors that leads to optimal power to then distribute them among the agents. This kind of solution resembles time division schemes in computer networks or cellular systems [28]. However, as discussed before, in the absence of a centralized controlling entity, the agents only have a limited knowledge about other agents. In any case, the agents have some information about the state of the system based on their own power consumption and the decision done. At time-step t, the power each agent consumes Pirts with i P A " t1, ..., Nu and is given by: airts µ paavgrts ` µq2 , Pirts " Ptyp (1) 5 Fig. 2. On the left: change in the average cooperation cavg depending on the system size N. On the right: examples of a typical system behavior as a function of time t where the points (pixels) represent the agent state Sirts such that red means defection, white cooperation, and black = doing nothing for N " 1000 (top), N " 100 (middle) and N " 10 (bottom). The system parameters are: λmin " 0.0005, RV " 2 Ω, R0 " R{N " 200 Ω, perr " 0.01 and V " 1 V. The communication network is configured as Watts-Strogatz graph. RV RV , µ " R where Ptyp " V 2 , airts is the number of active resistors the agent i possesses, rirts is the number of active resistors in the system excluding the source resistor RV and the ones controlled by agent i, and aavgrts " pairts`rirtsq{N. The physical system is then described by its size N, the ratio µ of the resistance values and the power source V . The resistors are scaled so that the optimal average number of ? is independent of N while the voltage might resistors avg be scaled with N to have a constant ratio of power per agent, as explained later on. The gain that agent i experiences at time-step t is then defined as: a ` III. NUMERICAL RESULTS In this section, we present our main results. We first revisit the results introduced in [12] where no demand-side manage- ment signaling is considered and the agents try to achieve a global optimum solely based in their local information. We then present our new results evaluating the effects of such factors on the system dynamics in comparison with our baseline model. Once again, it is important to say that the numerical results to be presented only exemplify the impact of different ways of signaling on the multi-layer system dy- namics. Although qualitative changes are expected to happen consistently, the actual numerical values depend on the actual signal construction. λirts " Pirts ´ Pirt ´ 1s Pirt ´ 1s " ∆Pi Pirt ´ 1s . (2) A. Baseline model This implies that the agents only use the information about the previous time-step t ´ 1. If we expand (2) using (1), the resulting equation that determines λirts becomes more complicated. To make the analysis clearer, we choose to apply the following approximation: λirts « dPi Pirts « ∆airts airts ´ 2 N 1 aavgrts ` µ p∆rirts ` ∆airtsq , (3) such that the gain λirts is now a function of the variations in agent i's own number of resistors ∆airts " airts´airt´1s and in the number of resistors controlled by other agents ∆rirts " rirts ´ rirt ´ 1s, as well as the average number of resistors aavgrts and the system parameters N and µ. In Fig. 2 one can see the inherent dynamics of the system when no demand-side management is done and the communi- cation network topology is a random graph. We identify two phenomena that are important for our analysis, as explained next. First, we see that the average number of cooperators varies with the size of the system, showing a very high number of cooperation for larger systems. However, as we will see later in Fig. 2, this is not sufficient to understand the system dynamics. This high level of cooperation is not necessarily a universally good outcome. The second phenomena is a change in behavior of the system for different sizes. While for small systems we see an almost checkerboard like distribution of white (coopera- tion) and black (defection) dots, mid-sized systems show a strikingly different image, in this case we now see a wave- like pattern where a large front of cooperation establishes and then suddenly brakes down. This is again in contrast to the 101102103sizelog(N)0.50.60.70.80.91.0cavgN=1000N=100tN=106 Fig. 3. Typical behavior of very small (N " 5), small (N " 10), medium (N " 100) and large systems (N " 1000). The small systems can operate much closer to the optimum, while mid-size systems show big jumps after stable periods. The system parameters are λmin " 0.005, RV " 2 Ω, R0 " R{N " 200 Ω, perr " 0.01 and V " 1 V. The communication network is configured as a ring. picture for large systems where the behavior is dominated by cooperation with some sparsely distributed red dots. The resulting changes in the power delivery can be seen in Fig. 3, which shows that only very small systems of less then 10 agents are able to stay very close to the optimum, while systems of more then 10 agents start to deviate from it. These curves also reflect the wave-like behavior for mid- sized systems, where the number resistors drops due to almost global cooperation. This brings the system close to the optimal point, when the number of resistors rises once again. For large systems, it does not get close to the optimum. However, it is more stable and predictable than the other cases. The background of how this global behavior emerges from the local interactions is found in our previous work [12]. Here, we will view this intrinsic unstable performance as a undesirable outcome. In the next subsection, we will explore ways to prevent this behaviors by using demand-side signaling with different communication topologies. As we will see later, our goal is to provide a global indication via either direct signaling or pricing so the agents can use this information to make their individual decisions. B. Demand-side management and the micro-grid aggregator Let us first define the power utilization Putil as the fraction of power that is utilized by the system and the available power: Putil " 4 Ptyp Pi,avg, (4) ÿ iPA T´1ÿ t"0 where Pi,avg is the time average power of agent i computed as: Pi,avg " lim TÑ8 1 T Pirts. (5) The first demand-side management policy is based on a global signal that every agent receives and labeled signal in the figures. In Fig. 4, we can see the results of such a global signal that is sent to all agents when the system is beyond the optimal point. The results are shown in comparison to the base case, labeled "basic graph" as presented in the previous subsection. This signal might be interpreted as a way for a micro-grid operator to implement a demand response mechanism when the system is experiencing a peak in consumption, while supply is very limited. Such a system are already found in the reality. As discussed in [29], the Scottish Island Eigg uses a traffic light system to signal its inhabitants the state of supply in the micro-grid. 0.00.20.40.60.81.0PutilN=5t0100200300400500600700aavgaoptavgN=10taoptavgN=100taoptavgN=1000taoptavg7 Fig. 4. Power utilization for different communication strategies and topologies depending on the system. The green line shows the utilization for the basic model. The red and brown lines are from the simulations with global demand response signal. Fig. 5. Power utilization for step pricing and agents with utility function described in this section. The green line shows the utilization for the basic model. From the agents' view, this signal provides a global informa- tion about the system so they do not have to rely only on their neighbors. This global information is assumed to be reliable and is treated in the same way as the neighbor information: if the signal is present and the agent experiences a small gain, it starts cooperating. As we can see in Fig. 4, the signal enables the system to reach the optimum point seems independent from its size. This contrasts with the original system, where on average is not possible for any size to reach such high states of power utilization, regardless of the communication network topology. The second demand-side management policy is the intro- duction of a two-stage real time price for all agents. For this purpose, we have to adapt our model further. Instead of just maximizing its own power demand, the agents now have to maximize their utility value taking the price into account. The price function needs to reflect the state of the system as a signal of overuse (when necessary). In order to build such a function, we first need to build an utility function that reflects how the power demand is valued. We adopted the following utility from [30], [31]: UαpPirts, ωiq " ωPi ´ α # , 2 P 2 i if 0 ď Pi ă ω if Pi ě ω α α ω2 2α where Pi being the power consumption while α and ω determine how consumption is valued. The function is chosen so that there exists a linear marginal benefit up to a certain point, at which the benefit does not increase anymore with higher consumption. In our case, we keep α fixed at 0.2 while ωi is uniformly distributed between 2.05 0.05, representing different values of consumers about the importance of power consumption. The price function, meanwhile, is modeled as a simple step # function: prts " if nrts ď nopt if nrts ą nopt , p1 p2 assuming that p1 " 0.2 and p2 " 5. The cost for each user is then calculated by multiplying the consumption with the current price: cirts " prtsPirts. The agent decision process itself is similar, but now, in- stead of optimizing the received power, the agents are now optimizing their benefit (the utility minus the cost): birts " UαpPirts, ωiq ´ cirts. The results using the proposed scheme can be viewed in Fig. 5. We can once again see that the results are much better than the one presented in Section III-A. However, if we compare the case of a simple ring graph (brown curve, labeled "util ring") describing the neighborhood versus a more meshed topology from Watts-Strogatz graph (green, labeled "util graph"), the communication network (where local information about the neighbors are transmitted) has still a an influence on the final outcome. This effect seems to be more pronounced for small networks. However, mid-sized systems, which are the most critical size when no signaling is considered, still experience stable outcomes, comparable to the global signal policy. IV. DISCUSSION AND FINAL REMARKS In our previous contribution [12], we showed that a simple direct current micro-grid system composed by three con- stituent layers (physical, informational and regulatory) may present emergent dynamics at macro-level. This kind of very unpredictable behavior that arises is normally undesirable to manage the system and is related to a lack of global information. This paper shows that it is possible to counteract this behavior when some kind of global signaling based on the micro-grid aggregator. One strategy is to send direct signals to end-users when the micro-grid is overused, while the other is a indirect signal through a price function. The first strategy is a rather simple way of providing each agent with global information about micro-grid aggre- gate power consumption. This strategy enables a more stable outcome, which is much less influenced by other parameters of the simulation, like the communication network topology. The second strategy in its turn maps the original power op- timization into a cost-benefit optimization, where each agents wants to achieve the individual optimal power usage consid- ering price and utility value of the consumed power. While similar outputs could be demonstrated with this approach, it is necessary to point out that the choice of parameters to achieve a stable outcome in the scenario is much more complicated. In 100101102103104log(N)0.00.20.40.60.81.0Putilsignalringsignalgraphbasicgraph100101102103104log(N)0.00.20.40.60.81.0Putilutilgraphutilringbasicgraph8 [14] B. D. S. Callaway and I. A. Hiskens, "Achieving Controllability of Electric Loads," Proceedings of the IEEE, vol. 99, no. 1, 2011. [15] W. Zhang, Y. Xu, S. Li, M. Zhou, W. Liu, and Y. Xu, "A distributed dynamic programming-based solution for load management in smart grids," IEEE Systems Journal, vol. PP, no. 99, pp. 1 -- 12, 2016. [16] V. Pradhan, V. S. K. M. Balijepalli, and S. A. Khaparde, "An effective model for demand response management systems of residential electric- ity consumers," IEEE Systems Journal, vol. 10, no. 2, pp. 434 -- 445, June 2016. [17] Q. Li, Z. Feng, W. Li, T. A. Gulliver, and P. Zhang, "Joint spatial and temporal spectrum sharing for demand response management in cognitive radio enabled smart grid," IEEE Transactions on Smart Grid, vol. 5, no. 4, pp. 1993 -- 2001, 2014. [18] A. A. Khan, M. H. Rehmani, and M. Reisslein, "Cognitive radio for smart grids: Survey of architectures, spectrum sensing mechanisms, and networking protocols," IEEE Communications Surveys & Tutorials, vol. 18, no. 1, pp. 860 -- 898, 2016. [19] P. H. Nardelli, M. de Castro Tomé, H. Alves, C. H. de Lima, and M. Latva-aho, "Maximizing the link throughput between smart meters and aggregators as secondary users under power and outage constraints," Ad Hoc Networks, vol. 41, pp. 57 -- 68, 2016. [20] M. C. Tomé, P. H. Nardelli, H. Alves, and M. Latva-aho, "Joint sampling-communication strategies for smart-meters to aggregator link as secondary users," in 2016 IEEE Energy Conference (ENERGYCON) International. [21] A. S. Cacciapuoti, M. Caleffi, and L. Paura, "On the probabilistic deployment of smart grid networks in tv white space," Sensors, vol. 16, no. 5, p. 671, 2016. [22] A. S. Cacciapuoti, M. Caleffi, F. Marino, and L. Paura, "Enabling smart grid via tv white space cognitive radio," in 2015 IEEE International Conference on Communication Workshop (ICCW), pp. 568 -- 572. [23] A. S. Cacciapuoti et al., "Mobile smart grids: Exploiting the tv white space in urban scenarios," IEEE Access, vol. 4, pp. 7199 -- 7211, 2016. [24] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang, "Complex networks: Structure and dynamics," Physics reports, vol. 424, no. 4, pp. 175 -- 308, 2006. [25] P. H. Nardelli and F. Kühnlenz, "Why smart appliances may result in a stupid energy grid?" arXiv preprint arXiv:1703.01195, 2017. [26] V. Lepskiy, "Evolution of cybernetics: Phylosophical and methodological analysis," in 2017 World Organisation of Systems and Cybernetics, 2017, to appear. [27] M. Archetti and I. Scheuring, "Review: Game theory of public goods in one-shot social dilemmas without assortment," Journal of Theoretical Biology, vol. 299, pp. 9 -- 20, 2012. [28] I. Marsic, Computer networks: Performance and quality of service. Rutgers University, 2010. [29] A. Yadoo, A. Gormally, and H. Cruickshank, "Low-carbon off-grid electrification for rural areas in the united kingdom: Lessons from the developing world," Energy Policy, vol. 39, no. 10, pp. 6400 -- 6407, 2011. [30] M. Fahrioglu and F. L. Alvarado, "Designing cost effective demand management contracts using game theory," in IEEE Power Engineering Society 1999 Winter Meeting, vol. 1. IEEE, 1999, pp. 427 -- 432. [31] P. Samadi et al., "Advanced demand side management for the future smart grid using mechanism design," IEEE Transactions on Smart Grid, vol. 3, no. 3, pp. 1170 -- 1180, 2012. [32] F. Kühnlenz and P. H. J. Nardelli, "Agent-based model for spot and balancing electricity markets," in 2017 IEEE International Conference on Communications Workshops (ICC Workshops). the non-trivial cases, where the sum of the demanded power of all agents exceeds the maximum available power, the outcome is highly dependent on the choice of not only the price levels but also the valuation of the power given by the variables α and ωi. These facts suggest that real time pricing in combination with demand response, which is often proposed as the way to utilize changes in available power, needs to be finely tuned in order to achieve a stable outcome. In other words, pricing does not inherently guarantee optimal usage and stability for the micro-grid. Rather, the outcome of the first strategy indicates that a direct response to capacity constraints without the intermediate pricing layer is much easier to implement in a stable manner. In any case, our results reinforce the importance of the micro-grid aggregator as a signaler of the global system state information for a stable peer relation between prosumers in the micro-grid. We expect to further develop this simulation by adding generation capabilities in the prosumers together with a more complete market description, as indicated by our initial results [32]. ACKNOWLEDGMENTS We would like to acknowledge the computing facilities of CSC - IT Center for Science Ltd. (Finland) that was used to run the simulation scenarios. REFERENCES [1] S. Bera, S. Misra, and J. J. Rodrigues, "Cloud Computing Applications for Smart Grid: A Survey," IEEE Transactions on Parallel and Dis- tributed Systems, no. c, pp. 1 -- 1, 2014. [2] P. H. J. Nardelli et al., "Models for the modern power grid," The European Physical Journal Special Topics, vol. 223, no. 12, pp. 2423 -- 2437, 2014. [3] A. H. Mohsenian-Rad et al., "Autonomous demand-side management based on game-theoretic energy consumption scheduling for the future smart grid," IEEE Transactions on Smart Grid, vol. 1, no. 3, pp. 320 -- 331, Dec 2010. [4] Y. Wang et al., "A game-theoretic approach to energy trading in the smart grid," IEEE Transactions on Smart Grid, vol. 5, no. 3, pp. 1439 -- 1450, May 2014. [5] W. Saad, Z. Han, H. V. Poor, and T. Basar, "Game-theoretic methods for the smart grid: An overview of microgrid systems, demand-side management, and smart grid communications," IEEE Signal Processing Magazine, vol. 29, no. 5, pp. 86 -- 105, 2012. [6] J. Engels, H. Almasalma, and G. Deconinck, "A distributed gossip-based voltage control algorithm for peer-to-peer microgrids," in 2016 IEEE International Conference on Smart Grid Communications (SmartGrid- Comm. IEEE, 2016, pp. 370 -- 375. [7] H. Almasalma, J. Engels, and G. Deconinck, "Peer-to-peer control of microgrids," arXiv preprint arXiv:1711.04070, 2017. [8] S. Safdar, B. Hamdaoui, E. Cotilla-Sanchez, and M. Guizani, "A survey on communication infrastructure for micro-grids," in 2013 9th Inter- national Wireless Communications and Mobile Computing Conference (IWCMC), pp. 545 -- 550. [9] C. S. Bale, L. Varga, and T. J. Foxon, "Energy and complexity: New ways forward," Applied Energy, vol. 138, pp. 150 -- 159, 2015. [10] F. L. Alvarado, J. Meng, and C. L. DeMarco, "Stability analysis of interconnected power systems coupled with market dynamics," IEEE Transactions on Power Systems, vol. 16, no. 4, pp. 695 -- 701, 2001. [11] J. W. Simpson-Porco, F. Dörfler, and F. Bullo, "Synchronization and power sharing for droop-controlled inverters in islanded microgrids," Automatica, vol. 49, no. 9, pp. 2603 -- 2611, Sep. 2013. [12] F. Kühnlenz and P. H. Nardelli, "Dynamics of complex systems built as coupled physical, communication and decision layers," PloS one, vol. 11, no. 1, p. e0145135, 2016. [13] S. Gottwalt et al., "Demand side management: A simulation of house- hold behavior under variable prices," Energy Policy, vol. 39, no. 12, pp. 8163 -- 8174, 2011.
1412.1523	2	1412	2015-12-06T20:48:53	Information Exchange and Learning Dynamics over Weakly-Connected Adaptive Networks	[ "cs.MA", "cs.IT", "cs.LG", "cs.IT" ]	The paper examines the learning mechanism of adaptive agents over weakly-connected graphs and reveals an interesting behavior on how information flows through such topologies. The results clarify how asymmetries in the exchange of data can mask local information at certain agents and make them totally dependent on other agents. A leader-follower relationship develops with the performance of some agents being fully determined by the performance of other agents that are outside their domain of influence. This scenario can arise, for example, due to intruder attacks by malicious agents or as the result of failures by some critical links. The findings in this work help explain why strong-connectivity of the network topology, adaptation of the combination weights, and clustering of agents are important ingredients to equalize the learning abilities of all agents against such disturbances. The results also clarify how weak-connectivity can be helpful in reducing the effect of outlier data on learning performance.	cs.MA	cs	Information Exchange and Learning Dynamics over Weakly-Connected Adaptive Networks Bicheng Ying, Student Member, IEEE, and Ali H. Sayed, Fellow, IEEE 1 5 1 0 2 c e D 6 ] A M . s c [ 2 v 3 2 5 1 . 2 1 4 1 : v i X r a Abstract -- The paper examines the learning mechanism of adaptive agents over weakly-connected graphs and reveals an interesting behavior on how information flows through such topologies. The results clarify how asymmetries in the exchange of data can mask local information at certain agents and make them totally dependent on other agents. A leader-follower relationship develops with the performance of some agents being fully determined by the performance of other agents that are outside their domain of influence. This scenario can arise, for example, due to intruder attacks by malicious agents or as the result of failures by some critical links. The findings in this work help explain why strong-connectivity of the network topology, adaptation of the combination weights, and clustering of agents are important ingredients to equalize the learning abilities of all agents against such disturbances. The results also clarify how weak-connectivity can be helpful in reducing the effect of outlier data on learning performance. Index Terms -- Weakly-connected graphs, distributed strategies, information exchange, Pareto optimality, leader-follower relation- ship, limit points, outlier data, malicious agents. I. INTRODUCTION This work examines the distributed solution of inference problems by a collection of networked agents. An individual convex cost function Jk(w) : RM → R is associated with each agent k = 1, 2, . . . , N , and the objective is for the agents to cooperate locally in order to determine the global minimizer of the aggregate cost. Several useful techniques have been developed in the literature for this purpose, including the use of consensus strategies [2] -- [8] and diffusion strategies [9] -- [11]. In most prior studies (see, e.g., [2] -- [25]), it has been generally assumed that the network topology is strongly- connected, which means that a path can be found linking any pair of agents and, moreover, at least one agent has a self- loop. In this case, and under some mild technical conditions [9], it is known that all agents are able to approach the global minimizer within O(µmax), where µmax denotes the largest step-size parameter used by the adaptive agents. This means that strongly-connected agents are able to learn well, with all agents attaining a similar performance level despite possible variations in the signal-to-noise ratios across the agents. The main theme of the current article is to examine how the learning behavior of the agents is affected when the network topology is not necessarily strongly-connected. Specifically, we shall examine the effect of weak-connectivity This work was supported in part by NSF grants CIF-1524250 and ECCS- 1407712. A short version of this word appeared in the conference publication [1]. The authors are with Department of Electrical Engineering, University of California, Los Angeles, CA 90095. Emails: {ybc,sayed}@ucla.edu. on information flow, where a weakly-connected network is one that consists of multiple clustered agents with at least one cluster always feeding information forward but never receiving information back from any of the other clusters. Such graph settings arise in important situations. For example, they can arise as the result of intruder attacks by malicious agents who keep feeding their neighbors with inaccurate information but never use the information fed back to them by the other agents. This behavior results in a unilateral and asymmetric informa- tion exchange scenario, which can be modeled by weakly- connected networks. A second example arises in the context of interactions over social networks where weak connectivity can be used to model the behavior of stubborn agents that insist on their opinions regardless of the feedback by others, or to model the ability of organizations to control the flow of media information [26], [27]. A third example arises in the context of information dissemination over social platforms such as Twitter and Weibo, where celebrities may have thousands of followers while these special users may be following only a small subset of trusted acquaintances. All these examples lead to heavily biased asymmetric information exchange scenarios, which can be modeled reasonably well by weakly-connected networks. Motivated by these considerations, we examine the learning mechanism of adaptive agents over weakly-connected graphs in some detail and reveal an interesting behavior. We will find that a leader-follower relationship develops among the agents, with the performance of some agents being fully determined/controlled by the performance of other agents that are outside their domain of influence. This scenario does not only arise as the result of intruder attacks or highly asymmetric information exchanges, as explained before, but can also be the the result of failures by some critical links that render the network topology weakly-connected. Among other contribu- tions, the findings established in this article will help explain why strong-connectivity of the network topology [9], [10] adaptation of the combination weights [9], [11] and clustering of agents [28], [29] are important ingredients to safeguard against such pitfalls. The conclusion will also highlight one useful scenario where weak-connectivity is beneficial, namely, in reducing the effect of outlier data on network performance. Some useful related works on leader-follower relationships appear in [30] -- [33]. References [31], [32] are primarily mo- tivated by the behavior of fish schools and the switching be- tween leader and follower agents, while [33] is concerned with controllability issues in the context of a switching topology. The work [30] also deals with weakly-connected networks but differs critically from our formulation in that it considers a static scenario with anchor sensors and initial fixed state values. The purpose is to show that an average consensus iteration will make the other sensors converge to a combination of the anchor states. While this is a useful conclusion, it is not surprising that it holds in this scenario due to the static nature of the problem formulation. The key difference in relation to the work proposed in the current manuscript is that we will be dealing with a dynamic network scenario where data are constantly streaming into the agents. There are no anchor agents or fixed state values. Instead, the data are continuously changing. Besides, the probability distribution of the data is unknown and the agents need to rely on approximate gradient information, as opposed to actual state values, to carry out their learning. Under such circumstances, gradient noise will be constantly present in the network and will seep into the operation of the distributed strategies. The topology couples the agents together and these gradient noise sources end up diffusing through the network. In this case, it is much less trivial to identify the limit points of the various sub-networks and agents. This is because the presence of persistent gradient noise now forces agents to fluctuate around limits points. The challenge is to show that the dynamics over the network is stable enough to keep these noises under check and to lead to a steady-state scenario with well-defined (mean-square-error) limit points. The analysis in the paper will show that this is indeed the case and will derive two main results (Theorems 1 and 4), which characterize analytically and in closed-form the stability and performance of weakly-connected graphs under dynamic continuous learning. Notation: We use lowercase letters to denote vectors, up- percase letters for matrices, plain letters for deterministic variables, and boldface letters for random variables. We also use (·)T to denote transposition, (·)−1 for matrix inversion, Tr(·) for the trace of a matrix, λ(·) for the eigenvalues of a matrix, k·k for the 2-norm of a matrix or the Euclidean norm of a vector, and ρ(·) for the spectral radius of a matrix. Besides, we use A ≥ B to denote that A − B is positive semi-definite, and p ≻ 0 to denote that all entries of vector p are positive. II. STRONGLY-CONNECTED NETWORKS In preparation for the derivation of the main results in this work, we first review the setting of strongly-connected networks following [9], [10]. Thus, consider a network consisting of N separate agents connected by a topology. We assign a pair of nonnegative weights, {akℓ, aℓk}, to the edge connecting any two agents k and ℓ. The scalar aℓk is used by agent k to scale the data it receives from agent ℓ and similarly for akℓ. The weights {akℓ, aℓk} can be different, and one or both weights can also be zero. The network is said to be connected if paths with nonzero scaling weights can be found linking any two distinct agents in both directions. The network is said to be strongly -- connected if it is connected with at least one self-loop, meaning that akk > 0 for some agent k. In this way, information can flow in both directions between any two distinct agents and, moreover, some vertices possess self-loops with positive weights [9]. Figure 1 shows one example of a 2 strongly -- connected network. For emphasis in this figure, each edge between two neighboring agents is represented by two directed arrows. The neighborhood of any agent k is denoted by Nk and it consists of all agents that are connected to k by edges; we assume by default that this set includes agent k regardless of whether agent k has a self-loop or not. ()*+,-./,..01 .21(.0)111131 # 4)5265..71 ! & % " $ ' Fig. 1. Agents that are linked by edges can share information. The neighborhood of agent k is marked by the broken line and consists of the set Nk = {4, 7, ℓ, k}. For each agent k, the nonnegative weights it employs to scale data from its neighbors will be convex combination coefficients satisfying: aℓk = 1, aℓk = 0 if ℓ /∈ Nk (1) aℓk ≥ 0, Xℓ∈Nk Assume we collect the coefficients {aℓk} into an N ×N matrix A = [aℓk]. Then, condition (1) implies that A satisfies AT1 = 1 (2) so that A is a left-stochastic matrix. Additionally, the strong connectivity of the network implies that A is a primitive matrix [9]. One important property of left-stochastic primitive matrices follows from the Perron-Frobenius Theorem [34], [35]; it asserts that the matrix A will have a single eigenvalue at one, with all other eigenvalues lying strictly inside the unit circle. Moreover, if we let p denote the right-eigenvector of A corresponding to its single eigenvalue at one, and normalize its entries to add up to one, then all entries of p will be strictly positive, meaning that p satisfies: Ap = p, 1Tp = 1, p ≻ 0 (3) We refer to p as the Perron eigenvector of A. A. Aggregate Cost Function We associate with each agent, k, a twice-differentiable and convex cost function, denoted by Jk(w) ∈ R, with indepen- dent variable w ∈ RM . We assume at least one of these costs is strongly-convex. The agents run a collaborative distributed strategy of the following adapt-then-combine (ATC) diffusion form: where the M × M matrix quantities {Hk, Gk} correspond to the Hessian matrix of the cost function and to the covariance matrix of the gradient noise process, respectively, at agent k: 3 \∇wTJ k(wk,i−1) ψk,i = wk,i−1 − µk aℓk ψℓ,i wk,i = Xℓ∈Nk (4) (5) Hk Gk w Jk(w⋆) ∆= ∇2 ∆ = lim i→∞ E(cid:2) sk,i(w⋆)sT k,i(w⋆) F i−1(cid:3) (12) (13) the vectors where µk is a positive step-size at agent k, {ψk,i, wk,i} denote iterates at agent k at time i, and [∇wJ k(wk,i−1) represents an approximation for the true gradi- ent vector of Jk(w); an approximation is needed because, over adaptive networks, the true cost functions Jk(w) are generally not known in advance and their gradients need to be estimated continually from streaming data [9], [10]. The difference between the true gradient vector and its approximation is called gradient noise -- see (14) further ahead. Other strategies of the consensus or diffusion type are possible (e.g., [3], [4], [9], [10]), but it is sufficient to illustrate the results using (4) -- (5) [9], [36]. We represent the step-sizes as scaled multiples of the same factor µmax, namely, µk ∆= τk µmax, k = 1, 2, . . . , N where 0 < τk ≤ 1. We also introduce the vector: q ∆= diag{µ1, µ2, . . . , µN } · p (6) (7) whose individual entries, {qk, k = 1, 2, . . . , N }, are obvi- ously positive. Using the {qk}, we define the strongly-convex weighted aggregate cost: J glob,⋆(w) ∆= NXk=1 qkJk(w) (8) B. Mean-Square-Error Performance We denote the unique minimizer of (8) by w⋆ and measure the error at each agent relative to this limit point by means of the vectors: ∆= w⋆ − wk,i, k = 1, 2, . . . , N (9) It was shown in [9], [10], [37] -- [39] that w⋆ serves as a Pareto optimal solution for the network. Specifically, under some reasonable technical conditions on the cost functions and gradient noise process, it holds that (see Theorem 9.1 of [9]): lim sup i→∞ E kewk,ik2 = O(µmax) That is, all agents approach in the mean-square-error sense the limiting point w⋆ to within O(µmax) for sufficiently small step-sizes. We let MSDk denote the size of the mean-square (10) We also let MSDav denote the average mean-square deviation across all N agents. It was further shown in [9], [10], [39] that these measures are given by (see Lemma 11.3 of [9]): deviation, E kewk,ik2, in steady-state to first-order in µmax. kGk!(11) qkHk!−1 NXk=1 Tr NXk=1 MSDk = MSDav = 1 2 q2 ewk,i with the symbol sk,i used to denote the gradient noise process: sk,i(wk,i−1) ∆= \∇wTJ k(wk,i−1) − ∇wTJk(wi−1) (14) and where F i−1 denotes the filtration corresponding to all past iterates across all agents: F i−1 = filtration by {wℓ,j, j ≤ i − 1, ℓ = 1, 2, . . . , N } (15) III. WEAKLY-CONNECTED NETWORKS We now examine the same learning mechanism over weakly-connected networks. The objective is to verify whether some agents can end up having a dominating effect on the performance of other agents. Broadly, a weakly-connected network consists of a collection of sub-networks with con- straints on how information flows among them. Figure 2 illustrates one particular situation consisting of three sub- networks, with the number of their agents denoted by N1, N2, and N3, respectively. These numbers can be equal to each other or they can be different. Although Figure 2 is limited to three sub-networks, the results and analysis in the sequel are general and apply to any number of sub-networks. The figure is used for illustration purposes only. There can be many more R−type (receiving) sub-networks, as well as many more S−type (sending) sub-networks. Illustration of a weakly-connected network consisting of three Fig. 2. sub-networks. In the figure, each of the two sub-networks on top is strongly-connected and does not receive information from 1, w⋆ any other sub-network (self-loops are not indicated in the figure). Each of these sub-networks has at least one strongly- convex cost and its own combination policy, denoted by {A1 ∈ RN1×N1, A2 ∈ RN2×N2}, and its own Perron vector, denoted by {p1, p2}. Therefore, if each of these sub-networks were to run the diffusion strategy (4) -- (5), then each one of them will independently converge in the mean-square-error sense towards its own Pareto solution, denoted by {w⋆ 2}. The same figure shows a third sub-network in the bottom, and which appears at the receiving end relative to the other sub-networks. The figure shows two arrows emanating in one direction from the top sub-networks towards the bottom sub- network. Therefore, this third sub-network is influenced by the behavior of the top sub-networks, while it does not feed any information back to them. We would like to examine how the limiting behavior of this third sub-network is influenced by the two top sub-networks and whether it can still exhibit independent behavior. This is an important question with criti- cal implications for inference over networks. We will discover that in situations like these, where one or more sub-networks are at the receiving end of other sub-networks, a leader- follower relationship develops with the limiting behavior of the receiving sub-networks being completely dictated by the other sub-networks regardless of their own local information (or opinions). A. Network Model Thus, consider a network consisting of a collection of S stand-alone strongly-connected sub-networks. Each of these sub-networks does not receive information from any other sub- network and they can therefore run their diffusion strategy independently of the other sub-networks. We further assume that the network contains a second collection of R sub- networks where some agents in these sub-networks can receive information from agents in the first collection. The letters S and R are chosen to designate "send" and "receive" effects: sub-networks in group S have agents that send information to sub-networks in the receiving group R. Whenever necessary, we will use the small letters s and r as subscripts to refer to sub-networks or quantities from group S and to sub-networks or quantities from group R. If we refer to Figure 2, then S = 2 and R = 1. The total number of agents in the network is still denoted by N and it is equal to the sum of the number of agents across all sub-networks.1 We collect all weighting coefficients {aℓk} from across all edges into a large N × N combination matrix A = [aℓk]. Without loss of generality, we assume the agents are numbered with the agents from the union of all S strongly-connected sub-networks coming first, followed by the agents from the remaining R sub-networks. 4 A useful matrix decomposition result proven in [41, Ch. 8] shows that the combination matrix of every such weakly- connected network can be transformed, via a symmetric per- mutation transformation of the form P TAP , to an upper block- triangular structure of the following form (in other words, the assumed structure (16) is general enough to represent any weakly-connected graph): Subnetworks:1,2,...,S Subnetworks:S+1,S+2,...,S+R z 0 A1 0 A2 ... ... 0 0 0 ... 0 0 0 0 ... 0   0 0 ... { z ... A1,S+1 A2,S+1 } . . . . . . .. . . . . AS AS,S+1 . . . . . . .. . . . . 0 0 0 ... 0 ... 0 } A1,S+2 A2,S+2 ... AS,S+2 AS+2 ... 0 ... AS,S+R A1,S+R A2,S+R . . . . . . . .. . . . . . . AS+1,S+R . . . AS+2,S+R . .. . . . AS+R ... {   AS+1 AS+1,S+2 (16) In the above expression, the quantities {A1, . . . , AS} are the left-stochastic primitive matrices corresponding to the S strongly-connected sub-networks. We sometimes denote a generic As from this set by As,s to mean that is the combination matrix from sub-network s to itself. We denote the size of each matrix As by Ns. it Likewise, the matrices {AS+1, . . . , AS+R} in the lower right-most block contain the internal combination coefficients for the R collection of sub-networks. For example, AS+1 contains the coefficients that appear on the edges within sub- network S + 1; this matrix is not left-stochastic because it does not contain all the combination coefficients that are used by the agents within sub-network S + 1. For example, AS+1 does not contain any of the coefficients on the edges arriving to sub-network S + 1 from any of the first S sub-networks. The zero entries in the lower-left block corner refer to the fact that none of the S sub-networks receive information from the R sub-networks. The entries in the right-most upper block contain the combination weights from the edges that emanate from the S sub-networks towards the R sub-networks. For example, for the network shown in Figure 3, one possibility for A is: A =   0.2 0.5 0.3 0 0 0 0 0 0.2 0.4 0.4 0 0 0 0 0 0.8 0.1 0.1 0 0 0 0 0 0 0 0 0.4 0.6 0 0 0 0 0 0 0.3 0.7 0 0 0 0 0.2 0.1 0.3 0 0.2 0.1 0.1 0 0 0 0 0 0.3 0.5 0.2 0 0.4 0 0 0 0.2 0.3 0.1   (17) 1We remark that our definition of weakly-connected networks is more strict than the terminology used in graph theory. There, a directed graph is called weakly-connected if replacing all of its directed edges with undirected edges produces a connected (undirected) graph [40]. This definition would also include strongly-connected networks as special cases. Our definition is meant to focus on networks that are truly weakly-connected in that they induce an asymmetric flow of information among some of its components. IV. STEADY-STATE DYNAMICS We are interested in examining the steady-state behavior of each agent in the network. To do so, we will need to examine first the limit of An as n → ∞, for matrices A that have the structure (16). note that 1 TAr (cid:22) 1 T 5 (22) where the symbol (cid:22) denotes entry-wise comparisons. Actually, at least one of the entries of the row vector 1TAr must be strictly smaller than one. We then conclude that it must hold: 1 TArpr < 1 Tpr (23) with strict inequality. For this conclusion to hold, at least one entry of the vector Arpr must be smaller than the corresponding entry in pr; this fact contradicts the identity Arpr = pr. We therefore conclude that ρ(Ar) < 1 and, consequently, ρ(TRR) < 1. Furthermore, recall that each of the leading blocks {As, s = 1, 2, . . . , S} is primitive. Therefore, the powers of each of these matrices tends to [9], [34]: lim n→∞ An s = ps1 T Ns , s = 1, 2, . . . , S (24) Combining this fact with ρ(TRR) < 1, we conclude that the limit of An exists as n → ∞, and that it has the following generic form: A∞ =(cid:20) Θ 0 X 0 (cid:21) (25) for some matrix X that we would like to identify. Using the relationship A∞A = A∞ , it holds that (cid:20) Θ X 0 (cid:21)(cid:18)I −(cid:20) TSS 0 0 TSR TRR (cid:21)(cid:19) =(cid:20) 0 0 0 0 (cid:21) (26) from which we conclude that X = ΘTSR(I − TRR)−1, as claimed. We can provide a useful interpretation for the factor W that appears on the right-hand side of (19). We denote the total number of agents in group S by ∆= N1 + N2 + · · · + NS and the total number of agents in group R by NgS NgR ∆= NS+1 + NS+2 + · · · + NS+R (27) (28) Then, entries on each column of W add up to one, as can be verified from the following equivalent statements: NgS W = 1T 1T NgR (20) ⇐⇒ 1T ⇐⇒ 1T NgS TSR = 1T NgS TSR + 1T NgR(I − TRR) NgRTRR = 1T NgR ⇐⇒ 1T N(cid:20) TSR TRR (cid:21) = 1T NgR (29) where the last step holds because A is left-stochastic. More- over, since TRR is a stable matrix, we have W = TSR(I − TRR)−1 = TSR(I + TRR + T 2 = TSR + TSRTRR + TSRT 2 RR + . . .) RR + . . . (30) Recall that TSR represents the combination weights that scale the data emanating from group S and reaching group R. In the same token, TRR represents the combination weights that are internal to group R and models how the sub-networks within this group scale their internal data. As such, the first term in (30) represents the information that is transferred from group S into group R, while the second term in (30) represents how this information is transformed internally within group R after one step, and similarly for the subsequent terms in (30) Fig. 3. A weakly connected network consisting of three sub-networks and the corresponding combination policy (17). A. Limiting Power of Combination Matrix We denote the block structure of A from (16) by A ∆= (cid:20) TSS 0 TSR TRR (cid:21) (18) (19) (20) (21) where, for example, TSS is block diagonal and consists of the left-stochastic and primitive entries {A1, A2, . . . , As}, while TRR is block upper triangular. The block TSR represents the influence of the S sub-networks on the R sub-networks. Lemma 1 (LIMITING POWER OF A): Let the Perron eigen- vectors of the S strongly-connected sub-networks be denoted by {ps, s = 1, 2, . . . , S}. It holds that: A∞ △ = lim n→∞ 0 where the matrices Θ and W are defined by An =(cid:20) Θ 0 ΘW (cid:21) W ∆= TSR(I − TRR)−1 Θ ∆= blockdiag(cid:8)p11T N1, . . . , pS 1T NS(cid:9) and the notation "blockdiag" refers to a block diagonal matrix constructed from its arguments. Proof: We start by establishing that ρ(Ar) < 1 for the networks at the receiving end of information, i.e., for S < r ≤ S + R. Note that we are referring here to the block entries {Ar} on the diagonal of TRR. Using the fact that the spectral radius of any matrix is upper bounded by any matrix norm, we have that ρ(Ar) ≤ kArk1 ≤ 1 in terms of the 1−norm; the property kArk ≤ 1 follows from the fact that the columns of Ar are subsets of longer columns whose entries add up to one; refer to the example in (17). We next show, by contradiction, that strict inequality must hold, i.e., ρ(Ar) < 1. Thus, assume to the contrary that ρ(Ar) = 1. Since Ar has non-negative entries and since sub-network r is connected, it follows from the Perron-Frobenius Theorem [34], [35] that2 we can find a vector, pr, with positive entries such that Arpr = pr. Now 2Consider a matrix A consisting of nonnegative entries representing the scaling weights over the edges of a connected network. Then, in general, A may have multiple eigenvalues that attain ρ(A). The Perron-Frobenius Theorem ascertains that each of these eigenvalues will have multiplicity one, and that only one of them is real, say, denoted by λ. The theorem further ensures that there exists an eigenvector vector p with positive entries such that Ap = λp. If the graph happens to be strongly-connected, instead of only connected, then λ is the only eigenvalue that attains ρ(A). Let {w⋆ s , s = 1, 2, . . . , S} denote the Pareto optimal solutions for the strongly-connected sub-networks in group S. In order to characterize the group behavior more fully, we need to introduce a more explicit notation. Thus, consider sub-network s from this group. It has Ns agents and their individual step-sizes will be denoted by {µs,k}, with the first subscript referring to the sub-network and the second subscript referring to the agent. Likewise, the Perron vector of sub- network s will be denoted by ps with individual entries {ps,k}. The associated scaled weights are denoted by: qs,k ∆= µs,kps,k, k = 1, 2, . . . , Ns (31) According to (8), the Pareto solution, w⋆ s, that corresponds to sub-network s is the unique solution to the following algebraic equation: qs,k∇wTJs,k(w⋆ s ) = 0 (32) NsXk=1 where the {Js,k(w)} denote the cost functions that are as- sociated with agents k within sub-network s. Each agent in sub-network s will converge towards this same w⋆ s within O(µmax). This conclusion means that the limit point will be uniform within each sub-network s; though the limit points can be distinct across the sub-networks. Collecting the Pareto solutions {w⋆ s } from across the S sub-networks, we find that the limiting points for all agents within group S are described by the following extended vector: W ⋆ ∆=  1N1 ⊗ w⋆ 1 1N2 ⊗ w⋆ 2 ... 1Ns ⊗ w⋆ S  (33) Recall that the total number of agents in group S is NgS. C. Limit Points for Group R of Sub-Networks Now consider an arbitrary sub-network r from group R. It turns out that, contrary to the uniform behavior observed for group S, each agent within the sub-network r will converge to an individual limit point. This conclusion is established in the sequel (see Theorem 1) and motivated as follows. First, we recall that the total number of agents across all sub- networks in group R is NgR. We denote the limiting value for each agent k in sub-network r by w• r,k, for k = 1, 2, . . . , Nr. In this way, the collection of limit points for each sub-network r will be • ∆=  • 1 • 2 W W ... W • R  The arguments that follow will allow us to identify the limit • for the sub-networks from group R in terms vector W of the Pareto solutions of the sub-networks from group S. Specifically, it will hold that • = W T W ⋆ W (36) where W = W ⊗ IM and W is the matrix transformation defined by (20). Result (36) is established further ahead in (62) by showing that each agent in the network approaches the corresponding limit point in (37) below within a mean- square-error that is in the order of µmax. Now recall that each column of W adds up to one, which implies that if all S sub-networks happen to have the same limit point, say, w⋆, then it would follow from (36) that all agents in group R will also converge to this same limit point. More generally, if sub-networks within group S have different limit points, w⋆ s, then each agent in group R will usually converge towards a different limit point, w• r,k. One important conclusion that readily follows from (36) is that the limit points of all agents in group R are fully determined by the limit points of group S. We collect the limiting points for all agents across the network into the vector W∞ △ =(cid:20) W⋆ • (cid:21) −→ (for group S) −→ (for group R) W (37) Before justifying (36), we establish a crucial property for W∞, and which will play a key role in the convergence analysis in future sections. Lemma 2 (RELATING W∞ TO A): Introduce the extended combination policy A = A ⊗ IM . Then, it holds that W∞ is a right-eigenvector for AT corresponding to the eigenvalue at one, namely, W∞ = AT W∞ (38) Eigenvectors are non-unique. However, if we fix the top entries of the right-eigenvector of AT to W⋆, then the solution to the linear system of equations x = ATx is unique and equal to W∞. Proof: In view of (18), establishing (38) is equivalent to estab- lishing the validity of the following identity: (cid:20) W W ⋆ • (cid:21) =(cid:20) T T 0 SS SR T T T T RR (cid:21)(cid:20) W W ⋆ • (cid:21) (39) where TSS = TSS ⊗ IM , TSR = TSR ⊗ IM , and TRR = TRR ⊗ IM. Let As = As ⊗ IM for each sub-network s from group S. The top row of (39) is obviously true since the blocks on the diagonal of T T s } and each As is left-stochastic. Thus, using properties of Kronecker products, note that SS are the matrices {AT involving higher-order powers of TRR. and the collection of limit points for all R sub-networks is 6 B. Limit Points for Group S of Sub-Networks W (35) W • r = w• r,1 w• r,2 ... w• r,Nr  (sub-network r) (34) T T SS W ⋆ = blockdiagnAT s (1Ns ⊗ w⋆ s )o = blockdiag{(AT = blockdiag{(AT s s ⊗ IM )(1Ns ⊗ w⋆ 1Ns ⊗ w⋆ s )} s )} = blockdiag{1Ns ⊗ w⋆ s } = W ⋆ (40) group has a total of NgS agents, numbered 1 through NgS, and both groups have a total of N agents): 7 (51) (52) (53) (54) (55) (56) (57) (58) (group S) (50)  ew1,i ew2,i ... ewNgS ,i ewNgS +1,i ewNgS+2,i ... ewN,i ∇2 wT Jk (w⋆ (group R)  k − tewk,i))dt  NgS +1) ... ... ∇wTJ1(w⋆ 1) ∇wTJNgS (w⋆ NgS ) ∇wTJNgS +1(w⋆ ∇wTJN (w⋆ N )  ∆= diag{H 1,i, H 2,i, · · · , H NgS ,i} ∆= diag{H NgS+1,i, · · · , H N,i} ∆= diag{µ1IM , µ2IM , . . . , µNgS IM } ∆= diag{µNgS+1IM , . . . , µN IM } eWS,i eWR,i H k,i HS,i HR,i bS bR MS MR SS,i 0 ∆ =  ∆=  ∆= Z 1 ∆= − ∆= − ∆=  Using the mean-value theorem [9], it holds that: ∇wTJk(wk,i) = ∇wTJk(w⋆ sNgS ,i ∇2 (60) (59) k) − , SR,i  wTJk(w⋆ ... sN,i s1,i s2,i ... (cid:20)Z 1 sNgS +1,i sNgS +2,i ∆=   k − tewk,i)dt(cid:21)ewk,i MR · bR (cid:21) (61) MR · SR,i (cid:21) − AT(cid:20) MS · bS eWR,i−1 (cid:21) MRHR,i−1 (cid:21)(cid:19)(cid:20) eWS,i−1 0 as claimed. Next, observe that the second block row in (39) is equivalent to requiring the following equivalent conditions: W • = T T SRW ⋆ + T T RRW • ⇐⇒ T T SRW (36) ⇐⇒ T T SRW • ⋆ =(cid:16)I − T T ⋆ =(cid:16)I − T T RR(cid:17) W RR(cid:17) W T ⋆ W (41) (a) ⇐⇒ T T SRω⋆ = T T SRω⋆ where step (a) is because of equation (20). The last condition is a trivial equality and, therefore, we conclude that relation (38) holds as a result of (36). We now examine the uniqueness of the solution to (38). Suppose ∞, satisfying (38) with the same top ◦. ⋆ but with a possibly different bottom entry, denoted by W ◦ must satisfy: there exists another solution, W entry W Then, W ′ W ◦ = T T SRW ⋆ + T T RRW ◦ Subtracting this equation from the same equation satisfied by W find that (I − T T RR) (W ◦ − W •) = 0 •, we (42) However, we know from the proof of Lemma 1 that ρ(T T so that (I − T T claimed. RR) is nonsingular. We conclude that W RR) < 1 •, as ◦ = W V. STABILITY OF ERROR MOMENTS For a generic agent k, we let w⋆ k denote its limit point. We know from the discussion in the previous section that the value of this limit point will depend on whether agent k belongs to group S or group R. Specifically, we have that w⋆ k = (cid:26) w⋆ s , w• r,k, (when k ∈ sub-network s in group S) (when k ∈ sub-network r in group R) (43) We denote the error vectors relative to this limit point by: ∆= w⋆ ∆= w⋆ k − ψk,i k − wk,i (44) (45) eψk,i ewk,i If we now subtract w⋆ (4) -- (5), we get k from both sides of the diffusion strategy Substituting (46) and (49), we find that the error dynamics across the network evolves according to the following recur- sion: \∇wTJ k(wk,i−1) aℓk ψℓ,i (46) (47) eWR,i (cid:21) = AT(cid:20) MS · SS,i (cid:20) eWS,i + AT(cid:18)I −(cid:20) MSHS,i−1 It follows from property (38) that the limit point w⋆ eψk,i = ewk,i−1 + µk k − Xℓ∈Nk ewk,i = w⋆ k = Xℓ∈Nk ewk,i = Xℓ∈Nk w⋆ aℓkw⋆ ℓ aℓk eψℓ,i This result allows us to absorb w⋆ (47) so that k into the right-most term in To proceed, we recall the definition of the gradient noise process from (14) and introduce the following block quantities, which collect variables from across all agents in the network (from within both groups S and R of sub-networks; the first k satisfies: (48) (49) The next statement establishes that the entries in (37) are indeed the limit points of the agents in the network in that the agents converge towards them in the mean-square error sense. Theorem 1 (MEAN-SQUARE-ERROR STABILITY): functions Jk(w) are convex, for all Assume the cost agents k = 1, . . . , N , and that at least one of the cost functions in each of the S sub-networks is strongly convex. Assume also that the first and second-order moments of the gradient noise process (14) satisfy the conditions stated the weakly-connected in Assumption 8.1 from [9]. Then, network is mean-square stable for sufficiently small step-sizes, namely, for every k it holds that, further that the individual costs Jk(w) satisfy the following smoothness condition relative to their limits points: 8 lim sup i→∞ E kewk,ik2 = O(µmax) (62) k∇2 w Jk(w⋆ k + ∆w) − ∇2 w Jk(w⋆)k ≤ κdk∆wk (69) for small perturbations k∆wk ≤ ǫ and for some κd ≥ 0. Then, for sufficient small step-sizes, it holds that: Proof: Using the block-triangular structure of A from (18), the error recursion (61) decouples into two separate recursions with the evolution of the error vectors in group S being independent of the evolution of the error vectors in group R. In particular, it holds that SSMsSS,i − T T eWS,i = T T SS(I − MSHS,i−1)eWS,i−1 + T T SSMS bS (63) where TSS is left-stochastic. This recursion is a special case of the form studied by Theorem 9.1 in [9]. Therefore, for agents in group S, it follows from the result of that theorem that (62) holds. Moreover, the arguments used in the proof of the theorem establish that construction (33) indeed corresponds to the limiting points for the sub-networks in group S. We still need to examine the stability of the agents from group R. For these agents, we conclude from (61) that the error recursion evolves instead as follows: eWR,i = T T RR (I − MRHR,i−1) eW R,i−1 + SR (I − MSHS,i−1) eWS,i−1 + T T T T SRMS SS,i + T T T T SRMSbS − T T RRMRSR,i − RRMRbR (64) We introduce the Jordan canonical decomposition of the matrix TRR as: TRR = VǫJǫV −1 ǫ (65) where Jǫ consists of Jordan blocks except that the ones on the lower diagonal are replaced by a small positive scalar ǫ [34]. We recall from the proof of Lemma 1 that ρ(Jǫ) < 1. We also introduce the extended version of (65): TRR = VǫJǫVǫ −1 (66) where Vǫ = Vǫ ⊗ IM and Jǫ = Jǫ ⊗ IM , and define the transformed quantities ǫ eW R,i, ¯WR,i = V T ¯T T SR = V T ǫ T T SR, ¯T T RR ∆= V T ǫ T T RR (67) If we now multiply both sides of (64) from the left by V T the equivalent recursion: ǫ we obtain ¯W R,i = J T ǫ (cid:17) ¯WR,i−1 + ǫ MRHR,i−1V −T ǫ (cid:16)I − V T SR (I − MSHS,i−1) eWS,i−1 + ¯T T ¯T T SRMS SS,i + ¯T T ¯T T SRMS bS − ¯T T RRMRSR,i − RRMRbR (68) In Appendix A we start from this recursion and complete the argument to conclude that (62) also holds for all agents in group R. Besides the stability of the second-order moment of the network error dynamics, we can similarly establish the stability of the first and fourth-order moments of erros. These results are useful in evaluating the performance of the network in the next section. Theorem 2 (ERROR MOMENT STABILITY): Assume the cost functions Jk(w) are convex, for all agents k = 1, . . . , N , and that at least one of the cost functions in each of the S sub-networks is strongly convex. Assume also that the first and fourth-order moments of the gradient noise process (14) satisfy the conditions stated in Theorem 9.2 from [9]. Assume lim sup i→∞ lim sup i→∞ kEewk,ik = O(µmax) E kewk,ik4 = O(µ2 max) (70) (71) Proof: We again consider the error recursion (63), which is a special case of the forms studied in Theorems 9.2 and 9.6 in [9]. Therefore, for agents in group S, it follows from the results of these theorems that results (70) -- (71) hold. We still need to establish that (70) -- (71) hold for all agents from group R. For these agents, we start from their error recursion (68). The derivation in Appendices B and C complete the derivation and show that (70) -- (71) hold for these agents as well. VI. LONG-TERM NETWORK DYNAMICS We can be more specific and assess the size of the mean- square deviation (MSD) defined by (62) to first-order in µmax. If we examine recursion (61), we find that it is a nonlinear and stochastic difference recursion due to the dependence of the Hessian matrices HS,i and HR,i on the error vectors eWS,i and eWR,i. To address this difficulty and arrive at closed- form expressions for the MSD levels at the various agents, we replace (61) by a long-term model that will be shown to provide an accurate approximation for the behavior of the network as i → ∞ and for small step-sizes. For this purpose, we extend a construction from [9, Ch. 10] to the current setting and introduce the constant Hessian matrices: HS HR Hk ∆= diag{H1, H2, . . . , HNgS } ∆= diag{HNgS+1, HNgS+2, . . . , HN } ∆= ∇2 wJk(w⋆ k) We also define the error matrices: ∆= HS − HS,i ∆= HR − HR,i eHS,i eHR,i We then replace the original error-recursion (61) by the fol- lowing model where the Hessian matrices HS,i−1 and HR,i−1 are replaced by the constant values (72) -- (73): ′ S,i ′ (cid:20) eW eW R,i (cid:21) = AT(cid:20) MS · SS,i + AT(cid:18)I −(cid:20) MSHS MR · bR (cid:21) (77) MR · SR,i (cid:21) − AT(cid:20) MS · bS R,i−1 (cid:21) MRHR (cid:21)(cid:19)(cid:20) eW eW Observe that in this model we are using the prime notation to refer to its state variables. The next result explains why the above model provides a good approximation for the actual network performance in steady-state. ′ S,i−1 ′ Theorem 3 (ACCURACY OF APPROXIMATION): Under the same conditions of Theorem 1, it holds that, for sufficiently (72) (73) (74) (75) (76) small step-sizes the state variables of models (61) and (77) are close to each other, namely, : lim sup i→∞ E(cid:13)(cid:13)(cid:13)(cid:13)(cid:20) eWS,i eWR,i (cid:21) −(cid:20) eW eW eWR,i (cid:21)(cid:13)(cid:13)(cid:13)(cid:13) E(cid:13)(cid:13)(cid:13)(cid:13)(cid:20) eWS,i i→∞ 2 2 ′ S,i ′ R,i (cid:21)(cid:13)(cid:13)(cid:13)(cid:13) E(cid:13)(cid:13)(cid:13)(cid:13)(cid:20) eW eW = lim sup lim sup i→∞ = O(µ2 max) (78) 2 ′ S,i ′ R,i (cid:21)(cid:13)(cid:13)(cid:13)(cid:13) +O(µ3/2 max) (79) Proof: In Theorem 10.2 of [9], it was established that the error bounds in (78) and (79) hold for strongly-connected networks and, hence, for all agents of type S. Appendix D completes the argument and shows that the same conclusion holds for agents of type R. We remark that in a manner similar to the proof of Theorems 1 and 2, we can also establish that ′ k,ik = O(µmax) k,ik2 = O(µmax) k,ik4 = O(µ2 max) lim sup i→∞ lim sup i→∞ lim sup i→∞ ′ kEew E kew E kew′ (80) (81) (82) VII. MEAN-SQUARE-ERROR PERFORMANCE In this section, we evaluate the steady-state mean-square deviation (MSD), defined as the value of the error variance, E kewk,ik2, as i → ∞ and for sufficiently small step-sizes. The following result holds depending on whether agent k belongs to group S or group R. For agents in group R, we introduce the S × 1 column vector: ck = 1T N1 1T N2 ... 1T NS  [W ]:,k where the notation [W ]:,k refers to the column of W from (20) corresponding to agent k. One useful interpretation for the entries of the vector ck is as follows. Since each column of W adds up to one, therefore, the entries of ck will add up to one. In addition, as expression (85) below reveals, the square of the s−th entry of ck will measure the influence of sub-network s from group S on the performance of agent k from group R. Theorem 4 (MSD PERFORMANCE): Assume agent k be- longs to a sub-network s from group S. Then, all agents within sub-network s achieve the same MSD level given by MSDs = 1 2 Tr NsXk=1 qs,kHs,k!−1 NsXk=1 q2 s,kGs,k! (84) where we are using the notation {Hs,k, Gs,k}, with a subscript s, to denote the Hessian and covariance matrices (12) -- (13) for agent k within sub-network s, in the same manner that we used s in the definition of the weighting scalars {qs,k} in (31). Assume, on the other hand, that agent k belongs to a sub- network r from group R. Then, in this case, its MSD level will be given by MSDR(k) = SXs=1 c2 k(s) · MSDs (85) 9 where ck(s) denotes the s−th entry of vector ck. That is, the performance of any agent from group R is given by a weighted combination of the MSD performance levels of the sub-networks from group S. Proof: The argument is given in Appendix E. VIII. SIMULATION RESULTS We illustrate the theoretical results by considering several examples. Our first example relates to a pattern classification problem and illustrates how a group S of agents can dramatically influence the learning behavior of agents in another group R. Thus, consider initially a simple network consisting of N = 2 agents, where one node plays the role of the S agent and the second node plays the role of the R agent. We assume each agent is observing data that belong to different classes: data at node S belong to one of two classes {A, B}, while data at node R belong to one of two other classes {C, D}. If each agent were to learn independently of the other agent, then they would be able to determine their respective separating hyper- planes and classify their data with reasonable accuracy. Here, however, we would like to examine the influence of node S on node R when both agents are connected by means of a weak topology say, one with combination matrix -- see Fig. 4: 1 S 0.03 R 0.97 (83) Fig. 4. An example of a weakly-connected two-agent network. A =(cid:20) 1 0 0.03 0.97 (cid:21) (86) This choice for A means that agent R is careful in dealing with agent S and is only assigning weight 0.03 (or 3%) to its interaction with S. Still, despite this level of caution, the corresponding matrix W defined by (20) reduces to a scalar in this case and evaluates to W = 0.03(1 − 0.97)−1 = 1 (87) In other words, agent R will be completely misguided by agent S. This effect is evident in Fig. 5, where it is seen that agent S (top plot) is able to determine its separating hyperplane, while agent R (bottom plot) is driven to select the same hyperplane (which is obviously wrong for its data). In this simulation, both agents are running a logistic regression classifier [9], [42], [43] We consider next a more involved example based on the topology shown earlier in Fig. 3 to show that a weakly- connected network can sometimes achieve better performance than strongly-connected networks. The objective now is to determine an elliptical curve that separates the data into two classes: class +1 consists of data that are concentrated inside the curve and class -1 consists of data that are concentrated outside the curve. We assume about 10% of the data available to agents R are outliers. The outlier data belong to class +1 Agent S classification result Agent 1 classification result 10 class +1 class−1 Weakly−Connected Strongly−Connected −2 −1.5 −1 −0.5 0.5 1 1.5 2 0 x Agent 8 classification result class +1 class −1 outlier 2 1 0 −1 −2 2 1 0 −1 y y y y 2 1 0 −1 −2 −3 2 1 0 −1 −2 −3 3 Class A Class B −2 −1 0 x 1 2 Agent R classification result Class C Class D −2 −1 0 x 1 2 3 Fig. 5. Agents S and R use logistic regression to learn the sepa- rating hyperplanes for their data. The agents are weakly connected. The simulation result shows that the behavior of agent R is fully misguided by agent S even though the data received by agent R is sufficient for it to learn its own hyperplane. but are located away from the origin. Obviously, since we are dealing with a weakly-connected network, agents in group S will not be affected by these outliers since the outliers are only sensed by the R agents, which in turn do not feed information into the S agents. This situation corresponds to a scenario where weak connectivity is advantageous. We assume each agent employs the same logistic cost function: Jk(w) = ρkwk2 + Enln[1 + e−γk hT kw]o where γk represents the class label {+1, −1}, ρ is a regu- larization parameter, and the feature vector hk is chosen as follows in terms of the coordinates of the measurements [42], [43]: (88) hk(1) = 5, hk(4) = x2, hk(2) = x, hk(5) = y2, hk(3) = y hk(6) = xy The gradient vector is estimated by using the instantaneous approximation: \∇wT J k(wk,i−1) = − γ k(i)hk,i 1 + eγ k(i)hT k,iwk,i−1 + ρwk,i−1 (89) We compare the performance of the weakly-connected topol- ogy against a strongly-connected (actually, fully-connected) −2 Weakly−Connected −2 −1.5 −1 −0.5 Strongly−Connected 0.5 1 1.5 2 0 x Logistic classification result using an elliptic separation Fig. 6. curve. In the simulation, agents {k = 6, 7, 8} in Fig. 3 suffer from outlier data. The black curve represents the result obtained by a strongly-connected network, while the green curve represents the result obtained by a weakly-connected network; observe that the curve is larger in the former case due to the influence of the outliers. network with combination matrix: A = 1 8 181T 8 (90) In Fig. 6, we observe that the black elliptic curve, which is the result obtained by the strongly-connected network (90), is larger than the green elliptic curve obtained by the weakly- connected network. Comparing both boundary curves, we find that the black curve includes a larger proportion of -1 data, which will be mistakenly inferred as belonging to class +1. Obviously, weakly-connected networks do not always provide performance that is superior to fully-connected networks. There are however situations when this can occur and this is what to show in this example. We considered a case in which some R−agents are subject to outliers. In a fully-connected network, the outliers will influence the performance because their presence will be sensed by all agents. In a weakly-connected network, if these outliers are affecting only the R−agents, then the S−agents will be able to suppress their influence according to Theorem 4 and performance will therefore be improved. the simulation results are meant Our third example illustrates result (36) concerning limit points. We associate with each agent in Fig. 3 the quadratic cost: Jk(w) = E (dk(i) − uk,iw)2 (91) where agent k senses streaming data {dk(i), uk,i}; the data are related to some unknown model w◦ k via the linear regression model: dk(i) = uk,iw◦ k + vk(i) (92) where vk(i) is observation noise and assumed to be spatially and temporally white. It is sufficient for this example to assume one-dimensional limit points and to set w◦ k = 1 for the agents in sub-network 1, w◦ k = 1.5 for the agents in sub- k = 1.25 for the agents in sub-network 3. network 2, and wo It is obvious that the Pareto optimal solutions of the two S sub-networks will be w⋆ 2 = 1.5, respectively. Computing the corresponding matrix W from (17) we get: 1 = 1 and w⋆ (93) (94) W = 0 0 0 0.4046 0.5267 0.7099 0.1489 0.1183 0.0725 0.4466 0.3550 0.2176 0 0 0  ) subnetwork 1 o subnetwork 2 so that the limit points for the R agents are w• r,6 = 1.2233, w• r,7 = 1.1775, w• r,8 = 1.1088 These values are consistent with the simulation results in Fig. 7. We further illustrate in Fig. 9 the MSD performance for the agents of type S and R. Figure 8 shows the observation noise and regression power profile across the agents. In these simulations, we employed Gaussian distributed data to generate the observation noise and regression data. The step- size parameters were set to µk = 0.0005 for all agents. The theoretical MSD value for agents in sub-network 1 follows from (84) and is found to be −56.43dB. Likewise, the theo- retical MSD value for agents in sub-network 2 is found to be −52.05 dB. On the other hand, using (93), we get c7(1) = 0 + 0.5267 + 0.1183 = 0.6450 c7(2) = 0.3550 + 0 = 0.3550 (95) (96) and the theoretical MSD value for agent 7 in sub-network 3 is then found from (85) to be MSDR(7) = 0.64502 × MSD1 + 0.35502 × MSD2 = −58.49dB (97) which is consistent with the simulated result. agent 6 limit point 1.25 1.2 w 1.15 agent 7 limit point 1.1 1.05 agent 8 limit point 1000 2000 Number of iterations 3000 4000 5000 Fig. 7. Limit points for agents in group R. Noise Power for Each Node Regression Power for Each Node 11 ) B d ( r e w o P e s o N e g a r e v A i −4 −6 −8 −10 −12 −14 i r e w o P n o s s e r g e R e g a r e v A 2 4 Index of Node 6 8 5 4 3 2 1 0 2 4 Index of Node 6 8 Fig. 8. Observation noise and regression power profile for all agents. 10 0 −10 −20 −30 −40 −50 −60 ) i B d ( n o i t a v e D e r a u q S n a e M subnetwork 1 subnetwork 2 agent 7 −50 −54 −58 −62 Simulation theoretical result 4000 4200 4400 4600 4800 0 1000 2000 3000 Number of iterations 4000 5000 Fig. 9. Mean-square-deviation (MSD) performance levels. IX. CONCLUDING REMARKS In this article we examined the learning behavior of adaptive agents over weakly-connected networks. We showed that a leader-follower relationship develops among the agents, and identified the limit points towards which the individual agents converge. We also derived closed-form expressions to quantify how close the agents get to their limit points in the mean- square-error sense, and revealed how the MSD performance of follower agents is determined by the MSD performance of leader agents. We illustrated the results by considering examples dealing with pattern classification and distributed estimation. We also illustrated how weak topologies can be beneficial in reducing the impact of outlier data. APPENDIX A MEAN-SQUARE-ERROR STABILITY OF (68) If we equate the squared-norm of both sides of (68), and compute expectations, we get E k ¯WR,ik2 (a) = E(cid:13)(cid:13)J T ǫ MRHR,i−1V −T ǫ (cid:0)I − V T SR (I − MSHS,i−1) eWS,i−1 − RRMRbR(cid:13)(cid:13)2 ¯T T ¯T T SRMSbS − ¯T T SRMS SS,i + ¯T T E k ¯T T + RRMRSR,ik2 ǫ (cid:1) ¯WR,i−1 + = E(cid:13)(cid:13)(cid:13)(cid:13) 1 − t 1 − t J T ǫ (cid:0)I − V T ǫ MRHR,i−1V −T ǫ (cid:1) ¯WR,i−1 + t 1 tn ¯T T SR (I − MSHS,i−1) eWS,i−1 − 2 ¯T T SRMSbS − ¯T T E k ¯T T SRMS SS,i + ¯T T RRMRbRo(cid:13)(cid:13)(cid:13) + RRMR SR,ik2 (b) ≤ (c) ≤ ǫ 1 1 − t 2 t 2 t 2µ2 E k ¯T T SRMSbS + ¯T T ǫ MRHR,i−1V −T SRk2E kSS,ik2 + k ¯T T (cid:1) ¯WR,i−1(cid:13)(cid:13)2 + + E(cid:13)(cid:13)J T ǫ (cid:0)I − V T SR (I − MSHS,i−1) eWS,i−1(cid:13)(cid:13)2 E(cid:13)(cid:13) ¯T T RRMRbRk2(cid:17) + max(cid:0)k ¯T T RRk2E kSR,ik2(cid:1) E keWS,i−1k2 + max(cid:0)k ¯TSRk2kbSk2 + k ¯TRRk2kbRk2(cid:1) + max(cid:0)k ¯TSRk2E kSS,ik2 + k ¯TRRk2E kSR,ik2(cid:1) E k ¯WR,i−1k2 + 2k ¯TSRk2 ρ(JǫJǫ 1 − t µ2 4 t 2µ2 T) t (98) for any 0 < t < 1, where step (a) is because the gradient noise process is zero-mean and independent of all other random variables conditioned on F i−1, step (b) is because of Jensen's inequality and the sub-multiplicative property of norms, and in step (c) we used the equalities k ¯TRRk = k ¯TRRk, k ¯T12k = k ¯T12k (99) as well as for sufficiently small µmax. In the statement of the theorem, we are assuming that the gradient noise process satisfies the conditions stated in Assumption 8.1 in [9], from which we conclude that E kSS,ik2 ≤ β2 S E kSR,ik2 ≤ β2 R where S E keWS,i−1k2 + σ2 E keWR,i−1k2 + σ2 R β2 S = β2 R = σ2 S = σ2 R = max 1≤k≤NgS β2 k max NgS +1≤k≤N β2 k σ2 k NgSXk=1 NXk=NgS +1 σ2 k Now let v2 ∆= kV −T ǫ k2 and note that (102) (103) (104) (105) (106) (107) kI − V T ǫ MRHR,i−1V −T ǫ k ≤ 1 kI − MSHS,i−1k ≤ 1 (100) (101) so that 12 E k ¯WR,ik2 T) 1 − t + 2µ2 + 2µ2 maxk ¯TRRk2v2β2 R! E k ¯WR,i−1k2 + ≤ ρ(JǫJǫ (cid:18) 2k ¯TSRk2 S(cid:19) E keWS,i−1k2 + max(cid:0)k ¯TRRk2kbRk2 + k ¯TSRk2kbSk2(cid:1) + S(cid:1) max(cid:0)k ¯TRRk2σ2 For compactness of notation, we introduce the scalar coeffi- cients: R + k ¯TSRk2σ2 maxk ¯TSRk2β2 4 t 2µ2 (110) µ2 t RR = k ¯TRRk2 σ2 (111) SR = k ¯TSRk2 σ2 (112) b2 = k ¯TRRk2kbRk2 + kTSRk2kbSk2 = O(1)(113) c2 = k ¯TRRk2σ2 (114) R + kTSRk2σ2 S = O(1) Before proceeding, let us examine the spectral radius of the matrix JǫJ T ǫ . For that purpose, we note that ǫ ) ≤ kJǫJ T T) = ρ(JǫJ T ρ(JǫJǫ ǫ k1 (115) where the last inequality is because the spectral radius of a matrix is bounded by any of its norms; the last norm is the one-norm (maximum absolute column sum). We next verify that kJǫJ T ǫ k1 = ρ2(TRR) ρ(JǫJǫ T) ≤ ρ2(TRR) (116) (117) To establish (116), we proceed as follows. The eigenvalues of the matrix TRR are composed of the eigenvalues of the matrices {Ak, k = S + 1, · · · , S + R} that appear on its block diagonal. Each of these matrices is associated with a connected network and has nonnegative entries. It follows from the Perron-Frobenius Theorem [34], [35], [41] that the largest eigenvalue of each {Ak, k = S +1, · · · , S +R} is simple (i.e., its multiplicity is equal to one). Some of the largest eigenvalues among the Ak may coincide with each other, in which case the largest eigenvalue of TRR may have multiplicity larger than one. However, this largest eigenvalue of TRR cannot be defective, meaning that its algebraic multiplicity must coincide with its geometric multiplicity. This is because the matrices {Ak} occur at different block locations in TRR and the eigenvectors corresponding to their largest eigenvalues will not coincide with each other. We therefore conclude that the largest eigenvalue of TRR is semi-simple (its algebraic and geometric multiplicities are equal to each other). Thus, assume we order the eigenvalues of TRR from largest to smallest, say, as: λ1(TRR) > λ2(TRR) ≥ λ3(TRR) ≥ . . . (118) ¯WR,i−1k2 ≤ v2 E k ¯WR,i−1k2 (108) E keWR,i−1k2 = E kV −T Then, we can write ǫ E kSR,ik2 ≤ v2β2 R Substituting these expressions into (98) gives: E keWR,i−1k2 + σ2 R (109) Then, the Jordan matrix Jǫ will, for example, be of the following representative form (say, for 3 Jordan blocks of size 3 × 3 each): Jǫ = blockdiag{J1, J2, J3} (119) where It follows that J1 = diag{λ1, λ1, λ1} J2 =  J3 =  λ2 ǫ λ3 ǫ λ2 ǫ λ3 ǫ λ2 λ3   (120) (121) (122) JǫJ T ǫ = blockdiag{K1, K2, K3} (123) where, for example, K1 = diag{ρ2(TRR), ρ2(TRR), ρ2(TRR)} K2 =  λ2 2 ǫλ2 λ2 ǫλ2 2 + ǫ2 ǫλ2 2 + ǫ2  λ2 (124) (125) from which we conclude that, we can choose ǫ small enough such that kJǫJ T ǫ k1 = max{ρ2(TRR), λ2 2(TRR) + 2ǫλ2(TRR) + ǫ2} = max{ρ2(TRR), (λ2(TRR) + ǫ)2} = ρ2(TRR) as claimed. Now we can use inequality (117) to replace ρ(JǫJǫ (110) and choose t = 1 − ρ(TRR), to obtain (126) T) in E k ¯WR,ik2 ≤ SR (cid:0)ρ(TRR) + 2µ2 (cid:18) 2σ2 (cid:18) 1 − ρ(TRR) 1 − ρ(TRR) 4 maxσ2 RRv2β2 + 2µ2 maxσ2 SRβ2 b2 + 2c2(cid:19) µ2 R(cid:1) E k ¯WR,i−1k2 + S(cid:19) E keWS,i−1k2 + max Noting that ρ(TRR) < 1 is independent of µmax, we have 1 − ρ(TRR) − 2µ2 maxσ2 RRv2β2 R = O(1) (127) for sufficiently small µmax. Letting i → ∞ we conclude that lim sup i→∞ E k ¯WR,ik2 ≤ d lim i→∞ where the scalar d is defined by: SR ∆ d = (cid:16) 2σ2 1−ρ(TRR) + 2µ2 1 − ρ(TRR) − 2µ2 maxβ2 maxσ2 Since we know that lim supi→∞ we find that max) E keWS,i−1k2 + O(µ2 SR(cid:17) E keWS,i−1k2 = O(µmax), Sσ2 RRv2β2 R = O(1) (128) (129) lim sup i→∞ E k ¯WR,ik2 = O(µmax) Then, combining equation (108) with above result we conclude that lim sup i→∞ E keWR,ik2 = O(µmax) (130) 13 APPENDIX B MEAN-ERROR STABILITY OF (68) We start by taking the expectation of both sides of (68): E ¯W R,i = J T ǫ ǫ MRHR,i−1V −T E(cid:2)(cid:0)I − V T SR (I − MSHS,i−1) eWS,i−1] − E [ ¯T T ¯T T SRMSbS − ¯T T RRMRbR ǫ (cid:1) ¯WR,i−1(cid:3) + (131) where the terms involving the gradient noise are canceled out. are defined in (75) and (76). Next, we need the error quantities eHS,i−1 and eHR,i−1,which Using condition (69) and proceeding similarly to the argu- ment that established (9.280) in [9], we can verify that ǫ ǫ (132) (133) RRMRbR + ǫ MRHRV −T for some nonnegative constants κS and κR. Substituting (75) and (76) back into (131), we get: (cid:1) E ¯WR,i−1 + ¯WR,i−1(cid:17) + E ¯WR,i = J T ¯T T ¯T T SRMSbS − ¯T T J T ǫ ¯T T SR keHS,i−1k ≤ κSkeWS,i−1k keHR,i−1k ≤ κRkeWR,i−1k ǫ (cid:0)I − V T SR (I − MSHS) E eW S,i−1 − E(cid:16)V T ǫ MR eHR,i−1V −T E [MS eHS,i−1eWS,i−1] ¯WR,i−1(cid:17)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)J T ǫ kkMRkkE eHR,i−1V −T ≤ ρ(TRR)v′µmaxE keHR,i−1kkV −T ≤ ρ(TRR)vv′µmaxκRE keWR,i−1kE k ¯WR,i−1k ≤ ρ(TRR)v2v′µmaxκRqE k ¯WR,i−1k2qE k ¯WR,i−1k2 ǫ MR eHR,i−1V −T Returning to (134), we can now bound the norms of its last (a) ≤ kJ T (b) E(cid:16)V T two terms as follows: ǫ kk ¯WR,i−1k ¯WR,i−1k ǫ kkV T (e) ≤ ρ(TRR)v2v′µmaxκRE k ¯WR,i−1k2 (134) (135) (d) (c) ǫ ǫ ǫ where step (a) is due to the sub-multiplicative property of norms, step (b) uses Jensen's inequality, and introduces In addition, combining kJ T also can conclude that, v′ = kV T ǫ k ǫ k = pρ(JǫJ T (136) ǫ ) with (117), we kJ T ǫ k ≤ ρ(TRR) (137) Step (c) uses (133), step (d) uses (108), and step (e) uses (E kxk)2 ≤ E kxk2 for any random variable x. Similarly, we get: E(cid:16)MS eHS,i−1eWS,i−1(cid:17)(cid:13)(cid:13)(cid:13) ≤ σSRµmaxκS E keWS,i−1k2 For compactness of notation, we introduce the scalars: (cid:13)(cid:13)(cid:13) ¯T T (138) SR e f ∆= ρ(TRR)vv′µmaxκR = O(µmax) ∆= σSRµmaxκS = O(µmax) (139) (140) and return to equation (134) to find that: kE ¯WR,ik ≤ kJ T k ¯T T k ¯T T ǫ kkI − V T ǫ MRHRV −T ǫ kkE ¯WR,i−1k + ǫ kkE ¯WR,i−1k + ǫ kkI − V T RRMRbRk + ǫ MRHRV −T SRMSbS − ¯T T ≤ kJ T k ¯T T SRkkI − MSHSkkEeWS,i−1k + eE k ¯WR,i−1k2 + f E keWS,i−1k2 SRkkI − MSHSkkEeWS,i−1k + eE k ¯WR,i−1k2 + f E keWS,i−1k2 + µmax(cid:0)k ¯T T RRkkbRk(cid:1) ≤ ρ(TRR)kE ¯WR,i−1k + σSRkEeWS,i−1k + eE k ¯WR,i−1k2 + f E keWS,i−1k2 + µmax(cid:0)k ¯T T RRkkbRk(cid:1) SRkkbSk + k ¯T T SRkkbSk + k ¯T T (a) (141) In step (a), we used following inequality kI − V T ǫ MRHRV −T ǫ k ≤ 1 kI − MSHSk ≤ 1 (142) (143) for sufficiently small µmax and denoted the scalar coefficient here: b′ ∆= (cid:0)k ¯T T If we now let i → ∞: SRkkbSk + k ¯T T RRkkbRk(cid:1) = O(1) (144) lim sup i→∞ kE ¯WR,ik ≤ σSR 1 − ρ(TRR) e 1 − ρ(TRR) f 1 − ρ(TRR) b′ 1 − ρ(TRR) E k ¯WR,i−1k2 + kEeWS,i−1k + E keWS,i−1k2 + µmax (145) Notice that ρ(TRR) < 1 and calling upon (62), we conclude that lim sup i→∞ kE ¯WR,ik = O(µmax) (146) Lastly, similar to equation equation (108), we get following inequality, kEeWR,ik = kV −T ǫ Combining (146) and (147) we arrive at E ¯WR,ik ≤ kV −T ǫ kkE ¯WR,ik (147) lim sup i→∞ kEeWR,ik = O(µmax) APPENDIX C (148) STABILITY OF FOURTH-ORDER ERROR MOMENT OF (68) From recursion (68) we obtain: E k ¯WR,ik4 = E(cid:13)(cid:13)J T ¯T T ǫ MRHR,i−1V −T ǫ (cid:0)I − V T SR (I − MSHS,i−1) eWS,i−1 − ǫ (cid:1) ¯WR,i−1 + 14 E(cid:13)(cid:13) ¯T T 6 E(cid:13)(cid:13)J T ¯T T − ¯T T ¯T T SRMSbS − ¯T T SRMS SS,i + ¯T T + RRMRbR(cid:13)(cid:13)4 RRMRSR,i(cid:13)(cid:13)4 + ǫ ǫ MRHR,i−1V −T ǫ (cid:0)I − V T SR (I − MSHS,i−1) eWS,i−1 − ¯T T RRMRbR(cid:13)(cid:13)2(cid:13)(cid:13) ¯T T (cid:1) ¯WR,i−1 + RRMRSR,i(cid:13)(cid:13)2 SRMS SS,i + ¯T T SRMSbS (149) where we used the expansion ka + bk4 = kak4 + kbk4 + 6kak2kbk2 + 4bTakak2 + 4aTbkbk2 and where the expectation of the last two terms in the expansion are eliminated since the gradient noise is zero mean and independent of other random variables. Let us consider the last term in (149): last term T) E k ¯WR,i−1k2 + 2k ¯TSRk2 (98) ≤ 6n ρ(JǫJǫ 1 − t µ2 4 t × 2 µ2 t E keWS,i−1k2+ max max(cid:0)k ¯TSRk2kbSk2 + k ¯TRRk2kbRk2(cid:1)o E (cid:8)k ¯TRRk2kSR,ik2 + k ¯TSRk2kSS,ik2(cid:9) E keWS,i−1k2+ E keWS,i−1k2+ maxb2o × 2 µ2 E keW R,i−1k2k2 + k ¯TRRk2σ2 2σ2 SR t SRβ2 max(cid:8)σ2 E k ¯WR,i−1k2 + R + kTSRk2σ2 R S S(cid:9) (a) ≤ 6n ρ(TRR)2 1 − t µ2 4 t σ2 RRβ2 (b) ≤ 6nρ(TRR)E k ¯WR,i−1k2+ 1 − ρ(TRR)(cid:19) E keW S,i−1k2 + (cid:18) 2σSR max(cid:8)σ2 E keWS,i−1k2 + σ2 SRβ2 2µ2 S 1 − ρ(TRR) RRβ2 R 4 µ2 maxb2o× E keWR,i−1k2 + c2(cid:9)(150) In step (a), we expand the stochastic gradient noise similar to (102) -- (103), and in step (b) we choose t = 1 − ρ(TRR) and use the scalar defined (114). To emphasize the order of the coefficients, we rewrite the above equation as: + O(µ2 last term ≤nO(1)E k ¯WR,i−1k2 + O(1)E keWS,i−1k2 max)o max)E keWS,i−1k4 max)o ×nO(µ2 max)E keWS,i−1k2 + O(µ2 max)E k ¯WR,i−1k4 + O(µ2 max)E k ¯WR,i−1k2+ We consider next the second term using Jensen's inequality: = O(µ2 + O(µ4 O(µ2 max) (151) second term ≤ 8σ4 maxkSR,ik4 Referring to equation (8.122) in [9], we also conclude that maxkSS,ik4 + 8σ4 RRµ4 SRµ4 E kSS,ik4 ≤ β2 E kSR,ik4 ≤ β2 4,S 4,R where β2 4,S = max 1≤k≤NgS E keWS,i−1k4 + σ2 E keWR,i−1k4 + σ2 4,S 4,R β2 4,k = O(1) (152) (153) (154) β2 4,R = max NgS +1≤k≤N β2 4,k = O(1) σ2 4,k = O(1) σ2 4,S = σ2 4,R = NgSXk=1 NXk=NgS +1 σ2 4,k = O(1) zR,i (cid:21) = AT(cid:18)I −(cid:20) MSHS (cid:20) zS,i (156) (157) so that (155) Subtracting recursions (61) and (77), we get: 15 zR,i (cid:21) MRHR (cid:21)(cid:19)(cid:20) zS,i eHR,ieWR,i #(167) + ATM" eHS,ieWS,i SRMS eHS,ieWS,i(168) Using the eigen-decomposition (65), we can transform the above equation into. zR,i = T T RR (I − MRHR) zR,i−1 + T T SR (I − MSHS) zS,i−1 + T T ¯zR,i = J T ¯T T SR (I − MSHS) zS,i−1 + ¯T T ǫ MRHRV −T(cid:1) ¯zR,i−1 + RRMR eHR,ieWR,i + T T ǫ (cid:0)I − V T RRMR eHR,ieWR,i + ¯T T ¯zR,i ∆ = V T ǫ zR,i SRMS eHS,ieWS,i(169) (170) where we are using the notation from (67) as well as Now note that k ¯T T SRMS eHS,ieWS,ik2 ≤ (112) = (132) ≤ Similarly, we can verify that k ¯T T µ2 maxσ2 µ2 maxσ2 SRκ2 SRk2kMSk2keHS,ieWS,ik2 SRkeHS,ieWS,ik2 SkeWS,ik4 RkeWR,ik4(172) maxσ2 RRκ2 (171) k ¯T T RRMR eHR,ieWR,ik2 ≤ µ2 Therefore, following steps similar to the arguments in Appen- dices A and B we then conclude that lim sup i→∞ which is equivalent to: E kzR,ik2 = O(µ2 max) (173) i→∞ Finally, note that E keW (174) (175) ′ ′ ′ ′ max) lim sup E keW R,ik2 = E keW ≤ E keW 2E (eW E keWR,ik2 = lim sup R,i − eWR,ik2 = O(µ2 R,i − eWR,i + eWR,ik2 R,i − eWR,ik2 + E keWR,ik2 + R,i − eWR,i)∗eWR,i E keW i→∞ APPENDIX E R,ik2 + O(µ3/2 ′ ′ PROOF OF THEOREM 4 max) (176) Hence, we conclude that We know from Theorem 11.2 in [9] that the MSD expres- sion for a generic agent k in a connected network is given by: MSDk = Tr(EkX ) (177) where Ek is a block diagonal matrix with zero blocks except for an identity matrix of size M × M at the location corre- Hence, we find for the second term with simplified notation that, second term ≤ O(µ4 O(µ4 max)E keWS,i−1k4+ max)E k ¯WR,i−1k4 + O(µ4 max) (158) Lastly, we consider the first term using Jensen's inequality again: first term ≤ E k ¯WR,i−1k4 + ρ(TRR)4 (1 − t)3 8σ4 E keWS,i−1k4 + RR t3 max(cid:0)σ4 64 t3 µ4 RRkbRk4 + σ4 SRkbSk4(cid:1)(159) for any 0 < t < 1. Here, we select t = 1 − ρ(TRR) so that first term ≤ (ρ(TRR)) E k ¯WR,i−1k4 + (160) 8σ4 RR E keWS,i−1k4 + max(cid:0)σ4 (1 − ρ(TRR))3 (1 − ρ(TRR))3 µ4 64 RRkbRk4 + σ4 Recalling that ρ(TRR) < 1, we get: SRkbSk4(cid:1) first term ≤ ρ(TRR)E k ¯WR,i−1k4 + O(1)E keWS,i−1k4 + max)(cid:17)E k ¯WR,i−1k4 + E k ¯WR,ik4 = (cid:16)ρ(TRR) + O(µ2 Substituting (151), (158), and (161) back into (149), we find: O(µ4 max) (161) (162) max) Recalling ρ(TRR) < 1 again, we then let i → ∞ to find: O(1)E keWS,i−1k4 + O(µ4 lim sup i→∞ E k ¯WR,ik4 = O(1) lim sup i→∞ Since we already know that O(µ2 max), we conclude that lim supi→∞ max) E keWS,i−1k4 + O(µ4 E keWS,i−1k = (163) lim sup i→∞ E k ¯WR,ik4 = O(µ2 max) Combining the square of equation (108) with the above result we arrive at lim sup i→∞ lim sup i→∞ E keWR,ik4 = O(µ2 APPENDIX D max) (164) PROOF OF THEOREM 3 To simplify the notation, we introduce the differences zS,i zR,i ∆= eW ∆= eW ′ ′ S,i − eWS,i R,i − eWR,i (165) (166) sponding to agent k. Moreover, the matrix X is defined as: Comparing with (19), we conclude that it must hold (I − F )−1bvec(Ek) (186) We conclude that X = BnYBTn ∞Xn=0 Y = ATMSMA S = diag{G1, G2, . . . , GN } B = AT(I − MH) M = diag{MS, MR} H = diag{HS, HR} HS = diag{H1, H2, . . . , HNgS } HR = diag{HNgS+1, . . . , HN } (178) (179) (180) (181) (182) (183) (184) (185) Using the block Kronecker product notation, it can be verified that expression (177) can be equivalently expressed in the form: MSDk =(cid:0)bvec(Y T)(cid:1)T where F ∆= BT ⊗b BT (187) and bvec(Z) denotes the block vectorization operation; for a matrix Z with blocks of size M × M , this operation vectorizes each M × M submatrix of Z and the resulting vectors are subsequently stacked on top of each other. We now examine how expression (177) simplifies for both cases of agents in group S and agents in group R. To do so, we introduce the canonical Jordan decompensation of matrix A. Recall that A has S leading primitive left-stochastic matrices, {As, s = 1, 2, . . . , S}. Each of these matrices has all its eigenvalues strictly inside the unit circle with the exception of a single eigenvalue equal to one. The corresponding Perron vector is denoted by ps and has dimensions Ns × 1. We therefore conclude that the canonical Jordan decomposition of A has the form: 16 (192) (193) (194) (195) (196) (cid:20) Θ 0 so that we can set P = ΘW 0 (cid:21) = P QT p1 p2 ... 0 pS   where we are defining N1, 1T QT = (cid:2) LT LTW (cid:3) LT ∆= blockdiag(cid:8)1T ǫ (cid:21) V −1 A = V(cid:20) ISM = (cid:2) P Vx (cid:3)(cid:20) ISM J ′ N2, . . . , 1T NS(cid:9) Vy (cid:21) ǫ (cid:21)(cid:20) QT J ′ where (197) (198) (199) (200) (201) (202) Using the Jordan decomposition (196) for A we can now write: V = V ⊗ IM P = P ⊗ IM QT = QT ⊗ IM J ′ ǫ = J ′ ǫ ⊗ IM Vx = Vx ⊗ IM Vy = Vy ⊗ IM All eigenvalues of J ′ V in the form ǫ are strictly smaller than one. We partition where P is N × S and Vx collects the remaining right- eigenvectors of A. We also partition V −1 as ∆ J ′ A = V ·(cid:20) IS ǫ (cid:21) V −1 Vx (cid:3) = (cid:2) P V −1 ∆= (cid:20) QT Vy (cid:21) V (189) where (188) (190) where the matrices {QT, V −1 y } contain the corresponding left- eigenvectors and QT is S × N . We can use expression (19) for the limiting power of A to identify the matrices {P, Q} in the above decomposition. Indeed, if we compute the limit of An using (188) we conclude that BT = (I − MH)A = A − MHA = V(cid:26)(cid:20) ISM = V(cid:20) ISM − D11 −D21 ǫ (cid:21) − V −1MHV(cid:20) ISM ǫ − D22 (cid:21) V −1 −D12 J ′ J ′ D11 = QTMHP = blockdiag( N1Xk=1 = O(µmax) q1,kH1,k, . . . , NSXk=1 D12 = QTMHVxJ ′ D21 = VyMHP = O(µmax) D22 = VyMHVxJ ′ ǫ = O(µmax) ǫ = O(µmax) ǫ (cid:21)(cid:27) V −1 J ′ (203) qS,kHS,k) (204) (205) (206) (207) (208) lim n→∞ An = lim n→∞(cid:18)V ·(cid:20) IS n (cid:21) · V −1(cid:19) J ′ ǫ = P QT We then obtain that (191) F = (V ⊗b V)Z(V ⊗b V)−1 where we introduced the matrix Z ∆= (cid:20) ISM − D11 −D21 −D12 Jǫ − D22 (cid:21)⊗b(cid:20) ISM − D11 −D21 −D12 Jǫ − D22 (cid:21) (209) This relation allows us to determine a low-rank expansion for (I − F ) as follows. First note that (I − F )−1 = (V ⊗b V)(I − Z)−1(V ⊗b V)−1 (210) 17 because of the definition of Es It is therefore clear that the performance of agents in group S will be different from the performance of agents in group R. For agents in group S, from this point onwards, we can follow the same argument from Lemma 11.3 in [9] to arrive at their MSD expression. With regards to agents in group R we proceed as follows. Let (218) (219) Then, we can appeal to the derivation used in the proof of Lemma 9.5 from [9] to conclude that the entries of (I − Z)−1 are in the order of: k ⊗ IM )) or, equivalently, by using the definition of Z1: 1 bvec(ckcT x1 = Z −1 (I − Z)−1 =(cid:20) O(1/µmax) O(1) O(1) (cid:21) O(1) (211) where the size of the leading block is (SM )2 × (SM )2. Since O(1/µmax) dominates O(1) for sufficiently small µmax, we have that (I − Z)−1 =(cid:20) (cid:0)I ⊗ D11 + DT 11 ⊗ I(cid:1)−1 0 0 0 (cid:21) + O(1) (212) Noting that the Hessian matrices {Hs,k} are symmetric, we conclude that DT 11 = D11. Substituting (212) into (210), we arrive at the following low-rank approximation: (213) (I − F )−1 If we now substitute (213) into (186), we obtain: = (P ⊗b P) (I ⊗ D11 + D11 ⊗ I)−1(cid:0)QT ⊗b QT(cid:1) + O(1) (cid:0)bvec(Y T)(cid:1)T (I − F )−1bvec(Ek) = O(µ2 (P ⊗b P) Z −1 max)+ 1 (cid:0)QT ⊗b QT(cid:1) bvec(Ek) (214) (cid:0)bvec(Y T)(cid:1)T (I ⊗ D11)x1 + (D11 ⊗ I)x1 = bvec(cid:0)ckcT k ⊗ IM(cid:1) Let X1 = unbvec(x1) denote the SM × SM matrix whose block vector representation is x1. Then, the matrix X1 is the solution to the Lyapunov equation: D11X1 + X1D11 = ckcT k ⊗ IM (220) We can solve for the block diagonal entries of X1. For example, consider the s−th diagonal block of size M × M . It satisfies: NsXk=1 so that qs,kHs,k! [X1]s,s +[X1]s,s NsXk=1 k(s) NsXk=1 qs,kHs,k! = c2 qs,kHs,k!−1 [X1]s,s = 1 2 c2 k(s)IM (221) (222) Since this information is sufficient to obtain the MSD level for agent k, we do not need to compute the off-diagonal blocks of X1. where Z1 ∆= I ⊗ D11 + D11 ⊗ I (215) Returning to equation (214), and using properties of the block Kronecker product operation, we have (216) = Tr [unbvec{(P ⊗b P) bvec(X1)}Y] We can simplify the right-hand side of (214) as follows. Starting from the rightmost term we note that: (cid:0)QT ⊗b QT(cid:1) bvec(Ek) = bvec(cid:0)QTEkQ(cid:1) = bvec(cid:0)(cid:2)QTEkQ(cid:3) ⊗ IM(cid:1) = (bvec(Es ⊗ IM ) bvec(ckcT (a) (when k ∈ sub-network s in group S) k ⊗ IM ) (when k ∈ sub-network r in group R) where the matrix Ek is an S × S matrix with all zero entries except at location (k, k), where the entry is equal to one. Moreover, the vector ck is the same defined earlier in (83). In step (a), we expanded the quantity QTEkQ using the following identity: (179) [P⊗bP]bvec(X1) (cid:0)bvec(Y T)(cid:1)T = Tr(cid:2)PX1P TY(cid:3) = Tr(cid:2)X1P T(ATMSMA)P(cid:3) = Tr(cid:2)X1U TSU(cid:3) Tr SXs=1 k(s) NsXk=1 (b) = 1 2 c2 (a) where in step (a) we introduced qs,kHs,k!−1 NsXk=1 q2 s,kGs,k!(223) QTEkQ = (b) = (194) = Ek(i, j)[QT]:,i[Q]j,: NXj=1 NXi=1 ([LT]:,k[L]k,: [QT]:,k[Q]k,: [LTW ]:,k[W TL]k,: (when k ∈ group S) (when k ∈ group R) (217) U ∆= MAP (31) = q1 q2 ... 0 qS   where the notation [ · ]:,i and [ · ]j,: denotes here the the i−th column and j−th row of the matrix, respectively. Step (b) is In step (b), we exploited the fact that U TSU is block diagonal matrix. ⊗ IM (224) 18 [26] D. Acemoglu and A. Ozdaglar, "Opinion dynamics and learning in social networks," Dynamic Games and Applications, vol. 1, no. 1, pp. 3 -- 49, 2011. [27] E. Yildiz, D. Acemoglu, A. Ozdaglar, A. Saberi, and A. Scaglione, "Binary opinion dynamics with stubborn agents," ACM Trans. Econ. Comput., vol. 1, no. 4, pp. 19:1 -- 19:30, Dec. 2013. [28] X. Zhao and A. H. Sayed, "Distributed clustering and learning over networks," IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3285 -- 3300, July 2015. [29] J. Chen, C. Richard, and A. H. Sayed, "Multitask diffusion adaptation over networks," IEEE Trans. Signal Process., vol. 62, no. 16, pp. 4129 -- 4144, Aug. 2014. [30] U. Khan, S. Kar, and J. M. F. Moura, "Higher dimensional consensus: Learning in large-scale networks," IEEE Trans. Signal Process., vol. 58, no. 5, pp. 2836 -- 2849, 2010. [31] N. Abaid and M. Porfiri, "Leader -- follower consensus over numerosity- constrained random networks," Automatica, vol. 48, no. 8, pp. 1845 -- 1851, 2012. [32] D. S. Bassett, D. L. Alderson, and J. M. Carlson, "Collective decision dynamics in the presence of external drivers," Phys. Rev. E, vol. 86, pp. 036105, Sep 2012. [33] B. Liu, T. Chu, L. Wang, and G. Xie, "Controllability of a leader -- IEEE Trans. follower dynamic network with switching topology," Autom. Control, vol. 53, no. 4, pp. 1009 -- 1013, 2008. [34] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, 2003. [35] S. U. Pillai, T. Suel, and S. Cha, "The Perron -- Frobenius theorem: Some of its applications," IEEE Signal Processing Magazine, vol. 22, no. 2, pp. 62 -- 75, 2005. [36] S.-Y. Tu and A. H. Sayed, "Diffusion strategies outperform consensus IEEE strategies for distributed estimation over adaptive networks," Trans. Signal Process., vol. 60, no. 12, pp. 6217 -- 6234, 2012. [37] J. Chen and A. H. Sayed, "Distributed Pareto optimization via diffusion strategies," IEEE J. Sel. Topics Signal Process., vol. 7, no. 2, pp. 205 -- 220, April 2013. [38] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part I: Transient analysis," IEEE Trans. Inf. Thy., vol. 61, no. 6, pp. 3487 -- 3517, June 2015. [39] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part II: Performance analysis," IEEE Trans. Inf. Thy., vol. 61, no. 6, pp. 3518 -- 3548, June 2015. [40] J. A. Bondy and U. S. R. Murty, Graph Theory with Applications, Macmillan London, 1976. [41] C. D. Meyer, Matrix Analysis and Applied Linear Algebra, SIAM, PA, 2000. [42] C. M. Bishop, Pattern Recognition and Machine Learning, vol. 1, Springer, NY, 2006. [43] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Academic Press, 4th edition, 2008. REFERENCES [1] B. Ying and A. H. Sayed, "Learning by weakly-connected adaptive agents," in Proc. IEEE ICASSP, Brisbane, Australia, Apr. 2015, pp. 5788 -- 5792. [2] S. Kar, J. M. F. Moura, and K. Ramanan, "Distributed consensus algorithms in sensor networks: Link failures and channel noise," IEEE Trans. Signal Process., vol. 57, no. 1, pp. 355 -- 369, 2009. [3] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi- agent optimization," IEEE Trans. Aut. Control, vol. 54, no. 1, pp. 48 -- 61, 2009. [4] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, "Gossip algorithms for distributed signal processing," Proceedings of the IEEE, vol. 98, no. 11, pp. 1847 -- 1864, Nov. 2010. [5] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip algorithms," IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 2508 -- 2530, 2006. [6] K. I. Tsianos, S. Lawlor, and M. G. Rabbat, "Consensus-based distributed optimization: Practical issues and applications in large-scale machine learning," in Proc. Annual Allerton Conference on Commun. Control and Comput., Monticello, IL, Oct. 2012, vol. 51, pp. 1543 -- 1550. [7] S. Kar, J. M. F. Moura, and H. V. Poor, "Distributed linear parameter estimation: Asymptotically efficient adaptive strategies," SIAM Journal on Control and Optimization, vol. 51, no. 3, pp. 2200 -- 2229, 2013. [8] S. S. Stankovic, M. S. Stankovic, and D. M. Stipanovic, "Decentralized parameter estimation by consensus based stochastic approximation," IEEE Trans. Aut. Control, vol. 56, no. 3, pp. 531 -- 543, Mar. 2011. [9] A. H. Sayed, "Adaptation, learning, and optimization over networks," Foundations and Trends in Machine Learning, vol. 7, no. 4 -- 5, pp. 311 -- 801, NOW Publishers, Boston-Delft, 2014. [10] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014. [11] A. H. Sayed, S.-Y. Tu, J. Chen, X. Zhao, and Z. Towfic, "Diffusion IEEE Signal strategies for adaptation and learning over networks," Processing Magazine, vol. 30, no. 3, pp. 155 -- 171, May 2013. [12] S. Chouvardas, K. Slavakis, and S. Theodoridis, "Adaptive robust distributed learning in diffusion sensor networks," IEEE Trans. Signal Process., vol. 59, no. 10, pp. 4692 -- 4707, Oct 2011. [13] P. Braca, S. Marano, and V. Matta, "Running consensus in wireless in Proc. Int. Conf. Inform. Fusion (FUSION), sensor networks," Cologne, Germany, June-July 2008, IEEE, pp. 1 -- 6. [14] D. H. Dini and D. P. Mandic, "Cooperative adaptive estimation of distributed noncircular complex signals," in Proc. Asilomar Conference, Pacific Grove, CA, Nov. 2012, pp. 1518 -- 1522. [15] N. Takahashi and I. Yamada, "Link probability control for probabilistic diffusion least-mean squares over resource-constrained networks," in Proc. IEEE ICASSP, Dallas, TX, Mar. 2010, pp. 3518 -- 3521. [16] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, "Distributed detection and estimation in wireless sensor networks," in Academic Press Library in Signal Processing, vol. 2, R. Chellapa and S. Theodoridis, eds., pp. 329 -- 408, Elsevier, 2014. [17] A. Bertrand and M. Moonen, "Seeing the bigger picture: How nodes can learn their place within a complex ad hoc network topology," IEEE Signal Processing Magazine, vol. 30, no. 3, pp. 71 -- 82, May 2013. [18] O N. Gharehshiran, V. Krishnamurthy, and G. Yin, "Distributed energy- aware diffusion least mean squares: Game-theoretic learning," IEEE J. Sel. Topics Signal Process, vol. 7, no. 5, pp. 821 -- 836, Oct 2013. [19] K. I. Tsianos and M. G. Rabbat, "Distributed strongly convex optimiza- tion," in Proc. Allerton Conf., Allerton, IL, Oct. 2012, pp. 593 -- 600. [20] S. S. Ram, A. Nedi´c, and V. V. Veeravalli, "Distributed stochastic subgradient projection algorithms for convex optimization," J. Optim. Theory Appl., vol. 147, no. 3, pp. 516 -- 545, 2010. [21] J. B. Predd, S. R. Kulkarni, and H. V. Poor, "A collaborative training algorithm for distributed learning," IEEE Trans. Inf. Theory, vol. 55, no. 4, pp. 1856 -- 1871, Apr. 2009. [22] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging," Systems & Control Letters, vol. 53, no. 1, pp. 65 -- 78, Sep. 2004. [23] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, "Distributed asyn- chronous deterministic and stochastic gradient optimization algorithms," IEEE Trans. Autom. Control, vol. 31, no. 9, pp. 803 -- 812, Sep. 1986. [24] V. Saligrama, M. Alanyali, and O. Savas, "Distributed detection in sensor networks with packet losses and finite capacity links," IEEE Trans. Signal Proc., vol. 54, no. 11, pp. 4118 -- 4132, Oct. 2006. [25] P. Bianchi, G. Fort, and W. Hachem, "Performance of a distributed stochastic approximation algorithm," IEEE Trans. Inf. Theory, vol. 59, no. 11, pp. 7405 -- 7418, 2013. This figure "photo_sayed.jpeg" is available in "jpeg"(cid:10) format from: http://arxiv.org/ps/1412.1523v2 This figure "photo_ybc.jpg" is available in "jpg"(cid:10) format from: http://arxiv.org/ps/1412.1523v2
1904.05267	1	1904	2019-04-10T16:09:53	Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks	[ "cs.MA", "cs.CY" ]	Current demographic trends in the UK include a fast-growing elderly population and dropping birth rates, and demand for social care amongst the aged is rising. The UK depends on informal social care -- family members or friends providing care -- for some 50\% of care provision. However, lower birth rates and a graying population mean that care availability is becoming a significant problem, causing concern amongst policy-makers that substantial public investment in formal care will be required in decades to come. In this paper we present an agent-based simulation of care provision in the UK, in which individual agents can decide to provide informal care, or pay for private care, for their loved ones. Agents base these decisions on factors including their own health, employment status, financial resources, relationship to the individual in need, and geographical location. Results demonstrate that the model can produce similar patterns of care need and availability as is observed in the real world, despite the model containing minimal empirical data. We propose that our model better captures the complexities of social care provision than other methods, due to the socioeconomic details present and the use of kinship networks to distribute care amongst family members.	cs.MA	cs	MODELLING SOCIAL CARE PROVISION IN AN AGENT-BASED FRAMEWORK WITH KINSHIP NETWORKS A PREPRINT Umberto Gostoli MRC/CSO Social and Public Health Sciences Unit University of Glasgow, Glasgow, UK G2 3AX Eric Silverman MRC/CSO Social and Public Health Sciences Unit University of Glasgow, Glasgow, UK G2 3AX [email protected] [email protected] ABSTRACT Current demographic trends in the UK include a fast-growing elderly population and dropping birth rates, and demand for social care amongst the aged is rising. The UK depends on informal social care -- family members or friends providing care -- for some 50% of care provision. However, lower birth rates and a graying population mean that care availability is becoming a significant problem, causing concern amongst policy-makers that substantial public investment in formal care will be required in decades to come. In this paper we present an agent-based simulation of care provision in the UK, in which individual agents can decide to provide informal care, or pay for private care, for their loved ones. Agents base these decisions on factors including their own health, employment status, financial resources, relationship to the individual in need, and geographical location. Results demonstrate that the model can produce similar patterns of care need and availability as is observed in the real world, despite the model containing minimal empirical data. We propose that our model better captures the complexities of social care provision than other methods, due to the socioeconomic details present and the use of kinship networks to distribute care amongst family members. Keywords Social Care, Kinship Networks, Agent-Based Modelling 1 Introduction As human lifespans continue to lengthen and birth-rates drop throughout much of the developed world, many nations are experiencing an increase in demand for social care -- the provision of personal and medical care for people in need of assistance due to age, disability or other factors. In the UK, the elderly consume the largest share of social care, and lower birth-rates mean that the supply of available carers is decreasing over time even as demand is growing rapidly [1]. As a result, social care is a frequent topic in UK policy debates, with widespread concern that the country will be unable to afford the significant public investment needed to fulfill this growing care need. Age UK notes that nearly 50% of those aged 75 and over are living with a long-term limiting health condition -- and the fastest-growing age group in the UK is the over-85s, so the problem will only get worse in the coming decades [2]. As a result, social care need is rising at a pace that outstrips the growth of public and private care supply. Unmet social care need is therefore an increasingly important and widespread social problem in the UK. The 2017 Ipsos MORI report Unmet Need for Care showed that "more than half of those with care needs had unmet need for at least some of their needs" [3]. This means that less than half of those elderly individuals in need of assistance with activities of daily living (ADLs) were able to receive sufficient care. Large numbers of UK citizens are thus living without sufficient care; an estimated 1.2 million people did not receive sufficient care for their needs in 2017, an increase of 48% since 2010 [2]. The increased pressure on the social care system also has an impact on the health care system, with delayed discharges from hospital being a particularly expensive consequence of unmet care need. Patients in need of social care frequently stay in hospital longer than necessary due to a lack of care availability, leading to a shortage of beds for other patients and increased costs to hospitals. According to Age UK, between 2010 and 2016 the number of additional bed- days attributable a lack of available home-care packages increased by 181.7%, while wait times for residential care Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT placements increased by 40% [2]. In 2016 the National Audit Office estimated the total delays attributable to care shortages amounted to 2.7 million bed-days per year, resulting in an annual cost to the National Health Service (NHS) of approximately £820 million [4]. The UK is largely dependent on informal social care, or care provided free-of-charge by family members and loved ones, in order to meet the needs of the population. Informal care is enormously widespread in the UK and is much larger than the formal care infrastructure. The Family Resources Survey 2013/14 showed that there were 5.3 million informal carers in the UK [5], while projections indicate that the number of people receiving informal care will increase by 60% in the period 2015-2035 [6]. The increase in demand is such that Carers UK proposes that carer numbers would need to increase 40% over the next two decades to meet demand [7]. At the same time, the number of people using privately-funded social care is expected to rise by almost 50% by 2035, and private expenditure for social care is projected to increase from £6.8 billion to almost £20 billion in the same period, almost a three-fold increase [6]. Understanding how the need for social care evolves and the process through which informal and formal care is provided is a vital component in developing sustainable social care policies. In this regard, the importance of support and care- giving networks has long been recognized [8]. In particular, research has shown that informal care is provided mostly through care networks with an average of three to five members [9]. These networks are predominantly composed of an individual's close relatives. Aldridge and Huges [5] report that 72% of carers provide care to a member of their immediate family, whether a parent (40%), partner (18%), son or daughter (14%), while Petrie and Kirkup [10] show that 51% of carers provide care to someone in their household. Wettstein and Zulkarnain[11] show that, in 2011, 31% of informal care in the US was provided by spouses, 47% by children and 18% by other relatives (sons-in-law, daughters-in-law and grandchildren), with only 4% being provided by non-relatives. Moreover, empirical research has shown that the kind of social care provided is affected by socioeconomic status. Petrie and Kirkup [10] report that people working in routine occupations and those with lower qualifications are more likely to provide informal care. Moreover, Laing and Buisson [12] find that, while across the UK an estimated 24% of care home residents are funded through private 'top-ups', in North-East England only 18% are funded privately, as opposed to 54% in the wealthier South-East. Overall, these findings suggest that informal care becomes less common in higher socioeconomic status groups than in lower groups, while formal care becomes more common. Finally, the social care literature also reveals a significant gender gap in social care provision. Wettstein et al. [11] reported that in the US daughters are almost twice as likely (31%) to provide care as compared to sons (16%). These findings are confirmed by Petrie and Kirkup [10], who find that 59% of family carers in the UK are women. At the population level, 16% of women provide informal care as compared to 12% of men. In this paper, we propose an agent-based model of the UK informal and privately-funded formal social care system. This model reflects the complexity of a system where demographic, social and economic processes interact to deter- mine the dynamics of social care demand and supply. Our aim is to provide a theoretical framework that allows us to to improve our understanding of the mechanisms driving unmet social care need. Moreover, using ABMs enables us to model scenarios of economic and social policy change in virtual populations, providing a means to test complex policies targeted at multiple levels of society. Such models can allow policy-makers to experiment with potential pol- icy interventions and reveal any possible unintended side-effects of those policies prior to implementing them in the real world. Previous work has attempted to address the social care problem using agent-based simulation approaches [13] [14]. Here we present a simulation that significantly extends these previous efforts. The model provides a more compre- hensive simulation of social care provision behaviour, via the inclusion of a detailed socioeconomic model and the representation of care provision as not just a simple one-to-one exchange of resources, but a complex negotiation taking place between members of the care receivers' kinship networks. 2 The Model 2.1 Motivations Here we present an agent-based model as a formal representation of the complex demographic and social processes affecting informal social care demand and supply, and the dynamics of the social care outcomes resulting from their interaction over time. Our primary concern at this stage, was not the precise replication of empirical, real-world data but, rather, the development of a theoretical framework potentially capable of representing the full complexity of the social care system. In our results we aim for qualitative similarity to real-world social care trends, rather than precise 2 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT numerical replication, in order to determine if our modelling of the underlying processes is producing appropriate outcomes. With these motivations in mind, at this initial stage the behavioural assumptions made in this paper should not be considered necessarily accurate, but rather should be seen as a first approximation, i.e. a necessary point of departure allowing us to provide proof-of-concept results and demonstrate the model's potential as a policy-making tool. Further work and more specialized expertise will be needed to refine and revise the model's behavioural assumptions in order for the model to be used in this way. Ultimately, our aim is for this simulation to facilitate the development and evaluation of alternative social care poli- cies, including complex interventions aimed at behavioural change. Our addition of health care costs and a detailed economic model to the simulation can also clarify the impact of interventions on other key areas of policy, enable more robust policy evaluation, and reduce the incidence of unintended consequences derived from otherwise well-meaning policy prescriptions. Future versions of this framework can also be applied to nations other than the UK, simply by altering the simulated geography and population data. 2.2 Basics of the model The model itself is complex and contains many detailed economic and social processes and sub-processes; this section provides a high-level overview of the model's functionality. For details of the previous models which formed the initial basis for this simulation, please refer to Noble et al. [13] and Silverman et al. [14]. 1 The simulated agents occupy a space roughly based on UK geography. Agents live in houses which form towns containing clusters of up to 1225 houses. These clusters vary in size in rough proportion to real-world population density, which varies across the 8×12 grid composing the model's geography. The agent population is scaled down from real UK levels at a factor of roughly 1:10,000. The model updates in one-year time steps. Initial populations are generated and randomly distributed in the year 1860, and the model then runs until 2040 when final social care costs are calculated and recorded. We start the model in 1860 to ensure the population dynamics have time to settle before empirical population data (UK Census data) is integrated into the model in 1951. 2.3 Agent Life-Course Newborn agents are classed as dependent children until age 16, at which point they reach adulthood (the minimum working age in the UK). The agents then decide whether to continue to study or to start looking for a job. This choice is repeated every two years until the age of 24. This is a probabilistic choice that depends on the household's income level and on the parent's education level. When the agent decides to look for a job (or when it reaches age 24 at the latest), the agent then enters the workforce. During their working life, agents can be hired, fired and can change jobs. When an agent is unemployed, they find a job with a certain probability which depends on the unemployment rate (which is an input of the model), then start earning an income and paying tax. At a certain age, agents retire from the workforce (at age 65 when default parameter settings are used), at which point they stop paying tax and begin receiving a pension. This version of the simulation uses a Gompertz-Makeham mortality model to approximate mortality rates until 1951, as in Noble et al. [13]. From 1951 the simulation uses mortality rates from the Human Mortality Database [15]. After 2009 a Lee-Carter model is used to generate future mortality rates, as in Silverman et al. [14]. 2.3.1 Partnership Formation Upon reaching adulthood, agents can form partnerships with other agents 2. Employed male agents and adult female agents are randomly paired with probabilities based on the agents' socioeconomic difference, age difference and geographical distance (with the relative weight of these three factors depending on the model's parameters). Age- specific annual divorce probabilities determine whether a couple dissolves their partnership. Fertility rates follow the procedures outlined in Silverman et al. [14], in which data from [16] and the Office for National Statistics [17] are used from 1950 -- 2009, at which point Lee-Carter projections are used. 1For those who wish to examine the model more closely, or run it for yourself, you can find the annotated Python code available at https://github.com/UmbertoGostoli/ABM-for-social-care/releases/tag/v.09 2Hereafter we use 'partnership' as shorthand for relationships capable of producing children. 3 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT 2.3.2 Migration As in the real world, agents can migrate for a variety of reasons. An agent becomes independent when they leave the parental home. This can happen when they form a partnership, when they find a job in a different town from the family home, or with a certain probability when they find a job in the same town of their parental home. When a partnership dissolves, the male agent will move elsewhere on the map, while any dependent children resulting from that partnership will stay with the mother. A family will relocate to a new house if one of the two parents finds a job in another town or when the family needs a larger house due to the increased size of the family. Retired agents with social care needs may elect to move into the household of one of their children, with a probability directly proportional to the amount of care provided by that household. Rarely, dependent children will be orphaned before reaching adulthood; in that case the agent will be adopted by a household of their kinship network or by a randomly-selected couple if no such household is available. 2.4 Health status and care need Agents begin in a healthy state and have no need of additional care. They may enter different categories of care need depending on age- and gender-specific probabilities. Table 1 shows the five possible categories of care need and the amount of care required per week at each level of need. Agents who enter a state of care need do not recover and return to normal health, but instead continue to progress to higher levels of need over time. The probability of progressing to higher care need levels depends positively on the agent's age and on the discounted sum of the agent's unmet care need in past periods 3. Table 1: The different care need categories, with the number of hours of care required per week Care need category Weekly hours of care required None Low Moderate Substantial Critical 0 8 16 32 80 Social care provision is linked to informal care availability in an agent's kinship network. An agent's kinship network's nodes consist of the households of agents with a familial relationship with the agent; the degree of kinship is defined as the network distance between the household and the agent. If they have time or available income, agents will provide informal or formal care to anyone in their kinship network with care need. The amount of care agents are available to provide depends on their status (through their income), their kinship relationship with the receiver and their geographical distance from the receiver (see the Kinship Network sub-section below for details). 2.5 Model Enhancements This updated version of the Linked Lives model presented in Silverman et al. [14] has been rewritten from the ground up and substantially extended for greater detail and realism. The following features have been introduced: relationship to the agent). • The population is composed of five socioeconomic status groups. • The care supply is provided by the agent's kinship network, (a network of households which have a kinship • Households allocate part of their income to care provision, which can be in the form of both informal and • A salary function implying an inverse relationship between time taken off work to provide informal care and • Unmet social care needs affect the agents' hospitalization probability (and the associated health care costs). the agent's hourly wage. formal care. 2.5.1 Socioeconomic Status Groups The population is composed of 5 distinct socioeconomic status (SES) groups. These categories follow the Approxi- mated Social Grade, a socioeconomic classification produced by the Office for National Statistics, which is composed 3The assumption is that a prolonged period of unmet care will increase the agents' frailty and, therefore, the probability that their health will further deteriorate. 4 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT of 6 categories (A, B, C1, C2, D and E). For convenience, we redistributed category E (state pensioners, casual and lowest grade workers, unemployed with state benefits only) into the categories D (semi-skilled and unskilled manual workers) and C2 (skilled manual workers), to maintain a unimodal distribution. We initialised the groups' distribution to roughly reflect the 2016 UK distribution. The SES groups are characterized by different education levels, different career paths (represented by income growth curves) and different unemployment rates. The introduction of SES groups has a number of effects on the various stages of agent life-courses. A higher socioe- conomic position is associated with lower mortality and fertility rates, and to a lower probability of developing care need. In the marriage market, the probability that two opposite-sex individuals will form a couple depends on their SES distance (which determine the marriage probability together with the partners' geographical distance and age dif- ference)4. In the job market, a higher SES is associated with higher starting and maximum salaries (but a lower salary growth rate); a higher probability to find a job and a lower probability to be fired (probabilities which are reflected in a lower unemployment rate); and a higher probability to change jobs if the job offer comes from a different town from their current hometown. Given that in this model care supply and relocation decisions depend on income level, the socioeconomic position of an agent affects its behaviour (and that of the household it belongs to) through the agent's income, as a higher socioeconomic position is associated with a higher income level. For example, the share of income the household allocates to care supply depends positively on the household's per capita income (see the Formal Care section below). We included an inter-generational mobility process which allows agents to move to a different SES group from their parents'. Each SES group is associated with an education level. From the age of 16, an agent can decide whether to continue its studies (a choice that will allow the agent to reach a higher education level and therefore a higher SES group) or start searching for a job (in which case the agent is assigned the SES group associated with the education level reached). This choice is made by the agents every two years, until the age of 24 (i.e. at ages 16, 18, 20 and 22) 5. The probability that an agent keeps studying depends on three factors: the per capita available income (i.e, net of social care costs) of the agent's household; the difference between the maximum education level reached by the agent's parents and the agent's current education level; and the amount of time the agent allocates to informal social care supply. In this highly stylized inter-generational mobility process, an agent's SES group is determined by the education level the agent reaches6. 2.5.2 Kinship Networks Each agent is associated with a kinship network -- a network of households containing at least one agent with a kinship relation to them. This network includes: • the agent's household (distance 0); • the households of the agents' parents and of children who are not part of the agent's household (distance I); • the households of the agents' grandparents, grandchildren, brothers and sisters who are not part of the agent's household nor of the households at distance I (distance II); • the households of the agents' uncles, aunts, nephews and nieces who are not part of the agent's household nor of the households at distance I and II (distance III). The kinship network has two functions. First, the network defines the total care supply available for a particular agent in need. The care receiver's total care supply is the sum of the available care supply of all the members of all the households that are part of the kinship network which met certain conditions. These conditions represent kinship- and space-specific limitations in the supply of care. Kinship relationship and distance determine the quantity of informal care that each member of the receiver's kinship network is available to supply, as shown in Table 2: 4The relationship between SES distance and probability of forming a couple is asymmetric: the probability of getting married decreases less rapidly with the SES distance if the higher-status individual is male rather than female. 5We assume each education step lasts 2 years. The educational stages correspond to A-level, Higher National Diploma, Degree 6Meaning the SES group of agents ending their studies when they are 16, 18, 20, 22 and 24 year-old, will be, respectively, D, and Higher Degree. C2, C1, B and A. 5 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT Table 2: Amount of care agents can provide depending on their status and distance from the receiver Agent status Household (D-0) D-I D-II D-III Teenager Student Employed Unemployed 16 24 28* 32 48 0 16 20* 24 36 0 8 12* 16 20 0 4 8* 8 10 Retired *These are the minimum amounts the employed can offer, representing the informal care provided outside of the work- ing hours. If needed, additional informal care can be supplied by these agents taking time off their working hours. More- over they can use their income to pay for formal care. See the Formal Care section for details. In particular, we assume first that the amount of care which an agent is available to provide to another agent depends positively on their kinship's closeness. Agents living with the care receiver are an exception, as these agents' available care supply is assumed to be equal to that of the receiver's next of kin (i.e. spouse, children and parents), independently from the kinship degree (although most of the time the members of the receiver's household are his/her next of kin). The second factor affecting the informal care supply availability is the physical distance from the receiver's household. With regards to the physical distance we can distinguish three classes of care suppliers: • agents living in the receiver's household; • relatives living in another household in the same town as the care receiver; • relatives living in a different town. We assume that only households living in the same town of the care receiver can provide informal care. Moreover, in the case of formal care we assume that only the care receiver's household and the households at distance I (i.e., the care receiver's parents or children) are available to provide formal care. The care allocation process proceeds in a series of steps in which a 4-hour 'quantum' of care supply is transferred to an agent needing care from one of the households with available supply in its kinship network. The care alloca- tion function first samples a care receiver from the pool of people with unmet social care need who have a kinship network with available care supply. The probability of each care receiver being sampled is directly proportional to the care receivers' quantity of unmet care need. Then the allocation function samples a household from the selected care receiver's kinship network with a probability that is directly proportional to a distance-weighted measure of the household's available supply 7. Once the supplying household has been selected, a 'quantum' of care is transferred from one of the household's members with available supply to the care receiver. The member who is to provide care within the selected household is determined in two steps. First, one of the house- hold's six possible care sources is selected with a probability proportional to the residual available care of each source. The six care sources consist of the five groups that can provide informal care as shown in Table 2: teenager, student, retired, unemployed and employed; plus a sixth source which represents the amount of care which the household is available to supply by allocating part of the household income (a category we call out-of-income care). The house- hold's out-of-income care supply is the share of income that the household has available to allocate to care, either directly in the form of formal care, or indirectly in the form of informal care provided by employed household mem- bers (who provide care by taking unpaid time off work). If one of the first five sources is selected, the household member with the greater residual available care within that category will provide care. If the out-of-income category is selected, the household will provide formal care if the lowest hourly wage among the hourly wages of the employed household members is higher than the hourly cost of social care. Otherwise, the employed household member with the lowest hourly wage will provide the 'quantum' of care8. At the end of this step, both the care receiver's care need and the supplying household's availability are reduced by the amount of care transferred (i.e. the 4-hour 'quantum' of care). The allocation process is repeated until the set of care re- 7For an household at kinship network's distance x from the care receiver, the weight is the reciprocal of the exponential function of the product of x and a parameter. 8Given that the quantity of out-of-income care supply depends on the household's per-capita income, this care allocation mech- anism implies that the probability that members of the households will spend time providing informal care is inversely related to the household's per-capita income. In other words, the wealthier the supplying household, the more likely that it will provide formal care 6 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT ceivers with unmet care need and available care supply is empty (i.e. there are no more care receivers with outstanding care need and available care supply). The ultimate result of this care allocation function is that units of social care are distributed from potential providers to receivers across the receiving agents' kinship network, and decisions about who provides care and choosing informal or formal care provision are made according to the composition, socioeconomic position and employment status of potential care providers. We suggest this detailed decision-making process better represents the complexities of care decisions that families need to navigate. The specifics of this complex process can be adjusted further by incorporating insights received from qualitative and quantitative data on informal care-givers and their decision-making. The kinship network also allows households to compute the informal care attraction associated with each town, which represents a rough measure of the informal care that a household expects to get or supply in a given town. The households use this town-specific attraction to make relocation decisions 9. More precisely, for a particular household Hi, the informal care attraction of a town T , is a function of the sum of the members of the households with a kin relationship with H, living in the town T 10. The contribution of a household Hj to the social care attraction is weighted according to the degree of kinship between Hi and Hj. This weighted sum is then multiplied by the complement to one of the share of care provided by the government. Two assumptions characterise the towns' informal care attraction: first, the larger the local kinship network in a town (in terms of number of people who are part of it), the higher the amount of social care the agent can expect to receive (or to supply) and the higher the social care 'value' associated with a town; second, the higher the share of care supplied by the government, the lower the importance of the potential informal care in the relocation decision. Apart from this 'network size' factor, a household's probability of relocating depends on the relocation cost, which increases with the household's size and the number of years the household's members lived in the current town 11, and the town's homophily attraction, which depends on the town's share of people belonging to other SES groups. Towns with a larger share of unoccupied houses are more likely to be chosen for relocation. 2.5.3 Formal Care Both informal and formal care are allocated through the care recipient's kinship network. Each household allocates a share of its income to care, which increases with the household's per capita income 12. We assume that formal care can be provided only by the care receiver's own household or from the households in the care receiver's kinship network with distance equal to 113. As explained above, the choice between the informal and formal care is stochastic, with probabilities equal to the relative availability of the two kinds of care. In order for the household to supply formal care, the hourly wage of the working member with the lowest wage (and available time left) must be higher than the price of formal social care. If the hourly wage of this member is lower than the price of formal social care, they will take time off work to provide informal care (as in this case it is cheaper for the household to give up the agent's salary than pay for formal care). However, the supplying household can provide informal care only if it is in the same town of the care receiver, otherwise it will provide only formal care. 2.5.4 Salary function. We model the hourly salary that an agent receives (on average) when it finds a job using the Gompertz function. This function is a double exponential that takes the following three SES-specific arguments: 9In this model, agents face the choice of relocating to other towns either if they receive a job offer or partner with someone from another town. 10The household's kinship network is obtained by joining the kinship networks of the household's members. 11The relocation cost R can be thought of as a measure of the social capital developed by the household in their current town. This social capital would be lost by relocating to another town, so it acts as a barrier to relocation. Formally, relocation cost is computed as: n(cid:88) i=1 R = K yp i where n is the set of the household's members, yi is the number of years the household member i spent in the current town and p is a parameter with value smaller than 1 (as additional years will increase the member's social capital by increasingly smaller amounts). K is a scaling parameter. 12More precisely, with x being the household's per-capita income, the share allocated to care supply is the complement to one of the reciprocal of the exponential function of the product of x and a parameter. 13In other words, only the care receiver's parents and children are available to pay for formal care if they cannot provide informal care. 7 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT • the initial salary level; • the final salary level; • the salary growth rate. The salary growth rate is multiplied by the agent's cumulative work experience, which is the discounted sum of all the fractions of the working week allocated to work (if an agent works full time, this fraction is equal to 1) 14. This equation implies that if an agent takes time off work to provide informal care, this will result in less work experience and, therefore a lower hourly salary. Given the properties of the care allocation mechanism, the agents employed with the lowest hourly salary will be more likely to provide informal care in the future (if there are not retired/students/unemployed agents or if the household's retired/students/unemployed agents do not have available informal care supply). In this model new mothers devote their time to caring for their newborn, meaning that we see a gender pay gap emerge due to the interaction between: • the allocation of part of the household's income to care supply; • the choice of the care giver within the supplying household; • the salary function. Retired agents receive a pension which is proportional to their final income level. We assume that when an agent needs care they retire due to sickness and thereafter receive a pension. If their care need is low or moderate, their ill-health pension is their normal pension reduced in accordance with the ratio between lost working years and the maximum number of working years (in other words, the working years of a person in good health). If the agent's care need is substantial or critical, the ratio is computed by reducing the number of lost working years by 50%. 2.5.5 Hospitalisation We assume that agents with care need will spend additional time in the hospital, time whose duration is a function of the agents' care need level and the average discounted share of unmet care need. The higher the agent's care need level and the higher her average share of unmet care need, the greater the number of days the agent is expected to spend in the hospital each year. By multiplying the sum of the hospitalisation's duration by the hospitalisation cost per day, we can determine the cost of unmet care need for the public health service. 2.5.6 Simulation steps More specifically, the simulation unfolds through the following eleven steps: 1. deaths: with a given probability which depends on age, SES and care need level, some agents are removed from the population; 2. adoptions: children without parents are adopted; 3. births: with a probability which depends on age and SES, married woman give birth to new agents; 4. divorces: some couples dissolve (and the male relocates); 5. marriages: with a certain probability which depends on age, SES and geographical location, some singles get married and go to live together. 6. social care allocation: social care is transferred from agents with available hours (or income) to agents with social care needs; 7. age transitions: the age of the agents is incremented, and their age-related status is updated. 8. social transitions: students decide whether to start working or keep studying, depending on their family income per capita, their parents' SES and their care responsibilities. In case they start looking for a job, they are assiged the SES associated to the education level they have reached. 14Formally, the salary function is: where: w = F ece−rt c = ln I F and F is the maximum hourly wage, I is the initial hourly wage, r is the wage growth rate and t is the discounted cumulative work experience. 8 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT 9. job market: employed and unemployed agents receive new job offers (and eventually accept them) and em- ployed agents are fired, with probabilities which depend on the age and SES-specific unemployment rate. 10. relocations: agents who have accepted job offers from towns different from their current town, relocate. 11. care transitions: with a probability which depends on age, SES, current care need level and unmet care need, agents pass to higher care need levels or are hospitalised. 3 Social Policy Experiments ABM allows us to investigate the effects of virtual social or economic policies. This model is characterised by various policy-related parameters, the value of which can be directly or indirectly related to specific measures of social care policy. Therefore, by varying these parameters, we can examine 'what if' scenarios and simulate social care outcomes and costs under alternative social care policies. For illustrative purposes, we investigated the effect on social care of two policies: the introduction of tax-deductible social care expenses, and direct governmental funding of care for people above a certain level of care need. For the first policy, we assume that the carers are allowed to deduct 100% of social care expenses from their tax. In the second policy experiment, we assume that the government directly pays for the social care expenses of those agents with the highest social care need (i.e., care need level at 'critical', see Table 1). We assume that the two policies are implemented from simulation year 2020 and compare the outputs of these two policy scenarios with the benchmark no-policy scenario in the period 2020-2040. We present two results: the hours of unmet care need in each scenarios and the incremental cost-effectiveness ratio (ICER) of the two policies considered 15. Figure 1: Hours of informal and formal care received and of unmet care need in the simulated population. Figure 2: Hours of informal and formal care received and unmet care need per recipient. 4 Results As described above, this simulation is highly complex with numerous processes at play, and a number of parameters governing system behaviour. For these early-stage results, we present figures from a representative single run at default parameter values plus a comparison of this run, taken as the benchmark, with two simulations where we introduce two alternative policy interventions in the year 2020. The single-run charts displayed here were chosen to highlight the key features of this new simulation framework, namely the modelling of formal care, additional economic and labour 15In this paper we formally define the ICER as: e = Cp − Cb Ub − Up where C is the total additional cost of policy implementation, U is the total unmet care need and the subscripts p and b indicate two scenarios (i.e. the policy scenario and the benchmark scenario, respectively). In this case, being b the no-policy scenario, Cb is equal to 0. 9 19902000201020202030204001000020000300004000050000Hours per weekSocial Care by Type and Unmet Care NeedsInformal CareFormal CareUnmet Care Needs1990200020102020203020400.02.55.07.510.012.515.017.5Hours per weekDelivered and Unmet Social Care Per RecipientInformal CareFormal CareUnmet CareModelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT market details such as socioeconomic status groups, interaction between social care need and health care demand, and agent kinship networks. Figure 1 shows the evolution of hours of informal and formal care delivered and of unmet social care need per week for the whole population over the period 1990-2040. In our simulations, the total population reaches a peak of around 10,000 agents in the period 2025, and then decreases slowly to around 8000 by 2040. As in Bijak et al. [18], the simulation's population projections roughly match those of the Office for National Statistics in the UK if we assume no international migration. Figure 3: Hours of informal and formal care received and unmet care need per recipient by care need level. Figure 4: Hours of informal and formal care received and unmet care need per recipient by SES group (with I being the poorest SES group). Figure 5: Informal care supplied by women as share of total informal care supplied, for the population. Figure 6: Ratio of women's to men's income, for the population as a whole. In Figure 1 we can see that from 2010 informal care cannot keep pace with the growth in care need, with a consequent rapid increase of unmet care need, which reaches a peak in 2025. In Figure 2 we show the weekly hours of informal and formal care delivered and the unmet social care need per recipient. Here we can see that the negative trend of the unmet care need after 2025 in Figure 1 is entirely due to the decrease of population: in fact, the unmet care need per recipient keeps growing quite steadily up to 2033. Figure 3 shows the mean informal and formal care received and the unmet care need by care need level over the period 2020-2040 (note that the heights of the bars correspond to the weekly hours of care required as shown in Table 1). We can see that most of the unmet care need is due to the agents with the highest care need. Figure 4 shows the mean informal and formal care received and the unmet care need by SES group (with I being the poorest SES group). We can see that the relative weight of informal care with respect to formal care decreases from the poorest to the wealthiest SES group, a result which reflects empirical findings. As expected, by comparing the green bars in Figure 4, we can 10 NL 1NL 2NL 3NL 401020304050607080Hours per weekInformal/Formal/Unmet Social Care by Care Need LevelInformal CareFormal CareUnmet Care NeedsSES-ISES-IISES-IIISES-IVSES-V0510152025Hours per weekInformal, Formal and Unmet Social Care Need per RecipientInformal CareFormal CareUnmet Care Needs1990200020102020203020400.600.650.700.750.800.850.900.951.00Share of careShare of Informal Care supplied by Women1990200020102020203020400.850.900.951.001.05Income RatioWomen and Men Income RatioModelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT see that the poorest SES group has a somewhat higher level of unmet care need compared to the wealthiest group but also that unmet care need is a significant share of total care need across SESs. Figure 5 shows the share of total informal care provided by women. At the population level this figure starts from just below 75% in 1990, then decreases steadily until 2020, when it starts fluctuating just above 60%, a value which is in line with empirical findings. Moreover, there is a marked difference between SES groups, with the poorest SES group having the highest (and growing) gender inequality. As shown in Figure 5, this inequality in care provision affects income inequality between genders. At the population level, female agents' income is about 5% lower than the male's income. Therefore our model can explain at least part of the gender pay gap. Figure 7 shows the total informal and formal care provided in the period 2020-2040 according to the kinship distance between caregiver and receiver: household (distance 0); parents and children (distance 1); grandparents, grandchildren and brothers (distance 2); aunts, uncles and nephews (distance 3). We can see that most social care is supplied within the household, however a significant part comes from outside the household, especially formal care. Figure 8 shows annual per capita health care cost due to the hospitalization of people with care needs. We can see that although the total unmet care need levels off after 2020 (as shown in Figure 1), the per capita health care cost keeps growing until the late 2030s. This is due to the decrease in population after 2015, resulting in the per capita health care cost reaching its peak 15 years after the unmet social care need. Figure 7: Total informal and formal care supplied by different groups of receivers' suppliers: household, parents and children (I), grandparents, grandchildren and brothers (II) and aunts, uncles and nephews (III). Figure 8: Per capita health care cost due to the hospi- talisation of people with social care needs (per year). In the last two charts we investigate and compare the efficacy of two policy interventions. These two policy exper- iments are illustrative examples that show how this model may be used as a tool for the design and evaluation of social care policies. In these examples we consider a tax deduction policy, in which the households are allowed to deduct all the social care expenses from their tax base, and a direct funding policy, in which the social care needs of people with the highest need level (the 'critical' level in Table 1) are taken care of directly by the state. The policies are implemented starting in simulation year 2020, and the two scenarios are compared to the benchmark no-policy scenario. In Figure 9 we can see that the two policies have a positive impact on total unmet social care need, as expected. However, while the tax deduction policy has a marginal effect, decreasing the total unmet care need by about 11%, the direct funding policy has a drastic effect, reducing the total unmet care to 23% of the original level. Although the tax deduction policy has a smaller effect on the unmet social care need, Figure 10 shows that the tax deduction policy is more cost-effective than the direct funding policy. This is because, while the effect of direct funding on total unmet social care is seven times the effect of the tax deduction, the cost of the former policy compared to the latter is much higher, as shown in Figure 11. We note however that notwithstanding its cost-effectiveness, the effect of the tax deduction policy on unmet care demand is limited and thus it is not a solution to the growing social care demand on its own. Finally, Figure 12 shows the discounted value of the hospitalization costs in the period 2020-2040 in the three policy scenarios. We can see that, apart from the direct costs associated with each policy, our model allows us to estimate the policies' spillover effects (in this case, the effect 11 HouseholdD-ID-IID-III020004000600080001000012000Hours per weekInformal and Formal Care per Kinship LevelInformal CareFormal Care199020002010202020302040050100150200250300350Pounds per YearPer Capita Health Care CostModelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT Figure 9: Unmet care need per recipient in the no- policy scenario and with the two alternative social policies. Figure 10: Cost-effectiveness of 'tax deduction' and 'direct funding' policies. on hospitalization cost), effects which we need to take into account to assess the relative advantages and disadvantages of alternative policies. Figure 11: Unmet care need per recipient in the no- policy scenario and the two alternative social policy scenarios. Figure 12: Cost-effectiveness of 'tax deduction' and 'direct funding' policies. 5 Discussion While the results presented here are still early, the outcomes of these simulation runs suggest that this model can produce broadly realistic portraits of the coming trends in UK social care. The overall population dynamics largely mirror those in evidence in the real world, with the notable exception of international migration which is not modelled here. The simulation also produces inequalities in care provision by both SES and gender; while the data presently available is not sufficient to determine if the socioeconomic inequalities are accurate, the gender pay gap is reflective of current UK norms. The simulation results on unmet care need lend credence to the worries expressed by UK policy-makers. The simula- tion shows that unmet care need will continue to grow over time, and given that 1.2 million older people in England alone are not receiving sufficient care [2], further growth in these figures could lead to severe consequences for pub- lic health. The results also show marked inequalities in care provision, with the wealthy capable of supporting their aged relatives in need through privately-funded care while still staying in work and receiving high wages. Among the lower-income groups women are providing an overwhelmingly larger share of informal care as compared to men, 12 Tax DeductionDirect FundingBenchmark0100000200000300000400000500000Hours of Unmet CareTotal Unmet Social Care NeedTax DeductionDirect Funding0.02.55.07.510.012.515.017.520.0ICERICER based on unmet care needTax DeductionDirect FundingBenchmark0.00.20.40.60.8Pounds per week1e7Total public expenditure per weekTax DeductionDirect FundingBenchmark0.00.51.01.52.02.53.03.54.0Cost in Pounds1e7Total Discounted Hospitalization CostModelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT meaning that women at the lower end of the socioeconomic scale are more likely to be pushed out of work and toward unpaid care provision. Our use of kinship networks as the mechanism for distributing care illustrates that kinship distance impacts care provision behaviour, and thus is an important aspect of care decision-making that should be taken into account in future research. We observed a significant difference in caring behaviours between within-household and kinship distance I agents, with the latter providing much more formal care than informal. This suggests that future models in this area should consider kinship networks and their impact on caring behaviours. Understanding the negotiation process within families regarding care provision will be an important aspect for policy-makers to examine as well, as policies put into place to support and encourage informal care may need to take account of these complex social aspects of care. Finally, the speculative policy scenarios we provide here demonstrate the efficacy of this platform as an aid to policy- makers who wish to examine the impact and possible unintended side-effects (spillover effects) of their planned policy interventions. The scenarios showed that while some policies such as tax deductions may seem like an easy 'win' for the policy-maker, the actual impact on social care need is minimal. For significant reductions in care need and related health care costs, more expansive -- and expensive -- interventions are needed. The direct funding scenario here is a relatively simplistic example used to demonstrate this point; the model is capable of simulating the results of much more nuanced policy interventions, due to the detailed modelling of key social care mechanisms and related economic and social processes. 6 Future Work While the current simulation does broadly reflect the expected trends in social care provision, validating the results is difficult. Ongoing projects in Scotland are linking administrative data sources to develop a clearer picture of social care provision and receipt across the country. In future revisions of the model we intend to make use of these data by focusing the simulation on Scotland rather than the whole UK. As the framework matures and incorporates more real- world data, we will use Gaussian process emulators to perform detailed sensitivity analyses, following the example of Silverman et al. [14]. Apart from the validation process, this model can be expanded in various directions. First, according to the Office for National Statistics the proportion of informal childcare to GDP increased from 13.8% to 17.6% in the period 2005- 2014 [19]. Child care represents a significant part of the total household's informal care need and, therefore, it affects the care supply which can be allocated to social care needs. Our next planned update to the model will include a child care mechanism which will allow us to model the interaction with social care supply. Second, according to the Office for National Statistics, in 2016, 28.2% of births in England and Wales were to women who were not born in the UK [20]. Moreover, projections show that post-2016 immigration accounts for 77% of total population growth until 2041 [21]. The addition of international migration to the model will allow us to understand UK demographic dynamics in future decades. Following on from the simplistic policy comparison presented here, in future work we will consult with social care policy-makers in Scotland to model proposed social care policies in more detail. Social care policy is very complex, with the programmes available varying not only by region (England, Wales, Scotland and Northern Ireland), but by individual local councils. The model is capable of representing these details by varying relevant policy parameters across the simulated UK geography. Once we are able to replicate the current state of social care policy in the UK, we will then be able to develop detailed evaluations of the possible impacts of new policy interventions. Our model's ability to capture detailed individual-level care decision-making will enable policy-makers to examine policies aimed at behavioural change as well as broader economic and social policies. Finally, while the current simulation is very UK-centric, the core simulation engine can be altered easily to examine the situation in other countries. Ultimately we hope to produce a simulation framework that is capable of modelling informal and formal social care across a variety of cultural and economic contexts. 7 Data Availability The code is available on GitHub: https://github.com/UmbertoGostoli/ABM-for-social-care/tree/v.09 The output data used for the figures is available in this Dryad repository: 13 Modelling Social Care Provision in An Agent-Based Framework with Kinship Networks A PREPRINT https://datadryad.org/review?doi=doi:10.5061/dryad.6tm6183 8 Authors' Contributions Umberto Gostoli developed the model (based on a previous version, developed by Eric Silverman), run the simulations, gathered the data and produced the charts. Eric Silverman conceived the research and helped develop the model. Both authors contributed to draft the paper. All authors gave final approval for publication. 9 Funding The authors are part of the Complexity in Health Improvement Programme supported by the Medical Research Council (MC UU 12017/14) and the Chief Scientist Office (SPHSU14). 10 Acknowledgements We would like to thank Chris Patterson and Lauren White from the MRC/CSO Social and Public Health Sciences Unit for their helpful comments. References [1] Coleman DA. 2002. Replacement migration, or why everyone is going to have to live in Korea: a fable for our times from the United Nations. Philos Trans R Soc London B Biol Sci 357. [2] Age UK. 2017. The Health and Care of Older People in England 2017. London, UK: Age UK. [3] Ipsos MORI. 2017. Unmet Need for Care, Report 15-042098-01. London, UK: Ipsos Public Affairs. [4] National Audit Office. 2016. Discharging older patients from hospital. London, UK: National Audit Office. [5] Aldridge H, and Hughes C. 2016. Informal carers and poverty in the UK. London, UK: New Policy Institute. [6] Wittenberg R, Hu B. 2015. Projections of demand for and costs of social care for older people and younger adults in England, 2015 to 2035. London, UK: London School of Economics. [7] Carers UK. 2015. Facts About Carers - Policy Briefing. London, UK: Carers UK. [8] Keating N, Otfinowski P, Wenger C, Fast J, Derksen L. 2003. Understanding the caring capacity of informal networks of frail seniors: a case for care networks. Ageing Soc 23, 1. [9] Tennstedt SL, McKinlay JB, Sullivan LM. 1989. Informal care for frail elders: The role of secondary caregivers. Gerontologist 29, 5. [10] Petrie K, Kirkup J. 2018. Caring for carers. London, UK: The Social Market Foundation. [11] Wettstein G, Zulkarnain A, et al. 2017. How Much Long-Term Care Do Adult Children Provide? Issue in Brief. [12] Laing and Buisson. 2016. Care of Older People: UK market report. London, UK: LaingBuisson. [13] Noble J, Silverman E, Bijak J, Rossiter S, Evandrou M, Bullock S, Vlachantoni A, Falkingham J. 2012. Linked lives: the utility of an agent-based approach to modelling partnership and household formation in the context of social care. Proceedings of the 2012 Winter Simulation Conference. Eds: C. Laroque, J. Himmelspach, R. Pasupathy, O. Rose and J.M. Uhrmacher. Publisher: IEEE. [14] Silverman E, Hilton J, Noble J, Bijak J. 2013. Simulating the cost of social care in an ageing population. Proceedings of the 27th European Conference on Modelling and Simulation. Eds: W. Rekdalsbakken, R.T. Bye, H. Zhang. Publisher: Digitaldruck Pirrot. [15] Human Mortality Database 2011. url: http://www.mortality.org/cgi-bin/hmd. Accessed 26/07/2011 [16] Eurostat Statistics Database: Domain Population and Social Conditions. url: http://epp.eurostat.ec.europa.eu. Accessed 27/10/2011 [17] Office for National Statistics 1998. Birth Statistics, Series FM1 (27). London, UK: Office for National Statistics. [18] Bijak J, Hilton J, Silverman E, Cao VD. 2013. Reforging the Wedding Ring: Exploring a Semi-Artificial Model of Population for the United Kingdom with Gaussian process emulators. Demogr Res 29, 27 [19] Office for National Statistics 2016. Chapter 2: Home produced childcare services. London, UK: Office for National Statistics. [20] Office for National Statistics 2017. Births by parents country of birth, England and Wales: 2016. London, UK: Office for National Statistics. [21] Cangiano A. 2018. The Impact of Migration on UK Population Growth. Oxford, UK: University of Oxford. url: https://migrationobservatory.ox.ac.uk/resources/briefings/the-impact-of-migration-on-uk-population-growth/ 14
1905.13275	1	1905	2019-05-30T19:45:06	New Algorithms for Functional Distributed Constraint Optimization Problems	[ "cs.MA" ]	The Distributed Constraint Optimization Problem (DCOP) formulation is a powerful tool to model multi-agent coordination problems that are distributed by nature. The formulation is suitable for problems where variables are discrete and constraint utilities are represented in tabular form. However, many real-world applications have variables that are continuous and tabular forms thus cannot accurately represent constraint utilities. To overcome this limitation, researchers have proposed the Functional DCOP (F-DCOP) model, which are DCOPs with continuous variables. But existing approaches usually come with some restrictions on the form of constraint utilities and are without quality guarantees. Therefore, in this paper, we (i) propose exact algorithms to solve a specific subclass of F-DCOPs; (ii) propose approximation methods with quality guarantees to solve general F-DCOPs; and (iii) empirically show that our algorithms outperform existing state-of-the-art F-DCOP algorithms on randomly generated instances when given the same communication limitations.	cs.MA	cs	New Algorithms for Functional Distributed Constraint Optimization Problems Khoi D. Hoang1 , William Yeoh1 , Makoto Yokoo2 and Zinovi Rabinovich3 1Washington University in St. Louis, USA 2Kyushu University, Japan 3Nanyang Technological University, Singapore {khoi.hoang, wyeoh}@wustl,edu, [email protected], [email protected] Abstract tool formulation is a powerful The Distributed Constraint Optimization Problem (DCOP) to model multi-agent coordination problems that are distributed by nature. The formulation is suitable for problems where variables are discrete and constraint utilities are represented in tabular form. However, many real-world applications have variables that are continuous and tabular forms thus cannot accurately represent constraint utilities. To overcome this limitation, researchers have proposed the Functional DCOP (F-DCOP) model, which are DCOPs with continuous variables. But existing approaches usually come with some restrictions on the form of constraint utilities and are without quality guarantees. Therefore, in this paper, we (i) propose exact algorithms to solve a specific subclass of F-DCOPs; (ii) propose approximation methods with quality guarantees to solve general F-DCOPs; and (iii) empirically show that our algorithms outperform existing state-of-the-art F-DCOP algorithms on randomly generated instances when given the same communication limitations. 1 Introduction [Modi et al., 2005; Constraint Optimization The Distributed Problem Petcu and Faltings, 2005] (DCOP) formulation is a powerful tool to model cooperative multi- agent problems. DCOPs are well-suited to model many problems that are distributed by nature and where agents need to coordinate their value assignments to maximize the aggregate constraint utilities. This model is widely employed to model distributed problems such as meeting scheduling problems [Maheswaran et al., 2004], sensor and wireless [Farinelli et al., 2008; Yeoh and Yokoo, 2012], networks [Zivan et al., 2015], multi-robot [Kumar et al., 2009; Miller et al., 2012; smart Fioretto et al., 2017b], genera- tion [Ueda et al., 2010] and smart homes [Rust et al., 2016; Fioretto et al., 2017a]. coordination structure coalition teams grids However, the regular DCOP model assumes that the vari- ables are discrete and the constraint utilities are represented in tabular form (i.e., a utility is defined for every combina- tion of discrete values of variables). While these assumptions are reasonable in some applications where values of vari- ables correspond to a set of discrete possibilities (e.g., the set of tasks that robots can perform in multi-robot coordination problems or the set of coalitions that agents can join in coali- tion structure generation problems), they make less sense in applications where values of variables correspond to a con- tinuous range of possibilities (e.g., the range of orientations a sensor can take in sensor networks or the range of frequencies an agent can choose in wireless networks). These limiting assumptions have prompted Stranders et al. [2009] to extend the DCOP formulation to allow for con- tinuous variables. We refer to this extension as Functional DCOPs (F-DCOPs) in this paper.1 Additionally, as vari- ables can now take values from a continuous range, con- straint utilities are also similarly extended from tabular forms to functional forms in F-DCOPs. To solve such problems, Stranders et al. [2009] extended the discrete Max-Sum (MS) algorithm [Farinelli et al., 2008] to Continuous MS (CMS), where constraint utility functions are approximated by piece- wise linear functions. Voice et al. [2010] later proposed Hybrid CMS (HCMS), which combines the discrete MS al- gorithm with continuous non-linear optimization methods. Specifically, agents in HCMS approximate the utility func- tions with a number of samples that they iteratively improve over time. A key limitation of CMS and HCMS is that they both do not provide quality guarantees on the solutions found. The reason for this is that they rely on discrete MS as the un- derlying algorithmic framework, which do not provide qual- ity guarantees on general graphs. To this extend overcome limitation, Pseudo-tree we Optimization [Petcu and Faltings, 2005] the Procedure Distributed (DPOP) algorithm to three extensions -- Exact Functional DPOP (EF-DPOP); Ap- proximate Functional DPOP (AF-DPOP); and Clustered AF-DPOP (CAF-DPOP). EF-DPOP provides an exact approach to solve F-DCOPs with linear or quadratic utility functions and are defined over tree-structure graphs. Both AF-DPOP and CAF-DPOP solve F-DCOPs approximately without any restriction on the type of utility functions or graph structure. We also provide theoretical properties on the error bounds and communication complexities of AF-DPOP and CAF-DPOP and show that they outperform HCMS in randomly generated instances when given the same 1As Stranders et al. [2009] did not name their extension in their paper, we choose a name so that we can refer to it easily. communication limitations. 3 Functional DCOP Model 2 Background DCOPs: A Distributed Constraint Optimization Problem (DCOP) is a tuple hA, X, D, F, αi: A = {ai}p i=1 is a set of agents; X = {xi}n i=1 is a set of decision variables; D = {Dx}x∈X is a set of finite domains and each variable x ∈ X takes values from the set Dx; F = {fi}m i=1 is a set of utility functions, each defined over a set of decision variables: fi : Qx∈xfi Dx → R ∪ {−∞}, where infeasible configu- rations have −∞ utilities, xfi ⊆ X is the scope of fi, and α : X → A is a mapping function that associates each deci- sion variable to one agent. A solution σ is a value assignment for a set xσ ⊆ X of variables that is consistent with their respective domains. The utility F(σ) = Pf ∈F,xf ⊆xσ f (σ) is the sum of the utilities across all the applicable utility functions in σ. A solution σ is complete if xσ = X. The goal is to find an optimal complete solution x∗ = argmaxx F(x). A constraint graph visualizes a DCOP, where nodes in the graph correspond to variables in the DCOP and edges connect pairs of variables appearing in the same utility function. A pseudo-tree arrangement has the same nodes and edges as the constraint graph and satisfies that (i) there is a subset of edges, called tree edges, that form a rooted tree and (ii) two variables in a utility function appear in the same branch of that tree. The other edges are called backedges. Tree edges connect parent-child nodes, while backedges connect a node with its pseudo-parents and its pseudo-children. In this paper, we assume that each agent controls exactly one decision variable and thus use the terms "agent" and "variable" interchangeably. We also assume that all utility functions are binary functions between two variables. DPOP: Distributed Pseudo-tree Optimization Procedure (DPOP) [Petcu and Faltings, 2005] is a complete inference algorithm that is composed of three phases: • Pseudo-tree Generation: In this phase, all agents start building a pseudo-tree [Hamadi et al., 1998]. • UTIL Propagation: Each agent, starting from the leaves of the pseudo-tree, adds the optimal sum of utilities in its subtree for each value combination of variables in its sep- arator.2 It does so by adding the utilities of its functions with the variables in its separator and the utilities in the UTIL messages received from its children. The agent then projects out its variable by optimizing over it and sends the projected function in a UTIL message to its parent. • VALUE Propagation: Each agent, starting from the root of the pseudo-tree, determines the optimal value for its vari- able and then sends the optimal value as well as the opti- mal values of agents in its separator to its children. The root agent does so by choosing the values of its variables from its UTIL computations, and send them as VALUE messages. 2The separator of xi contains all ancestors of xi in the pseudo- The Functional DCOP (F-DCOP) model generalizes the reg- ular discrete DCOP model by modeling the variables as con- tinuous decision variables [Stranders et al., 2009]. More for- mally, an F-DCOP is a tuple hA, X, D, F, αi, where A, F, and α are exactly as defined in DCOPs. The key differences are as follows: • X = {xi}n i=1 is now a set of continuous decision variables controlled by the agents. • D = {Dx}x∈X is now a set of continuous domains of the decision variables. Each variable x ∈ X takes values from the interval Dx = [LBx, U Bx]. The objective of an F-DCOP is the same as that in DCOPs -- to find an optimal complete solution x∗ = argmaxx F(x). 4 F-DCOP Algorithms We now introduce three F-DCOP algorithms: Exact Func- tional DPOP (EF-DPOP), Approximate Functional DPOP (AF-DPOP), and Clustered AF-DPOP (CAF-DPOP). All three algorithms are based on the framework of DPOP, where they extend the capability of DPOP such that they can solve F-DCOPs with continuous variables and utility functions. 4.1 Exact Functional DPOP Exact Functional DPOP (EF-DPOP) is an exact algorithm for F-DCOPs with linear or quadratic utility functions and are defined over tree-structure graphs. It extends the two pri- mary operations of DPOP in the UTIL propagation phase -- addition and projection. Addition Operation: In EF-DPOP, each UTIL message contains a piecewise function and the addition of two piece- wise functions is done by adding their sub-functions that may have different domains. We will use the following two func- tions to illustrate our operations: 12 f a f b f c f d 12 12 12 23 23 f a f b f c f d 23 23 f12(x1, x2) =  f23(x2, x3) =  if x1 ∈ [0, 4], x2 ∈ [0, 6] if x1 ∈ [0, 4], x2 ∈ [6, 10] if x1 ∈ [4, 10], x2 ∈ [0, 6] if x1 ∈ [4, 10], x2 ∈ [6, 10] if x2 ∈ [0, 3], x3 ∈ [0, 7] if x2 ∈ [0, 3], x3 ∈ [7, 10] if x2 ∈ [3, 10], x3 ∈ [0, 7] if x2 ∈ [3, 10], x3 ∈ [7, 10] (1) (2) When adding two piecewise functions, we first identify the common variable between the two functions and create a new set of atomic ranges for the variable. For example, when adding the functions f12 and f23 above, the common vari- able is x2, and the new ranges for x2 are [0, 3], [3, 6], and [6, 10]. The ranges of the other variables remain unchanged from their original functions. We then take the Cartesian product of the range sets of all common variables and associate the appropriate function to that range. For example, the addition of f12 and f23 will be a new function f123: tree that are connected to xi or to one of its descendants. f 123(x1, x2, x3) 23 f a 12 + f a f a 12 + f b f c 12 + f a f c 12 + f b . . . 23 23 23 =   if x1 ∈ [0, 4], x2 ∈ [0, 3], x3 ∈ [0, 7] if x1 ∈ [0, 4], x2 ∈ [0, 3], x3 ∈ [7, 10] if x1 ∈ [4, 10], x2 ∈ [0, 3], x3 ∈ [0, 7] if x1 ∈ [4, 10], x2 ∈ [0, 3], x3 ∈ [7, 10] (3) Projection Operation: Projecting out a variable xi from a function f (xi, xi1 , . . . , xik ) means finding the piecewise function: g(xi1, . . . , xik ) = argmax xi f (xi, xi1 , . . . , xik ) (4) To find g, we solve the following for closed-form solutions: ∂f (xi, xi1 , . . . , xik ) ∂xi = 0 Let ¯xi be the solution to the equation above. Then: ¯xi = g ′(xi1 , . . . , xik ) ¯g(xi1 , . . . , xik ) = f (xi = ¯xi, xi1 , . . . , xik ) Aside from ¯g, there are two other candidate functions: g = f (xi = LBxi , xi1 , . . . , xik ) g = f (xi = U Bxi , xi1 , . . . , xik ) (5) (6) (7) (8) (9) Next, we need to find the intervals where each of the func- tions ¯g, g and g is the largest. Those intervals are the inter- sections between the three functions and, thus, we solve each of the equations below to find them: g(xi1 , . . . , xik ) = g(xi1 , . . . , xik ) g(xi1 , . . . , xik ) = ¯g(xi1 , . . . , xik ) g(xi1 , . . . , xik ) = ¯g(xi1 , . . . , xik ) (10) (11) (12) The result of this process is a set of intervals where either ¯g, g, or g is the largest. The projected function g is the piecewise function that consists of ¯g, g, or g with the intervals that they are the largest in. Unfortunately, it is not always possible to find closed-form solutions to the partial derivative in Equation (5). We discuss below two types of functions -- binary linear and quadratic functions -- where it is possible to find closed-form solutions. • Binary linear functions of the form f (xi, xi1 ) = axi + bxi1 + c. By following the monotonicity property of linear f (xi, xi1 ) at functions, we can find g(xi1 ) = argmaxxi the two extremes: g(xi1 ) =(f (xi = LBxi , xi1 ) f (xi = U Bxi , xi1 ) if a > 0 otherwise (13) • Binary quadratic functions of the form f (xi, xi1 ) = ax2 i + i1 + dxi1 + exixi1 + f . We first take the partial bxi + cx2 derivative and setting it to 0 to find the critical point: ∂f ∂xi = 2axi + b + exi1 = 0 ¯xi = −b − exi1 2a (14) (15) As ¯xi has to belong to the interval [LBxi, U Bxi], we solve the inequalities below to find the range xi1 as the domain of ¯g(xi1 ): LBxi ≤ −b − exi1 2a ≤ U Bxi (16) 4.2 Approximate Functional DPOP In general F-DCOPs, Eq. (5) may be a multivariate equation, and it is not always possible to find a closed-form solution to such functions. Therefore, an approximation approach is desired for F-DCOPs. In this section, we introduce Approximate Functional DPOP (AF-DPOP), which is an approximation algorithm that can solve F-DCOPs without any restriction on the functional form of the constraint utilities. AF-DPOP is similar to DPOP in that the algorithm has the same three phases: pseudo- tree generation, UTIL propagation, and VALUE propaga- tion phases. The pseudo-tree generation phase is identical to that of DPOP, and the UTIL and VALUE propagation phases share some similarities. We now describe how these two propagation phases work at a high level. In the UTIL propagation phase, like DPOP, agents in AF-DPOP also discretizes the domains of variables and sends up UTIL tables that contain utilities for each value combination of values of separator agents. However, un- like DPOP, agents in AF-DPOP perform local optimization of these values by "moving" them along the gradients of rele- vant utility functions in order to improve the overall solution quality. As such, the addition and projection operators have to be updated as well. In the VALUE propagation phase, like DPOP, agents in AF-DPOP also sends down their best value down to their children in the pseudo-tree. However, unlike DPOP, agents in AF-DPOP may receive values of ancestors that do not map to computed utilities. As such, the agents must perform local interpolation of the utilities value in this phase. We now describe the algorithm in more detail, where we focus on the UTIL and VALUE propagation phases of the algorithm. UTIL Propagation: In this phase, each leaf agent first dis- cretizes the domains of the agents in its separator (i.e., its par- ent and pseudo-parents) and then stores the Cartesian product of these discrete values in the set V . Therefore, each element v ∈ V is a tuple hvi1 , . . . , vik i, where each value vij is the value of separator agent aij . Then, for each tuple v ∈ V , the agent "moves" each value vij in the tuple along the gradient of each function that is relevant to agent aij . Specifically, the agent updates value vij according to the following equation for each separator agent xij of the leaf agent xi: vij = vij + α ∂fij (xi, xij ) ∂xij vij (cid:12) (cid:12) (cid:12) (cid:12) argmaxxi (17) fij (xij =vij ) where fij (xi, xij ) is the utility function between the leaf agent xi and the separator agent xij and α is the learning rate of the algorithm. The agent can "move" the values as many times as it like until they have either converged or a maxi- mum number of iterations is reached. The agent can "move" the values as many times as desired until they have either con- verged or a maximum number of iterations is reached. Then, the updated values in V and their corresponding utilities de- fine the UTIL table that is sent to the parent of the agent in a UTIL message. As in DPOP, each non-leaf agent will first wait for the UTIL messages from each of its children. When all the UTIL messages are received, the agent processes the UTIL tables in the UTIL message from each child. Note that in regular DPOP, the Cartesian product of the values of agents are con- sistent across the UTIL tables of all children (i.e., if the values of an agent a exists in the Cartesian products of two children, then those values are identical). The reason is because all agents agree on the discretization of the domain of agent a and do not update the value of that agent (such as through Eq. (17)). Therefore, each agent can easily add up the utili- ties in the UTIL tables received together with the utilities of constraints between the agent and its separator. In contrast, since the values of agents are updated accord- ing to Eq. (17) in AF-DPOP, these values may no longer be consistent across different UTIL tables received. To remedy this issue, each agent first adds additional tuples to each UTIL table received such that the Cartesian product of the values of agents are consistent across all the UTIL tables. Then, it approximates the utilities of the newly added tuples by inter- polating between the utilities of the existing tuples. Finally, since the UTIL tables are now all consistent, the agent adds up the utilities in the UTIL tables of children together with the utilities of constraints between the agent and its separator in the same way as DPOP. After the utilities are added up, similar to leaf agents, the agent xi will proceed to repeatedly update the values vij of the separator aij in the updated Cartesian product V using: vij = vij + α ∂fij (xi, xij ) ∂xij (cid:12) (cid:12) (cid:12) (cid:12) vij argmaxxi U T ILi(vi1 ,...,vik ) (18) where U T ILi is the utility table that is constructed from the summation of the children's utilities and the utilities of con- straints between the agent xi with its separator set. The key difference between this Eq. (18) and the Eq. (17) used by leaf agents is that the substitution of fij (xij = vij ) with U T ILi(vi1 , . . . , vik ). The reason for this substitution is that the utilities in the UTIL tables of leaf agents are only a function of constraints with their separator agents and the functional form of those constraints are known. Therefore, leaf agents can optimize exactly those functions to get accurate gradients. In con- trast, utilities in the UTIL tables of non-leaf agents are also a function of the constraints between its descendant agents and its separator agent, and the functional form of those con- straints are not known. They are only represented by samples within the UTIL tables received and are now integrated into the UTIL table of the non-leaf agent. Therefore, in Eq. (18), the agent approximates its maximum value xi by choosing the best value of under the assumption that the values of the other separator agents are exactly the same as in the tuple hvi1 , . . . , vik i that is being updated. After these values are all updated, the agent approximates their corresponding utilities by interpolating between known utilities and sends these utilities up to its parent in a UTIL message. These UTIL messages propagate up to the root agent, which then starts the VALUE phase. VALUE Propagation: The root agent starts this phase after processing all the UTIL messages received from its children in the UTIL phase. It chooses its best value based on its com- puted UTIL table and sends this value down to its children. Like in DPOP, each agent will repeat the same process after receiving the values of its parent and pseudo-parents. However, unlike DPOP, an agent may receive the informa- tion that its parent or pseudo-parent is taking on a value that doesn't correspond to an existing value in the agent's UTIL table due to the values being moved during the UTIL propa- gation phase. As a result, the agent will need to approximate the utility for this new value received and it does so by inter- polating between known utilities in its UTIL table. Once all the leaf agents receive VALUE messages from their parents and choose their best values, the algorithm ter- minates. 4.3 Clustered Approximate Functional DPOP A possible limitation of AF-DPOP is that the number of tu- ples in the Cartesian product V that is propagated in the UTIL messages can be quite large, especially if additional tuples are added to maintain consistency between the UTIL tables of children. In communication-constrained applications, it is preferred that the number and size of messages transmitted between agents to be as small as possible. With this motivation in mind, we extend AF-DPOP to Clustered AF-DPOP (CAF-DPOP), which bounds the num- ber of tuples sent in UTIL messages to limit the message size. CAF-DPOP is identical to AF-DPOP in every way except that agents choose k representative tuples and their correspond- ing utilities to be sent up to their parents in UTIL messages. To choose these k representative tuples, we use the k-means clustering algorithm [MacQueen, 1967] to cluster the tuples and then approximate the utilities of those tuples through in- terpolation. This approach assumes that tuples that are close to each other will have similar values. Note that while only k tuples are sent between agents in UTIL messages, each agent still maintains the original un- clustered set of tuples in their memory. Thus, when they perform interpolation during the VALUE propagation phase, they will use the utilities of the unclustered set of tuples since they are more accurate than the utilities of the clustered set of tuples. 5 Theoretical Properties We now provide some theoretical properties of some of our algorithms as well as that of (discrete) DPOP and Hybrid Continuous Max-Sum (HCMS). For each reward function f (xi, xi1 , . . . , xik ) of an agent xi and its separator agents xi1 , . . . , xik , assume that agent xi discretizes the domains the reward function into hyper- cubes of size m (i.e., the distance between two neighboring discrete points for the same agent xij is m). Let ∇f (v) de- note the gradient of the function f (xi, xi1 , . . . , xik ) at v = (vi, vi1 , . . . , vik ): ∇f (v) = ( ∂f ∂xi (vi), ∂f ∂xi1 (vi1 ), . . . , ∂f ∂xik (vik )) (19) A HCMS DPOP 10 20 30 40 50 129k 306k 436k 636k 832k 220k 541k 766k 1104k 1456k 5 330k 795k 1128k 1587k 2109k AF-DPOP 10 356k 870k 1230k 1728k 2316k 15 374k 947k 1331k 1876k 2533k 20 404k 1008k 1414k 1980k 2687k EF-DPOP 518k -- -- -- -- Table 1: Experimental Results Varying the Number of Agents on Random Trees with Three Initial Discrete Points Furthermore, let ∇f (v) denote the sum of magnitude: ∇f (v) = ∂f ∂xi (vi) + ∂f ∂xi1 (vi1 ) + . . . + ∂f ∂xik (vik ) (20) Assume that ∇f (v) ≤ δ holds for all utility functions in the DCOP and for all v. Theorem 1. The error bound of discrete DPOP is Fmδ. Proof. First, we prove that the magnitude of the projection of function f is also bounded from above by δ. Let xi = vi be the point where: g(xi1 , . . . , xik ) = f (xi = vi, xi1 , . . . , xik ) f (xi, xi1 , . . . , xik ) = max xi (21) (22) Then, assume that ∇g(v) > δ for all v. Let v′ = (vi, vi1 , . . . , vik ) and v′ ∇f (v′) = ∂f ∂xi (vi) + −i = (vi1 , . . . , vik ), then: ∂f ∂xik (vi1 ) + . . . + ∂f ∂xi1 ≥ ∂f ∂xi1 (vi1 ) + . . . + ∂f ∂xik (vik ) −i) = ∇g(v′ > δ (vik ) (23) (24) (25) (26) This contradicts with assumption that ∇f (v) ≤ δ for all v. The error bound of each function is then mδ because each hypercube is of size m and the magnitude of the gradient within each hypercube is at most δ. As the error may be ac- cumulated each time an agent sums up utility functions, the total error bound for a problem is thus Fmδ, where F is the number of utility functions in the problem. Theorem 2. The error bound of AF-DPOP is F(m + Akαδ)δ, where k is the number of times each agent "moves" values of its separator by calling Eqs. (17) or (18). Proof. After each "move" of an agent by calling either Eqs. (17) or (18), the maximum size of the hypercubes in- creases by αδ, where α is the learning rate. Since each agent performs this update only k times, the largest increase in the size of the hypercube is kαδ. Finally, since the value of an agent can be updated by any of its children or pseudo- children, the total increase in the size of the hypercube is thus Akαδ, where A is the number of agents in the problem. Therefore, this combined with the proof of the bound for dis- crete DPOP, the error bound is thus F(m + Akαδ)δ. Theorem 3. In a binary constraint graph G = (X, E), the number of messages of HCMS with k iterations is 4kE. The number of messages of discrete DPOP, AF-DPOP, and CAF- DPOP is 2X. Proof. HCMS has the same number of messages as the Max- Sum algorithm [Farinelli et al., 2008]. Every edge of the con- straint graph has two variable nodes and one function node and, thus, it takes 4 messages per edge in one iteration. The total number of messages in HCMS is thus 4kE. The number of messages required by AF-DPOP and CAF- DPOP is identical to that of DPOP -- each agent sends one UTIL message to its parent and one VALUE message to each of its children in the pseudo-tree. Since pseudo-trees are spanning trees, the number of messages is thus 2X. Theorem 4. The message size complexity of discrete DPOP, AF-DPOP and CAF-DPOP is O(dw), O((dX)w), and max{A, k}, respectively, where d is the number of points used by each agent to discretize the domain of its separator agents, w is the induced width of the pseudo-tree, and k is the number of clusters used by CAF-DPOP. size the message [Petcu and Faltings, 2005]. complexity is Proof. For DPOP, O(dw) For AF-DPOP, as the values of an agent are "moved" by their children and pseudo-children, in the worst case, all the values are unique and the maximum number of such values is O(dX). The message sizes are then similar to discrete DPOP with O(dX) values per agent. Therefore, its message size com- plexity is O((dX)w). For CAF-DPOP, the message size complexity of UTIL messages is O(k) since only the utilities of the centroids of k clusters are sent. And the message size complexity of VALUE messages is O(A), such as in a fully-connected graph where an agent sends the values of every agent from the root of the pseudo-tree down to itself in a VALUE message to its child. Therefore, the message complexity of the algorithm is the O(max{A, k}). 6 Experimental Results We empirically evaluate EF-DPOP, AF-DPOP, and CAF- DPOP against (discrete) DPOP and HCMS on both random trees and random graphs. We adapt the (discrete) DPOP al- gorithm to solve F-DCOPs by discretizing the continuous do- main into discrete representative points. We measure the quality of solutions found by the algo- rithms as well as the number of messages taken by the al- gorithm. Since HCMS is an iterative algorithm that may take a long time and a large number of messages before converg- ing, in order for fair comparisons, we initially planned to ter- minate the algorithm after it sends as many messages as the DPOP-variants. However, even a single iteration of HCMS requires more messages than the DPOP-variants. We thus let HCMS terminate after one iteration. We did not report the A HCMS DPOP 15 20 25 30 265k 345k 439k 506k 522k 865k -- -- 5 710k 1171k -- -- AF-DPOP 10 763k 1285k -- -- 15 824k 1334k -- -- CAF-DPOP 20 891k 1407k 5 639k 1006k -- 1040k -- 1513k 10 697k 1017k 1101k 1682k 15 715k 975k 1024k 1597k 20 787k 973k 1027k 1656k Table 2: Experimental Results Varying the Number of Agents on Random Graphs with p1 = 0.2 and Three Initial Discrete Points (a) Random Trees with A = 20 (b) Random Graphs with A = 20 and p1 = 0.2 #points HCMS DPOP 0 541k 990k 0 306k 554k 1 3 9 AF-DPOP 254k 870k 1133k HCMS DPOP AF-DPOP 428k 1285k -- 8 345k 706k 15k 865k -- CAF-DPOP 431k 1017k 1272k Table 3: Experimental Results Varying the Number of Discretized Points actual number of messages since they could be trivially com- puted via Theorem 3. Tables 1 and 2 show the various algorithms' solution qual- ities on random trees and graphs, respectively, where we vary the number of agents A and every algorithm discretizes the domains of variables into three points. We also vary the number of times AF-DPOP and CAF-DPOP agents "move" a point (by calling Eqs. (17) or (18)) from 5 to 20. Tables 3(a) and 3(b) show the various algorithms' solution qualities on random trees and graphs, respectively, where we set the num- ber of agents A to 20 and vary the number of initial points used by the algorithms to discretize the domains from 1 to 9. In all our experiments, we set the domain of each agent to be in the range [−100, 100]. We generate utility functions that are binary quadratic functions, where the signs and co- efficients of the functions are randomly chosen. Our exper- iments were performed on a 2.10GHz machine with 100GB of RAM. Results are averaged over 20 runs. Random Trees: We omit the results of CAF-DPOP from Table 1 since it finds identical solutions to AF-DPOP on trees -- there is no need to perform any clustering on trees since an agent does not receive utilities for value combinations of its parent from its children since there are no backedges in the pseudo-tree. Not surprisingly, EF-DPOP finds the best solution since it is an exact algorithm. However, it could only solve the smallest of instances -- due to memory limitations, the agents could not store the necessary number of piecewise functions to accurately represent the utility functions after additions and projections. In general, AF-DPOP finds better solutions than DPOP, which finds better solutions than HCMS. The reason is because AF-DPOP updates the value of representative points before propagating up the pseudo-tree. In contrast, the val- ues chosen by DPOP is fixed from the start. Finally, HCMS performs poorly because a single iteration is insufficient for it to converge to a good solution. Additionally, as expected, the quality of solutions found by AF-DPOP improves with increasing number of times points are "moved" by the algo- rithm. We omit the results of EF-DPOP from Table 3(a) as it failed to solve these instances and we omit the results of CAF-DPOP because it finds identical solutions to AF-DPOP on trees. Not surprisingly, the quality of solutions found by all the three algorithms increase with increasing number of points. The reason is that the agents can more accurately rep- resent the utility function with more points. Random Networks: The trends in Table 2 are similar to those in random trees, except that CAF-DPOP finds solu- tions with qualities between that of AF-DPOP and DPOP. The reason is that CAF-DPOP clusters the points into k clusters and only propagate a representative point from each cluster. Therefore, the k points represent the utility functions less ac- curately than the full number of unclustered points that AF- DPOP uses. However, this reduced number of points prop- agated also improves the scalability of CAF-DPOP, where it is able to solve problems larger problems than AF-DPOP and DPOP. The trends in Table 3(b) are again similar to that in random trees, except that both AF-DPOP and DPOP ran out of mem- ory with 9 points. Interestingly, CAF-DPOP also finds better solutions than AF-DPOP when they use only 1 point. 7 Conclusions In many real-world applications, agents usually choose their values from continuous ranges. Researchers have thus pro- posed the F-DCOP formulation to model continuous vari- ables. However, existing methods suffer from the limitation that they do not provide quality guarantees on the solutions found. In this paper, we remedy this limitation by introducing (1) EF-DPOP, which finds exact solutions on F-DCOPs with linear or quadratic utility functions and are defined over tree- structure graphs; (2) AF-DPOP, which finds error-bounded solutions on general F-DCOPs; and (3) CAF-DPOP, which limits the message size of AF-DPOP to a user-defined pa- rameter k. Experimental results show that AF- and CAF- DPOP both find better solutions that HCMS, an existing state- of-the-art F-DCOP algorithm, when given the same commu- nication limitations. Therefore, these algorithms combined extend the applicability of DCOPs to more applications that require quality guarantees on the solutions found as well as those that require limited communication capabilities. problems. In Proceedings of the International Joint Con- ference on Artificial Intelligence (IJCAI), pages 468 -- 474, 2016. [Stranders et al., 2009] Ruben Alessandro Farinelli, Alex Rogers, and Nicholas Jennings. De- centralised coordination of continuously valued control parameters using the Max-Sum algorithm. In Proceedings of AAMAS, pages 601 -- 608, 2009. Stranders, [Ueda et al., 2010] Suguru Ueda, Atsushi Iwasaki, and Makoto Yokoo. Coalition structure generation based on distributed constraint optimization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 197 -- 203, 2010. [Voice et al., 2010] Thomas Voice, Ruben Stranders, Alex Rogers, and Nicholas Jennings. A hybrid continuous max- sum algorithm for decentralised coordination. In Proceed- ings of the European Conference on Artificial Intelligence (ECAI), pages 61 -- 66, 2010. [Yeoh and Yokoo, 2012] William Yeoh and Makoto Yokoo. Distributed problem solving. AI Magazine, 33(3):53 -- 65, 2012. [Zivan et al., 2015] Roie Zivan, Harel Yedidsion, Steven Okamoto, Robin Glinton, and Katia Sycara. Dis- tributed constraint optimization for teams of mobile sens- ing agents. Journal of Autonomous Agents and Multi- Agent Systems, 29(3):495 -- 536, 2015. References [Farinelli et al., 2008] Alessandro Farinelli, Alex Rogers, Adrian Petcu, and Nicholas Jennings. Decentralised coor- dination of low-power embedded devices using the Max- Sum algorithm. In Proceedings of the International Con- ference on Autonomous Agents and Multiagent Systems (AAMAS), pages 639 -- 646, 2008. [Fioretto et al., 2017a] Ferdinando Fioretto, William Yeoh, and Enrico Pontelli. A multiagent system approach to scheduling devices in smart homes. In Proceedings of the International Conference on Autonomous Agents and Mul- tiagent Systems (AAMAS), pages 981 -- 989, 2017. [Fioretto et al., 2017b] Ferdinando Fioretto, William Yeoh, Enrico Pontelli, Ye Ma, and Satishkumar Ranade. A DCOP approach to the economic dispatch with demand response. In Proceedings of the International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 999 -- 1007, 2017. [Hamadi et al., 1998] Youssef Hamadi, Christian Bessi`ere, and Joel Quinqueton. Distributed intelligent backtracking. In Proceedings of the European Conference on Artificial Intelligence (ECAI), pages 219 -- 223, 1998. [Kumar et al., 2009] Akshat Kumar, Boi Faltings, and Adrian Petcu. Distributed constraint optimization with structured resource constraints. In Proceedings of the In- ternational Conference on Autonomous Agents and Multi- agent Systems (AAMAS), pages 923 -- 930, 2009. [MacQueen, 1967] James MacQueen. Some methods for classification and analysis of multivariate observations. In Proceedings of the Berkeley Symposium on Mathematical Statistics and Probability, pages 281 -- 297, 1967. [Maheswaran et al., 2004] Rajiv Maheswaran, Milind Tambe, Emma Bowring, Jonathan Pearce, and Pradeep Varakantham. Taking DCOP to the real world: Efficient complete solutions for distributed event scheduling. In the International Conference on Au- Proceedings of tonomous Agents and Multiagent Systems (AAMAS), pages 310 -- 317, 2004. [Miller et al., 2012] Sam Miller, Sarvapali Ramchurn, and Alex Rogers. Optimal decentralised dispatch of embedded generation in the smart grid. In Proceedings of the Interna- tional Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 281 -- 288, 2012. [Modi et al., 2005] Pragnesh Modi, Wei-Min Shen, Milind Tambe, and Makoto Yokoo. ADOPT: Asynchronous dis- tributed constraint optimization with quality guarantees. Artificial Intelligence, 161(1 -- 2):149 -- 180, 2005. [Petcu and Faltings, 2005] Adrian Petcu and Boi Faltings. A scalable method for multiagent constraint optimization. In Proceedings of the International Joint Conference on Ar- tificial Intelligence (IJCAI), pages 1413 -- 1420, 2005. [Rust et al., 2016] Pierre Rust, Gauthier Picard, and Fano Ramparany. Using message-passing dcop algorithms to solve energy-efficient smart environment configuration
0911.0753	1	0911	2009-11-04T09:19:37	An XML-based Multi-Agent System for Supporting Online Recruitment Services	[ "cs.MA" ]	In this paper we propose an XML-based multi-agent recommender system for supporting online recruitment services. Our system is characterized by the following features: {\em (i)} it handles user profiles for personalizing the job search over the Internet; {\em (ii)} it is based on the Intelligent Agent Technology; {\em (iii)} it uses XML for guaranteeing a light, versatile and standard mechanism for information representation, storing and exchange. The paper discusses the basic features of the proposed system, presents the results of an experimental study we have carried out for evaluating its performance, and makes a comparison between the proposed system and other e-recruitment systems already presented in the past.	cs.MA	cs	IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 1 An XML-based Multi-Agent System for Supporting Online Recruitment Services Pasquale De Meo, Giovanni Quattrone, Giorgio Terracina, and Domenico Ursino 9 0 0 2 v o N 4 ] A M . s c [ 1 v 3 5 7 0 . 1 1 9 0 : v i X r a Abstract -- In this paper we propose an XML-based multi- agent recommender system for supporting online recruitment services. Our system is characterized by the following features: (i) it handles user profiles for personalizing the job search over the Internet; (ii) it is based on the Intelligent Agent Technology; (iii) it uses XML for guaranteeing a light, versatile and standard mechanism for information representation, storing and exchange. The paper discusses the basic features of the proposed system, presents the results of an experimental study we have carried out for evaluating its performance, and makes a comparison between the proposed system and other e-recruitment systems already presented in the past. Index Terms -- Multi-Agent Systems, XML, E-Services, Re- cruitment Services. I. INTRODUCTION I N the last decade Internet services emerged as a significant, both social and cultural, phenomenon; in fact, presently, many organizations provide their customers with the possibil- ity to access their services also on the Internet. In this scenario, online recruitment services, supporting both individuals looking for a job and companies looking for employees, are assuming a prominent role. In such a context, generally, companies populate a database of job proposals and individuals are supported in their job search by an engine based on classical Information Retrieval (IR) techniques1. Several studies show that the number of users accessing these services is dramatically increasing [1]; as a consequence, it is possible to foresee that, in the next years, a huge amount of job proposals will be handled by means of the Internet. In such a situation, classical IR techniques used by the present recruitment services could provide an individual with an excessively large number of job proposals not interesting for him. This could result in a low user perceived quality of service and, ultimately, in his decision to not access these services. A way for solving this problem consists of realizing person- alized search engines, which combine classical IR techniques and user modeling methodologies [2]. In the literature some approaches based on recommender systems [3] have been proposed to personalize search engines in various applica- tion contexts, such as Web browsing [4], e-commerce [5], e-learning [6], and so on. In addition, in some proposals, P. De Meo, G. Quattrone and D. Ursino are with DIMET (Dipartimento di Informatica, Matematica, Elettronica e Trasporti) of University Mediterranea of Reggio Calabria, Reggio Calabria, Italy. G. Terracina is with the Depart- ment of Mathematics of University of Calabria, Rende (CS), Italy. Email: {demeo,quattrone,ursino}@unirc.it, [email protected]. example, http://www.jobfinder.com. for 1See, hotjobs.yahoo.com or recommender systems have been combined with the agent technology for making them autonomous, reactive and pro- active [7]. This paper provides a contribution in this setting; in fact, it proposes an XML-based multi-agent recommender system that exploits rich user profiles to support personalized recruitment services. Our system is based on the agent technology. In fact, a User Agent is associated with each user and manages his profile, as well as the interaction with him; a Recruitment Agent supports each User Agent in the selection of those job proposals appearing to be the most adequate for the corresponding user. The exploitation of the intelligent agent technology allows our system to improve its autonomy and adaptiveness, as well as to easily partition the various tasks among several entities, thus improving its scalability. Finally, it allows an easy management of existing legacy systems by means of suitable wrappers; as a consequence, our system can easily interact with pre-existing recruitment Web sites and this contributes to enhance the completeness of obtained results. Our system is XML-based. XML is a standard language for representing and exchanging information. It embodies both representation capabilities, typical of HTML, and data management features, typical of DBMSs. XML information bases are plain text documents and, consequently, they are light, versatile, easy to be exchanged and can reside on different, both hardware and software, platforms. In spite of this simplicity, the information representation rules embodied in this language are powerful enough to allow the management of sophisticated queries. As for the contribution of our paper in the field of e- recruitment systems, we observe that: • In order to select the job proposals that best satisfy user requirements, a traditional e-recruitment system considers only the query of the user (that specifies his preferences and constraints), along with the features of available job proposals. On the contrary, our system takes into account not only the query of the user but also his reaction to previous job proposal recommendations it made to him. Moreover, it analyzes similitudes and/or correlations among job proposals to produce its suggestions; in this way, it can highlight some proposals that are usually ignored by traditional e-recruitment systems and that might be appealing for the user. • In a traditional e-recruitment system, a user can query the corresponding (internal) database. As a consequence, if the system's answer to a query is insufficient, the user must access other e-recruitment systems and re- submit the query; after this, he must compose the results IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 2 provided by each system to construct his final list of job proposals. This activity might be time consuming, since the user is obliged to contact and query several e-recruitment systems. In addition, it might be prone to errors, since two e-recruitment systems might exploit different methodologies and use different terminologies to both select and represent job proposals. Specifically, as for methodologies, it could happen that the same job proposal for a user ui receives contrasting scores by different e-recruitment systems. As a consequence, when ui compares the results provided by the various systems, he might be confused and, therefore, incapable of making a decision about its suitability for him. As for termi- nologies, two independent e-recruitment systems might present to a user the same job proposal by exploiting different terms. As an example, a job proposal regarding the design, the development and the validation of software programs might be called "Program Application Support Specialist" by an e-recruitment system, and "Software Test and Software Design Engineer" by another one. In situations like that previously described, since the two systems are totally independent each other, there is no form of coordination and, consequently, no way to inform a user that two or more terms represent the same job proposal. Our system is capable of facing all these problems; specif- ically, as it will be clarified in the following, it continuously monitors existing online recruitment services and collects job proposals coming from them; as a consequence, user search is not restricted to the database associated with an e-recruitment system but it spreads across disparate sources. This system organization allows a user to pose his queries in a transparent way; in other words, in order to obtain an answer, he could be unaware of the models and the languages of the involved databases, as well as of their internal structure. As for the Recommender Systems research area, the system we propose in this paper provides the following novelties: • A traditional Recommender System usually identifies the "Top K" recommendations, i.e., the K information items (in our case, job proposals) having the highest relevance for the user. The coefficient K is usually constant and pre-defined. In line with current advances [8], our system can analyze user reactions to past proposals to assess if the number of job proposals it presents to users is too low (resp., too high) and, in the affirmative case, it can dynamically augment (resp., reduce) this number. • In content-based Recommender Systems, a user profile is generally represented as a vector of pairs hki, wii, where ki is a keyword and wi is a weight denoting the impor- tance of ki for the user. On the contrary, in our system, user profiles are richer and take into account various characteristics and preferences of involved users. Even if this implies a higher complexity in their management, the improvement of results that can be obtained is very satisfying, as it will be clarified in Section V-E. • Our approach applies in the Recommender Systems field some mathematic tools generally used in other appli- Fig. 1. General architecture of our system cation contexts, such as manufacturing engineering [9], bioinformatics [10] and econometrics [11]. Interestingly enough, the computational effort it requires is not partic- ularly heavy, as it will be demonstrated in Section IV-B.1. II. GENERAL ARCHITECTURE OF OUR SYSTEM Our system consists of two main agent typologies, namely: (i) a User Agent (U A), that manages the profile of a user, as well as the interaction with him, and (ii) a Recruitment Agent (RA), that handles both job proposals and recommendation activities. Moreover, a Job Proposal Database (J P D) is used to store all available job proposals. J P D can be populated by companies who can insert, remove or modify job proposals by means of a Company Agent, i.e., a suitable Interface Agent, analogous to that described in [12]. In addition, J P D can be automatically enriched by suitable Wrapper Agents, analogous to that described in [13], that continuously monitor existing online recruitment services, find new job proposals and store them in J P D, if they are not already present. J P D can be described by an XML document; the corresponding XML Schema is shown in the Appendix available at the address http://www.ing.unirc.it/ursino/tsmc05/Appendix.pdf. Specifically, J P D stores is job proposals; a by tuple hJ IDl, J U RLl, J T opicSetl, J CharacteristicSetli, where: proposal a set of described job J Pl a • J IDl is an identifier; • J U RLl is the url where the complete description of the job proposal is available; • J T opicSetl = {J T opicl1, J T opicl2 , . . . , J T opiclm} is a set of topics describing J Pl; a topic is characterized by its name; • J CharacteristicSetl is a dynamic set of characteristics associated with J Pl. Each characteristic is described by a pair hF eature, V aluei; some of the possible features are the salary associated with the job proposal, the city/country of the job proposal, the foreign languages, the skills, the years of experience and the academic title(s) required from the candidate. Observe that we have chosen to store available job proposals into an independent knowledge base and not in the ontology IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 3 of a specific agent; independent and specialized agents can operate in our system to automatically and continuously update and refine them. in this way, several The general architecture of our system is shown in Figure 1. Its behaviour is as follows. When a user ui wants to perform a job search, he submits a query to the associated User Agent U Ai. After this, U Ai contacts RA and provides it with both the query and the profile of ui. RA selects from J P D the available job proposals satisfying user query, ranks them on the basis of the presumed user preferences and selects those ones that best fit the past interests of ui, as well as other interests, someway related to those ui considered in the past and that he still disregarded. After this, RA sends the selected proposals to U Ai that presents them to ui; U Ai monitors ui and, when he performs his choices, it updates his profile accordingly. As previously pointed out, the role of XML in our system is crucial. In fact: • Agent ontologies are stored as XML documents; as a con- sequence, they are light, versatile, easy to be exchanged and can reside on different hardware and software plat- forms. In spite of this simplicity, the information repre- sentation rules embodied in XML are powerful enough to allow a sophisticated information management. • The agent communication language adopted in our sys- tem is ACML [14]; this is the XML encoding of the standard Agent Communication Language. • The extraction of information from the various data structures is carried out by means of XQuery [15]. This is becoming the standard query language for the XML environment. Since XQuery is based on the XML framework, it can handle a large variety of data. It has capabilities typical of database query languages as well as features typical of document management systems. • The manipulation of agent ontologies is performed by means of the Document Object Model (DOM) [16]. DOM makes it possible for programmers to write applications working properly on all browsers and servers as well as on a large variety of both hardware and software platforms. The architecture of our system can be considered mixed, i.e., partially centralized and partially distributed. In this sense it follows the ideas expressed in [17], where an architecture for handling telecommunication networks, based on the presence of an auctioneer agent that carries out most of the negotiation activities, is described. It can be considered centralized for the presence of the Recruitment Agent that performs most of the activities concerning the selection of the job proposals best fitting user exigencies. On the other hand, it can be assumed to be distributed for the presence of many User Agents, Wrapper Agents and Company Agents that continuously cooperate with the Recruitment Agent for performing the whole recruitment task. Observe that, in our system, there exists only a central Recruitment Agent. Such a choice is justified by the following motivations: • Network Congestion. In our architecture, each User Agent is in charge of forwarding both the User Profile and the user query to the Recruitment Agent that processes this query and sends the corresponding results to it; as a consequence, the two agents exchange a small number of (generally simple) messages and this prevents network overload. In a distributed implementation of the Recruit- ment Agent activities, various Recruitment Agents would cooperate for processing user queries. As a consequence, they should continuously communicate and exchange messages; this would generate a high network traffic and the whole system would quickly be overwhelmed with messages among agents. This problem is particularly in our reference scenario where the number relevant of job seekers, and consequently of User Agents, is very high; this would cause an overwhelming amount of exchanged messages and, ultimately, significant latencies and users dissatisfaction. • Task Assignment. In a distributed implementation of the Recruitment Agent activities, a query is propagated through several Recruitment Agents that spend memory and CPU resources to process it. In this scenario, it could happen that some Recruitment Agents would be overwhelmed with requests, whereas other ones would be almost unoccupied. In order to optimize the system per- formances, it would be necessary to define protocols for intelligently assigning tasks to the various Recruitment Agents. These protocols appear very complex to be real- ized because the potential job seekers (and, consequently, the potential User Agents) might be very numerous and geographically sparse. • Message Duplication. In a distributed implementation of the Recruitment Agent activities, multiple copies of a query might be received by the same Recruitment Agent (we call these copies duplicated messages). To better clarify this concept and its consequences, consider four Recruitment Agents RAd1, RAd2, RAd3 and RAd4, and assume that: (i) a user submits a query Q to RAd1; (ii) RAd1 is not capable of answering Q; (iii) RAd1 asks RAd2 and RAd3 to answer Q; (iv) RAd2 and RAd3 are only capable of producing a partial answer to Q (e.g., the number of job proposals retrieved by them is lesser than that required by the user); (v) RAd2 and RAd3 might independently forward Q to RAd4 that would receive and, in its turn, could process and possibly transmit Q twice. This example clearly shows that duplicated messages are pure overhead; they increase the network congestion and determine a waste of CPU and memory resources for the Recruitment Agents that are often obliged to process the same query more than once. In spite of the resource waste caused by them, they do not improve system accuracy since they do not contribute to increase the chance of finding job proposals relevant for user. These problems are emphasized in an e-recruitment sce- nario, where the number of job seekers contemporarily accessing our system could be very high; in this case, the probability of duplicating messages might be extremely high. As a consequence, suitable mechanisms for iden- tifying duplicated messages and avoiding to process and forward them are required; however, these mechanisms IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 4 are generally difficult to be designed and implemented. Interestingly enough, these motivations are in accordance also with the ideas illustrated in [18]. Finally, we would like to point out that our system does not provide for any additional layer between the User Agent and the Recruitment Agent. This choice is justified by the following motivations: • Ability to cooperate with other e-recruitment systems. One of the most important capabilities of our system consists of its ability to effectively cooperate with other e- recruitment systems (see below). A layered architecture does not appear particularly suitable for pursuing such an objective. In fact, in a layered architecture, system functionalities are associated with different abstraction levels, each represented by a layer; the abstraction degree generally increases from the lowest levels to the highest ones. The presence of different abstraction levels has the following consequences: -- The interaction between two layered systems re- quires each of them to have a deep knowledge of the architectural features of the other one. In fact, each system should know the abstraction level associated with each layer of the other system to properly ex- change information and cooperate. Such a knowledge is often missing and, hence, the mapping from a layer in a system to a layer in another one might be not obvious and, in some cases, not possible. -- Many systems cannot be easily structured in a lay- ered fashion. As a consequence, if our system would be layered, we could have a scenario in which layered and non-layered architectures coexist; in this case, suitable communication protocols between lay- ered and non-layered systems are compulsory; their realization, however, is often a hard and expensive activity. • Performance Issues. Several studies point out that the number of users accessing e-recruitment systems is rapidly increasing [1]. As a consequence, an e- recruitment system should be capable of quickly process- ing user queries. Such a requirement cannot be easily satisfied in a layered system; in fact, in this case, each layer supplies services to the layer above it and operates as a client for the layer below it. Therefore, in a layered system, a continuous message exchange between low- level layers and the high-level ones is required. Since a message often has to pass through many layers, the overhead associated with message exchange is often significant and might determine a performance decay. Various techniques have been proposed to boost the performances of a layered system (e.g., it is possible to couple non-adjacent layers) but they seem too difficult to be realized in practical contexts, especially in our application case where the various layers could belong to different owners. • Reliability Issues. The fast identification of faults and the quick reaction to them cannot be easily performed in a layered architecture since a failure in a layer might determine the crash of the whole system. Even the (pos- sibly non fast) fault identification might be difficult to be carried out since the functionalities of the low-level layers are often hidden to the high-level ones and, consequently, an application running on high-level layers has many difficulties in identifying where a certain problem has occurred. In the following sections we provide a detailed description of the two main agents involved in our system, namely the User Agent and the Recruitment Agent. III. THE USER AGENT A. Ontology The ontology of the User Agent U Ai, associated with a user ui, stores the profile of ui. It contains: • The code U IDi identifying ui. • A set T opicSeti storing the preferences of ui; each preference corresponds to a topic which ui looked for in the past. Each element T opicij of T opicSeti is represented by a tuple hT opicN ameij , Countij , F irst- T imeStampij i, where: (i) T opicN ameij indicates the is an access counter name of T opicij ; (ii) Countij denoting how many times T opicij has been specified in a query submitted by ui; (iii) F irstT imeStampij indicates the first time instant in which T opicij has been specified in a query of ui. As it will be clarified in the following, these last two coefficients allow the relevance of T opicij for ui to be measured. • A dynamic set ConstraintSeti of constraints that ui fixes for the desired job. Each constraint is described by a pair hF eature, V aluei; some of the possible con- straints are the minimum salary ui would like to earn, the city/country where he desires to work, the foreign language(s) he knows, his skills, the years of experience he has attained, and his academic title(s). • A set P astQueriesi, storing information about i i. i , αk the queries that ui posed to the system in the past. Each element of P astQueriesi consists of a pair hσk is the satisfaction coefficient; it belongs to the real σk i interval [0, 1] and indicates how much ui appreciated the recommendations provided by the system as a response to his kth query. In the literature many proposals have been presented to measure the satisfaction of a user for a set of recommendations (see, for example, [19], [20]). In this paper we follow the ideas illustrated in [20] and define σk i be the set of job proposals recommended by the system as a response to the kth query of ui and let AcceptedJ P k i be the set of job proposals accepted by ui; σk i can be defined as i = AcceptedJP k i σk JP Listk i is the audacity coefficient; it belongs to the real αk i interval [0, 1]; the higher αk i is, the higher the number of job proposals returned by the system as its answer to the query Qk is computed by the Recruitment i will be. αk Agent; the various strategies for its computation are shown in Section IV-B.1; they have been designed in i as follows. Let J P Listk . i IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 5 such a way to make our system adaptive against the past behaviour of ui. C. Influences of job market specifics in the design of the User Agent architecture The XML Schema associated with the ontology of U Ai is shown in the Appendix. B. Behaviour i , QT Setk i by means of a suitable wizard. Qk i i. SelDegreek The behaviour of U Ai consists of the following steps. • Step 1. When ui wants to perform a job search, U Ai supports him in the construction of a corresponding query i consists of a pair Qk i , belonging to the hSelDegreek real interval [0, 1], is the Selectivity Degree; it indicates how much the system should be selective in the search of proposals (see below); the wizard associated with U Ai proposes to ui some possibilities and he chooses the most suitable one for him in a friendly, guided manner. iq } is the set of QT Setk topics describing the job proposals currently interesting for ui. Moreover, if ui wants to update the constraints for a desired job, U Ai provides him with a suitable graphical interface that allows him to view the constraints he has defined, as well as to add, drop or modify some of them. • Step 2. U Ai updates T opicSeti; this task is performed by inserting in it those topics of Qk i not already present therein and by increasing the access counters of the topics corresponding to QT opick i = {QT opick i1, . . . , QT opick i1, . . . , QT opick iq . that best fit • Step 3. U Ai contacts RA and supplies it both Qk i and the profile of ui. RA selects the job proposals answering interests of ui, as well as Qk i other interests, someway related to those ui considered in the past and that he still disregarded. Then, it sends to U Ai both the selected job proposals and the audacity coefficient αk i the past it used. • Step 4. U Ai presents to ui the job proposals provided by RA. ui accepts those ones appearing to be the closest to his exigencies and rejects the other ones. After this, U Ai computes the satisfaction degree σk i that ui showed for the provided recommendations, and stores the pair hσk i i into P astQueriesi. In order to avoid an excessive growth of T opicSeti, U Ai executes a pruning activity on it. Such a task is carried out by computing the relevance for ui of each topic T opicij stored in T opicSeti and by remov- ing from T opicSeti those topics presenting a relevance smaller than a certain threshold. The relevance of T opicij at the time instant T is computed by means of the formula . This formula is r(T opicij , T ) = justified by considering that the more a user is interested in a topic, the more frequently he asks information about it. T −F irstT imeStampij Countij i , αk • Step 5. In order to show the exploitation of ACML and XQuery in our system, in the Appendix, we present the ACML message that U Ai sends to RA for requiring a set of recommendations and the query that U Ai executes for identifying those topics of Qk i not already present in T opicSeti. The design of our User Agent model has been influenced by various job market specifics. In the following we examine the most relevant of them. 1) Construction of a network of job seekers: Several studies point out the importance of constructing a job seeker network (JSN), i.e., a community of job seekers that share and ex- change information about their past job experiences. A JSN is an effective channel for disseminating job information and can successfully support its users in their job search: as an example, a study presented in [21] shows that the probability of a user to find a job grows with the increase of the size of the network he belongs to. Actually, the construction of a J SN is a delicate activity. In fact, job seekers, due to both selfishness and rivalry, might be unwilling to cooperate. To overcome this drawback, it is necessary to create a free-of-competition environment based on trustful relationships among its members. In order to carry out this task, it is necessary to enhance the cooperation among job seekers belonging to the same job context for whom a rivalry might not exist; think, for example, to job seekers having different experience degrees (e.g., a senior software programmer who helps a newly graduated person to find his first job). The experience shows also that information sharing among members of a JSN increases their capability of posing precise queries. In fact, users accessing e-recruitment services are ex- tremely heterogeneous: job seekers having a deep knowledge of a job domain often coexist with other seekers having a su- perficial knowledge of it. An expert is capable of formulating a query that really captures his needs; on the contrary, a novice might ignore the information content of the database storing available job proposals and, consequently, might be unable to formulate precise queries. As previously pointed out, a JSN might support novices to compose their queries; specifically, a novice might ask the experts of a JSN to support him in his query composition. Our User Agent model has been designed in such a way to facilitate the creation of a JSN; specifically, a User Agent U Ai, on behalf of ui, could contact other User Agents for finding job seekers having a profile similar to ui but being more expert than him; such a task can be carried out by taking into account the lists of topics specified in user profiles and by comparing them by means of the Jaccard Coefficient [22]. When the list of job seekers similar to ui is available, ui can ask U Ai to activate a contact with one or more of them in such a way to require a support to formulate his queries. 2) Integration with e-learning systems: Several studies point out that the capability of a job seeker to continuously update his know-how has a significant implication on the success of his job search, as well as on the advancement of his career; as a consequence, learning activities should be regarded as a lifelong process. For this reason a strict integration between e-recruitment and e-learning systems might provide a job seeker with a significant competitive advantage. A key component of most e-learning systems is the so- called student profile, storing information about the student IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 6 background knowledge and information needs. In this way, it is possible to diversify the learning process on the basis of user skills and necessities. Our User Agent model can make the interaction between our system and an e-learning one easier; specifically, a User Agent could transfer the corresponding User Profile to an e-learning system that might construct an initial student profile starting from it. Conversely, the profile of a student that acquired some skills with the support of an e-learning system could be exploited to construct an initial user profile for the User Agent of our e-recruitment system. 3) Efficient retrieval of curricula: Presently, a large number of curricula vitae is available on the Internet. Generally, a cur- riculum consists of a title and a set of additional information provided by the job seeker (e.g., his past experience). These curricula can be consulted by businesses; a common way to do this activity is the so called "curricula spidering", i.e., the exploitation of a software agent that tracks down curricula and finds those of interest for businesses. The search of curricula is generally performed by means of a traditional search engine; in this case the business manager specifies some words to the engine and this last searches for those curricula containing these words. These techniques might lead to irrelevant and spurious results; in fact, a cur- riculum might be too vague; it might be not updated; last, but not the least, the words appearing in a curriculum might not belong to the vocabulary of terms used by business managers in their searches. Our User Agent model is capable of enhancing the curricula search capabilities of businesses; in fact: • A User Agent is in charge of monitoring the behaviour of its user and this allows it to maintain the corresponding profile always updated. • A User Agent learns the profile of its user on the basis of the job proposals that he searches for and, eventually, accepts or rejects. In this way, the topics specified in a User Profile derive from the job proposal features directly inserted by businesses; such a policy avoids ambiguity, that could arise when a term present in a traditional curriculum could be interpreted in more than one way by a business manager, and protects the job seeker against his lack of knowledge of the terms that are currently used by businesses. IV. THE RECRUITMENT AGENT A. Ontology set of The ontology of a Recruitment Agent RA consists in Sec- is described by a tuple of a tion II, a Job Proposal J Pl hJ IDl, J U RLl, J T opicSetl, J CharacteristicSetli. job proposals; as pointed out The XML Schema associated with the ontology of RA is shown in the Appendix. B. Behaviour RA is activated by U Ai when ui wants to perform a query i for searching job proposals. It receives Qk i and the profile Qk of ui; it returns to U Ai: (i) the job proposals answering Qk i that best fit the past interests of ui, as well as other interests, someway related to those ui considered in the past and that i it used to he still disregarded; (ii) the audacity coefficient αk select the job proposals. In order to perform its activity, RA carries out the following steps: • Step 1. It queries J P D for retrieving the job proposals (keyword-based filtering) and satisfying matching Qk i user constraints (feature-based filtering); to this purpose it applies classical Information Retrieval techniques [23]. • Step 2. It ranks the selected proposals on the basis of user preferences, as stored in T opicSeti. Such a task is carried out as follows: let J Pl = hJ IDl, J U RLl, J T opicSetl, J CharacteristicSetli be a job proposal and let T Sil = T opicij ∈ T opicSeti, T opicN ameij ∈ {T opicij J T opicSetl} be the set of topics of J Pl appearing to be interesting for ui. RA assigns an interest degree to J Pl, for the user ui at the current time instant T , according r(T opicij , T ). to the formula ρ(T Sil, T ) = PT opicij ∈T Sil Here, the relevance r has been defined in Section III-B. Clearly, the higher the value of ρ(T Sil, T ) is, the higher the end of the relevance of J Pl for ui will be. At this phase, RA constructs J P T empListk i , obtained by sorting the selected job proposals on the basis of their interest degrees. Observe that a more complex formula could be adopted for ρ. However, in this way, the complexity of the Recruit- ment Agent would unnecessarily increase. In fact, since, during the selection of job proposals, our approach takes into account also user past preferences and satisfaction, a simple ranking function is sufficient. This is confirmed also by the experimental results shown in Section V. is already a good candidate to be pre- J P T empListk i sented to ui. However, two further improvements could be performed by RA to make it more adequate. Firstly, we observe that J P T empListk i ranks the job proposals on the basis of their relevance for ui, according only to his past preferences; however, it does not consider topics, someway related to those ui considered in the past and that he still disregarded. Secondly, J P T empListk i could still contain quite a large number of proposals. The next steps performed by RA are devoted to carry out these improvements. of • Step 3. RA constructs the set SeedP roposalsk i obtained by selecting the first ⌈SelDegreek i ⌉ of proposals J P T empListk i are used by RA as seeds to choose SeedP roposalsk other job proposals containing topics, someway related to those ui considered in the past and that he still disregarded. i × J P T empListk i . The elements ∈ i = SeedP roposalsk • Step 4. RA constructs the final list J P Listk i of recom- mended job proposals as J P Listk i ∪ {J Pl ∈ J Pl i , J Pm i }. Here αk SeedP roposalsk δ(J Pl, J Pm) measures "dissimilarity degree" between J Pl and J Pm. It is defined as δ(J Pl, J Pm) = 1 − 2JT opicSetl∩JT opicSetm JT opicSetl+JT opicSetm , where J T opicSetl and topics associated J T opicSetm represent i , δ(J Pl, J Pm) the J P T empListk the sets of ≤ IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 7 with J Pl and J Pm, respectively, and the symbol S represents the cardinality of a set S. Interestingly enough, the value 2S1∩S2 S1+S2 is known as Dice's coefficient in the literature. It is easy to show that δ ranges between 0 (if the two job proposals coincide) and 1 (if the two job proposals are completely different). The audacity coefficient αk i is dynamically computed by RA on the basis of the feedback that ui showed for past recommendations. In this paper we propose three different strategies for the computation of this coefficient; they are extensively described in Section IV-B.1. i , obtained at the end of this step, contains at least J P Listk the seed proposals; moreover, it could include also some job proposals relative to topics not particularly relevant for ui in the past, but that could be of interest for him in the future (since they are sufficiently similar to those he judged relevant in the past). In the selection of this last category of proposals δ plays a key role; in fact, it measures the dissimilarity degree of two job proposals on the basis of their topics, i.e., of their meaning and seman- tics, without considering the relevance that ui assigned to them in the past. i to U Ai. i and αk • Step 5. RA sends J P Listk 1) Strategies for the audacity computation: In our system we have chosen to dynamically update the audacity coefficient after each query performed by the user ui. This allows it to be very sensitive to the user judgements about its recom- mendations and significantly improves its accuracy. However, this makes the manual update of the audacity coefficient a very difficult in our opinion, asking the user to manually update the audacity coefficient after each query is obtrusive. In addition, even with the exploitation of suitable graphic interfaces, the user could be not capable of understanding the role of the audacity coefficient and to choose the most correct value for it. task. In fact, However, in our prototype, in order to consider the possible presence of very smart users, we have inserted a module allowing a smart user to personally modify the values of the audacity coefficient determined by the system. i that RA associates with Qk In this section we illustrate three different strategies that may be used for automatically computing the audacity coef- ficient αk i . Recall that the higher this coefficient is, the higher the number of job proposals returned by the system for answering Qk i will be; this strictly depends on the satisfaction that ui showed for the recommen- dations of RA in the past. As a consequence, any strategy for the computation of αk i must evaluate the satisfaction of ui w.r.t. the previous recommendations of RA. As previously pointed out, the choice to make audacity dependent on satisfaction allows our system to be adaptive against the past behaviour of ui. a) Strategy 1: Positive and Negative Feedback (PNF). In this strategy, αk i is dynamically updated on the basis of the feedback provided by ui for the last recommendation; this is encoded in the satisfaction coefficient σk−1 relative to the recommendations associated with the query Qk−1 . Recall that σk−1 is the fraction of job proposals accepted by ui w.r.t. i those recommended by the system when it answered Qk−1 . i i i PNF works as follows: • If, during the execution of Qk−1 i i > 1 , the number of accepted recommendations has been greater than the number of the rejected ones (i.e., σk−1 2 ), then it is possible to argue that ui has appreciated the system recommendations, since he accepted most of them. As a consequence, if the system would send further recommendations, it is extremely probable that ui would receive other proposals interesting for him. Now, since in our system the amount of proposals sent to the user as the answer to the query Qk i i , increasing is strictly related to the audacity coefficient αk this coefficient would imply an increase of the number of proposals recommended by the system when answering i . Qk Clearly, the key issue in this reasoning is the correct estimation of the increase of αk i . In fact, if such an increase is too high, then ui would receive several use- less proposals and his perceived quality of the overall recommendations could decrease. On the contrary, if the increase is too low, the system could miss to recommend some proposals interesting for ui. The previous reasoning led us to define the increase εk i to apply on the audacity αk−1 i as a linear function of the user for obtaining αk i = σk−1 2 . This formula shows satisfaction; formally, εk that the more ui is satisfied, the more σk−1 is higher i is increased w.r.t. αk−1 . Clearly, than 1 other functions could be adopted for εk i , e.g., logarithmic, quadratic, cubic or exponential functions; however, in our opinion, the linear function is the most suitable one to satisfy the contrasting exigencies mentioned above and to provide the system with the correct balance between stability and readiness in adapting its behaviour to user satisfaction (see the Appendix for an experimental anal- ysis of this topic). 2 and the more αk i − 1 i i i • If, during the execution of Qk−1 i i < 1 , the number of rejected recommendations has been greater than the number of the accepted ones (i.e., σk−1 2 ), then it is possible to argue that ui has disliked system recommendations, since he refused most of them, and that he is interested to examine fewer proposals during the next queries. As previously the number of system recommendations pointed out, depends on αk i ; therefore, in order to make the system to recommend fewer proposals to ui, it is necessary that αk i is lower than αk−1 . For the same reasoning introduced in the previous case, the most suitable function for defining i appears the decrease εk for obtaining αk the linear one. More formally, the decrease εk i can be defined as εk i to apply to αk−1 . i = 1 2 − σk−1 i i i • The last possible situation happens when the number of accepted proposals has been equal to the number of the rejected ones (i.e., σk−1 2 ). This situation can be seen as a specific case of both the first and the second situations described above. If it is consid- ered as a particular case of the first (resp., second) situation, the increase (resp., decrease) εk is equal to i 0 and, consequently, it is possible to conclude that the number of recommendations should be neither increased = 1 i IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 8 nor decreased; this implies that the audacity coefficient should be maintained constant. The previous reasoning leads to the following formula for computing the new value of αk i : min{1, αk−1 αk−1 max{0, αk−1 αk i =   i i if σk−1 if σk−1 if σk−1 i i > 1 2 = 1 2 < 1 2 i + εk i } i − εk i } This formula is valid when ui exploits the system at least for the second time. During the first query α1 i is set equal to a constant value α. An analysis of the impact of α on the system performance is reported in the Appendix. b) Strategy 2: 2-Least Square Error (2-LSE). This strategy is based on the Least Square Error tech- nique (LSE) that is largely exploited in several research areas, such as manufacturing engineering [9], bioinformatics [10], econometrics [11], and so on. LSE considers two sets X = {x1, x2, . . . , xN } and Y = {y1, y2, . . . , yN } and a function y = f (x, a0, a1, . . . , ap) depending on p+1 unknown parameters but having a predefined form (e.g., f might be a polynomial, a linear combination of exponentials, a sinusoid, and so on). The goal of LSE is to identify the values a∗ 0, a∗ 1, . . ., a∗ p of the parameters a0, a1, . . . , ap such that the residual N p)(cid:1)2 is minimum. In such a Pl=1(cid:0)yl − f (xl, a∗ p) is the function that best 0, a∗ case it is said that f (x, a∗ "fits" the sets X and Y , in the least squares sense. 1, . . . , a∗ 0, a∗ 1, . . . , a∗ R = i , this strategy: In order to compute αk • Applies the LSE technique for determining the values a∗ 0, . . . , a∗ p of the parameters a0, . . . , ap of the fitting it exploits the sets X = function. To this purpose, }. Ob- {α1 serve that, as a consequence of this choice for the sets X and Y , the fitting function represents the user satisfaction against the system audacity. } and Y = {σ1 i , . . . , σk−1 i , . . . , αk−1 i , α2 i , σ2 i i • Sets αk i to the value of the audacity corresponding to the maximum value2 of the fitting function and, consequently, to the maximum value of the user satisfaction. A crucial problem in the implementation of the LSE strategy is the choice of both the shape and the parameters of the fitting function. Such a choice must minimize the residual R; as a consequence, it requires to compute the partial derivatives of R and to solve the set of equations ∂R ∂aj = 0, ∀j = 0, . . . , p. If we use, as fitting function, a combination of sinusoids, logarithms and exponentials, these equations generally define a non-linear system that, often, cannot be exactly solved; in this case, even the calculation of an approximate solution requires an heavy computational effort. On the contrary, if the fitting function has a polynomial shape, i.e., f (x, a0, a1, . . . , ap) = a0xp + a1xp−1 + . . . + ap, it is possible to show that these equations are equivalent to the linear system [24] Ax = b, where A is a matrix called Vandermonde matrix. It is possible to show that at least an approximate solution of this system can be computed in a 2Observe that, since the independent variable (i.e., the audacity) belongs to the compact interval [0, 1], the fitting function has always a maximum within this interval. polynomial time; generally, the exact solution can be computed in O(p2) steps [24]; as a consequence, a polynomial shape for the fitting function appears well suited for our purposes. The next step of our activity consists of determining the order of the polynomial to be used as fitting function. To this purpose observe that LSE must compute the audacity value that maximizes the fitting function f (x); to compute this value we need to determine the derivative f ′(x) of f (x) and to solve the equation f ′(x) = 0. Now, if f (x) is a pth order polynomial, f ′(x) will be a (p − 1)th order polynomial; as a consequence, we have that: (i) if p = 2 (i.e., if the fitting function is a parabola), it is possible to directly obtain an exact solution of f ′(x) = 0 by solving a first order equation; (ii) if p = 3 (i.e., if the fitting function is a cubic), it is possible to obtain an exact solution of f ′(x) = 0 by solving a second order equation; (iii) if p > 3, the computation of the roots of f ′(x) can be performed by means of numerical (and approximate) techniques; with regard to this, it was shown that the time necessary for finding all the roots of f ′(x) = 0 with 2b is O(p log p4 + p log2 p log b) a maximum error equal to 1 [25]. This analysis shows that, for obtaining a correct (i.e., not approximate) solution of the equation f ′(x) = 0, the fitting function must be either a parabola or a cubic. Since the accuracy ensured by the polynomial of degree 2 (i.e., the parabola) is high [26], and since the computation time it requires for solving f ′(x) = 0 is lower than that necessary if the fitting function is a cubic, we have adopted the parabola as fitting function; in this case the LSE technique is called 2-LSE technique. Therefore, in our case, the fitting function is y = a0x2 + a1x + a2; this choice implies that the 2-LSE strategy is valid only when ui exploits the system at least for the fourth time. During the first three queries α1 i are set equal to three constant values α′, α′′, α′′′, respectively. An analysis of the impact of α′, α′′, α′′′ on the system performance is reported in the Appendix. i , α3 i , α2 c) Strategy 3: Weighted Sum (WS). i i i i and αk,2−LSE ; specifically, αk i = γ × αk,P N F In order to compute αk i , . The overall audacity value αk i this strategy applies the PNF and the 2-LSE strategies and obtains two audacity values αk,P N F is obtained by computing a suitable weighted mean of αk,P N F and αk,2−LSE + (1 − γ) × αk,2−LSE . Here, γ is a coefficient ranging in the real interval [0, 1]. Specifically, if γ = 1, the WS strategy coincides with the PNF strategy; on the contrary, if γ = 0, the WS strategy coincides with the 2-LSE strategy. In our experiments we have considered different values for γ and we have studied their impact on the system performance (see the Appendix for more details); in addition, we have examined how to dynamically compute γ for maximizing the system performances. i i V. EXPERIMENTAL RESULTS In this section we describe in detail the various experiments we have conducted for testing the performance of our system. Specifically, in Section V-A we describe the characteristics of both the users and the job proposals considered in our IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 9 tests; Section V-B presents the accuracy measures we have chosen for testing our approach. Section V-C is devoted to make a comparison of the three strategies for the audacity computation. In Section V-D we evaluate the performance of our system in different application domains. Finally, Section V-E is devoted to present an experimental comparison of our system with other e-recruitment ones. In order to perform our tests, we have designed a pro- totype implementing our approach. This prototype has been developed in JADE3 (Java Agent DEvelopment framework), a software framework conceived for supporting the implemen- tation of agent-based applications in compliance with FIPA specifications [27]. A. Characteristics of the users and job proposals In our tests we have considered a set of users U Set = {u1, u2, . . . , u200} consisting of 200 volunteers. As for their distribution against their age we have that: (i) 21.00% of them was 18-24 years old; (ii) 39.00% of them was 25-34 years old; (iii) 25.50% of them was 35-44 years old; (iv) 13.50% of them was 45-54 years old; (v) 1.00 % of them was more than 54 years old. As for their distribution against their past usage to com- mercial e-recruitment systems we have that: (i) 11.00% of them had never used a commercial e-recruitment system; (ii) 10.00% of them exploited it very rarely; (iii) 10.00% of them used it rarely; (iv) 25.50% of them exploited it sometimes; (iv) 31.00% of them exploited it often; (v) 12.50% of them used it very often. job The proposals, available sites extracted http://www.jobpilot.co.uk about from and the http://www.careerbuilder.com were 3000. They belonged to different domains; specifically, reference domains where: (i) "Information Technology" for 19.29% of them; (ii) "Health Care Management" for 17.02% of them; (iii) "Finance" for 15.84% of them; (iv) "Engineering" for 12.07% of them; (v) "Banking" for 10.04% of them; (vi) "Legal" for 7.02% of them; (vii) "Real Estate Management" for 5.26% of them; (viii) "Others" for 13.46% of them. B. Evaluation Measures In our tests we have computed two widely accepted eval- uation measures, namely Precision and Recall. Precision is defined as the share of job proposals accepted by ui among those recommended by the system; Recall is the share of job proposals suggested by the system among those of interest for ui. These parameters have been computed as follows: • Each user ui ∈ U Set was asked to submit a query Qk i i was obtained (see and the corresponding J P T empListk Section IV). • ui was asked to identify the subset U serListk i ⊆ i of job proposals that he considered in- J P T empListk teresting. 3JADE is an open source platform; it can be downloaded from http://sharon.cselt.it/projects/jade. AvgPre PNF AvgPre 2−LSE 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 n o s i i c e r P e g a r e v A 0 0 5 10 15 Number of Queries 20 25 Fig. 2. Average Precision for the PNF and the 2-LSE strategies • Our prototype was run for obtaining the set J P Listk i ⊆ i of job proposals to be recommended to J P T empListk ui. • The Precision P rek i and the Recall Reck i relative to the kth query have been obtained by applying the formulas: P rek i = i ∩U serListk JP Listk i i = JP Listk . i ∩U serListk i JP Listk i , Reck U serListk i • Finally, the Average Precision AvgP rek and the Average Recall AvgReck relative to the kth query have been obtained as AvgP rek = Pi=1..U Set Pi=1..U Set , AvgReck = P rek i U Set Reck i . U Set Observe that Precision, Recall, Average Precision and Av- erage Recall belong to the real interval [0, 1]; specifically, the higher these coefficients are the better the system works. C. Experimental comparison of the three strategies imple- mented in our approach In this section we experimentally compare the strategies for the audacity computation described in Section IV-B.1. Actually, it is necessary to examine only the PNF and the 2-LSE strategies; in fact, the WS strategy is a weighted mean of the two other ones. In the test of the PNF strategy we have set α = 0.55 whereas in the test relative to the 2-LSE strategy we have set α′ = 0.5, α′′ = 0.6, α′′′ = 0.4 (see the Appendix). For each strategy we have computed the Average Precision and the Average Recall against the number of queries. The obtained results are shown in Figures 2 and 3. From the analysis of these figures we can conclude that: • If the number of queries carried out by users is low (i.e. lesser than 10), the PNF strategy shows a better performance than the 2-LSE strategy; in fact, its Average Precision and its Average Recall increase up to 0.75 and 0.69, respectively. This behaviour can be motivated by considering that, before 10 queries have been performed, the 2-LSE strategy has still not completed the audacity "tuning" and, hence, the corresponding results are not particularly satisfactory. • If the number of queries carried out by users is high (i.e. greater than 15), the performance improvements of the PNF strategy are small. On the contrary, the 2-LSE IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 10 AvgRec PNF AvgRec 2−LSE 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 l l a c e R e g a r e v A 0 0 5 10 15 Number of Queries 20 25 Fig. 3. Average Recall for the PNF and the 2-LSE strategies strategy shows a very good performance; as an example, after 15 queries, the Average Precision and the Average Recall are 0.85 and 0.76, respectively. This behaviour can be motivated by considering that the 2-LSE strategy "tunes" the audacity coefficient on the basis of the user feedback during all the k − 1 previous queries, whereas the PNF strategy considers only the last one. As a consequence, a high number of queries allows the 2-LSE strategy to more precisely predict user expectations. • During the initial queries, both Average Precision and Average Recall have an oscillatory behaviour; this de- pends on the fact that, initially, both user profiles and user feedbacks are relatively poor. • Finally, we observe that the Average Precision is gen- erally greater than the Average Recall; this confirms the results obtained in [28]. As pointed out in this paper, such a feature is to be considered a positive characteristic for a recruitment system; in fact, in this application context, Precision is generally assumed to be more important than Recall since the user could be frustrated by many irrelevant proposals. Finally, it is worth pointing out that the results of this experiment confirm the results of the experiments about the tuning of γ described in the Appendix. In fact, in that case, we have obtained that, for a small number of queries (i.e., lesser than 10), the value of γ should be high (i.e., the WS strategy should tend to coincide with the PNF strategy); on the contrary, when the number of queries is high, the value of γ should be low (i.e., the WS strategy should tend to coincide with the 2-LSE strategy). D. Performance of our system in different application domains This series of experiments was devoted to determine the performance of our system in several application domains; some of them were very generic, other ones were more specialized. In all existing e-recruitment systems, job proposals can be hierarchically organized on the basis of the application domains they refer to; specifically, the most generic domains (e.g., "Health-Care" or "Information Technology") are located on the top of the hierarchy; on the contrary, the most spe- cialized ones (like "Biochemical Scientist") are located at the bottom. This hierarchy can be graphically represented by means of a tree; each node of the tree, with the exception of the root, represents a domain; a fragment of this tree is graphically shown in Figure 4. With regard to this classification tree, we say that the specialization level of a domain D is j if the depth of the node associated with D is j. As an example, the specialization level of the domain "Application Developer" is 3, since the depth of the node representing this domain is 3. In this test we have computed the Average Precision and the Average Recall obtained by our system when applied to four domains, namely "Information Technology", "Pharmacy", "Software Support" and "Biomedical Scientist", character- ized by different specialization levels. During this compu- tation we have used the WS strategy by setting γ(k) = max(cid:8)0, 1 − (cid:0) k−1 25 (cid:1)(cid:9), α = 0.55 and α′ = 0.5, α′′ = 0.6, α′′′ = 0.4. The obtained results are shown in Table I. From the analysis of this table it is possible to observe that, in general domains (i.e., domains at a specialization level 1 or 2), our system rapidly (i.e., after a few number of queries) obtains good results; however, this performance slightly worsens when the number of queries increases. On the contrary, in specific domains, our system needs several queries for achieving good results but, after this initial phase, it maintains, and even improves, its performance. This behaviour can be motivated by the following reasoning: during the first queries, users interested in general domains do not have a precise idea about their needs and, consequently, many of the provided recommendations appear interesting to them. On the contrary, users interested in specific domains have a precise idea of their desires already during the initial phase; as a consequence, it is more difficult to satisfy them during the first queries. When the number of queries increases, users of general domains "ripen" a more precise idea of their needs, but the application domain they are interested in is too generic to precisely satisfy their requirements; however, Table I shows that, even in these conditions, our system achieves quite good results. On the contrary, after many queries, the profiles of the users of specific domains are quite rich and this allows our system to precisely identify their needs and to significantly reduce the job proposals search space. E. Experimental comparison of our system with other e- recruitment prototypes Job Search - www.jobsearch.co.uk, system with We have experimentally compared our other, already available, ones. Specifically, the systems that we have taken into account were: (i) CareerBuilder - www.careerbuilder.com, (ii) 1job - www.1job.co.uk, (iii) (iv) - www.jobpilot.co.uk, and (v) Fish4Jobs - JobPilot www.fish4.co.uk/iad/jobs. These systems share various similarities; specifically, they operate on the Internet and manage an internal, relational database storing all available job proposals. In addition, they provide job seekers with a graphical interface to retrieve the job proposals they are interested in; a job seeker can use this interface to specify some constraints (e.g., the place where he would like to work), as well as a list of keywords describing the job typology of his interest. Finally, some of these systems allow each IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 11 ROOT Health Care Information Technology Research- Development Application Developer Network Administration ... ... Technical Support Software Support ... ... Medical Pharmacy Surgeon Paramedical ... ... Nurse ... Sales Consultant ... Research ... Biomedical Scientist ... Fig. 4. A fragment of a tree classifying job proposals Finance ... Financial Analyst Risk Management ... Investment Consultant ... ... PERFORMANCE OF OUR SYSTEM IN DIFFERENT APPLICATION DOMAINS TABLE I Domain Specialization Level Query 5 Query 10 Query 15 Query 20 Query 25 Information Technology Pharmacy Software Support Biomedical Scientist 1 2 3 4 Avg Pre Avg Rec Avg Pre Avg Rec Avg Pre Avg Rec Avg Pre Avg Rec Avg Pre Avg Rec 0.73 0.68 0.54 0.49 0.67 0.63 0.52 0.46 0.76 0.72 0.61 0.60 0.70 0.69 0.58 0.56 0.74 0.79 0.80 0.81 0.68 0.71 0.72 0.73 0.72 0.78 0.85 0.86 0.65 0.70 0.73 0.76 0.68 0.77 0.86 0.90 0.61 0.69 0.75 0.82 user to construct a simple profile by filling a questionnaire; each profile stores user skills (e.g., the foreign languages he knows) as well as a brief curriculum vitae. Since the queries submitted by a job seeker consist of a list of keywords, each system must transform them into SQL queries; for this reason, the underlying DBMS must incorporate Information Retrieval (hereafter, IR) algorithms [29]. Our activity aimed at comparing the accuracy of our system against that of the other systems mentioned previously. To this purpose we asked each user of U Set to submit some queries; these were processed by the systems into examination and the corresponding Average Precision and Average Recall were computed. As for our system we have chosen the WS strategy and we have set γ(k) = max(cid:8)0, 1 − (cid:0) k−1 25 (cid:1)(cid:9), α = 0.55 and α′ = 0.5, α′′ = 0.6, α′′′ = 0.4. The results we have obtained are shown in Table II. From the analysis of this table, we can observe that the accuracy of our system is quite satisfying. This interesting result can be justified by the following reasoning: • In order to score available job proposals, the ranking function adopted by our system considers several aspects of the profile of a user, as well as his reaction to past proposals (see Sections III-B and IV-B); on the contrary, the ranking functions adopted by the other systems into examination consider only some constraints specified by the user (e.g., the salary he would like to earn). Taking into account a large number of issues makes our system more sensible to the real user exigencies; this justifies the improvement of the Average Precision w.r.t. the other systems into examination. • Our system suggests to users not only the job proposals exactly matching their queries but also those ones some- way semantically related to them. Hence, it is capable of suggesting potentially interesting job proposals that cannot be revealed by traditional IR algorithms; this implies an improvement of the Average Recall w.r.t. the other systems into examination. • Our system is capable of considering job proposals avail- able in several other systems; this contributes to improve both the Average Precision and the Average Recall w.r.t. the other systems into examination. However, neither Precision nor Recall alone are capable of completely capturing the "quality" of the proposals provided by e-recruitment systems. As an example, a key aspect of these systems regards their capability of correctly scoring their job proposals. To better clarify this concept, consider two e-recruitment systems, ERS1 and ERS2, that receive the same query from a user Ui and return the same set of results but in a reverse order: in other words, ERS1 returns the list hJ P1, J P2, . . . , J Pn−1, J Pni whereas ERS2 returns the list hJ Pn, J Pn−1, . . . , J P2, J P1i; this implies that ERS1 (resp., ERS2) rates J P1 (resp., J Pn) as the most relevant proposal for Ui and J Pn (resp., J P1) as the least relevant one. Finally, assume that Ui fully agrees with the suggestions provided by ERS1. In this case ERS1 and ERS2 return the same proposals and, consequently, obtain the same Average Precision and the same Average Recall; however, from the previous reasoning, it is possible to conclude that the "quality" of suggestions provided by ERS2 is lower than the quality of suggestions provided by ERS1: in fact, Ui must browse the whole list of suggestions provided by ERS2 to find the job proposal of his maximum interest and this might be both a boring and a time-consuming activity. To measure the ranking capability of an e-recruitment sys- tem we adopted the Newell Distance [30], a parameter defined in User Modelling theory. To define the Newell Distance associated with a user Ui and a system to evaluate Sj we need to consider a set J P Set of test job proposals and a pair of functions, sys and usr; sys (resp., usr) takes a job proposal J Pk ∈ J P Set as input and returns an integer p belonging to the set {1, 2, . . . , J P Set} as output; p indicates IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 12 TABLE II AVERAGE PRECISION AND AVERAGE RECALL OF THE SYSTEMS INTO EXAMINATION System Query 5 Query 10 Query 15 Query 20 Query 25 Avg Pre Avg Rec Avg Pre Avg Rec Avg Pre Avg Rec Avg Pre Avg Rec Avg Pre Avg Rec Our System CareerBuilder 1job Job Search JobPilot Fish4Jobs 0.62 0.50 0.52 0.54 0.44 0.47 0.61 0.48 0.46 0.44 0.51 0.48 0.73 0.61 0.63 0.64 0.47 0.50 0.66 0.52 0.57 0.60 0.53 0.52 0.81 0.68 0.71 0.69 0.59 0.60 0.74 0.57 0.66 0.64 0.54 0.55 0.86 0.72 0.68 0.70 0.67 0.63 0.76 0.63 0.67 0.65 0.60 0.62 0.87 0.73 0.70 0.73 0.71 0.69 0.79 0.65 0.63 0.68 0.66 0.63 AVERAGE NORMALIZED NEWELL DISTANCE ACHIEVED BY THE SYSTEMS TABLE III INTO EXAMINATION System Query 5 Query 10 Query 15 Query 20 Query 25 Our System CareerBuilder 1job Job Search JobPilot Fish4Jobs 0.27 0.40 0.38 0.33 0.39 0.38 0.20 0.31 0.36 0.29 0.33 0.31 0.13 0.19 0.22 0.21 0.29 0.25 0.09 0.15 0.14 0.14 0.24 0.19 0.06 0.12 0.11 0.09 0.15 0.13 14 12 10 8 6 4 2 ) s e t y b K ( e z i S e g a r e v A that, according to Sj (resp., Ui), J Pk has the pth highest score among all job proposals existing in J P Set. 0 0 5 10 15 Number of Queries 20 25 The Newell Distance Nij associated with Ui and Sj is defined as Nij = PJ Pk∈J P Set w(usr(J Pk)) × usr(J Pk) − w(sys(J Pk)) × sys(J Pk) . Here w(i) is a weighting function defined as: w(i) = (cid:16) JP Set−i ; it assumes its maximum value when i = 1 and decreases when i increases; this implies that it gives a different importance to the errors possibly made by the system to evaluate; specifically, if a system is unable to detect the most important job proposal, it commits a more serious error than if it is not capable of identifying the least relevant ones. (cid:17)2 i The Newell Distance can be properly normalized in such a way to lie in the real interval [0,1]; specifically, if we indicate by N max the maximum value of the Newell Distance measured over all users and systems into consideration, the normalized Newell Distance can be defined as Nij = Nij N max . Observe that the lesser Nij is, the better a system works. In Table III we report the normalized Newell Distance, aver- aged on all users, obtained for the systems into examination. From the analysis of this table we observe that the Newell Distance obtained by our system is smaller than that obtained by the other systems into evaluation. In our opinion, this result is motivated by considering that: • The parameters adopted by our system for scoring avail- able job proposals are strictly related to the profile of a user, as well as to his reaction to past proposals; this enhances its capability of correctly identifying the most relevant job proposals. • Our system continuously monitors user behaviour and is capable of detecting if the relevance of a job proposal for a given user changes over time. On the contrary, the other systems we have considered in our experiments do not consider the possible modifications of user desires over time. From the previous experiments we can conclude that the recommendations provided by our system are more accurate than those supplied by the other systems into examination; this Fig. 5. Average size of user profiles against the number of queries carried out by users improvement is mainly due to quite a sophisticated manage- ment of information about user profiles and past behaviour. This information is combined with classical IR techniques exploited also by the other systems into consideration and allows the most appropriate answers for user needs to be found. On the other hand, our system requires an additional amount of space for storing user profiles; this is a problem affecting all systems handling and exploiting user profiles. In order to quantify the space requirements of our system for storing user profiles we have performed a final experiment. Specifically, we have computed the average size of user profiles (expressed in Kbytes) against the number of queries carried out by users. The obtained results are shown in Figure 5. From the analysis of this figure we can observe that, initially, user profiles are generally "poor" and, consequently, the storage space they require is negligible; when a user poses his queries, the system enriches the corresponding profile by inserting new topics. After a "reasonably large" number of queries (i.e., after about 10 queries) user profiles become rich and occupy a certain amount of space; however, this space occupation does not increase during the next queries; in fact, when user profiles become excessively large, the system activates a pruning task in such a way that the number of new topics inserted in a profile is approximately equal to the number of topics removed from it; as a consequence, after 15 queries, the average size of user profiles remains quite constant (specifically, it is about 10 Kbytes). The previous analysis shows that the quantity of space our system needs for storing user profiles is limited and acceptable; moreover, since user profiles are implemented in XML, it is possible to apply the methodologies illustrated in [31] for handling their space occupation in a more efficient way. IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 13 VI. RELATED WORK Even if e-recruitment is quite a novel research area, sev- eral systems devoted to handle such an activity have been already presented in the literature. In this section we aim at positioning our system amongst other related ones. In order to carry out this activity, we have considered three terms of comparison, namely: (i) Purpose: e-recruitment systems can be classified as company oriented, if they aim at supporting companies in the selection of new candidates, and seeker oriented, if they support users looking for new job proposals; (ii) Architecture: e-recruitment systems could be centralized, if a single computational entity is in charge of managing all activities, distributed, if tasks are partitioned among several computational entities, and mixed, if some tasks are performed in a centralized fashion and other ones are carried out in a distributed way; (iii) User Profile Construction: e-recruitment systems might be obtrusive, if they interact with the user for constructing his profile, or unobtrusive, if they learn the profile of a user by monitoring his accesses to the system. On the basis of these terms of comparison our system can be considered seeker oriented, mixed and unobtrusive. In the following we compare our approach with other related ones and highlight similarities and differences existing among them. Table IV provides a summary of the in-depth comparison we carry out in the following subsections. A. Approach of [32] [32] proposes a recommender system devoted to support users looking for new job proposals. The system operates as follows: (i) it classifies each user according to his personal traits (e.g., shy/sociable or talktive/taciturn); (ii) a user queries it to retrieve new job proposals; (iii) it ranks retrieved job proposals according to their similarity to the personal traits of the user. There are some similarities between our approach and [32]; in fact, both of them: (i) are seeker oriented; (ii) can cooperate with existing e-recruitment Web sites; (iii) classify job proposals according to their relevance for the user. As for differences, we may notice that: (i) in [32] user profiles are constructed by means of psychological tools (e.g., questionnaires) whereas our system obtains information about users unobtrusively, by watching their interactions with it; (ii) in [32] no mechanisms for user profile update are provided; (iii) the architecture of [32] is centralized whereas our system is mixed, since the Recruitment Agent performs most of the activities connected with the selection of the job proposals best fitting user exigencies, but User Agents, Wrapper Agents and Company Agents continuously cooperate with the Re- cruitment Agent for performing the whole recruitment task (see Section II). Interestingly enough, our system might inherit some features typical of the approach of [32]; specifically, the user profiles managed in our system might include information about user personal traits, analogous to that handled in [32], to enhance the job search process. B. Approach of [33] In [33] the authors present a probabilistic e-recruitment system that uses both content-based and collaborative filter- ing techniques to produce recommendations. There are some similarities between our approach and that described in [33]; more specifically, both of them: (i) provide a mechanism for automatically and unobtrusively learning user preferences; (ii) take into account the user feedback that influences the system behaviour: specifically, in our approach, the knowledge of the job proposals accepted/rejected by the user is exploited for automatically tuning the system audacity; analogously, in [33], user preferences influence the coefficients of the probabilistic model adopted for deriving recommendations; (iii) are seeker oriented. As for differences, we may notice that: (i) [33] exploits a probabilistic model for deducing new recommendations; such an approach is very refined but it might be excessively time-consuming; (ii) our system is mixed whereas the approach of [33] is centralized; (iii) the approach of [33] is hybrid, whereas our own is content-based. C. CASPER In [28], [34] the system CASPER is described. CASPER is a client-server system that separately exploits collaborative filtering and content-based techniques to provide new job recommendations. There are some similarities between our approach and CASPER; in fact, both of them: (i) are seeker oriented; (ii) construct user profiles by unobtrusively monitor- ing user activities; (iii) follow similar criteria for estimating user preferences; (iv) are mixed. As for differences we may notice that: (i) CASPER can operate either as a content- based or as a collaborative filtering recommender system; on the contrary, our system is content-based; (ii) CASPER does not provide tools for modelling the semantics of a job proposal; on the contrary, our approach uses an XML Schema for representing the main features of each job proposal; these features play a key role during the recommendation process. D. Personnel Mall In [35] the Personnel Mall system is proposed. Personnel Mall is an agent based platform conceived for matching job seekers with companies in a distributed (and electronic) job marketplace. Specifically, Personnel Mall defines a set of rules to represent and manage the preferences/exigencies of job seekers and companies and uses them during the matching process. There are some similarities between our approach and Personnel Mall; in fact, both of them: (i) use the agent technology to perform e-recruitment activities; (ii) model job marketplace as a complex system in which heterogeneous components (i.e., companies and individuals) incessantly in- teract. As for differences, we may observe that: (i) Personnel Mall is company oriented; (ii) in Personnel Mall the matching between a job seeker and a job proposal depends not only on subjective variables (e.g., the personal interests of a user) but also on economic parameters (e.g., the quantity of labor that companies would like to hire for a certain wage and the quantity of labor that a job seeker would like to supply); (iii) Personnel Mall aims at maintaining the conditions of equilibrium in the job marketplace: as an example, it is capable of dynamically raising/lowering wages in response to labor surpluses/shortages; such a feature is not present in our system; IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 14 TABLE IV A COMPARISON OF SOME E-RECRUITMENT SYSTEMS System Purpose Architecture User Profile construction Additional Features Our approach Supjarerndee et al. [32] Farber et al. [33] CASPER [28], [34] Personnel Mall [35] Dafoulas et al. [36] CommOnCV [37] e-SRS [38] Seeker Oriented Seeker Oriented Seeker Oriented Seeker Oriented Company Oriented Company Oriented Seeker Oriented Company Oriented Mixed Centralized Centralized Mixed Distributed Centralized Centralized Distributed Unobtrusive Obtrusive Unobtrusive Unobtrusive Obtrusive Obtrusive Obtrusive Obtrusive It exploits Agent Technology and XML It considers the personal traits of a user It is an hybrid recommender system It separately exploits content based and collaborative filtering techniques It exploits the agent technology and a matching algorithm It allows online interviews to be performed It exploits Semantic Web languages to model a curriculum vitae It exploits both psychometric techniques and clustering algorithms to recruit candidates (iv) Personnel Mall does not provide mechanisms for updating the preferences and the exigencies of job seekers; (v) Personnel Mall is obtrusive; (vi) Personnel Mall is distributed; on the contrary, our system is mixed. each interest. However, it should be taken into account that the exploitation of a network, instead of a list, for representing user interests, would cause an increase of the computational complexity of our system. E. Approach of [36] G. e-SRS In [36] an agent based approach for supporting e- recruitment activities is proposed. This system embodies the functionalities typical of a traditional e-recruitment Web site (e.g., a user can insert/update his profile, a company can post new job proposals, and so on); moreover, it allows companies to perform an on-line interview of candidates; this interview is useful for both companies (that can perform a preliminary screening of candidates) and job seekers (that can enrich their profiles by storing their past interviews). It is possible to identify only few similarities between our approach and [36]; specifically, both of them: (i) define and manage user profiles; (ii) use the agent technology. As for differences we observe that: (i) [36] is company oriented; (ii) [36] provides a layered and centralized architecture whereas our approach is mixed; (iii) [36] is obtrusive. F. CommOnCV In [37] the authors propose CommOnCV, a system capable of representing and managing a curriculum vitae by means of Semantic Web tools. The approach implemented by Com- mOnCV is based on the concept of competency that is used to represent the knowledge/skills owned by a job seeker; competencies are defined and described by means of suitable ontologies. In addition, CommOnCV defines a curriculum vitae as a network of competencies and represents it by means of Se- mantic Web languages, like RDF/RDFs or DAML+OIL. There are some similarities between our approach and CommOnCV; in fact, both of them: (i) exploit ontologies to allow companies and job seekers to have a common reference for representing competencies and tasks; (ii) represent the contents of a job proposal or a curriculum vitae by means of suitable tools (i.e., XML in our approach, and Semantic Web languages in [37]); (iii) are seeker oriented. As for differences, we may observe that: (i) CommOnCV is centralized and obtrusive whereas our approach is mixed and unobtrusive; (ii) CommOnCV is not a recommender system. Our system might inherit some features from CommOnCV. Specifically, analogously to CommOnCV, it might represent a user profile as a network of interests, instead of a list of topics; such a choice would allow a better comprehension of the relationships existing among user interests, as well as a better evaluation of the relevance of In [38] a multi-agent platform, called e-SRS (e-Sales Re- cruitment System), for recruiting and benchmarking sales persons, is proposed. e-SRS exploits psychometric tools (e.g., questionnaires) for classifying each user in a specific category; it considers four user categories representing specific behav- ioral models (e.g., if a user rarely/frequently trusts others, if a user is/is not aggressive, and so on). This categorization can be improved by applying a suitable clustering algorithm (fuzzy K-means). The obtained results can be considered by a human expert for selecting the best candidates. There are some simi- larities between our approach and e-SRS. In fact, both of them: (i) exploit the agent technology and artificial intelligence tools to overcome the limitations of existing recruitment services; (ii) construct, handle and exploit accurate user profiles. As for differences, we may observe that: (i) e-SRS is company oriented; (ii) it is obtrusive; (iii) it is distributed, whereas our system is mixed; (iv) it is very specific, since it has been conceived for managing the recruitment of sales persons; on the contrary, our approach is generic, since it can support the recruitment activities in several professional areas; due to this reason, e-SRS produces accurate and refined results in its application area but its extension to other professional categories seems to require a valuable effort. e-SRS might be combined with our approach; specifically, in both the recruitment and the benchmark of sales persons, it might take into account not only the behavioural aspects of a job candidate but also his preferences and "constraints". VII. CONCLUSIONS AND HINTS FOR THE FUTURE In this paper we have presented an XML-based, multi- agent system for supporting e-recruitment services. We have seen that our system is characterized by various interesting capabilities, namely: (i) it allows a uniform management of heterogeneous job proposals; as a consequence, a job seeker can pose his queries across different, and presumably hetero- geneous, job databases; (ii) it is agent- and XML-based; as a consequence, it can easily cooperate with company informa- tion systems; (iii) the job proposal selection algorithm takes into account information about possible interests that the user did not consider in the past but that appear to be potentially interesting for him in the future; (iv) it is capable of handling IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 15 different recommendation strategies; (v) its recommendation strategies encompass some mathematic tools already applied in many fields. In the paper we have illustrated our system in detail, we have shown the obtained experimental results and, finally, we have compared it with other related systems already presented in the literature. Various application contexts might benefit from our research efforts. Some of the most promising ones are the following: • Team Building. In many organizations it is necessary to match individuals for composing a team working on the same project. Such an activity, often called team building or partnership building [39], is becoming an increasingly popular strategy for encouraging and boosting production. Team building has been studied in a large variety of scientific areas: for instance, Human Resource based approaches consider the skills and the abilities owned by each individual as a leading criterium to compose teams; analogously, sociological and psychological ap- proaches focus on the role of trust among individuals as a key element for the success of a team. Team building activities generally consist of two phases, namely: (i) identification of potential members of a team, and (ii) choice of individuals that really compose a team, among the previously defined potential candidates. As observed in [39], while various Web sites offer tools for effectively carrying out phase (i), decision support for phase (ii) is still rudimentary. In our opinion, our system might offer a concrete contribution in this area. Specifically, since our approach manages a large volume of data stored in user profiles, it can elaborate this information to compute the "affinity degree" existing among two individuals. In other words, an analysis of the personal features of an individual, as well as of its preferences and exigencies, allows managers: (i) to assign each individual to a team; (ii) to predict if trust can be established among the potential members of a team, even if they do not share a common work history. • Task Assignment. A classical activity of Human Resource managers consists of associating individuals to tasks; this activity should be carried out in such a way to exalt the skills and the capabilities of individuals. Since this activ- ity has a great impact on the business success, an exten- sive research on it has been performed in the past [40]. In our opinion our system could be used to effectively solve the problem of assigning individuals to tasks. Specifically, information stored in user profiles could be exploited for capturing user aptitudes/preferences whereas the XML- based model describing job proposals could be used for representing task features. A matching between a user profile and a job proposal description could be adopted for computing the so called Person-Job Fit (PJF), a pa- rameter used by psychologists to measure the suitability of a candidate for a specific job. • Outsourcing. In recent past businesses have outsourced a large variety of activities for improving the quality of services and products, for reducing the duration of the production cycle and for lowering costs. However, an excessive reliance on outsourcing may negatively affect business performance: for instance, it might lead to a reduced technological innovation [41] or to a limited ca- pability of competing with outsourcing partners [42]. Our system could be adopted to support business managers to decide about the effectiveness of an outsourcing activity; specifically, it might analyze the information stored in user profiles for assessing the capability of the internal components of an organization of executing a given task and to evaluate if it is/is not convenient its outsourcing. As for future work, we plan to contribute to the imple- mentation of the ideas mentioned previously. In addition, we plan to enrich our system by providing it with more powerful algorithms that try to predict the future interests of a user, as well as by designing and implementing modules for supporting companies looking for candidates. This last feature would allow the realization of a hybrid e-recruitment system, embodying the functionalities of both seeker oriented and company oriented systems. Finally, we plan to study the possibility to integrate our system with an e-learning framework for realizing a system capable of suggesting (and, next, teaching) to a user the skills he should acquire for accessing new, more appealing, job proposals. [1] "Bureau of labor REFERENCES statistics, U.S. Department of Labor." http://stats.bls.gov/opub/ted/2002/nov/wk1/art03.htm. [2] F. Liu, C. Yu, and W. Meng, "Personalized Web search for improving retrieval effectiveness," IEEE Transactions on Knowledge and Data Engineering, vol. 16, no. 1, pp. 28 -- 40, 2004. [3] P. Resnick and H. Varian, "Recommender systems," Communications of the ACM, vol. 40, no. 3, pp. 56 -- 58, 1997. [4] H. Lieberman, "Letizia: An agent that assists Web browsing," in Proc. of the International Joint Conference on Artificial Intelligence (IJCAI' 95). Montreal, Quebec, Canada: Morgan Kaufmann, 1995, pp. 924 -- 929. [5] M. He, N. Jennings, and H. Leung, "On agent-mediated electronic commerce," IEEE Transactions on Knowledge and Data Engineering, vol. 15, no. 4, pp. 985 -- 1003, 2003. [6] O. Zaiane, "Building a recommender agent for e-learning systems," in Proc. of the International Conference on Computers in Education (ICCE 2002). Auckland, New Zealand: IEEE Press, 2002, pp. 55 -- 59. [7] M. Wooldridge, An Introduction to MultiAgent Systems. Chichister: John Wiley & Sons, 2002. [8] R. Lawrence, G. Almasi, V. Kotlyar, M. Viveros, and S. Duri, "Person- alization of supermarket product recommendations," Data Mining and Knowledge Discovery, vol. 5(1-2), pp. 11 -- 32, 2001. [9] D. Gilsinn, H. Bandy, and A. Ling, "A spline algorithm for modeling cutting errors on turning centers," Journal of Intelligent Manufacturing, vol. 13, no. 5, pp. 391 -- 401, 2002. [10] S. Sadray, S. Rezaee, and S. Rezakhah, "Non-linear heteroscedastic regression model for determination of methotrexate in human plasma by high-performance liquid chromatography," Journal of Chromatography B, Analytical Technologies in the Biomedical and Life Sciences, vol. 787, no. 2, pp. 293 -- 302, 2003. [11] L. Yikang, "Low-pass filtered least squares estimators of cointegrating vectors," Journal of Econometrics, vol. 85, no. 2, pp. 289 -- 316, 1998. [12] A. Cesta and D. D'Aloisi, "Building interfaces as personal agents: a case study," ACM SIGCHI Bullettin, vol. 28, no. 3, pp. 108 -- 113, 1996. [13] C. Hsinchun, C. Yi-Ming, M. Ramsey, and C. Yang, "An intelligent per- sonal spider (agent) for dynamic Internet/intranet searching," Decision Support System, vol. 23, no. 1, pp. 41 -- 58, 1998. [14] B. Grosof and Y. Labrou, "An approach to using XML and a rule-based content language with an agent communication language," in Proc. of the IJCAI-99 Workshop on Agent Communication Language, Stockolm, Sweden, 1999, pp. 96 -- 117. Web [15] "World Consortium Wide - XML query," http://www.w3.org/XML/Query. [16] "Document Object Model," http://www.w3.org/DOM/. IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 37, NO. 4, 2007 16 [40] M. O'Connell, D. Doverspike, A. Cober, and J. Philips, "Forging work-teams: Effects of the distribution of cognitive ability on team performance," Applied Human Resource Management Research, vol. 6, pp. 115 -- 128, 2001. [41] M. Kotabe, Global sourcing strategy: R&D, manufacturing, and mar- keting interfaces. New York, United States: Quorum, 1992. [42] K. Gilley, C. Greer, and A. Rasheed, "Human resource outsourcing and organizational performance in manufacturing firms," Journal of Business Research, vol. 57, no. 3, pp. 232 -- 240, 2004. [17] N. Haque, N. Jennings, and L. Moreau, "Resource allocation in com- munication networks using market-based agents," International Journal on Knowledge Based Systems, vol. 18, no. 4 -- 5, pp. 163 -- 170, 2005. [18] R. Dash, D. Parkes, and N. Jennings, "Computational mechanism design: A call to arms," IEEE Intelligent Systems, vol. 18(6), pp. 40 -- 47, 2003. [19] R. Carreira, J. Crato, D. Goncalves, and J. Jorge, "Evaluating adaptive user profiles for news classification," in Proc. of the International Conference on Intelligent User Interface (IUI 2004). Funchal, Madeira, Portugal: ACM Press, 2004, pp. 206 -- 212. [20] A. Thor and E. Rahm, "AWESOME - a data warehouse-based system for adaptive website recommendations," in Proc. of the International Conference on Very Large Databases (VLDB 2004). Toronto, Canada: Morgan Kaufmann, 2004, pp. 384 -- 395. [21] J. Wahba and Y. Zenou, "Density, Social Networks and Job Search Methods: Theory and Application to Egypt," The Research Insti- tute of Industrial Economics, Tech. Rep. 629, 2004, available at http://ideas.repec.org/p/hhs/iuiwop/0629.html. [22] J. Han and M. Kamber, Data Mining: Concepts and Techniques. Morgan Kaufmann Publishers, 2001. [23] R. Baeza-Yates and B. Ribeiro-Neto, Modern Information Retrieval. Addison Wesley Longman, 1999. [24] G. Golub and C. V. Loan, Matrix Computations, third edition. Hopkins University Press, 1996. Johns [25] V. Pan, "Solving a Polynomial Equation: Some History and Recent Progress," SIAM Review, vol. 39, no. 2, pp. 187 -- 220, 1997. [26] D. Bertsekas, Nonlinear Programming. Belmont, Massachusets, USA: Athena Scientific, 1999. [27] FIPA, "Foundation for Intelligent Physical Agents Specifications," http://www.fipa.org. [28] K. Bradley and B. Smyth, "Personalized information ordering: a case study in online recruitment," Knowledge-Based Systems, vol. 16, no. 5-6, pp. 269 -- 275, 2003. [29] V. Hristidis, L. Gravano, and Y. Papakonstantinou, "Efficient IR-style keyword search over relational databases," in Proc. of the International Conference on very large Databases (VLDB 2003), Berlin, Germany, 2003, pp. 850 -- 861. [30] S. Newell, "User models and filtering agents for improved internet information retrieval," User Modeling and User-Adapted Interaction, vol. 7, no. 4, pp. 223 -- 237, 1997. [31] H. Jagadish, S. Al-Khalifa, A. Chapman, L. Lakshmanan, A. Nierman, S. Paparizos, J. Patel, D. Srivastava, N. Wiwatwattana, Y. Wu, and C. Yu, "TIMBER: A native XML database," The VLDB Journal, vol. 11, no. 4, pp. 274 -- 291, 2002. [32] S. Supjarerndee, Y. Temtanapat, and U. Phalavonk, "Recruitment filter- ing with personality-job fit model," in Proc. of the IEEE International Conference on Information Technology: Coding and Computing (ITCC 2002). Las Vegas, Nevada, United States: IEEE Computer Society, 2002, pp. 46 -- 53. [33] F. Farber, T. Keim, and T. Weitzel, "An automated recommendation approach to personnel selection," in Proc. of the Americas Conference on Information Systems (AMCIS'2003). Tampa, Florida, United States: Association for Information System Electronic Library, 2003, pp. 2329 -- 2339. [34] R. Rafter, K. Bradley, and B. Smyth, "Automated collaborative filtering applications for online recruitment services," in Proc. of the Interna- tional Conference on Adaptive Hypermedia and Adaptive Web-Based Systems (AH 2000). Trento, Italy: Lecture Notes in Computer Science, Springer Verlag, 2000, pp. 363 -- 368. [35] W. Gates and M. Nissen, "Designing agent-based electronic employment markets," Electronic Commerce Research, vol. 1, no. 3, pp. 239 -- 263, 2001. [36] G. Dafoulas, A. Nikolaou, and M. Turega, "E-Services in the Internet Job Market," in Proc. of the Hawaii International Conference on System Sciences (HICSS '03). Big Island, Hawaii, United States: IEEE Computer Society, 2003. [37] M. Harzallah, M. Leclere, and F. Trichet, "CommOnCV: modelling the competencies underlying a curriculum vitae," in Proc. of the Interna- tional Conference on Software Engineering and Knowledge Engineering (SEKE'02). Ischia, Italy: ACM Press, 2002, pp. 65 -- 71. [38] R. Khosla and T. Goonesekera, "An online multi-agent e-sales recruit- ment system," in Proc. of the IEEE/WIC International Conference on Web Intelligence (WI 2003). Halifax, Canada: IEEE Computer Society Press, 2003, pp. 111 -- 117. [39] T. Keim and T. Weitzel, "An integrated approach to online partnership building," in Proc. of the Hawaiian International Conference on System Sciences (HICSS '05). Big Island, Hawaii, United States: IEEE Computer Society Press, 2005.
1902.01024	1	1902	2019-02-04T03:43:47	Cooperative Driving at Unsignalized Intersections Using Tree Search	[ "cs.MA" ]	In this paper, we propose a new cooperative driving strategy for connected and automated vehicles (CAVs) at unsignalized intersections. Based on the tree representation of the solution space for the passing order, we combine Monte Carlo tree search (MCTS) and some heuristic rules to find a nearly global-optimal passing order (leaf node) within a very short planning time. Testing results show that this new strategy can keep a good tradeoff between performance and computation flexibility.	cs.MA	cs	Cooperative Driving at Unsignalized Intersections Using Tree Search Huile Xu, Yi Zhang, Member, IEEE, Li Li, Fellow, IEEE, and Weixia Li 1 9 1 0 2 b e F 4 ] A M . s c [ 1 v 4 2 0 1 0 . 2 0 9 1 : v i X r a Abstract -- In this paper, we propose a new cooperative driv- ing strategy for connected and automated vehicles (CAVs) at unsignalized intersections. Based on the tree representation of the solution space for the passing order, we combine Monte Carlo tree search (MCTS) and some heuristic rules to find a nearly global-optimal passing order (leaf node) within a very short planning time. Testing results show that this new strategy can keep a good tradeoff between performance and computation flexibility. Index Terms -- Connected and Automated Vehicles (CAVs), cooperative driving, unsignalized intersection, Monte Carlo tree search (MCTS). I. INTRODUCTION C ONNECTED and Automated Vehicles (CAVs) are be- lieved to be a key role in the next-generation transporta- tion systems [1]. With the aid of vehicle-to-vehicle (V2V) communication, CAVs can share their driving states (position, velocity, etc.) and intentions with adjacent vehicles [2], [3] to better coordinate their motions to alleviate traffic congestion and improve traffic safety. In the last decade, various strategies had been proposed to make optimal coordination for CAVs at a typical driving scenario: unsignalized intersection. It is pointed out in [4] and [5] that the key problem is to determine the optimal order of CAVs that passed the intersection. As summarized in [6], there are two kinds of cooperative driving strategies, planning based and ad hoc negotiation based, for determining the passing order. Planning based strategies aim to enumerate all possible passing orders to find the globally optimal solution [7]. There are two equivalent formulations of the problem. Most state-of- the-art studies formulate the problem as a mixed integer linear programming problem of vehicles' passing time scheduling [1], [8]. The objective is usually set to minimize the total delay of all CAVs. Li et al. showed that we can also view this problem as a tree search problem. Each tree node indicates a Manuscript received in February 4, 2019; This work was supported in part by National Natural Science Foundation of China under Grant 61673233, 61790565, 61603005, and the Beijing Municipal Commission of Transport Program under Grant ZC179074Z. (Corresponding author is Li Li). H. Xu is with Department of Automation, Tsinghua University, Beijing 100084, China. (E-mail: [email protected]) Y. Zhang is with Department of Automation, BNRist, Tsinghua University, Beijing 100084, China and also with Berkeley Shenzhen Institute (TBSI), Tower C2, Nanshan Intelligence Park 1001, Xueyuan Blvd., Nanshan District, Shenzhen 518055, China. (E-mail: [email protected]) L. Li is with Department of Automation, BNRist, Tsinghua University, Bei- jing 100084, China. (Tel:+86(10)62782071, E-mail: [email protected]). W. Li is with Department of Automation, Tsinghua University, Beijing 100084, China. (E-mail: [email protected]) special (partial) passing order. The equivalent objective is to find the leaf node corresponds to the minimum total delay of all CAVs [4], [5]. It was shown in [9] that some planning based strategies work well for ramp metering scenarios. However, the time to enumerate all the nodes increases sharply as the number of vehicles increases, especially for unsignalized intersection scenarios. This problem hinders their applications in practice. Ad hoc negotiation based strategies aim to find an accept- able passing order using some heuristic rules within a very short time. For example, Stone et al. proposed autonomous intersection management (AIM) cooperative driving strategy which divides the intersection into grids (resources) and assigns these grids to CAVs in a roughly First-In-First-Out (FIFO) manner [10], [11]. This strategy has several variations, including reservation strategy [12]. However, as shown in [6], the passing orders found by ad hoc negotiation based strategies were not good enough in many situations. To keep a good tradeoff between performance and computa- tion flexibility, we propose a new cooperative driving strategy based on the tree representation of the solution space for the passing order. Its key idea is to use the limited planning time to explore the nodes that are potential to be the optimal solution. To this end, we combine Monte Carlo tree search (MCTS) and some heuristic rules to accelerate the searching process, since the solution space of this problem has special structures to be exploited. Testing results show that we can find a nearly global-optimal passing order within a short enough planning time. To give a better presentation of our finding, the rest of this paper is arranged as follows. Section II formulates the problem and briefly reviews the existing strategies. Section III presents the new strategy. Section IV validates the effectiveness of the proposed strategy via numerical testing results. Finally, Section V gives concluding remarks. II. PROBLEM FORMULATION Fig. 1 shows a typical intersection scenario with multiple lanes in each leg. The area within the circle is called the control zone, and the shadow area is called the conflict zone where lateral collisions might happen. According to the geometry of the intersection, the conflict zone can be further divided into several conflict subzones. We assign each vehicle that enters the control zone a unique to denote the conflict identity Vi. We also use the set Zi subzones that Vi will pass through. For example, Zi = {4, 1} means Vi will pass through Conflict Subzone 4 and Conflict Subzone 1 in sequence. 2 Fig. 2. An intersection scenario with 5 vehicles. indicates the first vehicle in a special passing order. The nodes in the third layer refer to one string consisting of two indices symbols that indicate the first two vehicles in a special passing order. Similarly, the child nodes expand their child nodes, and all possible passing order are generated as leaf nodes in the bottom layer of the solution tree as shown in Fig. 3. Fig. 3. The solution tree stemmed from the intersection scenario shown in Fig. 2. The leaf nodes in the bottom layer represent the complete passing orders for all vehicles. If a (partial) passing order is given, the desired arrival times for all the vehicles that has been covered in this (partial) passing order can be directly derived by the following Passing Order to Trajectory Interpretation Algorithm. Our objective turns to seek the leaf node that corresponds to the shortest total delay. Moreover, the total delay values of leaf nodes can be used to evaluate the potential of their parent nodes in a backpropagation way. This method provides us a chance to find a nearly global-optimal leaf node but only search a small part of the whole tree. In Algorithm 1, P (i) is the ith element in the input (partial) passing order, tmax,z is the largest arrival time that the subzone Fig. 1. A typical intersection scenario. To simplify the problem, we adopt the following assump- tions: • Each vehicle instantly and thoroughly shares its driving states (position, velocity, etc.) and intentions with other vehicles via vehicle-to-vehicle (V2V) communication. • Changing lane maneuver is prohibited in the control zone to ensure vehicle safety. • Similar to [12] and [13], the velocities of vehicles are constant when passing through the conflict zone. The cooperative driving strategy aims to minimize the total delay of vehicles by scheduling the velocity and acceleration profiles of all vehicles [14]. So, we can get the following optimization problem n(cid:88) (tassign,i,Zi(1) − tmin,i,Zi(1)) min J = (1) i=1 where tassign,i,z is the desired arrival time to the conflict subzone z for Vi, tmin,i,z is the minimum arrival time to the conflict subzone z when Vi travels at the maximum velocity and the maximum acceleration, Zi(1) is the first element in the set Zi, n is the number of vehicles in the control zone. To directly attack Problem (1) often leads to a mixed integer linear programming (MILP) problem whose computation time increases exponentially with the increase of the number of vehicles [8], [9]. Noticing that the traffic efficiency mainly depends on the passing order of vehicles [6], we can formulate the whole problem as a tree search problem in the solution space that consists of all possible passing orders. Each leaf node repre- sents a passing order of vehicles which can also be denoted as a string [5]. For example, string CAB means vehicle C, vehicle A, and vehicle B enter the conflict zone sequentially. Let us take the intersection scenario shown in Fig. 2 as an example to explain how to build the tree representation of the solution space gradually . At first, we set the passing order in the root node to be empty. Then, each direct child node of the root node (in the second layer) refers to one index symbol that Lane 1Lane 2Lane 3Lane 4Lane 5Lane 6Lane 7Lane 8Lane 9Lane 10Lane 11Lane 12Control ZoneConflict ZoneADCEB1234RootABCDC AC BC DC EC B AC B DC B EC B D AC B D EC B D A EC B D E A Algorithm 1 Passing Order to Trajectory Interpretation Input: A (partial) passing order P Output: The total delay J of the covered vehicles and their arrival times tassign for each z ∈ Zi do 1: for each i ∈ [1, length(P )] do 2: 3: 4: 5: 6: Vj is the last vehicle that passed through subzone z tassign,P (i),z = max(tmin,P (i),z, tmax,z + ∆j,a) end for Adjust tassign,P (i),z according to the constraint: the velocity of Vi in the conflict zone is constant. for each z ∈ Zi do tmax,z = tassign,P (i),z 7: 8: 9: 10: end for end for 11: J =(cid:80)length(P ) i=1 tassign,P (i),Zi(1) z has been occupied. ∆j,a is the minimum safety gap between two consecutive vehicles passing through the same subzone. the time complexity of Algorithm 1 is O(n). Obviously, A detailed explanation of Algorithm 1 can be found in our previous report [7]. III. MCTS BASED COOPERATIVE DRIVING STRATEGIES It is usually impossible to expand all the nodes of the solution tree within the limited computation budget, when there are lots of vehicles in the control zone. In this paper, we use MCTS + heuristic rules to select nodes with the potential to be the optimal solution. The recent success of the MCTS method in the game of Go shows it is an effective way to deal with such problems [15], [16]. A. The Classical MCTS Based Strategy In MCTS, each node in the formulated tree will be assigned a score to evaluate its potential. The score of a leaf node is equal to the total delay of its corresponding passing order. MCTS uses these scores to determine which branch of tree to explore. Generally, MCTS gradually builds a search tree in an iteration way. One iteration consists of four steps: selection, expansion, simulation, and backpropagation [17]; see Fig. 4. 1) Selection: Starting at the root node, we select the most urgent expandable node based on the following policy [18] arg max i Qi + C (2) where Qi is the score of child node i and the value of Qi is within [0, 1]. n is the number of times the current node has been visited, ni is the number of times child node i has been visited, and C is a weighting parameter. The child node with the largest total score is selected. (cid:114) ln n ni 3 Here, an expandable node refers to a node that is not a leaf node and has unvisited child nodes. This child node selection policy is suggested in the field of computer Go and is called UCB1 [18]. The first term in the equation encourages to select the child node that is currently believed to be optimal, while the second term encourages to explore more child nodes. 2) Expansion: We randomly select one unvisited child node of the most urgent expandable node to be a new node that is added to the tree. 3) Simulation: We run several rollout simulations to de- termine a complete passing order based on the partial passing order represented by the current new node to evaluate the potential of the new node. The classical MCTS randomly samples and adds the uncovered vehicles into the passing order string one by one, until we find a complete passing order string and reach the maximum depth of the tree from the current new node without branching [16]. For example, when we apply random sampling policy to the node CB shown in Fig. 4, we can randomly expand a direct child node in its next layer; say node CBA. The node CBA will be further expanded by repeating such a process until a leaf node (e.g., node CBADE) is reached. Finally, the partial passing order will be evaluated by all its simulated off-spring leaf nodes (passing orders). Sometimes, the generated passing order is not invalid, because it may violate the prohibition of lane change, such solutions will be discarded after check. After simulation, we update the scores of the current new node as follows: i Apply Algorithm 1 to calculate the total delay ¯Ji of the partial order corresponds with the current new node. ii Apply Algorithm 1 to calculate the total delay Ji of the partial order corresponds with the best off-spring node of the current new node via simulation. iii Calculate the score Qi of the current new node as Qi = ω ¯Ji + (1 − ω) Ji (3) where ω is a weighting parameter. Since Qi ∈ [0, 1], we normalize ¯Ji and Ji into [0, 1] before updating Qi. 4) Backpropagation: The simulation result is backpropa- gated through the selected nodes to update the scores of all its parent nodes. During the building process of the search tree, the state-of- the-art best passing order is continuously updated. As soon as the computation budget is reached, the search terminates and returns the state-of-the-art best passing order. The planned ar- rival times of vehicles can be determined by using Algorithm 1. The velocity and acceleration profiles of each vehicle plan will be finally calculated by using the motion planning method proposed in [14]. We can see that the performance of the proposed strategy is influenced by the choice of the parameters including the maximum search time and two weighting parameters C and 4 Algorithm 2 Heuristic Simulation Policy Input: Locations and velocities of all vehicles Output: A possible passing order 1: Among all uncovered vehicles, we select the vehicles which are the closest to the conflict zone in each lane as candidate vehicles and calculate their arrival times to all conflict subzones. 2: If there exists a candidate vehicle whose arrival times to all conflict subzones are all the smallest, we add it into the passing order string. If not, we randomly select one vehicle among all candidate vehicles and add it into the passing order string. 3: Then we repeat the steps 1 and 2 until a complete passing order string is generated. 4: The objective value (1) of the generated passing order can be easily derived by Algorithm 1 and denoted as q2,i. is assumed to be a Poisson process. We vary the mean value of this Poisson process to test the performance of the proposed strategy under different traffic demands. The vehicles arrival rates at all lanes are the same unless otherwise specified. It should be pointed out that we had tested other intersections with different road geometries and various vehicle arrival patterns, but the conclusions remain unchanged. To accurately describe the total delays of vehicles, we adopt the point-queue model in the simulation [6], [19]. The model assumes vehicles travel in free flow state until it gets to the boundary of the intersection we study. If the preceding vehicle leaves enough spaces, the first vehicle in the point-queue will dequeue and enter the intersection. Otherwise, it will stay in the virtual queue. Each lane has an independent point-queue. In this paper, we reschedule the passing order of all the vehicles within the control zone every 2 seconds. As sug- gested in [6], we set the minimum safety gap between two consecutive vehicles passing through the same subzone as a slightly enlarged constant as 1.5s a = 1 2s a = 2 1.5s a = 3 ∆j,a = (go straight) (turn left) (turn right) (4) to avoid the collisions caused by position measurement errors and communication delay. Fig. 4. One iteration of the MCTS based cooperative driving strategy. ω. We will discuss how to choose these parameters in Section IV below. B. The MCTS + Heuristic Rules As aforementioned, the classical MCTS strategy uses ran- dom sampling to generate a leaf node (a passing order) in the simulation step. However, because of the huge number of possible passing orders, the passing orders generated by random sampling cannot help us quickly capture the real potential of a node during simulation. Thus, we propose the following heuristic rules to help decide which nodes (vehicles) should be expanded (added into the candidate passing order string) during simulation. Heuristic rule 1 helps to quickly prune the invalid passing order [5]. Heuristic rule 2 determines the vehicle among the candidates to be chosen. 1) For the vehicles on the same lane, the vehicle which is the closest to the conflict zone should be added earlier than other vehicles since changing lane maneuver is prohibited. 2) For the vehicles passing through the same conflict sub- zone, the vehicle with a less desired arrival time should be added earlier. The simulation step can be summarized as Algorithm 2. We can see that the classic MCTS applies random sampling in both expansion and simulation steps; while our MCTS + heuristic rules applies random sampling only in expansion step. The Ad hoc negotiation based strategies organize all the vehicles according to the FIFO principle. In contrast, the new simulation policy tends to organize just a part of vehicles (the vehicles uncovered in the current partial passing order) according to the FIFO strategy. This trick helps to avoid the convergence to a over-greedy solution. IV. SIMULATION RESULTS A. Simulation Settings We design three experiments to determine the best parame- ter set for the new cooperative driving strategy and compare it with some classical ones. These experiments are conducted for the intersection with three lanes in each leg shown in Fig. 1. The mandatory signs stipulate the permitted directions for each lane. According to the geometry of the intersection, the conflict zone is further divided into 36 subzones. The vehicles arrival RootABCDC ARootABCDC AC BRootABCDC AC BA Leaf Node(A Passing Order)RootABCDC AC BSelectionExpansionSimulationBackpropagationSimulation PolicyB. The Choice of Parameters 5 In this paper, we consider two performance indices: the delay J of the given n vehicles and the traffic throughput (the number of vehicles that has passed the intersection control zone) within a given time interval to compare different coop- erative driving strategies. Specially, we highlight the decreased ratio of the total delay if being compared withe baseline solution that is gotten by the FIFO strategy JF IF O − JM CT S (5) η = JF IF O where JF IF O is the objective value of the FIFO passing order, and JM CT S is the objective value of the best passing order from the MCTS based strategy. To determine the best parameter setting of the new MCTS + heuristic rules, we first fix the time budget as 0.1 s and vary ω and C from 0 to 1. To better understand the performance of the strategy under different traffic conditions, we vary the vehicle arrival rate to generate a series of intersection scenarios with different number of vehicles. Fig. 6. The results of the improvement rates with respect to the number of searched nodes for the intersection shown in Fig.1. through searching 1000 nodes. For most intersection scenarios, 1000 nodes can be searched within 0.1 s in our personal computer, so we set the maximum search time as 0.1 s for the following experiments. C. Comparisons of Different Cooperative Driving Strategies To further clarify the difference between the FIFO strategy and our new strategy, we study a typical intersection scenario with single lane in each leg and 20 vehicles. We calculate the objective values for all the valid solutions (passing orders) and plot them in a histogram manner; see Fig. 1. Fig. 5. The improvement rate of the MCTS based strategy with different parameter settings for the intersection shown in Fig.1. Fig. 5 gives the improvement rates for the intersection shown in Fig.1 with 30 vehicles. We can see that a significant improvement can be achieved even with the worst parameter setting. The parameter C and ω are not so critical but may still influence the balance between exploitation and exploration, partly because we use heuristic rules in simulation step to reduce the influence of random sampling. We further study the scenarios with other numbers of vehicles and the results are all similar. Thus, in the rest of this paper, we set ω = 0.85 and C = 0.05. Then, to determine an appropriate time budget, we vary the time limits of tree search. To eliminate the influence of the computing power of the device, we examine the improvement rates with respect to the number of nodes that has been searched. It can be seen from Fig. 6 that the improvement rate increases significantly when the number of searched nodes increases from 10 to 1000. However, the improvement rate soon becomes saturated after that. Thus, we believe that the proposed strategy can obtain a good enough passing order Fig. 7. The histogram of all solution values for a single lane intersection scenario with 20 vehicles. It is clear that the solution found by the MCTS based strategy is nearly the same as the global optimal solution found by the enumeration based strategy, while the computation time of the MCTS based strategy is much less. For the FIFO based strategy, the computation time is the least, but the solution is far away from the optimal solution. The solution found by the MCTS based strategy ranks 648th in the nearly 10 billion solutions; while the solution of the FIFO based strategy ranks 4563421793th. We then carry out another comparison for the intersection shown in Fig. 1, where the average arrival rate is varied to 0.15510.160.1650.5010.90.80.70.60.50.40.30.20.10Best Parameter Setting3003504004505005500.20.40.60.811.21.41.61.8Probability (%)31032033034000.050.10.15explore the influence of different traffic demands. For each arrival rate, we simulate 20-minute. It is obvious that our new strategy further reduces the average delay and improves the traffic throughput in all situations. COMPARISON RESULTS OF DIFFERENT COOPERATIVE DRIVING TABLE I STRATEGIES scenario and find a nearly optimal solution with a short time. Although we only discuss the schedule of vehicles at unsignalized intersections, this method can be easily adapted to other scenarios (e.g., ramping areas and working zones). We are currently building several automated vehicle prototypes so that we can test our new strategy in field studies in the near future. 6 Arrival rate veh/(lane*h) Strategies Average delay (s) Traffic Throughput 150 (veh) 589 605 1095 1168 1205 1766 ∗ The computation time of the MCTS based strategy is 0.1s. 1.3053 0.4499 39.8313 1.1407 41.6996 4.8743 FIFO MCTS FIFO MCTS FIFO MCTS 300 450 D. A Further Look into the Structure of the Obtained Search Tree Fig. 8 shows the formulated search tree of our new strategy for an intersection scenario with 50 vehicles. Similar to the classical MCTS strategy, our new strategy tends to first find some promising branches (partial passing orders) of the tree and spends most search time to further explore these branches. However, the search tree generated by the classical MCTS strategy contains much more unnecessary leaf nodes. In contrast, when the heuristic rules are introduced, only a very small number of leaf nodes will be finally reached. For this case, although there are more than 1046 possible passing orders, only about two thousands passing orders are explored by our new strategy within 0.1 s. This difference explains why the classical MCTS needs much more time to find a good enough passing order. Fig. 8. The structure of a search tree for an intersection scenario with 50 vehicles. V. CONCLUSION In this paper, we propose a cooperative driving strategy that combines Monte Carlo simulation and heuristic rule simulation to accelerate the search of the passing order. This new method can quickly learn the tree structure knowledge of the given REFERENCES [1] P. T. Li and X. Zhou, "Recasting and optimizing intersection automa- tion as a connected-and-automated-vehicle (cav) scheduling problem: A sequential branch-and-bound search approach in phase-time-traffic hypernetwork," Transportation Research Part B: Methodological, vol. 105, pp. 479 -- 506, 2017. [2] L. Li, D. Wen, and D. Yao, "A survey of traffic control with vehicular communications," IEEE Transactions on Intelligent Transportation Sys- tems, vol. 15, no. 1, pp. 425 -- 432, 2014. [3] T. Sukuvaara and P. Nurmi, "Wireless traffic service platform for com- bined vehicle-to-vehicle and vehicle-to-infrastructure communications," IEEE Wireless Communications, vol. 16, no. 6, 2009. [4] S. I. Guler, M. Menendez, and L. Meier, "Using connected vehicle technology to improve the efficiency of intersections," Transportation Research Part C: Emerging Technologies, vol. 46, pp. 121 -- 131, 2014. [5] L. Li and F. Wang, "Cooperative driving at blind crossings using in- tervehicle communication," IEEE Transactions on Vehicular technology, vol. 55, no. 6, pp. 1712 -- 1724, 2006. [6] Y. Meng, L. Li, F. Wang, K. Li, and Z. Li, "Analysis of cooperative driving strategies for nonsignalized intersections," IEEE Transactions on Vehicular Technology, vol. 67, no. 4, pp. 2900 -- 2911, 2018. [7] H. Xu, S. Feng, Y. Zhang, and L. Li, "A grouping based coop- erative driving strategy for cavs merging problems," arXiv preprint arXiv:1804.01250, 2018. [8] E. R. Muller, R. C. Carlson, and W. K. Junior, "Intersection control for automated vehicles with milp," IFAC-PapersOnLine, vol. 49, no. 3, pp. 37 -- 42, 2016. [9] L. Chen and C. Englund, "Cooperative intersection management: a survey," IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 2, pp. 570 -- 586, 2016. [10] K. Dresner and P. Stone, "Multiagent traffic management: A reservation- based intersection control mechanism," in Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems-Volume 2. IEEE Computer Society, 2004, pp. 530 -- 537. [11] -- -- , "A multiagent approach to autonomous intersection management," Journal of artificial intelligence research, vol. 31, pp. 591 -- 656, 2008. [12] M. Choi, A. Rubenecia, and H. H. Choi, "Reservation-based cooperative traffic management at an intersection of multi-lane roads," in Information Networking (ICOIN), 2018 International Conference on. IEEE, 2018, pp. 456 -- 460. [13] Y. Zhang, A. A. Malikopoulos, and C. G. Cassandras, "Decentralized optimal control for connected automated vehicles at intersections includ- ing left and right turns," arXiv preprint arXiv:1703.06956, 2017. [14] A. A. Malikopoulos, C. G. Cassandras, and Y. J. Zhang, "A decentralized energy-optimal control framework for connected automated vehicles at signal-free intersections," Automatica, vol. 93, pp. 244 -- 256, 2018. [15] M. Enzenberger, M. Muller, B. Arneson, and R. Segal, "Fuegołan open- source framework for board games and go engine based on monte carlo tree search," IEEE Transactions on Computational Intelligence and AI in Games, vol. 2, no. 4, pp. 259 -- 270, 2010. [16] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton et al., "Mastering the game of go without human knowledge," Nature, vol. 550, no. 7676, pp. 354 -- 359, 2017. [17] C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P. I. Cowling, P. Rohlfshagen, S. Tavener, D. Perez, S. Samothrakis, and S. Colton, "A survey of monte carlo tree search methods," IEEE Transactions on Computational Intelligence and AI in games, vol. 4, no. 1, pp. 1 -- 43, 2012. [18] L. Kocsis and C. Szepesv´ari, "Bandit based monte-carlo planning," in European conference on machine learning. Springer, 2006, pp. 282 -- 293. [19] X. J. Ban, J. Pang, H. X. Liu, and R. Ma, "Continuous-time point-queue models in dynamic network loading," Transportation Research Part B: Methodological, vol. 46, no. 3, pp. 360 -- 380, 2012.
1609.08116	2	1609	2017-03-06T23:07:01	Predictive Positioning and Quality Of Service Ridesharing for Campus Mobility On Demand Systems	[ "cs.MA" ]	Autonomous Mobility On Demand (MOD) systems can utilize fleet management strategies in order to provide a high customer quality of service (QoS). Previous works on autonomous MOD systems have developed methods for rebalancing single capacity vehicles, where QoS is maintained through large fleet sizing. This work focuses on MOD systems utilizing a small number of vehicles, such as those found on a campus, where additional vehicles cannot be introduced as demand for rides increases. A predictive positioning method is presented for improving customer QoS by identifying key locations to position the fleet in order to minimize expected customer wait time. Ridesharing is introduced as a means for improving customer QoS as arrival rates increase. However, with ridesharing perceived QoS is dependent on an often unknown customer preference. To address this challenge, a customer ratings model, which learns customer preference from a 5-star rating, is developed and incorporated directly into a ridesharing algorithm. The predictive positioning and ridesharing methods are applied to simulation of a real-world campus MOD system.A combined predictive positioning and ridesharing approach is shown to reduce customer service times by up to 29% and the customer ratings model is shown to provide the best overall MOD fleet management performance over a range of customer preferences.	cs.MA	cs	Predictive Positioning and Quality Of Service Ridesharing for Campus Mobility On Demand Systems Justin Miller and Jonathan P. How 7 1 0 2 r a M 6 ] A M . s c [ 2 v 6 1 1 8 0 . 9 0 6 1 : v i X r a Abstract- Autonomous Mobility On Demand (MOD) systems can utilize fleet management strategies in order to provide a high customer quality of service (QoS). Previous works on autonomous MOD systems have developed methods for rebalancing single capacity vehicles, where QoS is maintained through large fleet sizing. This work focuses on MOD systems utilizing a small number of vehicles, such as those found on a campus, where additional vehicles cannot be introduced as demand for rides increases. A predictive positioning method is presented for improving customer QoS by identifying key locations to position the fleet in order to minimize expected customer wait time. Ridesharing is introduced as a means for improving customer QoS as arrival rates increase. However, with ridesharing perceived QoS is dependent on an often unknown customer preference. To address this challenge, a customer ratings model, which learns customer preference from a 5-star rating, is developed and incorporated directly into a ridesharing algorithm. The predictive positioning and ridesharing methods are applied to simulation of a real-world campus MOD system. A combined predictive positioning and ridesharing approach is shown to reduce customer service times by up to 29%. and the customer ratings model is shown to provide the best overall MOD fleet management performance over a range of customer preferences. I. INTRODUCTION Mobility On Demand (MOD) systems have the potential to revolutionize transportation systems in urban settings by providing commuters access to vehicles without requiring private ownership. In such systems, a fleet of shared vehicles continually services multiple customers by transporting them from their requested on demand pickup location to their desired destination. It is estimated that by 2030, as much as 26% of all global miles traveled will be from customers using shared vehicles [1]. A fundamental challenge for MOD systems is providing a high customer quality of service (QoS) in order to minimize any drawbacks that customers may experience by relying on the shared resources. There are many factors that can affect a customer's QoS such as cost, comfort, safety and convenience [2], many of which can be improved through the use of autonomous MOD systems composed of self-driving vehicles [3]. Several autonomous MOD fleet management strategies have been introduced as a means of improving QoS. These approaches attempt to find a "rebalancing" policy that redistributes vehicles within the system based on customer demand using either a fluid model approach [4], a Markov transition model approach [5], a queueing-theoretical model approach [6], or a model predictive control approach [7]. The approaches Laboratory of Information and Decision Systems, Massachusetts Institute of Technology, 77 Massachusetts Ave., Cambridge, MA, USA {justinm, jhow}@mit.edu (a) MIT MOD fleet (b) Rating app Fig. 1: (a) A fleet of three electric shuttles operate as MOD vehicles on MIT campus. (b) MIT MOD customers have the ability to provide app-based feedback in the form of a 5-star rating. were developed to operate on city-wide scales and assume that an appropriately large fleet of autonomous vehicles is available, an assumption that does not yet reflect the current state of real-world autonomous MOD systems. In practice, autonomous MOD systems are being deployed with limited sized fleets [8]–[10]. This work is further motivated by our own MOD system operating on MIT campus, where autonomous fleet management strategies are applied to a fleet of only three human-driven vehicles, shown in Figure 1a. Previous rebalancing methods do not scale well when applied to such small fleets. For example, the policies are designed to redistribute multiple excess vehicles to relatively few network arrival locations. The methods breakdown when the reverse is true and fleet sizes are relatively small compared to the number of network locations such that there generally aren't enough vehicles to cover all arrival locations. This work seeks to address this challenge by instead identifying key locations in the MOD system which minimize expected customer wait times regardless of fleet size. Additionally, those previous works only consider the use of single capacity vehicles and do not address the benefits and challenges of utilizing ridesharing. In ridesharing, multiple customers may share a ride in an MOD vehicle at the same time. Newly arrived passengers can be picked up before onboard passengers have been dropped off, allowing for more customers to be serviced with fewer vehicles. However, the reduced wait time for requested passengers comes at the expense of increased ride time of onboard passengers. With ridesharing, perceived QoS becomes dependent on customer preference (i.e. how much customers prefer one service metric over another). A large body of work has studied a form of ridesharing known as the Dial-A-Ride Problem (DARP), which is a specialization of the Vehicle Routing Problem formulated specifically for the transportation of customers. DARP problem formulations typically take either the form of an integer program or a scheduling problem. Integer program formulations encode each customer QoS metric as a decision variable and use heuristic methods such as genetic algorithms [11], simulated annealing [12], and tabu search [13] to minimize an objective, such as a cost function composed of a weighted sum of the metrics. Scheduling problem formulations enumerate the possible ways of inserting new passengers into vehicle schedules, and encode customer QoS metrics as feasibility constraints [14]–[16]. The main challenge with all of these approaches is that the weights or constraint thresholds that define customer preference may be chosen incorrectly and even the structure that encodes the QoS metrics could be wrong. In general determining customer preference can be difficult. However, many MOD systems such as Uber, Lyft, and the MIT MOD system are able to query passengers for feedback in the form of a 5-star rating using a ride request app, as shown in Figure 1b. This work takes advantage of this available information through a ridesharing algorithm which does not encode customer preference directly, but rather utilizes a learned customer ratings model when solving the DARP scheduling problem. The contributions of this work are: 1) a predictive posi- tioning approach that minimizes expected customer wait time in MOD networks with fewer vehicles than customer arrival locations; 2) a customer rating model which learns from 5- star rating feedback and serves as a customer QoS-focused cost function; and 3) a schedule-based ridesharing framework that accommodates customer QoS metrics without the need for encoding customer preference constraints. II. PROBLEM FORMULATION 1) Customer Arrival Model: An MOD network is modeled as a directed network graph denoted by G = (N ,L), where N = {n1, . . . , nNn} is a set of Nn nodes, and L = {l1, . . . , lNl} is a set of Nl directed link edges each taking the form of an ordered pair of neighbor nodes l = (ni, nj) ∈ N 2. A route r(ni, nj) is defined as a sequence of directed links Lr ⊆ L which corresponds to a unique minimum-travel-time path between a pair of nodes (ni, nj). Customers are assumed to arrive at nodes in the network graph according to a Poisson process with arrival rate parameter λn. In many MOD systems, customer arrival rates will be time-varying and may feature large fluctuations on short-time scales. In this work, a discrete- time approximation is used where Poisson arrival rates are static for the operating duration, with short-term fluctuations averaged out across the time period. 2) Ridesharing: Upon arrival, customers send ride requests consisting of a pickup node p ∈ N and a drop off node e ∈ N . Let C of size Nc be the set of customers who have requested rides, let O of size No = 2Nc be the set of requested customer pickup and drop off nodes, and let V of size Nv be the set of vehicles in the MOD fleet. Vehicle v will service a subset of the customers Cv ⊆ C by visiting their pickup and drop off nodes. All customer nodes are inserted into a schedule sv = {s1, . . . sNo}, si ∈ {∅,N}, which the vehicle traverses in order using a sequence of routes. Let Q be the maximum capacity of each vehicle. The ridesharing problem is that of finding both the customer to vehicle assignments as well as the vehicle schedules such that the cost of scheduling all customers is minimized. A four-index ILP formulation for the DARP is proposed as an extension of the models presented in [17], with emphasis placed on the ordering of customers within schedules. The cij ∈ {0, 1} are equal to 1 if vehicle v is decision variables xv assigned customer c, with pc and ec respectively sequenced at si and sj; and zero otherwise. The objective of the ridesharing problem formulation is to minimize the total customer QoS cost, that is, ∀c ∈ C ∀v ∈ V, c ∈ C ∀ v ∈ V, i, j ∈ {1, . . . , No} (1) (2) (3) (4) argmin xv cij s.t. No(cid:88) gv cijxv cij i=1 j=1 xv cij = 1 v∈V v∈V c∈C No(cid:88) (cid:88) (cid:88) (cid:88) No(cid:88) No(cid:88) i(cid:88) No(cid:88) (cid:88) (cid:88) j=1 xv cij = 1 k(cid:88) No(cid:88) c∈C j=1 i=1 i=1 xv cij = 0 i=1 j=k+1 c∈C cij ∈ {0, 1} xv cij ≤ Q ∀v ∈ V, k ∈ {1, . . . , No} xv (5) ∀v ∈ V, c ∈ C, i, j ∈ {1, . . . , No}, where (2) enforces that all customers are assigned, (3) enforces that a customer is picked up before being dropped off, (4) enforces that only one customer can occupy a given schedule position, and (5) enforces that the capacity of the vehicle is not exceeded. The cost, further described in Section III-B.2, is chosen to be a function of the QoS metrics mc that a customer experiences, that is gv cij = g(mvij c ) where the metrics themselves are a function of the customer's position in a vehicle's schedule. 3) QoS Metrics: Customer QoS is quantified using a set of transportation metrics, similar to those in [2]. First, customer requests are either accepted by the MOD sys- tem, in which case the customers will be assigned to a vehicle and given a ride, or rejected by the system and the customers will walk. Let 1rej be a rejection indicator variable which takes value 1 if the customer is rejected and 0 otherwise. The remaining customer QoS metrics are: ride time tride, wait time twait, service time tservice, ratio of ride time to direct time tratio, excess ride time texcess ride, maximum number of stops while user is onboard N stops, the notification time tnotify, the total traveled distance while onboard dtraveled, the time it would have taken the customer to walk, twalk, and service time in excess of walk time texcess walk. Let mc ∈ R11 be the QoS metrics for customer c, that is mc = {1rej , N stops , , texcess ride c , tride , tservice , tratio , twait c c c c c c Algorithm 1: Predictive Positioning 1 Input: customer node arrival rates {λ1, . . . λNn} 2 Output: vehicle locations k∗ 3 enumerate vehicle placement options K 4 enumerate possible arrival locations A 5 for a ∈ A do 6 7 wk,a ← computeTotalWaitTime(k, a) for k ∈ K do pa ← computeProbability(a) (cid:80) a∈A wk,apa 8 9 k∗ = argmin k∈K 10 return k∗ , dtraveled tnotify c is further detailed in Appendix A. , twalk, texcess walk c c }. Evaluation of these metrics III. APPROACH This section present two new approaches for utilizing and solving the ridesharing problem, (1) to (5), in order to improve customer service under different operating regimes. The first approach manages vehicles in the absence of ride requests by using a predictive positioning algorithm, while the second approach manages vehicles to accommodate ride requests using a ridesharing algorithm. Additionally, two QoS focused cost functions are presented, a traditional weighted cost function, and a customer rating model which learns customer QoS preference from customer rating feedback. A. Predictive Positioning The predictive positioning algorithm uses known customer arrival rates {λ1, . . . λNn} to find the key predictive nodes within the network graph to place unassigned vehicles such that the expected wait time for arriving customers is minimized. A special form of the ridesharing problem, (1) to (5), is solved for a sequence of Na predicted customer arrivals, where 1) vehicles are only assigned to at most one predicted customer (Q = 1); 2) the number of considered customers is set to be the number of vehicles (Na = Nv); and 3) the cost is set to be the customer wait time. Let the vector k ∈ {0, 1, . . . , Nv}Nn denote the number of vehicles located at each node and a ∈ {0, 1, . . . , Na}Nn be the number of i=1 ki = Nv} is the set of all possible vehicle placement options, and i=1 ai = Na} is the set of all combinations of possible arrivals. The predictive node locations k∗ for which to place vehicles is determined using Algorithm 1. customer arrivals on each node. K = {k (cid:80)Nn A = {a (cid:80)Nn The wait time cost wk,a in Line 7 is determined by using a greedy solution to the ridesharing problem. The wait times are assumed to be dependent only on the structure of the network graph, and therefore are computed and stored offline. To handle the cases where some vehicles are serving customers while others need to be predictively positioned, wait times are computed for all numbers of free vehicles from 1 to Nv. The probability of a set of arrivals pa in Line 8 is determined using decomposition of the total network arrival Algorithm 2: Ridesharing 1 Input: request {pc, ec}, previous customer allocations {C1, . . .CNv}, schedules {s1, . . . , sNv} (cid:80) 2 Output: schedule for assigned vehicle sv∗ 3 for v = 1 : Nv do Cv ← {c,Cv} 4 Sv ← enumerateInsertions(sv, pc, ec) 5 s∗ g(h(sv, ic, jc)) v = argmin 6 sv∈ Sv 7 M∗ v ← computeNewMetrics(s∗ v) 8 Mv ← computeBaselineMetrics(sv) 9 M∗ Nv+1 ← computeRejectionMetrics(pc, ec) 10 v∗ ← assignToLowestBid({M,M∗}1:Nv ,M∗ 11 sv∗ ← s∗ v∗ 12 return sv∗ c∈ Cv Nv+1) Poisson process. Given that an arrival occurs, the probability of that arrival occurring at a node ni is given by P (ai = 1 Na = 1) = λni Λ . The probability of a set of arrivals occurring according to a follows a multinomial distribution, (cid:18) λ1 (cid:19)a1 ··· (cid:18) λNn (cid:19)aNn P (a) = Na! . (6) a1! . . . aNn ! Λ Λ Arrival probabilities are computed online whenever customer arrival rates change. B. Ridesharing 1) Ridesharing Algorithm: The ridesharing algorithm uses an insertion method as a heuristic solution to the ridesharing problem. Algorithm 2 presents the method for assigning a new customer request to a vehicle such that the total QoS cost to the system is minimized. The algorithm is executed online whenever a new customer ride request is submitted. Line 4 temporarily assigns the new customer to each vehicle generating a temporary customer allocation Cv. Line 5 enumerates all feasible ways of inserting the request into the schedule, where Sv is the set of all feasible schedules sv, and where feasibility is met by ensuring that pc is inserted before ec and that vehicle capacity is not exceeded. Lines 6 and 7 find the best feasible schedule s∗ v and corresponding new customer QoS metrics M∗ v for each vehicle, where v = {m¯c ¯c ∈ Cv}. For comparison, Line 8 computes the M∗ original customer QoS metrics Mv for each vehicle, where Mv = {m¯c ¯c ∈ Cv}. Rather than imposing constraints on customer QoS metrics, the algorithm uses a virtual "rejection vehicle", v = Nv + 1 to make bids that consider the case where the customer is not serviced by the MOD system. Line 9 computes the rejection vehicle customer metrics M∗ . While seemingly counter-intuitive, rejections are in fact important for improving customer QoS. For example, if a customer's wait time is significantly longer than the time it would take for them to walk, then they may prefer to be rejected rather than to wait to use the service. In this work, rejected customers are prevented from making additional requests, although this Nv+1 = mrejected c could be adapted to allow customers to resubmit with QoS relaxations. Line 10 finds the combination of baseline and new metrics that includes all customers and has the lowest overall customer QoS cost, and then returns the vehicle v∗ that contains c. If v∗ is the rejection vehicle, then the customer is rejected, otherwise the schedule for v∗ is updated to accommodate the request. The primary benefit of the ridesharing algorithm is the ability to evaluate customer QoS without having to encode the customer preference structure into the algorithm. For example, other methods [14]–[16] encode feasibility constraints on customer metrics such as wait time or ride time, where customers are rejected if these are not met. Instead, a more general approach is taken in Algorithm 2 where a competing bid is made to reject a customer, and the rejection is made only when the overall QoS of the system would be improved by doing so. This approach opens the door for a ratings based cost function where bids are made without constraining the customer metrics directly. 2) Ridesharing Cost Functions: Two ridesharing cost functions are presented which evaluate the customer QoS cost from a set of customer ride metrics. First, a cost function composed of a linear weighted combination of the customer metrics is proposed as mc(cid:88) i=2 g(mc) = wrej i 1rej c mc,i + wacpt i (1 − 1rej c )mc,i, (7) i i and wacpt where wrej are weights for metric i that are used to allow for differentiating between rejected and serviced customers. For example, a service time focused cost function would be g(mc) = 1rej , where the cost is the service time if the customer receives a ride and the walk time if they are rejected. This form of cost function requires that the weights be properly chosen to reflect customer preference, and can result in poor customer QoS if the weights are wrongly chosen. c + (1 − 1rej c twalk c )tservice c To overcome the need to choose weights, a second ratings based cost function is presented which learns and uses customer preference through feedback from 5-star ratings. The rating model utilizes a random forest of classification decision trees, based on the work of [18]. Random forest algorithms tend to prevent overfitting and have been demonstrated to perform well empirically [19]. To train the random forest model, a dataset D from ND customers is collected in the form of 5-star ratings Ytrain = {y1, . . . , yND} and a ride metrics feature vector Xtrain = {m1, . . . , mND} such that Ytrain = RF (Xtrain) where RF (X) is the trained random forest. The trained random forest then serves as the ridesharing cost function such that g(mc) = −RF (mc) where the minus sign is included to maximize customer rating. It is assumed that a customer's 5-star rating is given purely to reflect their ride metrics and not factors such as driver interactions which are not necessary for autonomous MOD systems. Fig. 2: Pedestrian traffic network on MIT campus with overlaid pedestrian trajectories. The network graph is composed of 27 nodes, 106 directed links, and 1056 precomputed routes. IV. EXPERIMENTS The predictive positioning and ridesharing methods are tested using simulation of the MIT MOD system. Specially, there are two motivating test cases: 1) evaluating how service times for customers are affected under different fleet management strategies as customer arrival rates grow; and 2) evaluating how customer QoS is affected by various ridesharing strategies operating under a range of unknown customer preference models. The MIT MOD system is used to provide simulation parameters that reflect a realistic operating environment for vehicles and customers. A. Simulation Setup Pedestrians and vehicles operate within a network graph for the MIT campus. The network graph, shown in Figure 2, is generated using pedestrian trajectory data collected from sensors onboard the MOD vehicles following the method presented in [20]. A two hour time period is simulated; during which time a subset of 10 randomly chosen nodes are assigned non-zero pedestrian arrival rates in order to reflect that not every campus location experiences arrivals at all times. Customer arrival rates are static for the time period and take values between 0 and 0.45 ped/min at each node. There are 3 vehicles in the MIT MOD system each with a maximum capacity of 3 passengers. Vehicles travel between nodes according to the schedule generated by the ridesharing algorithm. A vehicle picks up its assigned customers upon reaching a scheduled node. If a vehicle's schedule is empty, the vehicle will travel to nodes prescribed by the predictive positioning algorithm. Vehicle link speeds are either 11m/s for links corresponding city streets or 4m/s for campus pathways. Predictive Positioning The predictive positioning algorithm finds key predictive node positions to place unallocated MOD vehicles based on customer arrival rate parameters. Figure 3 shows the predictive nodes chosen for either 1, 2, or 3 unallocated vehicles for a particular set of arrival rates. The predictive nodes take into Fig. 3: Relative customer arrival rates and computed predictive nodes. The numbered predictive nodes illustrate where unassigned MOD vehicles should be located to minimize the expected wait time for customers. The radius of the circles indicates the relative magnitudes of the node arrival rates. The numbers indicate the locations to place either 1, 2, or 3 vehicles depending on the current number of unallocated vehicles. account both the probability of the arrivals occurring and the vehicles' travel time to reach each node. If only a single vehicle is unallocated, it will tend to be positioned centrally within the network, but skewed towards large arrival rates. When more vehicles are unallocated, the predictive nodes are further spread out for better coverage. The performance of the predictive positioning algorithm is evaluated through comparison against a baseline unmanaged MOD strategy, where vehicles respond to pickup requests but are not repositioned after dropping off customers. The two methods are first compared without the use of ridesharing (vehicle capacities are 1), with assignments evaluated using a minimum service time cost function. Figure 4 shows that predictive positioning is able to reduce customer service times. When arrival rates are lower than 0.35 ped/min per vehicle, there is often time between arrivals for vehicles to reposition to the predictive nodes and service times can be reduced by up to 20%. As arrival rates increase, however, the benefits of predictive positioning are reduced as vehicles are continually allocated to requests. Under high arrival rates, it would be desirable to add more vehicles to shift to the lower portion of the curve. However, with a fixed-sized fleet, that option is not available so ridesharing is used as an alternative. B. Ridesharing The performance of the ridesharing algorithm is evaluated by comparing the single capacity predictive positioning and unmanaged MOD methods with ridesharing versions where the maximum vehicle capacity is increased to 3. The ridesharing algorithm is applied using a minimum service time cost function. Figure 4 also shows that ridesharing reduces customer service times. When arrival rates are lower than 0.35 ped/min per vehicle, the ridesharing methods perform similarly to their single capacity counterparts since current customers can be serviced before new customers arrive. But Fig. 4: Service times for unmanaged and predictive positioning methods, with and without ridesharing, over a range of customer arrival rates. The figure shows that predictive positioning reduces service times when arrival rates are low and ridesharing reduces service times when arrival rates are high. The service times are normalized by the direct time so as not to penalize service times for customers with longer routes. Lower service times are better. The arrival rates are normalized by the number of vehicles. The mean and standard deviation from 100 iterations are shown. at higher arrival rates, the ridesharing algorithm begins to utilize the excess vehicle capacity and new customers are inserted into vehicle schedules before previous customers have finished their ride. Through the use of a combined predictive positioning and ridesharing approach, the MOD system is able achieve a better customer QoS across all arrival rates, resulting in as much as a 29% reduction in service time compared to the single capacity unmanged MOD strategy. C. Customer Preference In the previous analysis, the customer preference was assumed to be focused on service time. However, this assumption can be wrong and customers may give poor ratings if the the true customer preference lies elsewhere in other QoS metrics. To evaluate the rating performance of an MOD system, a simulated rating model is used to assign a 5-star ratings to customers based on a set of customer preference weights. The details of the simulated rating model are provided in Appendix B, which also specifies how customer preference is encoded using weight concentration parameters. Six customer preference modes are analyzed, where the weight concentration parameters are 90% skewed towards either wait time, ride time, service time, the number of stops while onboard, ride distance, or a combined weight between service time and ride distance. Five fleet management strategies are considered. First, a minimum vehicle distance strategy is considered, where assignments are not made based on customer ratings but rather based on the traditional minimum vehicle travel distance metric which would minimize fuel consumption. Next, three focused strategies based on ride time, service time, and wait time are included, where each strategy is given access to the underlying ground truth ratings function but chooses rating weights according to its focus. Finally, the presented random TABLE I: MOD fleet management performance as measured by average customer rating. Customer Preference Strategy Wait Ride Service # Stops Distance Combined Average 4.53 4.70 Distance 4.42 4.97 Ride Time Service Time 4.67 4.88 4.69 4.42 Wait Time 4.67 4.96 4.22 4.43 4.64 4.09 4.63 4.17 4.70 4.63 3.82 4.60 Ratings 3.91 4.50 4.30 3.64 4.95 4.10 4.97 4.62 3.51 4.96 4.27 4.66 4.62 4.03 4.80 forest ratings model strategy is included where the ground truth ratings function is not available but rather customer preference is learned from 5-star customer feedback ratings. The random forest model is implemented using [21], which is trained separately under each customer preference mode for 10 runs. To further test the ratings model, all ride distance metrics are removed from the random forest feature vector in order to see if performance can be learned using only non-corresponding, but correlated metrics. The arrival rate is fixed at 0.35 ped/min per vehicle so that the MOD fleet is operating under the ridesharing regime. Figure 5 shows how the performance, in terms of average received rating, for each strategy depends on the underly- ing customer preferences. Table I summarizes the average customer rating over all 20 iterations. The results show that wait time, ride time, and service time each perform best when the customer preference mode matches. Additionally, the ride time metric performs best under the excess ride distance and combined customer preference modes because ride distance and ride time are correlated. However, each of the focused fleet management strategies performs relatively poorly under at least one customer preference mode, and the minimum distance strategy always performs poorly because customer preference is not considered. In contrast, the ratings model demonstrates robust performance across all customer preference modes. The ratings model is within 0.5% of the focused strategies under their respective modes, demonstrating that it was able to learn which customer metrics were important under each mode. The excess ride distance mode illustrates how the ratings model is able to learn customer preference using elements in its feature vector that are only correlated with the customer preference metric and not included directly. Finally, when considering the average performance over all customer preference modes, the ratings model performed best. V. CONCLUSION This work demonstrated that predictive positioning and ridesharing can be utilized in an MOD system in order to improve customer QoS. A predictive positioning algorithm was presented which uses customer arrival rate information to position vehicles at key nodes in the MOD network graph which minimize the expected customer wait time. In a simulated campus setting, the predictive positioning method was shown to reduce customer service times by as much as 20% when customer arrival rates are low. To improve QoS as arrival rates increase, a ridesharing approach was presented which utilizes a customer QoS based cost function. (a) Wait Time (b) Ride Time (c) Service Time (d) Number of Stops (e) Excess Ride Distance (f) Combined Fig. 5: Average customer rating for several fleet management strategies under customer preferences that are skewed towards either (a) wait time, (b) ride time, (c) service time, (d) the number of stops while onboard, (e) the excess ride distance between the received ride and a direct ride, or (f) a combined skew between service time and excess ride distance. The figures show that the performance of each strategy is dependent on the underlying customer preference, with the exception of the ratings strategy which performs well in all cases. Each figure shows the results of 20 runs. Performance is measured by the average customer rating where higher ratings are better. A combined predictive positioning and ridesharing approach was shown to reduce customer service times by as much as 29%. A customer ratings model was introduced as a means for learning customer preference through feedback in the form of a 5-star rating. The customer ratings model is shown to provide the best overall MOD fleet management performance over a range of customer preferences. The predictive positioning and ridesharing methods which were evaluated for the campus setting could also be applied to larger networks. To improve the scalability of Algorithm 1, predictive node locations could represent larger regions within an MOD service area, where customer wait times include both the travel time between regions and the expected travel time within a region. The ridesharing method can be applied to larger fleets by parallelizing the scheduling computations 05101520Run Number33.544.55Rating05101520 Run Number33.544.55Rating05101520Run Number33.544.55Rating05101520Run Number 33.544.55Rating05101520Run Number33.544.55Rating05101520Run Number33.544.55Ratingfor each vehicle across multiple machines which bid to a centralized machine for the customer allocations. Future work will apply these methods to the physical MIT MOD system. Methods for learning and predicting customer arrival rate trends will be studied so that vehicles can be pre- dictively positioned to match real time demand. Additionally, ratings and service metrics from actual customers will be used to improve and evaluate rating models based on random forest and other machine learning techniques. ACKNOWLEDGMENT Research supported by the Ford Motor Company through the Ford-MIT Alliance. REFERENCES [1] "Morgan StanleyVoice: Shared Mobility On The Road Of The Future," Forbes, 2016. [2] J. Paquette, J.-F. Cordeau, and G. Laporte, "Quality of service in dial- a-ride operations," Computers & Industrial Engineering, vol. 56, no. 4, pp. 1721–1734, May 2009. [3] M. Pavone, "Autonomous Mobility-on-Demand Systems for Future Urban Mobility," in Autonomes Fahren, M. Maurer, J. C. Gerdes, B. Lenz, and H. Winner, Eds. Springer Berlin Heidelberg, 2015, pp. 399–416. [4] M. Pavone, S. L. Smith, E. Frazzoli, and D. Rus, "Robotic load balancing for mobility-on-demand systems," The International Journal of Robotics Research, vol. 31, no. 7, pp. 839–854, June 2012. [5] M. Volkov, J. Aslam, and D. Rus, "Markov-based redistribution policy model for future urban mobility networks," in 2012 15th International IEEE Conference on Intelligent Transportation Systems (ITSC), Sept. 2012, pp. 1906–1911. [6] R. Zhang and M. Pavone, "Control of robotic mobility-on-demand systems: A queueing-theoretical perspective," The International Journal of Robotics Research, vol. 35, no. 1-3, pp. 186–203, 2016. [7] R. Zhang, F. Rossi, and M. Pavone, "Model predictive control of autonomous mobility-on-demand systems," in 2016 IEEE International Conference on Robotics and Automation (ICRA), May 2016, pp. 1382– 1389. [8] S. Abuelsamid, "Singapore, Delphi and nuTonomy To Launch Pilot Of Autonomous, On-Demand Car Service," Forbes, 2016. [9] P. E. Ross, "Helsinki Tries Self-Driving Buses in Real Traffic," IEEE Spectrum: Technology, Engineering, and Science News, Aug. 2016. [10] P. Ross, "Uber Will Start Driverless Service in Pittsburgh-This Month," IEEE Spectrum: Technology, Engineering, and Science News, Aug. 2016. [11] R. M. Jorgensen, J. Larsen, and K. B. Bergvinsdottir, "Solving the Dial-a-Ride Problem Using Genetic Algorithms," The Journal of the Operational Research Society, vol. 58, no. 10, pp. 1321–1331, 2007. [12] G. R. Mauri, L. Antonio, and N. Lorena, "Customers' satisfaction in a dial-a-ride problem," IEEE Intelligent Transportation Systems Magazine, vol. 1, no. 3, pp. 6–14, Fall 2009. [13] J. Paquette, J.-F. Cordeau, G. Laporte, and M. M. B. Pascoal, "Combining multicriteria analysis and tabu search for dial-a-ride problems," Transportation Research Part B: Methodological, vol. 52, pp. 1–16, June 2013. [14] J.-J. Jaw, A. R. Odoni, H. N. Psaraftis, and N. H. M. Wilson, "A heuristic algorithm for the multi-vehicle advance request dial-a- ride problem with time windows," Transportation Research Part B: Methodological, vol. 20, no. 3, pp. 243–257, June 1986. [15] L. Coslovich, R. Pesenti, and W. Ukovich, "A two-phase insertion technique of unexpected customers for a dynamic dial-a-ride problem," European Journal of Operational Research, vol. 175, no. 3, pp. 1605– 1615, Dec. 2006. [16] D. J. Fagnant and K. M. Kockelman, "Dynamic Ride-Sharing and Optimal Fleet Sizing for a System of Shared Autonomous Vehicles," Transportation Research Board 94th Annual Meeting, 2015. [17] J.-F. Cordeau and G. Laporte, "The dial-a-ride problem: Models and algorithms," Annals of Operations Research, vol. 153, no. 1, pp. 29–46, May 2007. [18] L. Breiman, "Random forests," Machine Learning, vol. 45, no. 1, pp. 5–32, Oct. 2001. [19] R. Caruana and A. Niculescu-Mizil, "An empirical comparison of supervised learning algorithms," in Proceedings of the 23rd International Conference on Machine Learning, ser. ICML '06. ACM, pp. 161–168. [Online]. Available: http://doi.acm.org/10.1145/1143844. 1143865 [20] J. Miller, A. Hasfura, S.-Y. Liu, and J. P. How, "Dynamic Arrival Rate Estimation for Campus Mobility on Demand Network Graphs," IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). [21] A. Jaiantilal, "Randomforest-matlab," https://code.google.com/archive/ p/randomforest-matlab/, 2010. A. Computing Customer Ride Metrics APPENDIX This section provides details on how customer metrics are evaluated using vehicle schedule information. Customer c makes a request at the point in time t request and is assigned to v currently located at node nv. The pickup pc and drop off ec nodes for c occur at respective nodes si and sj in the vehicle schedule sv. The vehicle travels between any adjacent nodes sk and sk+1 in its schedule using route r(sk, sk+1). The travel distance and travel time between the nodes are c d(sk, sk+1) = t(sk, sk+1) = (cid:88) (cid:88) l∈Lr(sk ,sk+1) l∈Lr(sk ,sk+1) dl, dl ul , (8) (9) where dl and ul are the length and average travel speed of link l, respectively. The metrics mvij are computed as follows: c t pickup c = t(nv, s1) + t dropoff c = t(nv, s1) + t(sk, sk+1), t(kk, sk+1), i−1(cid:88) j−1(cid:88) k=1 k=1 − t pickup − t request , c , c c tdirect = t(pc, ec), c ddirect = d(pc, ec), c twalk c = ¯t(pc, ec), tride c = t dropoff twait c = t pickup = twait tservice c + tride c c = tride tratio c /tdirect , c − tdirect texcess ride = tride c c = k − j, N stops j−1(cid:88) tnotify = t assigned c c , c c c c dtraveled c = , − t request , c d(sk, sk+1), k=i texcess walk c = tservice c − twalk c , c c , twalk c , tnotify = }. If the customer is given a ride, then the met- , N stops , not apply and the customer metrics are set to be mrejected {1rej rics are mc = {1rej c tnotify , twalk, texcess walk c B. Simulated Rating Model , twait , tservice }. , texcess ride , dtraveled c , tride , tratio c c c c c c c c This section presents a simulated rating model, which is used as ground truth in simulation to assign 5-star ratings to MOD customers. The values and functional forms are chosen based on an assumed customer preference and are kept hidden from the 5-star rating model. If a customer is rejected, then they give one of the two lowest ratings based on how long they waited to be notified of their rejection. A rejected customer's rating is rrejected = if tnotify c twalk c 2, 1, otherwise. ≤ 0.1 (23) (cid:40) If a customer is given a ride, then the rating will be a weighted sum of 5 aggregate ratings based on wait time, ride time, service time, number of stops, and ride distance computed as raccepted c c +w3rservice c +w2rride +w4rstops = w1rwait c c +w5rdistance , (24) c with rwait c = Range rride c = Range (cid:19) (cid:19) (cid:19) , 0, 1 , 0, 1 c twait c (cid:18) (cid:18) tride (cid:18) tservice (cid:18) dtraveled c − tdirect twalk c − tdirect c − tdirect twalk − tdirect c c − tdirect twalk c ) − ddirect c c c c c , 0, 1 (cid:19) = Range rservice c c = max(1, 6 − N stops rstops c rdistance c = Range ddirect c , 0, 0.5 , where Range(α, β, γ) maps α to the i-th interval of 5 exponentially spaced values between β and γ and assigns the value 6-i as the rating. Range values were chosen to reflect a set of possible expected customer satisfaction levels for each metric. For example, setting the γ value for rdistance to 0.5 reflects that customers would give the lowest rating once their journey distance exceeded the nominal distance by a factor of 0.5. Exponential spacing is used to cause ratings to drop off more quickly as metrics worsen for customers. c The majority of customers follow the same set of weights w, but some customers do not. To accommodate this, the weights are drawn from a Dirichlet distribution such that w ∼ Dir( w), where the concentration parameters w represent the nominal weights for the population. (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) c c c is the point in time c is picked up, t dropoff where t pickup is the point in time c is dropped off, tdirect is the time it would take to drive directly from pc to ec, ddirect is the direct route distance between pc and ec, twalk is the time it would take to walk from pc to ec, ¯t(ni, nj) is (9) evaluated with r(ni, nj) and vl, as the respective route and speed of the pedestrian instead of a vehicle, and t assigned is the point in time when the ridesharing algorithm assigns the customer to the vehicle. Note, if a customer is rejected (1rej c = 1), then many of the metrics do c c c
1905.12083	1	1905	2019-04-24T09:36:29	EasySched: a multi-agent architecture for the predictive and reactive scheduling of Industry 4.0 production systems based on the available renewable energy	[ "cs.MA" ]	Industry 4.0 is concerned with sustainable development constraints. In this context, we propose a multi-agent architecture, named EasySched, aiming at elaborating predictive and reactive scheduling as the result of a coordination between systems producing goods and systems producing renewable energy. The validation of this architecture is original, and was conducted in a completely and physically distributed way, using networked embedded systems. This validation was done on a series of instances inspired by the literature. The results showed that EasySched succeeds in adapting the production of goods according to the available renewable energy	cs.MA	cs	EasySched : une architecture multi-agent pour l'ordonnancement prédictif et réactif de systèmes de production de biens en fonction de l'énergie renouvelable disponible dans un contexte industrie 4.0 Maroua Nouiri, Damien Trentesaux, Abdelghani Bekrar LAMIH UMR CNRS 8201, Université Polytechnique Hauts-de-France Le Mont Houy, 59313, Valenciennes, France. Corresponding author: [email protected] Résumé : L'industrie 4.0 s'accompagne de la prise en compte de contraintes de développement durable. Dans ce contexte, nous proposons une architecture multi-agent pour l'ordonnancement prédictif et réactif coordonné entre des systèmes de production de biens et des systèmes de production d'énergie renouvelable, appelée EasySched. La validation de cette architecture est originale, elle est menée de manière complètement et physiquement distribuée en utilisant des systèmes embarqués en réseau. Cette validation est menée sur une série d'instances inspirées de la littérature. Les résultats montrent que les mécanismes proposés permettent d'adapter la production selon l'énergie renouvelable disponible. : production, ordonnancement prédictif, ordonnancement réactif, énergies Mot clés renouvelables, modélisation multi-agent, Optimisation par essaims particulaires, systèmes embarquées 1. Introduction Les enjeux environnementaux, tels que la réduction de la pollution ou des émissions de dioxyde de carbone (CO2) et la maîtrise des changements climatiques prennent une importance croissante. L'industrie 4.0 se doit d'assurer un équilibre entre un accroissement des besoins énergétiques, et la réduction de l'impact environnemental des ressources énergétiques utilisées (Cardin et al., 2017), (Trentesaux et al., 2017). Le système énergétique mondial subit des transformations en profondeur de par l'avènement des énergies renouvelables. Néanmoins, la disponibilité de ce type d'énergie est fortement variable et difficilement prévisible (Prabhu et al., 2015). Afin de pallier ce problème, il est primordial de mettre en place et d'optimiser une coopération entre la source (les fournisseurs d'énergies) et le client (les usines de production) (Trentesaux et al., 2016). Cette coopération doit être réactive pour pouvoir faire face à des événements difficilement prévisibles, de fréquence et d'importance variées. 1 Dans ce contexte, nous proposons EasySched, une architecture multi agent pour l'ordonnancement prédictif et réactif de systèmes de production de biens (consommateur d'énergie) en fonction des contraintes issues de systèmes de production d'énergie renouvelables (fournisseurs) dans un contexte Industrie 4.0. L'ordonnancement est en effet une fonction de la production qui doit désormais être abordée en tenant compte d'aspects relatifs au développement durable et à la gestion de l'énergie (Giret et al., 2015). Pour ce faire, un mécanisme de coopération entre ces systèmes de production est proposé. Il concerne le calcul (en prédictif) et l'ajustement (en réactif) de l'ordonnancement de la production de biens vis-à-vis de la consommation d'énergie. La production d'énergie de type renouvelable est difficilement prédictible car fortement dépendante de conditions non contrôlables (par exemple, les conditions météorologiques). L'originalité de cette architecture réside dans le fait que l'élaboration de l'ordonnancement se fait en tenant compte des besoins en production et du niveau dynamique d'énergie renouvelable disponible. La validation de cette architecture a été faite de manière complètement distribuée, en intégrant à la fois du matériel et du logiciel. Dans cet article, et pour illustrer nos propos, deux sources d'énergie (éolienne et photovoltaïque) sont considérées et les systèmes de production considérés sont de type job-shop, la variable d'ajustement sur la consommation étant la vitesse des machines de production. Notre approche ambitionne ainsi d'adapter en temps réel la production, et donc la consommation d'énergie dynamiquement à l'apport énergétique. Dans ce qui suit, nous présentons en premier lieu un état de l'art relatif à la production sous contrainte énergétique ou en environnement perturbé. Les détails de notre architecture seront présentés dans la section 3. La section 4 présentera les divers tests réalisés et les résultats d'expérimentation. Une conclusion et des perspectives seront fournies dans la section 5. 2. État d'art Nos travaux relèvent de la problématique scientifique suivante : une production de biens doit être ordonnancée de manière prédictive et réactive en tenant compte de l'évolution prévisionnelle et en temps réel de l'énergie renouvelable dont la disponibilité est incertaine. Il existe de nombreuses approches relevant de ce contexte. Sans vouloir être exhaustif, nous présentons dans cette partie un ensemble de références qu'il nous semble important d'étudier et que nous avons regroupées en trois catégories. Une première catégorie traite de cette problématique de manière complètement statique. Par exemple, Mikhaylidi et al., (2015) ont considéré un problème de contrôle des opérations de fabrication avec des prix de l'électricité connus variant dans le temps sur un horizon de planification fini. L'objet est de déterminer la date à laquelle chaque opération doit être effectuée dans l'horizon temporel donné afin de minimiser la consommation totale d'électricité et les coûts liés aux pénalités de retard des opérations. Les auteurs ont proposé une solution de type programmation dynamique. Gahm et al., (2016) ont défini un ordonnancement économe en énergie, avec un objectif d'amélioration de l'efficacité énergétique. Pour évaluer leur travail, ils 2 ont pris en compte non seulement les processus et les systèmes de production, mais également les processus et les systèmes d'approvisionnement en énergie. Fang et al., (2011) ont présenté un modèle de programmation mathématique du problème d'ordonnancement d'ateliers qui prend en compte la charge de pointe, la consommation d'énergie et l'empreinte carbone associée, en plus du temps de cycle. Le modèle est illustré à l'aide d'une étude de cas simple : un atelier de fabrication où deux machines sont utilisées pour produire une variété de pièces. Le problème d'ordonnancement proposé considère la vitesse de fonctionnement comme une variable indépendante, qui peut être modifiée pour affecter la charge de pointe et la consommation d'énergie. Gonzalez et al., (2017) ont suggéré une méta-heuristique hybride associant un algorithme génétique à une nouvelle méthode de recherche locale et une approche de programmation linéaire pour améliorer le coût énergétique d'un ordonnancement d'atelier de type job shop. He et al., (2015) ont conçu une méthode d'optimisation économe en énergie en tenant compte de la sélection de la machine et de la séquence de fonctionnement. Tonelli et al., (2016) ont proposé un modèle centralisé et distribué pour trouver l'ordonnancement prédictif hors ligne. Plitsos et al., (2017) ont défini un système d'aide à la décision (DSS) composé d'un algorithme de recherche local qui offre une optimisation hiérarchique (basée sur plusieurs objectifs) de la planification de la production. L'énergie est considérée sous la forme de contraintes. Lamy., (2017) a élaboré un modèle mathématique et une méta-heuristique Greedy Randomized Adaptive Search Procedure (GRASP) pour résoudre le problème de job-shop avec contrainte énergétique. Dans cette catégorie, tous les travaux présentés ont porté sur des modèles d'ordonnancement statique avec optimisation de la consommation d'énergie. Plusieurs sortes d'aléas peuvent arriver en cours de la production et peuvent être de différentes origines (Chaari et al., 2014). Ceci complexifie les problèmes d'ordonnancement. La solution produite doit en effet prendre en compte les perturbations de l'environnement, tout en assurant de bonnes performances. Dans ce contexte, une seconde catégorie de méthodes a été développée pour prendre en compte la résolution du problème d'ordonnancement. Chaari et al., (2014) ont suggéré une classification qui positionne les différentes méthodes d'ordonnancement sous incertitudes. Celle-ci comporte trois types : les approches proactives, réactives et hybrides. Cette dernière regroupe les approches prédictives- réactives et les approches proactives-réactives. Les approches prédictives réactives sont constituées de deux phases : la première phase consiste à construire un ordonnancement déterministe hors-ligne ne prenant pas en considération les événements imprévisibles. Par exemple, les opérations prévues sont toutes connues dès le départ, les temps de traitement ont été préalablement déterminés et temps durant l'ordonnancement. Durant la deuxième phase, cet ordonnancement est utilisé en ligne. Il est adapté en temps réel pour tenir compte des perturbations (Chaari et al., 2014). Au delà du mécanisme de réactivité qui doit être fourni par la méthode d'ordonnancement, un objectif de durabilité doit aussi être visé dès la phase de conception. Par exemple, Salido et al., (2016) se sont concentrés sur le problème de ré-ordonnancement d'ateliers dynamiques où les machines peuvent fonctionner à des vitesses différentes. Les auteurs ont proposé une nouvelle technique les machines sont disponibles tout l'incertitude lors de le 3 basée sur un algorithme génétique pour trouver un ordonnancement avec un temps assurant la minimisation de la consommation d'énergie. Zhang et al., (2013) ont défini un nouveau modèle mathématique pour résoudre le problème de job-shop flexible qui prend en compte simultanément la consommation d'énergie et l'efficacité de la planification. Ils ont également proposé une méthode de ré-ordonnancement basée sur un algorithme génétique. Les travaux précédemment cités proposent des méthodes de ré-ordonnancement en tenant compte de la consommation d'énergie face à une classe de perturbations de type « panne machine » ou « attribution d'une tâche aléatoire ». Cependant, ils ne tiennent pas compte de la variation dynamique de la disponibilité énergétique. La troisième et dernière catégorie de travaux relève de cet aspect. Par exemple Pach et al., (2015) ont proposé une méthode qui considère la variation de la consommation d'énergie comme contrainte mais aussi comme étant une source d'incertitude. Les auteurs ont proposé une approche basée sur les champs de potentiel pour activer / désactiver les ressources de manière réactive en fonction de la situation du système de fabrication flexible (FMS) afin de réduire la perte d'énergie. Le contrôle de la consommation électrique globale a alors été introduit durant la fabrication afin de respecter un seuil d'énergie disponible déterminé dynamiquement et a priori inconnu. Dans (Trentesaux et al., 2017), une architecture multi-agent prédictive et réactive intégrant consommateurs d'énergie et producteurs d'énergie a été spécifiée, cependant aucune modélisation n'a été proposée. En se basant sur ce travail, Nouiri et al., (2018) ont élaboré un modèle MA-EAPSRS (Multi Agent-Energy Aware Production Scheduling and Rescheduling System) en décrivant les différents agents et leurs comportements. Des protocoles de négociation ont été proposés et validés pour calculer de manière prédictive une production durable. Cependant, aucun mécanisme de décision réactive n'a été proposé. De notre point de vue, quelle que soit la catégorie de contribution concernée, les travaux restent à un niveau de validation très conceptuel, au mieux, en simulation et considèrent très peu une variabilité dans la disponibilité de l'énergie alors que les énergies renouvelables se développent de plus en plus. Il nous semble ainsi important de proposer des algorithmes pour ordonnancer de manière prédictive et réactive des centres de production de bien, consommateurs d'énergie, en tenant compte de contraintes issues de centres de production d'énergie renouvelables. Une validation sur des expérimentations plus réalistes de ces algorithmes est également souhaitable. Ces deux points font l'objet de cet article. Plus précisément, notre contribution porte sur l'enrichissement, au travers d'une architecture multi-agent expérimentale qu'il est possible de valider de manière complètement distribué, intégrant matériel et logiciel, de l'architecture précédemment proposée MA-EAPSRS. Cet enrichissement porte sur l'intégration de mécanismes réactifs permettant d'adapter la production à l'énergie disponible. Afin de valider cette architecture multi-agent, des expérimentations seront proposées pour planifier en temps réel, la consommation énergétique des usines connectées à un réseau intelligent tenant compte d'une fluctuation de production énergétique. La validation aura ainsi lieu sur les parties prédictives et réactives. 4 Nous présentons dans la suite tout d'abord le modèle multi-agent support de l'architecture et la contribution scientifique concernant la dimension réactive avant de présenter sa réalisation logicielle et matérielle qui nous permettra de tester et valider les mécanismes de coopération proposés. 3. Proposition d'un modèle multi-agent support de l'architecture 3.1 Hypothèses de travail Les décisions prises par les systèmes de production que nous étudions sont des décisions d'ordonnancement prédictif et réactif. Il s'agit de programmer l'exécution d'un ensemble de tâches à des dates spécifiques, tenant compte d'une contrainte énergétique communiquée en temps réel. Les contraintes considérées lors de l'établissement de l'ordonnancement sont présentées figure 1. Ces contraintes peuvent être classées selon leur nature déterministe (contrainte de précédence par exemple) ou dynamique, prenant en considération la disponibilité des machines au fil du temps, par exemple. Figure 1. Contraintes dans un problème d'ordonnancement. Notre objectif est d'optimiser la date de réalisation de la production (makespan) tout en assurant l'efficience de celle-ci par la minimisation de la consommation énergétique des ressources de fabrication (soit usine à l'échelle macroscopique, soit machines à l'échelle microscopique intra-usine). Nous ne considérons pas les autres consommateurs d'énergie (systèmes de transport, chauffage, etc.). La méthode proposée est hybride et comporte deux phases. Elle est basée sur une coopération entre les systèmes consommateurs d'énergie et les systèmes producteurs d'énergie. La phase prédictive de la méthode proposée consiste à calculer hors ligne un ordonnancement pour une consommation d'énergie déterminée (cette phase a été décrite dans (Nouiri et al., 2018)). La deuxième phase, réactive, est détaillée dans cet article. Elle a pour objectif l'ajustement, en temps réel, des plans de planification en tenant compte des évènements non 5 prévus. Ces évènements peuvent êtres internes comme les pannes de machines ou externes (ordres urgents). La figure 2 illustre les différents types de fournisseurs et consommateurs d'énergie qu'il est possible de considérer, que ce soit au niveau des systèmes de production consommateurs d'énergie (en production manufacturière : single machine, flow shop, ateliers flexibles, etc.) ou au niveau des fournisseurs d'énergie renouvelable (énergie éolienne, photovoltaïque, etc.). Figure 2. Différents types de fournisseurs et consommateurs d'énergie. Dans le cadre de cet article, nous considérons que les usines sont composées de systèmes de production de type job-shop avec pour chacun, la possibilité de faire varier les vitesses de fonctionnement des machines (ce qui modifie leur consommation énergétique) et deux sources d'énergie uniquement de type renouvelable, éolien et photovoltaïque. Le changement de vitesse des machines fait référence à la variation de leur cadence et non à un arrêt total pendant une période de temps. Dans ce qui suit, nous détaillons les types d'agents, leurs comportements et décrivons les types de perturbations considérées. 3.2 Types d'agents et comportements L'approche multi-agent permet de mettre en oeuvre des mécanismes de type prédictif (en intégrant par exemples des algorithmes d'optimisation) et réactif (par interaction dynamique) (Whitrook et al., 2018). C'est cette capacité qui a guidé notre choix d'une approche multi- agent. Il existe de nombreuses manières de définir, structurer et articuler les interactions entre agents dans un contexte de couplage entre une phase prédictive et une phase réactive (Jimenez et al., 2017). Dans le cadre de nos travaux, l'architecture multi-agent proposée est complètement distribuée afin de faciliter la réactivité face à une variété d'événements imprévus. Elle a été conçue selon la méthodologie de conception Go-Green ANEMONA (Giret et al., 2017). Elle est constituée d'agents spécialisés qui représentent d'une part les systèmes de production de bien et d'autre part les fournisseurs d'énergie. Les rôles de ces agents sont articulés selon ces deux phases (prédictives et réactives). Nous associons un agent à chaque usine de production : chaque 6 usine est ainsi représentée par un agent nommé "Agent Ordonnanceur d'usine" (AOU). Chaque fournisseur d'énergie est représenté de manière symétrique par un agent nommé "Agent Ordonnanceur d'énergie" (AOE). Un agent nommé " Agent Contrôleur Energie" (ACE) est également proposé. Son rôle est de contrôler en temps réel les consommations d'énergie des usines connectées. La figure 3 schématise les rôles des différents agents de l'architecture proposée appelée EasySched (Energy-Aware production SYstem SCHEDuling). Nous détaillerons dans ce qui suit les agents AOU et AOE. Figure 3. Rôles des agents dans l'architecture EasySched 3.2.1. Agent Ordonnanceur de fournisseur d'énergie (AOE) L'Agent Ordonnanceur de fournisseur d'énergie a pour rôle de :  Valider les ordonnancements prédictifs et les demandes en énergie envoyées par les usines connectées. (hors ligne)  Contrôler la disponibilité de l'énergie renouvelable. (en ligne) Le rôle en mode hors ligne consiste à valider les ordonnancements prédictifs. La validation porte sur les demandes en consommation en énergie et leurs capacités à être satisfaites ou non (elles sont dès lors refusées). Le rôle du mode en ligne est relatif au contrôle et au suivi, en temps réel de la disponibilité de l'énergie renouvelable. Il s'agit d'un comportement cyclique qui permet de collecter les données des capteurs spécifiques à chaque type de ressource (par exemple, de température, d'humidité, de luminosité, du vent, etc.). Les données ainsi collectées seront utilisées pour alimenter les processus de prise de décision de ré-ordonnancement. Des boucles de rétroaction plus complexes 7 peuvent être créées sur la base des données sensorielles obtenues. Les données ainsi collectées sont exploitées pour contrôler la consommation d'énergie des usines connectées. Si aucune perturbation n'est détectée, l'agent AOE envoie un message indiquant l'absence de la fluctuation de l'énergie aux agents AOU. Si une variation de l'énergie est détectée, l'AOE calcule tout d'abord le taux d'énergie tauxEnergy % à réduire. Un ordre Reschedule qui précise le temps de ré-ordonnancement et la contrainte énergétique à respecter est alors envoyé. La réception de ce type de message implique que les AOU doivent exécuter une technique de ré-ordonnancement permettant d'obtenir un nouvel ordonnancement adapté aux changements imposés par les fournisseurs d'énergie. 3.2.2 Agent Ordonnanceur d'usine (AOU) L'Agent Ordonnanceur d'usine a pour rôle de :  Calculer un ordonnancement prédictif moyennant une méthode d'optimisation. (hors ligne)  Exécuter une technique de ré-ordonnancement suite aux événements aléatoires perçus moyennant des stratégies de négociation. (en ligne)  Contrôler et détecter localement les perturbations de production. (en ligne) Dans ce travail, l'ordonnancement prédictif est déterminé par un algorithme de type PSO repris de (Nouiri et al., 2018) où une particule représente une solution potentielle au problème. Chacune de ces particules est dotée par un vecteur de position, une vitesse qui permet à la particule de se déplacer et un voisinage c'est-à-dire un ensemble de particules (Neighbours) qui interagissent directement sur la particule, en particulier celle qui a le meilleur critère. La fonction objective eq. (1) utilisée pour évaluer les solutions combine à la fois le makespan ( ) et l'énergie totale ). Comme les deux objectives ne sont pas de même grandeur ni unité, une phase de normalisation est exigée. Un facteur de pondération permet de fixer le degré d'importance de chaque objectif. (1) Si, durant la phase prédictive, l'AOU reçoit un message de l'AOE indiquant le refus de sa demande d'énergie issue du calcul initial, l'algorithme PSO est exécuté à nouveau mais avec une nouvelle valeur . Ce facteur de pondération sera réduit par une valeur α favorisant ainsi la réduction de la consommation énergétique au détriment de l'efficacité de l'ordonnancement. La figure suivante illustre le pseudo code du comportement en ligne de l'agent AOU. Il s'agit d'un comportement cyclique lancé périodiquement, la période étant notée p3. Si le message reçu de la part de l'AOE ne fait référence à aucune situation perturbante, l'ordonnancement est appliqué selon le plan initial. Si ce n'est pas le cas, une technique de ré-ordonnancement est exécutée. 8 Figure 4. Pseudo code du comportement en ligne de l'agent AOU Nous détaillons dans ce qui suit les méthodes de ré-ordonnancement utilisées par les AOU en cas de réception d'une alerte de la part de l'AOE indiquant une fluctuation de l'énergie disponible. Plusieurs méthodes de ré-ordonnancement ont été proposées. Par exemple, pour le problème de type job shop classique, nous avons identifié la méthode de décalage à droite, le ré- ordonnancement total et le ré-ordonnancement partiel (Nouiri et al., 2017). Avant de présenter la technique de ré-ordonnancement proposée, nous détaillons le type de perturbation considéré. 3.3 Type de perturbation considéré La figure 5 illustre les différents paramètres composant une perturbation. Générer une perturbation revient à déterminer la ressource affectée, la date d'occurrence de la perturbation, la durée et son type (par exemple, réparable ou non réparable). La ressource affectée peut être soit renouvelable soit consommable. Dans notre travail, nous nous intéressons à la variation de la disponibilité de l'énergie renouvelable comme source d'évènement perturbant dans le système. Figure 5. Types de perturbation 9 3.4 Les techniques de ré-ordonnancement proposées l'efficience lors de tiennent rarement compte de Plusieurs techniques de ré-ordonnancement ont été proposées dans la littérature pour résoudre le problème de job shop dynamique. Vieira et al., (2003) ont classé ces méthodes en trois groupes : méthode de décalage à droite, ré-ordonnancement total et ré-ordonnancement partiel. Ces méthodes la sélection du nouvel ordonnancement. Dans ce travail, nous avons proposé deux techniques de ré-ordonnancement ayant comme objectif d'assurer l'efficience en respectant un seuil énergétique communiqué en temps réel par les agents AOE. Les techniques de ré-ordonnancement sont exécutées par les agents AOU. Nous rappelons que nous supposons dans cet article que les systèmes de production de ces usines sont de type job-shop. La première étape de la technique de ré-ordonnancement proposée consiste à déterminer les opérations affectées par la panne à partir de l'ordonnancement prédictif. La figure 6 illustre l'organigramme de la première méthode de ré-ordonnancement proposée. Figure 6. La technique 1 de ré-ordonnancement proposée La méthode est basée sur le changement de vitesse des machines des opérations affectées qui est, nous le rappelons, la variable d'ajustement choisie dans cet article. L'ordre de passage des opérations sur les machines reste identique à celui qui est fixé par l'ordonnancement prédictif afin de maintenir la solution la plus stable possible. À chaque itération, la vitesse de chaque opération affectée est changée et la nouvelle consommation énergétique est alors calculée. Si la nouvelle consommation respecte le taux envoyé par les AOE alors la solution est considérée comme la meilleure. Sinon, ce processus est répété sur toutes les autres opérations affectées. Si à la dernière itération, le taux énergétique demandé n'est pas respecté, alors une pénalité est associée à l'ordonnancement réactif trouvé. Dans ce cas la deuxième méthode de ré- ordonnancement est utilisée. Cette dernière se base sur le changement de séquencements, 10 jusque-là invariants. La solution obtenue risque dès lors d'être très différente de la solution initiale, le résultat risque ainsi d'être moins stable en terme de changements dans la séquence. La figure 7 illustre l'organigramme de la deuxième méthode. L'ordre de passage des opérations aux machines est ainsi ici changé. Dans cette méthode, la recherche s'effectue à partir de permutations entre opérations et les vitesses des machines. À chaque fois, la consommation d'énergie est calculée pour choisir la meilleure solution respectant au mieux la contrainte énergétique. Figure 7. La technique 2 de ré-ordonnancement proposée 4. Implémentation physiquement distribuée de l'architecture multi-agent Comme expliqué, l'architecture EasySched est composée de plusieurs AOU, plusieurs AOE et un ACE. La validation par simulation sur un seul calculateur, de manière centralisée est possible. Toutefois, ce type de validation reste partiel. Réaliser une validation réellement distribuée, sur différents systèmes, intégrant les éléments physiques (hardware) et logiciels et capteurs connectés (IOT) est proposé dans cet article. Ces éléments sont connectés en réseaux, ce qui permet un retour d'expérience plus pertinent, notamment dans l'objectif d'appliquer l'architecture sur un système réel, tel que la cellule AIP-PRIMECA de l'Université Polytechnique Hauts-de-France. La figure 8 schématise cette implémentation. L'évolution d'une implémentation centralisée vers l'implémentation proposée dans cet article y est mise en évidence. Dans un contexte multi agent, la vie d'un agent exige la présence d'un conteneur. Un conteneur est une classe abstraite qui contient tous les services nécessaires à l'hébergement et la gestion des agents durant l'exécution de leur programme. Ces conteneurs d'agents peuvent être répartis sur le réseau. Les agents créés sont par défaut localisés dans le même "Main container". Avant d'être déployé sur le système physiquement distribué composé de systèmes embarqués connectés entre eux, un conteneur est créé pour chaque agent distant. « Usine Container i » et « 11 Energy Provider Container i » sont ainsi des conteneurs alloués respectivement aux agents AOU, AOE, ACE (voir la figure 8). Des systèmes embarqués ARM A7 sont utilisés pour embarquer les containers AOE1, AOE2, AOU1, AOU2 et ACE. Figure 8. Implémentation physiquement distribuée de EasySched La quantité d'énergie renouvelable reste totalement simulée. Cependant, pour prendre en compte sa forte imprévisibilité, nous avons choisi d'utiliser des capteurs ambiants de température, d'humidité et de luminosité. Les sorties de ces capteurs (qu'il est facile de modifier) représentent le suivi de la variation des grandeurs dont dépend la puissance délivrée par les deux ressources de production d'énergie. Par exemple, la variation de l'éclairement qui influe systématiquement sur la variation de la puissance délivrée par toute centrale photovoltaïque est facilement simulée en masquant plus ou moins le capteur de lumière. De même la puissance délivrée par une centrale éolienne est dépendante de la température et de l'humidité. Elle est modifiable facilement en influençant sur le capteur de température et d'humidité. La figure suivante contient le pseudo code qui permet aux AOE de détecter une perturbation sur la puissance fournie. Il s'agit d'un comportement cyclique qui se répète chaque période fixée p1. 12 Figure 9. Pseudo code du comportement « Acquisition données » d'AOE P(t) et P(0) sont les puissances calculées à partir de valeur de température et d'humidité collectés. Dans le contexte de cet article, nous avons choisi d'utiliser le modèle proposé par (Marshall and Plumb., 2008), où la puissance électrique délivrée par une éolienne peut être calculée comme suit : Avec S est la surface de l'éolienne, V vitesse de vent et est la masse volumique de l'air calculée comme suit : avec T et sont la température et l'humidité récupéré à partir des capteurs et , la pression de l'air. Si le rapport P(t) / P(0) est inferieur à 1, la production de l'énergie diminue. La variable "Alarm" déterminant l'état de système est alors mise à jour. Afin de limiter le bruit de capteur sur les signaux et ainsi éviter de considérer chaque évolution de la valeur des signaux comme une perturbation, nous avons ajouté un algorithme artificiel de filtrage. Pour cela, nous avons contraint le choix de la période de comportement cyclique (p2) plus grande que celle de l'acquisition des données des divers capteurs. La figure suivante contient le pseudo code de ce comportement. La période p2 est alors modifiée quand l'AOE détecte une anomalie (voir figure 10). Ceci a pour premier objectif d'éviter le lancement successif de signaux de ré-ordonnancement et pour second objectif d'espacer les horizons de ré- ordonnancement. La figure suivante 11 illustre la différence entre les différentes périodes. Figure 10. Pseudo code du comportement « vérification Fausse Alarme » d'AOE 13 Figure 11. Valeurs initiales des périodes des comportements cycliques 5. Expérimentations et interprétation Pour tester l'efficacité et l'architecture au la performance de travers de son implémentation distribuée, nous avons exécuté différents jeux de test construits à partir de trois différents benchmarks de type job-shop composé de machines avec des vitesses variables avec des tailles d'instance (machines x Vmax x Pi) (3x3x10), (3x25x100) et (7x10x100). Les détails des instances sont fournis dans (Salido et al., 2016). Chaque instance est exécutée 5 fois. L'implémentation distribuée est constituée d'un PC et de cinq systèmes embarqués ARM A7 (AOE1, AOE2, AOU1, AOU2 et ACE). Le PC dispose d'un processeur Intel "Core2Duo" cadencé à 2, 4 GHz et 8 Go de RAM. La carte intégrée fonctionne à 900 Mhz avec 1Go de RAM. Le système embarqué cible exécute la version Debian de Linux. Toutes les machines exécutent Java SE Embedded. Un LAN est utilisé pour la communication entre tous les composants de l'implémentation distribuée. Pour nos expérimentations, les périodes p1, p2, p3 sont initialement fixées à 3, 6 et 8 s respectivement. La figure 12 contient différentes photos prises de notre implémentation distribuée (au sein du LAMIH). L'expérimentation est composée de deux étapes : la première permet de valider la partie prédictive de EasySched tandis que la deuxième permet de valider la partie réactive. Figure 12. Implémentation distribuée de EasySched 14 5. 1 Evaluation de la partie prédictive de l'architecture EasySched Dans cette partie, nous calculons et évaluons la solution fournie dans la partie prédictive. Le tableau 1 contient le résultat des essais. La première colonne représentant l'instance à tester, la deuxième la qualité de la solution fournie par l'AOU1, AOU2 en termes de Makespan (Mk, en s) et l'énergie consommée (E en wh), tout en précisent la valeur du facteur de pondération et le message envoyé par le AOE1 et 2 (deux états : oui ou non) qui permettra d'arrêter ou non la recherche prédictive de la solution. Chaque AOU envoie une demande de besoin énergétique après avoir calculé un ordonnancement prédictif en utilisant le PSO. Si l'AOE dispose de suffisamment d'énergie, celui-ci va envoyer un message « Oui » indiquant qu'il est possible de procéder avec cet ordonnancement. Si non, il envoie un message « Non » à l'agent AOU pour forcer le calcul d'une autre solution. Dans ce cas, l'AOU va exécuter le PSO de nouveau mais en réduisant la valeur du facteur de pondération, ce qui favorisera une réduction de la consommation d'énergie jusqu'à l'obtention d'une solution satisfaisante, correspondante à l'envoi du message « Oui ». Ces résultats illustrent le bon fonctionnement du mécanisme de coopération qui se met en place entre les consommateurs et les fournisseurs d'énergie pour converger vers des ordonnancements acceptables. Bien évidemment, d'autres tests sont à mener pour conclure de manière plus crédible sur ce bon fonctionnement. Des comparaisons avec les méthodes existantes doivent également être menées. Tableau 1 : Résultats- partie prédictive Instance AOU 1 AOE1 AOU 2 Mk E Message envoyé Mk E AOE1 Message envoyé 3x5x10 1 0.9 0.8 0.7 41 41 42.2 45.3 145.8 Non 133.7 Non 123.5 Non 112.6 Oui 7x10x100 1 625.9 2773.4 Non 3x25x100 0.9 626 2560.7 Oui - 1 0.9 0.8 0.7 - - - 1711.8 6797.2 Non 1732.2 6311 Non 1791.4 5726.2 Non 1943.7 5097.4 Non 1 0.9 - - 1 0.9 0.8 1 0.9 0.8 0.7 41 41 - - 145.8 Non 133.2 Oui - - - - 625.9 2773.4 Non 626 2560.7 Non 642.4 2418.9 Oui 1711.8 6797.2 Non 1732.2 6311 Non 1791.4 5726.2 Non 1886.3 6107.7 Non 15 0.6 0.5 2118.3 4617.4 Non - - - 0.6 0.5 2197.6 4547.9 Non 2317.1 4257.3 Oui 5. 2 Evaluation de la partie réactive de l'architecture EasySched Dans cette partie, nous testons le comportement réactif de l'architecture EasySched face à une fluctuation de l'énergie renouvelable disponible chez les AOE. La figure suivante illustre la solution prédictive délivrée par l'AOU1 connecté avec le fournisseur d'énergie que nous utiliserons. Le makespan trouvé est égal à 44 s avec une consommation d'énergie égale à 133 Wh. La particule (PSO) correspondante à cette solution ainsi que le diagramme de Gantt sont donnés dans la figure 13. Figure 13. Diagramme de Gantt de l'ordonnancement prédictif trouvé par AOU1 La figure 14 contient la courbe de la variation de température et d'humidité collectée par l'AOE à partir des capteurs nous permettant de simuler en temps réel et de manière complètement imprévisible les conditions météorologiques. Par exemple, à l'instant 22 s, la valeur de la température est « 28 » et celle de l'humidité « 33 ». La valeur de la puissance à cet instant est alors plus petite que celle à l'instant initial (T=19, H=36). Le rapport de P(t)/P(0) est égal à 0,711. Par conséquent, une alerte est délivrée aux AOU indiquant qu'il faut recalculer une nouvelle solution présentant une consommation d'énergie réduite de 26%. 16 Figure 14. Diagramme de Gantt de l'ordonnancement réactif trouvé par AOU1 La consommation d'énergie de l'ordonnancement réactif trouvé est désormais réduite pour être égale à 94Wh après la perturbation alors que le makespan se dégrade et vaut désormais 53 s. Dans ce scenario, l'adaptation de la consommation d'énergie des usines pour respecter le seuil énergétique 26% a engendré une dégradation de performance sous la forme d'une augmentation de la date de réalisation de la production. Il faut noter que même si l'on peut s'attendre à ce que ce type de comportement se reproduise dans d'autres situations (dégradation du makespan pour tenir compte d'une plus faible disponibilité énergétique), la quantification de cette dégradation dépend des modèles utilisés pour la production et la consommation d'énergie. Cette solution est trouvée en utilisant la première méthode de ré-ordonnancement proposée qui cherche à préserver la stabilité et ne changer que les vitesses des machines jusqu'à satisfaire la contrainte énergétique. Les vitesses des machines M1, M2, M3 sont ainsi changées à partir de l'instant de ré-ordonnancement. Le tableau suivant contient d'autres résultats de simulation. Les taux d'énergie délivrés par les AOE sont précisés. Dans notre cas, le pourcentage envoyé par l'AOE photovoltaïque est fixe et égal à 10 %. L'expérimentation se conclut ainsi favorablement, la réactivité de notre architecture multi-agent étant mise en évidence sur ce cas d'étude. Tout comme pour la partie prédictive, des études complémentaires doivent être menées au niveau réactif pour pouvoir conclure de manière plus générale sur l'efficacité de la méthode proposée. Des protocoles expérimentaux rigoureux sont en cours de développement. Des tests statistiques devront être menés. Cependant, les résultats obtenus et présentés dans cet article nous confortent dans cette efficacité. 17 Tableau 2 : Résultats- partie réactive Instance AOU 1 AOE1 AOU 2 AOE1 Initial Mk, E Mk 3x5x10 44, 133 53 E 94 7x10x100 626,2773 703 2099 3x25x100 1711, 6797 2056 4742 Taux E% Initial Mk, E Mk E Taux E% 26 22 28 41,145.8 42 123 626,2773 642 2418 1711,6797 1854 5466 10 10 10 6. Conclusion travail nous avons proposé EasySched, une architecture multi-agent pour Dans ce l'ordonnancement prédictif et réactif de la production en fonction d'une énergie renouvelable disponible difficilement prévisible. Nous avons proposé une implémentation distribuée originale de l'architecture qui permet de valider ces mécanismes sur des instances issues de la littérature. Une première perspective concerne l'intégration d'un mécanisme permettant de basculer entre les différentes sources d'énergie disponible au travers d'un pourcentage déterminé par négociation entre agents. Une deuxième perspective consiste à proposer une architecture holonique multicouche pour généraliser EasySched. Dans ce cas, l'agent ordonnanceur représentant l'usine connecté serait un holon récursivement composé par un ensemble d'holons représentant chacun une machine ou un produit intelligent dans l'industrie. Remerciements Le projet ELSAT2020 est cofinancé par l'Union Européenne avec le Fond européen de développement régional (FEDER), par l'Etat et la Région Hauts de France. Références Cardin, O., Ounnar, F., Thomas, A., Trentesaux, D., Future industrial systems: best practices of the intelligent manufacturing and services systems (IMS2) French Research Group, IEEE Transactions on Industrial Informatics, 13(2), 704-713, 2017. Carlier, J., Chrétienne, P., Problèmes d'ordonnancement Algorithmes, Éditions Masson, 1988. : Modélisation, Complexité, Esquirol, P., Lopez, P., L'ordonnancement, Collection Gestion, série : Production et techniques quantitatives appliquées à la gestion, Economica, 1999. Fang, K., Uhan, N., Zhao, F., Sutherland, J. W., A new approach to scheduling in manufacturing for power consumption and carbon footprint reduction, Journal of Manufacturing Systems, 30(4), 234-240, 2011. 18 Gahm, C., Denz, F., Dirr, M., Tuma, A., Energy-efficient scheduling in manufacturing companies: a review and research framework, European Journal of Operational Research, 248(3), 744-757, 2016. Giret, A., Trentesaux, D., Prabhu, V., Sustainability in manufacturing operations scheduling: A state of the art review, Journal of Manufacturing Systems, 37, 126-140, 2015. Giret, A., Trentesaux, D., Salido, M. A., Garcia, E., Adam, E., A holonic multi-agent methodology to design sustainable intelligent manufacturing control systems, Journal of cleaner production, 167, 1370-1386, 2017. González, M. Á., Oddi, A., Rasconi, R., Multi-objective optimization in a job shop with energy costs through hybrid evolutionary techniques, In Twenty-Seventh International Conference on Automated Planning and Scheduling, 2017. He, Y., Li, Y., Wu, T., Sutherland, J. W., An energy-responsive optimization method for machine tool selection and operation sequence in flexible machining job shops, Journal of Cleaner Production, 87, 245-254, 2015. Jimenez, J. F., Bekrar, A., Zambrano-Rey, G., Trentesaux, D., Leitão, P., Pollux: a dynamic hybrid control architecture for flexible job shop systems, International Journal of Production Research, 55(15), 4229-4247, 2017. Marshall, J., Plumb, R. A., Atmosphere, ocean and climate dynamics: an introductory text International Geophysics series, a series of monographs and textbooks, 93, 2008. Lamy D., Méthodes et outils pour l'ordonnancement d'ateliers avec prise en compte des contraintes additionnelles : énergétiques et environnementales, thèse de doctorat, Université Clermont Auvergne, 2017. Mikhaylidi, Y., Hussein, N., Yedidsion. L., Operations scheduling under electricity time-varying prices, Journal of Production Research, 53(23), 7136-7157, 2015. Nouiri, M., Bekrar, A., Jemai, A., Trentesaux, D., Ammari, A. C., Niar, S., Two stage particle swarm optimization to solve the flexible job shop predictive scheduling problem considering possible machine breakdowns, Computers and Industrial Engineering, 112, 595-606, 2017. Nouiri, M., Trentesaux, D., Bekrar, A., Giret, A., Salido, M. A., Cooperation Between Smart Manufacturing Scheduling Systems and Energy Providers: A Multi-agent Perspective, In International Workshop on Service Orientation in Holonic and Multi-Agent Manufacturing, Studies in Computational Intelligence, 803, Springer, 197-210, 2018. Pach, C., Berger, T., Sallez, Y., Bonte, T., Adam, E., Trentesaux, D., Reactive and energy-aware scheduling of flexible manufacturing systems using potential fields, Computers in Industry, 65(3), 434-448, 2015. 19 Pach, C., Berger, T., Sallez, Y., Trentesaux, D., Reactive control of overall power consumption in flexible manufacturing systems scheduling: A Potential Fields model, Control Engineering Practice, 44, 193-208, 2014. Plitsos, S., Repoussis, P. P., Mourtos, I., Tarantilis, C. D., Energy-aware decision support for production scheduling, Decision Support Systems, 93, 88-97, 2017. Prabhu, V., Trentesaux, D., Taisch, M., Energy-aware manufacturing operations, Journal of Production Research, 53(23), 6994-7004, 2015. Raileanu, S., Anton, F., Iatan, A., Borangiu, T., Anton, S., Morariu, O., Resource scheduling based on energy consumption Intelligent Manufacturing, 28(7), 1519-1530, 2017. for sustainable manufacturing, Journal of Salido, M. A., Escamilla, J., Barber, F., Giret, A., Rescheduling in job-shop problems for sustainable manufacturing systems, Journal of cleaner production, 162, S121-S132, 2016. Trentesaux, D., Giret, A., Tonelli, F., Skobelev, P., Emerging key requirements for future energy-aware production scheduling systems: a multi-agent and holonic perspective, In International Workshop on Service Orientation in Holonic and Multi-Agent Manufacturing, Studies in Computational Intelligence, Springer, 694, 127-141, 2016. Trentesaux, D., Borangiu, T., Thomas, A., Emerging ICT concepts for smart, safe and sustainable industrial systems, Computers in Industry, 81, 1-10, 2016. Tonelli, F., Bruzzone, A. A. G., Paolucci, M., Carpanzano, E., Nicolò, G., Giret, A., Salido, M., Trentesaux, D., Assessment of mathematical programming and agent-based modeling for off-line scheduling: application to energy aware manufacturing, CIRP Annals Manufacturing Technology, 65(1), 405-408, 2016. Vieira, G. E., Herrmann, J. W., Lin, E., Rescheduling manufacturing systems: a framework of strategies, policies, and methods, Journal of scheduling, 6(1), 39-62, 2003. Whitbrook, A., Meng, Q., Chung, P. W., Reliable, distributed scheduling and rescheduling for time-critical, multiagent systems, IEEE Transactions on Automation Science and Engineering, 15(2), 732-747, 2018. Zhang, L., Li, X., Gao, L., Zhang, G., Dynamic rescheduling in FMS that is simultaneously considering energy consumption and schedule efficiency, The International Journal of Advanced Manufacturing Technology, 87(5-8), 1387-1399, 2013. 20
1801.07747	1	1801	2018-01-23T19:51:58	Quantified Degrees of Group Responsibility (Extended Abstract)	[ "cs.MA", "cs.AI" ]	This paper builds on an existing notion of group responsibility and proposes two ways to define the degree of group responsibility: structural and functional degrees of responsibility. These notions measure the potential responsibilities of (agent) groups for avoiding a state of affairs. According to these notions, a degree of responsibility for a state of affairs can be assigned to a group of agents if, and to the extent that, the group has the potential to preclude the state of affairs.	cs.MA	cs	Quantified Degrees of Group Responsibility (Extended Abstract)1 Vahid Yazdanpanah Mehdi Dastani Utrecht University, The Netherlands 1 Introduction This paper builds on an existing notion of group responsibility in [2] and proposes two ways to define the degree of group responsibility: structural and functional degrees of responsibility. These notions measure the potential responsibilities of (agent) groups for avoiding a state of affairs. According to these notions, a degree of responsibility for a state of affairs can be assigned to a group of agents if, and to the extent that, the group has the potential to preclude the state of affairs. 2 Preliminaries In this work, the behaviour of the multi-agent system is modelled in a Concurrent Game Structure (CGS) [1] which is a tuple M = (N, Q, Act, d, o), where N = {1, . . . , k} is a set of agents, Q is a set of states, Act is a set of actions, function d : N × Q → P(Act) identifies the set of available actions for each agent in N at each state q ∈ Q, and o is a transition function that assigns a state q(cid:48) = o(q, α1, . . . , αk) to a state q and an action profile (α1, . . . , αk) such that all k agents in N choose actions in the action profile respectively. Finally, a state of affairs refers to a set S ⊆ Q and ¯S denotes the set Q \ S. In the rest of this paper, we say C ⊆ N is (weakly) q-responsible for S iff it can preclude S in q (see [2] for formal details). Let M be a multi-agent system, S a state of affairs in M, C ⊆ N an arbitrary group, and C be a (weakly) q-responsible for S in M. Definition 1 (Power measures) We say that the structural power difference of C and C in q ∈ Q with ( C, C), is equal to cardinality of C\C. Moreover, we say that C respect to S in M, denoted by ΘS,M has a power acquisition sequence (cid:104) ¯α1, . . . , ¯αn(cid:105) in q ∈ Q for S in M iff for qi ∈ Q, o(qi, ¯αi) = qi+1 for 1 ≤ i ≤ n such that q = q1 and qn+1 = q(cid:48) and C is (weakly) q(cid:48)-responsible for S in M. q 3 Structural Degree of Responsibility In our conception of Structural Degree of Responsibility (SDR), we say that any (agent) group that shares members with the responsible groups, should be assigned a degree of responsibility that reflects its proportional contribution to the responsible groups. Accordingly, the relative size of a group and its share in the responsible groups for the state of affairs are substantial parameters in our formulation of the structural responsibility degree. We would like to emphasize that this concept of responsibility degree is supported by the fact that beneficiary parties, e.g., lobbyists in the political context, do proportionally invest their limited resources on the groups that can play a role in some key decisions. Definition 2 (Structural degree of responsibility) Let WS,M denote the set of all (weakly) q-responsible groups for state of affairs S in multi-agent system M, and C ⊆ N be an arbitrary group. In case q 1The full version of this paper appears in [4]. = ∅, the structural degree of q-responsibility of any C for S in M is undefined; otherwise, the WS,M structural degree of q-responsibility of C for S in M denoted SDRS,M (C), is defined as follows: q q SDRS,M q (C) = max C∈WS,M q ({i i = 1 − ΘS,M q ( C, C) C }) Intuitively, SDRS,M q (C) measures the highest contribution of a group C in a (weakly) q-responsible C for S. Hence, structural degree of responsibility is in range of [0, 1]. 4 Functional Degree of Responsibility Functional Degree of Responsibility (FDR) addresses the dynamics of preclusive power of a group of agents (in the sense of [3]) with respect to a given state of affairs. We deem that a reasonable differentiation could be made between the groups which do have the chance of acquiring the preclusive power and those they do not have any chance of power acquisition. This notion addresses the eventuality of a state in which a group possesses the preclusive power regarding the state of affairs. This degree is formulated based on the notion of power acquisition sequence (Definition 1) by tracing the number of necessary state transitions from a source state, in order to reach a state in which the group in question is responsible for the state of affairs. Definition 3 (Functional degree of responsibility) Let PS,M (C) denote the set of all power acquisi- tion sequences of C ⊆ N in q for S in M. Let also (cid:96) = min ({i i = length(k)}) be the length k∈PS,M of a shortest power acquisition sequence. The functional degree of q-responsibility of C for S in M, denoted by FDRS,M (C), is defined as follows: (C) q q q FDRS,M q (C) = (C) = ∅ if PS,M otherwise q (cid:26) 0 1 ((cid:96)+1) The notion of FDRS,M q (C) is formulated based on the minimum length of power acquisition se- quences, which taken to be 0 if C is a (weakly) q-responsible for S. Hence, the functional degree of q-responsibility of such a C for S is equal to 1. If there exists no power acquisition sequence for C, then the minimum length of a power acquisition sequence is taken to be ∞ and the functional degree of q-responsibility of C for S becomes 0. In other cases FDRS,M 5 Conclusion (C) is strictly between 0 and 1. q The proposed notions can be used as a tool for analyzing the potential responsibility of agent groups towards a state of affairs. In our approach, the structural degree of responsibility captures the respon- sibility of an agent group based on the accumulated preclusive power of the included agents while the functional degree of responsibility captures the responsibility of a group of agents due to the potentiality of reaching a state in which it has the preclusive power. In the full version of the paper, we specify pertinent properties of the notions and consider additional semantics. References [1] Rajeev Alur, Thomas A. Henzinger, and Orna Kupferman. Alternating-time temporal logic. J. ACM, 49(5):672 -- 713, 2002. [2] Nils Bulling and Mehdi Dastani. Coalitional responsibility in strategic settings. In Proceedings of 14th International Workshop on Computational Logic in Multi-Agent Systems (CLIMA-2013), pages 172 -- 189, 2013. [3] Nicholas R Miller. Power in game forms. Springer, 1982. In Power, voting, and voting power, pages 33 -- 51. [4] Vahid Yazdanpanah and Mehdi Dastani. Quantified degrees of group responsibility. In Coordina- tion, Organizations, Institutions, and Normes in Agent Systems XI - COIN 2015, Revised Selected Papers, pages 418 -- 436, 2015.
1402.4303	2	1402	2016-03-02T17:01:37	Finding Preference Profiles of Condorcet Dimension $k$ via SAT	[ "cs.MA", "cs.AI", "cs.LO" ]	Condorcet winning sets are a set-valued generalization of the well-known concept of a Condorcet winner. As supersets of Condorcet winning sets are always Condorcet winning sets themselves, an interesting property of preference profiles is the size of the smallest Condorcet winning set they admit. This smallest size is called the Condorcet dimension of a preference profile. Since little is known about profiles that have a certain Condorcet dimension, we show in this paper how the problem of finding a preference profile that has a given Condorcet dimension can be encoded as a satisfiability problem and solved by a SAT solver. Initial results include a minimal example of a preference profile of Condorcet dimension 3, improving previously known examples both in terms of the number of agents as well as alternatives. Due to the high complexity of such problems it remains open whether a preference profile of Condorcet dimension 4 exists.	cs.MA	cs	Finding Preference Profiles of Condorcet Dimension k via SAT Christian Geist Technische Universität München Munich, Germany [email protected] 6 1 0 2 r a M 2 ] A M . s c [ 2 v 3 0 3 4 . 2 0 4 1 : v i X r a ABSTRACT Condorcet winning sets are a set-valued generalization of the well-known concept of a Condorcet winner. As supersets of Condorcet winning sets are always Condorcet winning sets themselves, an interesting property of preference profiles is the size of the smallest Condorcet winning set they admit. This smallest size is called the Condorcet dimension of a preference profile. Since little is known about profiles that have a certain Condorcet dimension, we show in this paper how the problem of finding a preference profile that has a given Condorcet dimension can be encoded as a satisfiability problem and solved by a SAT solver. Initial results include a minimal example of a preference profile of Condorcet dimen- sion 3, improving previously known examples both in terms of the number of agents as well as alternatives. Due to the high complexity of such problems it remains open whether a preference profile of Condorcet dimension 4 exists. 1. INTRODUCTION The contribution of this paper is twofold. Firstly, we pro- vide a practical implementation for finding a preference pro- file for a given Condorcet dimension by encoding the prob- lem as a boolean satisfiability (SAT) problem [2], which is then solved by a SAT solver. This technique has proven use- ful for a range of other problems in social choice theory (see, e.g., [9, 7, 4, 3]) and can easily be adapted. For instance, only little needs to be altered in order answer similar questions for dominating sets rather than Condorcet winning sets. Sec- ondly, we give an answer to an open question by Elkind et al. [5] and provide a minimal example of a preference profile of Condorcet dimension 3, which we computed using our im- plementation. This profile involves 6 alternatives and agents only, improving the size of previous examples both in terms of agents and alternatives.1 The formalization in SAT turns out to be efficient enough, not only to discover this partic- ular profile of Condorcet dimension 3, but also to show its minimality. 1For instance, the example in Elkind et al. [5] required 15 alternatives and agents. Appears in: Proceedings of the 13th International Confer- ence on Autonomous Agents and Multiagent Systems (AA- MAS 2014), Lomuscio, Scerri, Bazzan, Huhns (eds.), May, 5 -- 9, 2014, Paris, France. Copyright c(cid:13) 2014, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. 2. PRELIMINARIES Let A be a set of m alternatives and N = {1, . . . , n} a set of agents. The preferences of agent i ∈ N are represented by a linear (i.e., reflexive, complete, transitive, and antisym- metric) preference relation Ri ⊆ A × A. The interpretation of (a, b) ∈ Ri, usually denoted by a Ri b, is that agent i values alternative a at least as much as alternative b. A preference profile R = (R1, . . . , Rn) is an n-tuple containing a preference relation Ri for each agent i ∈ N . Let R be a preference profile. As introduced by Elkind et al. [5], we now define the notion of a Condorcet win- ning set through an underlying covering relation between sets of alternatives and alternatives: A set of alternatives X θ-covers an alternative y (short: X ≻θ R y) if {i ∈ N ∃x ∈ X such that x Ri y} > θn. A set of alternatives X is called a Condorcet winning set if for each alternative y /∈ X the set X 1 2 -covers y. The set of all Condorcet winning sets of R will be denoted by C(R). The Condorcet dimension dimC(R) is defined as the size of the smallest Condorcet winning set the profile R admits, i.e., dimC (R) := min{k ∈ N k = S and S ∈ C(R)}. Example 1. Consider the preference profile R depicted in Figure 1. As R does not have a Condorcet winner dimC (R) ≥ 2. It can easily be checked that {a, b} (like any other two-element set in this case) is a Condorcet winning set of R and, thus, dimC(R) = 2. 1 1 1 c a b b c a c a b Figure 1: A preference profile of Condorcet dimension 2. In this work, we address the computational problem of finding a preference profile of a given Condorcet dimension. To this end, we define the problem of checking whether for a given number of agents n and alternatives m there exists a preference profile R with dimC (R) = k. Name: Check-Condorcet-Dimension-k Instance: A pair of natural numbers n and m. Question: Does there exist a preference profile R with n agents and m alternatives that has Condorcet dimension of at least k? Preference profiles n = 3 n = 5 n = 6 n = 7 n = 10 n = 15 m = 5 m = 6 m = 7 m = 10 ∼ 1.7 · 106 ∼ 2.5 · 1010 ∼ 3.0 · 1012 ∼ 3.6 · 1014 ∼ 6.2 · 1020 ∼ 1.5 · 1031 ∼ 3.7 · 108 ∼ 1.9 · 1014 ∼ 1.4 · 1017 ∼ 1.0 · 1020 ∼ 3.7 · 1028 ∼ 7.2 · 1042 ∼ 1.3 · 1011 ∼ 3.3 · 1018 ∼ 1.6 · 1022 ∼ 8.3 · 1025 ∼ 1.1 · 1037 ∼ 3.4 · 1055 ∼ 4.8 · 1019 ∼ 6.3 · 1032 ∼ 2.3 · 1039 ∼ 8.3 · 1045 ∼ 4.0 · 1065 ∼ 2.5 · 1098 Table 1: Number of objects involved in the Check-Condorcet-Dimension-3 problem. For k = 3 the subsets of size 2 are the candidates for Condorcet winning sets. Note that the following simple observation can be used to prune the search space in terms of the number of alterna- tives. Observation 1. If there is a preference profile R of Con- dorcet dimension dimC (R) involving m alternatives, then there is also one of the same dimension involving m + 1 alternatives. Proof. Let R be a preference profile on a set of m alter- natives A with dimC(R). We need to construct a preference profile R′ on a set of m + 1 alternatives A′ = A ∪ {a′} with a′ /∈ A such that dimC (R′) = dimC (R). For each i, define R′ i := Ri ∪ {(x, a′) x ∈ A}, i.e., add a′ in the last place of agent i's preference ordering. It is then immediately clear that C(R) ⊆ C(R′), which establishes dimC(R) ≥ dimC (R′). On the other hand, if we assume dimC (R) > dimC (R′), then there exist a Condorcet winning set S′ for R′ of size k := S′ < dimC(R). This set, however, must -- by the con- struction of R′ -- also be a Condorcet winning set for R; a contradiction. 3. METHODOLOGY The number of objects potentially involved in the Check- Condorcet-Dimension-k problem are given in Table 1 for k = 3. It is immediately clear that a naıve algorithm will not solve the problem in a satisfactory manner. This section describes our algorithmic efforts to solve this problem for reasonably large instances. 3.1 Translation to propositional logic (SAT) In order to solve the problem Check-Condorcet- Dimension-k for arbitrary k ∈ N, we follow a similar ap- proach as Brandt et al. [4]: we translate the problem to propositional logic (on a computer) and use state-of-the-art SAT solvers to find a solution. At a glance, the overall solv- ing steps are shown in Algorithm 1. Generally speaking, the problem at hand can be under- stood as the problem of finding a preference profile that satisfies certain conditions -- here: having a Condorcet di- mension of at least k). Thus, a satisfying instance of the propositional formula to be designed should represent a pref- erence profile. To capture this, a formalization based on two types of variables suffices. The boolean variable ri,a,b rep- resents a Ri b, i.e., agent i ranking alternative a at least as high as alternative b; and the variable cS,y stands for the set S covering alternative y. In more detail, the following conditions/axioms need to be formalized:2 2The further axiom for neutrality is not required for correct- ness, but speeds up the solving process. It is discussed in Section 3.2. Input: positive integers n and m Output: whether there exists a preference profile R with n agents and m alternatives and dimC(R) ≥ k /* Encoding of problem in CNF File cnfFile; foreach agent i do / cnfFile += Encoder.reflexivePreferences(i); cnfFile += Encoder.completePreferences(i); cnfFile += Encoder.transitivePreferences(i); cnfFile += Encoder.antisymmetricPreferences(i); foreach set S ⊆ A with S = k − 1 do cnfFile += Encoder.noCondorcetWinningSet(S); / Symmetry breaking cnfFile += Encoder.neutrality(); /* SAT solving satisfiable = SATsolver.solve(cnfFile); if instance is satisfiable then return true; else return false / / Algorithm 1: SAT-Check-Condorcet-Dimension-k 1. All n agents have linear orders over the m alternatives as their preferences (short: linear preferences) 2. For each set S ⊆ A with S = k − 1, it is not the case that S is a Condorcet winning set (short: no Condorcet set) For the first axiom, we encode reflexivity, completeness, transitivity, and anti-symmetry of the relation Ri for all agents i. The complete translation to CNF (conjunctive normal form, the established standard input format for SAT solvers) is given exemplarily for the case of transitivity; the other axioms are converted analogously. In formal terms transitivity can be written as (∀i)(∀x, y, z) (x Ri y ∧ y Ri z → x Ri z) ≡ (∀i)(∀x, y, z) (ri,x,y ∧ ri,y,z → ri,x,z) ≡ ^i ^x,y,z ≡ ^i ^x,y,z (¬ (ri,x,y ∧ ri,y,z) ∨ ri,x,z) (¬ri,x,y ∨ ¬ri,y,z ∨ ri,x,z) , which then translates to the pseudo code in Algorithm 2 for generating the CNF file. The key in the translation of the inherently higher order axioms to propositional logic is (as pointed out by Geist and Endriss [7] already) that because of finite domains, all quantifiers can be replaced by finite conjunctions or disjunctions, respectively. In all algorithms, a subroutine r(i, x, y) takes care of the foreach agent i do foreach alternative x do foreach alternative y do foreach alternative z do variable not(r(i, x, y)); variable not(r(i, y, z)); variable(r(i, x, z)); newClause; Algorithm 2: Encoding of transitivity of individual pref- erences compact enumeration of variables.3 The axiom "no Condorcet set" can be formalized in a sim- ilar fashion, but requires further subroutines to avoid an exponential blow-up of the size of the formula in CNF. In short, the axiom can be written as (∀S ⊆ A) (S = k − 1 → S /∈ C(R)) ≡ (∀S ⊆ A)(cid:16)S = k − 1 → (∃y /∈ X)S ⊁θ ≡ ^S⊆A S=k−1 _y /∈X ¬cS,y. R y(cid:17) It remains as part of this axiom to define a sufficient con- dition for S ≻θ R y. In the following, we denote the small- est number of agents required for a strict θ-majority by m(n) := ⌊θk⌋ + 1. In formal terms, we write for each set S ⊆ A with S = k − 1 and each alternative y /∈ X: S ≻θ R y ← ((∃M ⊆ N )M = m(n)∧ (∀i ∈ M )(∃x ∈ S)x Ri y) ≡ S ≻θ R y ∨ ((∀M ⊆ N )M = m(n) → (∃i ∈ M )(∀x ∈ S)¬x Ri y) ≡ cS,y ∨  ^M ⊆N M =m(n) _i∈M ^x∈S ¬ri,x,y  . In order to avoid an exponential blow-up when converting this formula to CNF, variable replacement (a standard pro- cedure also known as Tseitin transformation) is applied. In our case, we replaced Vx∈S ¬ri,x,y by new variables of the form hS,y,i and introduced the following defining clauses:4 S=k−1 ^y∈A ^i∈N hS,y,i → ^x∈S ^S⊆A S=k−1 ^y∈A ^i∈N ¬hS,y,i ∨ ^x∈S ≡ ^S⊆A S=k−1 ^y∈A ^i∈N ^x∈S ≡ ^S⊆A ¬ri,x,y! ¬ri,x,y! (¬hS,y,i ∨ ¬ri,x,y) . In this case, the helper variables even have an intuitive meaning as hS,y,i enforces that for no alternative x ∈ S it is the case that agent i prefers alternative y over alternative x, i.e., agent i does not contribute to S θ-covering y. Note that the conditions like S = k − 1 can easily be fulfilled during generation of the corresponding CNF formula on a computer. For enumerating all subsets of alternatives of a given size we, for instance, used Gosper's Hack [8]. The corresponding pseudo code for the "no Condorcet set" axiom can be found in Algorithm 3. foreach set S ⊆ A with S = k − 1 do foreach alternative y /∈ S do variable not(c(S, y)); newClause; /* Definition of variable cS,y foreach set M ⊆ N with M = m(n) do / variable(c(S, y)); foreach agent i ∈ M do variable(h(S, y, i)); newClause; / Definition of auxiliary variable hS,y,i foreach agent i ∈ N do */ foreach x ∈ S do variable not(r(i, x, y)); variable not(h(S, y, i)); newClause; Algorithm 3: Encoding of the axiom "no Condorcet set" With all axioms formalized in propositional logic, we are now ready to search for preference profiles R of Condorcet dimension dimC(R) ≥ k. Before we do so, however, we de- scribe a (standard) optimization technique called symmetry breaking, which speeds up the solving process of the SAT solver. 3.2 Optimized computation Observe that from a given example of a preference profile R with dimC(R) ≥ k we can always generate further exam- ples simply by permuting the (names of the) alternatives. One could say that all positive witnesses to the SAT-Check- Condorcet-Dimension-k problem are invariant under per- mutations of the alternatives. Therefore, we implemented a standard technique in SAT solving called symmetry break- ing; here in the form of setting agent 1's preferences to a fixed preference ordering, for instance to lexicographic pref- erences. This trims the search space for the SAT solver and therefore reduces the runtime of the solving process. An en- coding can be achieved simply by adding a subformula of the form r(n1, x, y), ^x<y which sets the first agents preferences to lexicographic or- dering. 3The DIMACS CNF format only allows for integer names of variables. But since we know in advance how many agents and alternatives there are, we can simply use a standard enumeration method for tuples of objects. 4Note that one direction of the standard bi-implication suf- fices here. 4. INITIAL RESULTS All computations were run on a Intel Core i5, 2.66GHz (quad-core) machine with 12 GB RAM using the SAT solver plingeling [1]. 1 a b c d e f 1 1 1 c d b d f c b e f e d a a e f a b c 1 1 f e e a d a b b c c f d Figure 2: A smallest preference profile of Condorcet dimen- sion 3 (with n = 6 agents m = 6 alternatives). When called with the parameters n = m = 6, our im- plementation of SAT-Check-Condorcet-Dimension-k re- turns the preference profile Rdim3 within about one second. Rdim3 is a smallest preference profile of Condorcet dimension 3 and is shown in Figure 2.5 Furthermore, it turns out that this preference profile is a smallest profile of Condorcet dimension 3. All strictly smaller profiles (i.e., with less agents and at most as many alternatives, or with less alternatives and at most as many agents) can be shown to have a Condorcet dimension of at most 2 via SAT-Check-Condorcet-Dimension-3.6 An overview of further (preliminary) results can be found in Table 2. Model (decoding of satisfying assignment) found: Agent 0: 0 > 1 > 2 > 3 > 4 > 5 Agent 1: 2 > 3 > 5 > 0 > 4 > 1 Agent 2: 5 > 4 > 0 > 1 > 2 > 3 Agent 3: 3 > 5 > 1 > 4 > 0 > 2 Agent 4: 4 > 0 > 3 > 1 > 2 > 5 Agent 5: 1 > 2 > 4 > 3 > 5 > 0 does not have a Condorcet winning set of size 2 (6 agents and 6 alternatives). Witnesses: {0, 1} does not cover alternative(s): 5 {0, 2} does not cover alternative(s): 4 {1, 2} does not cover alternative(s): 0 {0, 3} does not cover alternative(s): 4 {1, 3} does not cover alternative(s): 0 {2, 3} does not cover alternative(s): 0 {0, 4} does not cover alternative(s): 5 {1, 4} does not cover alternative(s): 5 {2, 4} does not cover alternative(s): 1 {3, 4} does not cover alternative(s): 2 {0, 5} does not cover alternative(s): 3 {1, 5} does not cover alternative(s): 3 {2, 5} does not cover alternative(s): 1 {3, 5} does not cover alternative(s): 2 {4, 5} does not cover alternative(s): 3 Figure 3: Output of SAT-Check-Condorcet-Dimension- 3 for n = 6 agents and m = 6 alternatives. 5The witnesses for all sets S ⊆ A with S = 2 not being Condorcet winning sets are also returned by SAT-Check- Condorcet-Dimension-3 and can be obtained from the output in Figure 3. That there is a larger set (e.g., {a, b, c}) which forms a Condorcet winning set can easily be con- firmed manually (or by calling SAT-Check-Condorcet- Dimension-4). 6The running time to check all cases again is only a few seconds. m\n 1 1 2 3 4 5 6 7 8 9 10 -- -- -- -- -- -- -- -- -- -- 2 -- -- -- -- -- -- -- -- -- -- 3 -- -- -- -- -- -- -- -- -- -- 4 -- -- -- -- -- -- -- 5 6 7 -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- + -- -- + -- -- + + + 8 -- -- -- -- -- 9 10 11 12 -- -- -- -- -- -- -- -- -- -- -- -- -- + + + + -- -- -- -- -- -- -- -- -- -- -- + + + + + Table 2: Preliminary collection of results obtained with SAT-Check-Condorcet-Dimension-3 for different num- bers of alternatives m and voters n. A plus (+) stands for a preference profile found; a minus ( -- ) for the fact that all preference profiles have a Condorcet winning set of size 2. 5. OUTLOOK AND FUTURE WORK Our implementation might be useful to find preference profiles of Condorcet dimension 4, a problem that has been raised by Elkind et al. [5]. Even though with the current formalization the solving process did not terminate within a reasonable amount of time, we intend to further pursue this direction in future work. Adding further symmetry break- ing clauses (which make use of anonymity in addition to neutrality) could be a first step in this direction. Furthermore, one could extend the notion of Condorcet dimension to other individual preferences, e.g., with agents having weak (i.e., ties are allowed) or even incomplete pref- erences. Because of the high flexibility of our SAT formal- ization, one can easily apply the same method to analyze these related concepts and questions.7 A formalization with other solving techniques, e.g., ASP [6], might be another way to achieve the desired perfor- mance. Acknowledgments This material is based upon work supported by Deutsche Forschungsgemeinschaft under grant BR 2312/9-1. The au- thor thanks Felix Brandt and Hans Georg Seedig for helpful discussions and their support. REFERENCES [1] A. Biere. Lingeling, Plingeling and Treengeling entering the SAT competition 2013. In Proceedings of the SAT Competition 2013, pages 51 -- 52, 2013. [2] A. Biere, M. Heule, H. van Maaren, and T. Walsh, edi- tors. Handbook of Satisfiability, volume 185 of Frontiers in Artificial Intelligence and Applications. IOS Press, 2009. [3] F. Brandt and C. Geist. Finding strategyproof social choice functions via SAT solving. Journal of Artificial Intelligence Research, 55:565 -- 602, 2016. [4] F. Brandt, C. Geist, and H. G. Seedig. Identify- In Proceed- ing k-majority digraphs via SAT solving. ings of the 1st AAMAS Workshop on Exploring Beyond 7For the two suggested variants, deleting axioms from the formalization suffices. the Worst Case in Computational Social Choice (EX- PLORE), 2014. [5] E. Elkind, J. Lang, and A. Saffidine. Choosing collec- tively optimal sets of alternatives based on the condorcet criterion. In Proceedings of the 22nd International Joint Conference on Artificial Intelligence (IJCAI), pages 186 -- 191. AAAI Press, 2011. [6] M. Gebser, R. Kaminski, B. Kaufmann, and T. Schaub. Answer set solving in practice. Synthesis Lectures on Artificial Intelligence and Machine Learning, 6(3):1 -- 238, 2012. [7] C. Geist and U. Endriss. Automated search for impossi- bility theorems in social choice theory: Ranking sets of objects. Journal of Artificial Intelligence Research, 40: 143 -- 174, 2011. [8] D. E. Knuth. Combinatorial Algorithms, volume 4A, part 1 of The Art of Computer Programming. Addison- Wesley, 2011. [9] P. Tang and F. Lin. Computer-aided proofs of Arrow's and other impossibility theorems. Artificial Intelligence, 173(11):1041 -- 1053, 2009.
1902.05943	1	1902	2019-02-15T18:52:26	Privacy of Existence of Secrets: Introducing Steganographic DCOPs and Revisiting DCOP Frameworks	[ "cs.MA", "cs.AI", "cs.CR" ]	Here we identify a type of privacy concern in Distributed Constraint Optimization (DCOPs) not previously addressed in literature, despite its importance and impact on the application field: the privacy of existence of secrets. Science only starts where metrics and assumptions are clearly defined. The area of Distributed Constraint Optimization has emerged at the intersection of the multi-agent system community and constraint programming. For the multi-agent community, the constraint optimization problems are an elegant way to express many of the problems occurring in trading and distributed robotics. For the theoretical constraint programming community the DCOPs are a natural extension of their main object of study, the constraint satisfaction problem. As such, the understanding of the DCOP framework has been refined with the needs of the two communities, but sometimes without spelling the new assumptions formally and therefore making it difficult to compare techniques. Here we give a direction to the efforts for structuring concepts in this area.	cs.MA	cs	Privacy of Existence of Secrets: Introducing Steganographic DCOPs and Revisiting DCOP Frameworks Viorel D. Silaghi and Marius C. Silaghi and Ren´e Mandiau Florida Institute of Technology University of Valenciennes [email protected], [email protected], [email protected] Abstract Here we identify a type of privacy concern in Dis- tributed Constraint Optimization (DCOPs) not pre- viously addressed in literature, despite its importance and impact on the application field: the privacy of ex- istence of secrets. Science only starts where metrics and assumptions are clearly defined. The area of Dis- tributed Constraint Optimization has emerged at the intersection of the multi-agent system community and constraint programming. For the multi-agent commu- nity, the constraint optimization problems are an ele- gant way to express many of the problems occurring in trading and distributed robotics. For the theoretical constraint programming community the DCOPs are a natural extension of their main object of study, the constraint satisfaction problem. As such, the under- standing of the DCOP framework has been refined with the needs of the two communities, but sometimes with- out spelling the new assumptions formally and there- fore making it difficult to compare techniques. Here we give a direction to the efforts for structuring concepts in this area. Introduction We identify the "privacy of existence of secrets" as a not yet formally explored issue in the area of the Dis- tributed Constraint Optimization problems (DCOPs). In order to scientifically draw conclusions about a con- cept or procedure, it is essential to formally define the involved concepts, metrics, and assumptions. Conse- quently, we also revisit some common concepts to sug- gest a direction for the possible structuring of DCOP frameworks. The distributed constraint optimization has emerged at the confluence of research in Multi-Agent Systems and research in Constraint Programming. The multi- agent community has found in the DCOP framework an elegant way to specify requirements existing in the area of cooperating robot teams, negotiation, and trad- ing. For the Constraint Programming community, the DCOP is a natural extension of a framework that is at the center of their field, the constraint satisfaction problem (CSP). The reasons and the expectations from the new DCOP framework are slightly different for the researchers in the involved areas, differences that for various reasons are sometimes not spelled formally in- side the DCOP definitions, making comparison and ref- erence difficult and error-prone for new researchers in this booming area. The Constraint Programming community has devel- oped an impressive body of knowledge and set of tech- niques focused on improving the efficiency of problems represented as constraint satisfaction problems (CSP) and for various extensions of such problems includ- ing the egalitarian Fuzzy CSPs (FCSPs), Probabilistic CSPs (PCSPs), dynamic CSPs (DCSPs), the utilitar- ian Weighted CSPs (WCSPs), and the general Valued CSPs (VCSPs) (Schiex et al. 1995) and Semiring based CSPs (Bistarelli et al. 1999). The Constraint Program- ming community is careful in formally defining each ex- tension to CSPs and clearly stating each framework by specifying formally and exactly the inputs and unam- biguously defining outputs and evaluation metrics. This allows for work in the Constraint Programming field to be largely unambiguously comparable, and technol- ogy transfer between extensions to be possible with- out need of repeating work. However, most research in that area concerns efficiency while privacy and se- crecy of constraints are generally not part of the CSP community criteria. An early set of extensions of the CSP family to the multi-agent world addressed paral- lelism (Kasif 1990), and continued to be focused solely on efficiency, as constraint were considered shared be- tween agents. The seminal work in (Yokoo et al. 1992) brought the concept of a distribution of CSPs where privacy was mentioned as a motivation, besides natural distribu- tions of the constraints and control of variables to agents, as a reason to not attempt sharing constraints except on a need basis during a distributed backtrack- ing effort. While this is also seen as an extension of the CSP family, the efficiency was no longer a sufficient criteria and direct comparison with traditional central- ized constraints was seen as irrelevant, despite a lack of formal specification of the impediments to additional constraint sharing in terms of either technical cost or quantified privacy constraints. This lack of specifica- tion gave birth to a flurry of works exploiting various amounts of constraint information without quantifica- tion and with efficiency gains whose trade-off was not al- ways clear (Silaghi, Sam-Haroud, and Faltings 2000a) in terms of the privacy or in terms of the technical cost of constraint data centralization. In a concern to make the field coherent and easily maintain scientific soundness, the constraint commu- nity recommended at the CP2000 distributed constraint satisfaction workshop panel to name the new family of extensions by appending distributed in front of each CSP extension, yielding: • distributed constraint satisfaction (DisCSP) • distributed fuzzy constraint satisfaction (DisFCSP) • distributed probabilistic constraint satisfaction (Dis- PCSP) • distributed dynamic constraint satisfaction (DisD- CSP or DisDyCSP) • distributed weighted constraint satisfaction (Dis- WCSP) • distributed valued constraint satisfaction (DisVCSP) • distributed semiring-based constraint satisfaction (DisSCSP) Works had started to confusingly use the acronym DCSP for distributed CSPs, conflicting with the acronym for dynamic CSPs. It was recommended for the new frameworks to use the prefix "Dis" rather than "D" as a way to differentiate from dynamic CSPs which were intensively studied at the time. It was suggested that the latter be referred as DyCSP to further reduce confusion. Later, additional types of problem emerged, highlighting the wisdom of the suggestion, namely: • dynamic distributed constraint satisfaction (Dy- DisCSP) • dynamic distributed dynamic constraint satisfaction (DyDisDyCSP) where besides a dynamism in constraint emergence there also exist a dynamism in agent involve- ment redistribu- tion (Silaghi et al. 2001; Silaghi and Faltings 2002b). in constraint (i.e., churning) or However, while privacy was already a stated motiva- tion for distributed CSPs, (and by extension assumed present in all these frameworks), it remained unspeci- fied formally, perpetuating the difficulty of correct com- parison and evaluation of proposed techniques. Meanwhile, the multi-agent community adopted the new direction and developed techniques coining the name "distributed constraint optimization" (DCOP) for a framework of distributed weighted constraints. For the early DCOPs definitions, privacy was also not formally stated, and the names DCOP and Dis- WCSP coexisted for a while until the community largely adopted the DCOP acronym. Extensions in this com- munity are built on the DCOP concept, such as in the case of Leximin DCOPs (Matsui et al. 2018). to fix the problem either Further research noticed the negative effects from lack of formal definitions and quantification of pri- vacy criteria in the distributed CSP family. At- tempts redefined the framework (Silaghi et al. 2005) without changing its name/acronym (in an effort to redeem or adopt ear- lier work that stressed privacy), or alternatively pro- posed new names and acronyms. New proposed names/acronyms appeared in the case of distributed Privacy CSPs (DisPrivCSPs) or distributed privacy DCOPs (DisPrivCOPs) (Doshi et al. 2008) and Util- itarian CSPs (UCSPs) or Utilitarian DCOPs (UD- COP) (Savaux et al. 2016; Savaux et al. 2017). Work has also addressed the classification of the nature of privacy concerns, as in (L´eaut´e and Faltings 2013). In this work, privacy is defined as a value: Definition 1 Privacy is the utility that agents draw from conserving the secrecy of their personal informa- tion. In the Background section, we review the original and main formal definitions of distributed constraint satisfaction (DisCSP) and distributed constraint op- timization problems (DCOPs/DisWCSPs), as well as the main informal privacy classifications in literature. We also introduce the concept of steganography, which is more intensively studied in the area of Cryptol- ogy. In the Steganographic DCOP section we then present motivation and define formally the Stegano- graphic DCOPs. Further in the Frameworks Disam- biguation section we propose a direction for the struc- ture of the nomenclature and definitions for various con- cepts in the DCOP areas. We conclude with a summary of the contributions. Background We start by reviewing the main constraint optimiza- tion frameworks as emerging from constraint satisfac- tion. Later we review the various classifications of pri- vacy found in literature, and the definitions we used for privacy and steganography. Formal Distributed Constraint Optimization Frameworks Several distributed CSPs are based on the CSP frame- work: Definition 2 (CSP (Yokoo et al. 1998)) A CSP consists of n variables x1, x2, ..., xn, whose values are taken from finite, discrete domains D1, D2, ..., Dn respectively, and a set of constraints on their values. A constraint is defined by a predicate. That is, the con- straint pk(xk1, ..., xkj ) is a predicate that is defined on the Cartesian product Dk1 × ... × Dkj. This predicate is true iff the value assignment of these variables satisfies this constraint. Solving a CSP is equivalent to finding an assignment of values to all variables such that all constraints are satisfied. The first distributed CSP definition from (Yokoo et al. 1992; Yokoo et al. 1998) is: 1, 2, ..., m. Definition 3 (DCSP (Yokoo et al. 1998)) There exist m agents: Each variable xj be- longs to one agent i (this relation is represented as belongs(xj, i)). Constraints are also distributed among agents. The fact that an agent l knows a constraint predicate pk is represented as known(pk, l). A Distributed CSP is solved iff the following condi- tions are satisfied: ∀i, ∀xj where belongs(xj, i), value dj, is the value of xj assigned to some and ∀l, ∀pk where known(pk, l), pk is true under the assignment xj = dj. As observed, even as (Yokoo et al. 1998) has a sec- tion dedicated to privacy motivations, the formal def- inition does not specify a way to quantify privacy re- quirements and costs. Further, the semantic of variable ownership specified with the belongs predicate is not further formally defined except as implied by the fact that in the proposed algorithms, new assignments for a variable are only proposed by the agents to which the variable belongs. that While equals this definition assumes the num- the number of variables, ber of agents the domain of with each agent keeping secret exactly one variable, other versions studied the case of unequal numbers of variables and agents with no con- straints (Silaghi, Sam-Haroud, and Faltings 2000b). secret domains but with secret Definition 4 (DisCSP (Silaghi and Faltings 2005)) Distributed constraint satisfaction problems (DisCSPs) is defined by: • a set of n variables X = {x1, ..., xn}, • a set of n domains, D = {d1, ..., dn}, for the vari- ables, • a set of t constraints, C = {c1 = (xi, xj, ...), ..., ct}, each of which is a subset of the set of variables, linked with a relation, and • a set of t relations, R = {r1, ..., rt}. ri gives the al- lowed value combinations for the corresponding con- straint ci. • a set of m independent but communicating agents A = {A0, .., Am} • an ownership mapping M : X ∪ C → P(A) that assigns each variable or constraint to the subset of agents that own it. P(A) is a common notation for the set of subsets of A. A solution to a CSP is an assignment of values from the corresponding domains to each variable such that for all constraints, the combination of assigned values is allowed by the corresponding relation. The Distributed Constraint Optimization Framework was introduced in (Modi et al. 2002). Definition 5 (DCOP (Modi et al. 2002)) A Dis- tributed Constraint Optimization Problem (DCOP) consists of n variables V = {x1, x2, ...xn}, each assigned to an agent, where the values of the variables are taken from finite, discrete domains D1, D2, ...Dn, respectively. Only the agent who is assigned a variable has control of its value and knowledge of its domain. The goal is to choose values for variables such that an objective function is minimized or maximized. The objective function described is addition over costs, but can be any associative, commutative, monotonic ag- gregation operator defined over a totally ordered set of valuations, with minimum and maximum element (described by Schiex, Fargier and Verfaillie as Valued CSPs (Schiex et al. 1995)). The definition in (Modi et al. 2002) makes a formal link to the Weighted CSP instance of the Valued CSP framework while claiming generality in terms of appli- cation to remaining CSP extensions. However, while further stressing the control of values, it limits input secrecy to domains and does not further quantify effects of privacy loss or penalties for discus- sions by agents about values that they do not "control". Despite the framework not quantifying privacy, measurements of privacy loss based on qualitative com- parisons were made in (Silaghi and Faltings 2002a), and based on using assumptions from informa- tion theory, standard DCOP def- (Freuder, Minca, and Wallace 2001; initions, Wallace and Silaghi 2004; Franzin et al. 2002; Maheswaran et al. 2006) in (Greenstadt, Pearce, and Tambe 2006; Faltings, L´eaut´e, and Petcu 2008). outside and the in An effort to redefine DisCSPs to formalize privacy requirements is in (Silaghi et al. 2005): Definition 6 (DisCSP (Silaghi et al. 2005)) A [Cryptographic] Distributed CSP (DisCSP) is defined by six sets (A, X, D, C, I, O) and an algebraic structure F . A = {A1, ..., An} is a set of agents. X, D, and the solution are defined like for CSPs. I = {I1, ..., In} is a set of secret inputs. Ii is a tuple of αi secret inputs (defined on F ) from the agent Ai. Each input Ii belongs to F αi. Like for CSPs, C is a set of constraints. There may exist a public constraint in C, φ0, defined by a predicate φ0(ε) on tuples of assignments ε, known to everybody. However, each constraint φi, i > 0, in C is defined as a set of known predicates φi(ε, I) over the secret inputs I, and the tuples ε of assignments to all the variables in a set of variables Xi, Xi ⊆ X. O = {o1, ..., on} is the set of outputs to the different agents. Let m be the number of variables. oi : D1 × ... × Dm → F ωi is a function receiving as parameter a solution and returning ωi secret outputs (from F ) that will be revealed only to the agent Ai. This definition uses the same name and acronym as previously encountered definitions of distributed CSPs, but formally specifies that solutions may only be revealed to specific agents (introducing a con- cept now referred in literature as privacy of deci- sion (L´eaut´e and Faltings 2013)). The formulation re- quires that there should not be any kind of privacy loss, being mainly appropriate for cryptographic techniques. The extension of such privacy formalization to DCOPs is reported in (Silaghi 2005a): Definition 7 (DisWCSP (Silaghi 2005a)) A dis- tributed constraint satisfaction problem (DisWCSP) is defined by six sets (A, X, D, C, I, O), an arithmetic structure F , and a set of acceptable solution qualities B, that can be often represented as an interval [B1, B2]. A = {A1, ..., An} is a set of agents. X = {x1, ..., xm} is a set of variables and D = {D1, ..., Dm} is a set of finite domains such that xi can take values only from Di = {vi 1, ..., vi di}. An assignment is a pair hxi, vi ki meaning that the variable xi is assigned the value vi k. A tuple is an ordered set. I = {I1, ..., In} is a set of secret inputs. Ii is a tuple of αi secret inputs (defined on a set F ) from the agent Ai. Each input Ii belongs to F αi. C = {φ0, ..., φc} is a set of constraints. A constraint φi weights the legality of each combination of assign- ments to the variables of an ordered subset Xi of the variables in X, Xi ⊆ X. φ0 is a public constraint de- fined by a function φ0(ε) on tuples of assignments ε, known to everybody. Each constraint φi, i > 0, in C is defined as a known function φi(ε, I) over the secret inputs I, and the tuples ε of assignments to all the vari- ables in a set of variables Xi, Xi ⊆ X. The projection of a tuple ε of assignments over a tuple of variables Xi is denoted εXi. A solution is: ε∗ = argminε∈D1×...×Dn φi(εXi), c Xi=1 i=1 φi(ε∗ Xi ) ∈ [B1...B2]. if Pc O = {o1, ..., on} is the set of outputs to the different agents. oi : I × D1 × ... × Dm → F ωi is a function receiving as parameter the inputs and a solution, and returning ωi secret outputs (from F ) that will be re- vealed only to the agent Ai. The problem is to generate O. The above DCOP framework was designed to be ad- dressed with cryptographic protocols and allows for chaining multiple DCOPs for solving complex problems such as Vickrey Auctions (Silaghi 2005a). It does not support non-cryptographic privacy aware solvers as it does not model costs for partial privacy leaks, for DCOP With Privacy for Frameworks Non-cryptographic quantifica- tion of privacy in distributed CSPs is introduced in (Savaux et al. 2016), to enable the evaluation of non-cryptographic solvers: Solvers The Definition 8 (UDCSP) A UDisCSP is formally de- fined as a tuple hA, V, D, C, U, Ri where: • A = hA1, ..., Ani is a vector of n agents • V = hx1, ..., xni is a vector of n variables. Each agent Ai controls the variable xi. • D = hD1, ..., Dni where Di is the domain for the vari- able xi, known only to Ai, and a subset of {1, ..., d}. • C = {C1, ..., Cm} is a set of interagent constraints. • U = {u1,1, ..., un,d} is a matrix of privacy costs where ui,j is the cost of agent Ai for revealing whether j ∈ Di. • R = hr1, ..., rni is a vector of rewards, where ri is the reward agent Ai receives if an agreement is found. An agreement as a set of assignments for all the vari- ables with values from their domain, such that all the constraints are satisfied. The state of agent Ai includes the subset of Di that it has revealed, as well as the achievement of an agree- ment. The problem is to define a set of communication actions and a policy for each agent such that their util- ity is maximized. Another framework that allows for privacy loss but quantifies it was proposed in (Doshi et al. 2008). It is defined formalizing the concept of revelation: Definition 9 (Revelation) Given a set of Boolean (propositional) secrets S and a set of agents A, a possi- ble revelation R(A, S) is a function R(A, S) : A × S → [0, 1] which maps each peer agent and a secret to the probability learned by that agent about the secret. Definition 10 (DPCOP (Doshi et al. 2008)) A (minimization) Distributed Private Constraint Optimization Problem (DPCOP) is defined by a tuple (A, X, D, C, P, U ). A is a set of agents {A1, ..., AK}. X is a set of variables {x1, ..., xn}, and D is a set of domains {D1, ..., Dn} such that each variable xi may take values only from the domain Di. The variables are subject to a set C of sets of weighted constraints {C0, C1, ..., CK}, where Ci = {φ1 i } holds the secret weighted constraints of agent Ai, and C0 holds the public constraints. Each weighted constraint is defined as a function φi : Xi → R+ where Xi ⊆ X. The value of such a function in an input point is called constraint entry, and each Ci can be seen as a set Ci of such constraint entries. P is a set of privacy loss cost functions {P1, ..., PK }, one for each agent. Pi defines the cost inflicted to Ai by each revelation r of its secrets, i.e., Pi(r) : R(A, Ci) → R+. U is a set {U1, ..., UK}. Ui is the reward received by Ai if a solution is found (used for deciding to abandon the search). i , ..., φci A solution is an agreement between agents in A on a tuple ǫ∗ of assignments of values to variables that min- imizes the total cost: ǫ = argminǫXi  Xj φj (ǫ) (ǫ)!  + Pi Yi where Qi(ǫ) is the revelation R(A, Ci) during the pro- cess leading to the agreement on the assignments ǫ. This definition quantifies privacy and drops the con- cept of control by agents on variables, which we de- scribe as unclear in previous definitions. A vari- ant related to UDCSP is described and evaluated in (Savaux et al. 2017). Privacy Classifications There concepts tion (L´eaut´e and Faltings 2013). distributed several efforts are in classify to constraint privacy optimiza- • Domain Privacy As Existence of Values The ex- planation in (Yokoo et al. 1998) concerning the first distributed CSP definition can be interpreted as de- scribing a privacy concerning the very existence of values in domains. This goes to the point where the constraints of other agents cannot be defined exten- sionally due to the impossibility to enumerate these domains (e.g., impossibility to enumerate all meeting places that an agent controlling the corresponding variable may think about). This kind of domain pri- vacy was highlighted in (Brito and Meseguer 2003). • Domain Privacy As Unary Constraint The first types of privacy mentioned in (Yokoo et al. 1998) are domain and constraint privacy. Domain pri- vacy is frequently understood as a unary constraint on the values of a variable (Yokoo et al. 1998). This interpretation is assumed in many subse- quent works, where agents' constraints are assumed to be extensionally defined using mutual knowl- edge of the values in domains of each others' variables (Silaghi, Sam-Haroud, and Faltings 2000b; Modi et al. 2002). • Constraint Privacy for focusing controlling Constraint privacy is linked to the knows pred- in the original distributed CSP defini- icate Later work propose tion (Yokoo et al. 1998). assumptions to drop the variable in exchange on constraint pri- (Silaghi, Sam-Haroud, and Faltings 2000a; vacy Silaghi, Sam-Haroud, and Faltings 2000b). Inter- mediary frameworks allowing for some constraint privacy by distribution in order to maintain the notion of control of variables have also been proposed by splitting binary constraints into private halves, but it is unclear if the under-defined control of variables offers a tradeoff for the reduced constraint privacy. • Privacy of Existence of Solution The fact that the problem in not solvable may in itself be a secret and impact on revelation of information. This privacy was addressed first in (Silaghi 2005b), based on stochastically discarding solutions even in problems found solvable by systematic solvers, to hide solution existence. • Privacy of Decision The decision, as to what value is assigned to an individual variable in the final so- lution, only needs to be revealed to a predefined set of agents, and this was introduced in (Silaghi 2004b; Silaghi et al. 2005). • Privacy of Assignments The individual values being tried during search by an agent can be secret and certainly reveal secrets about unary constraint of the proposer. This privacy is arguably at least partially achieved by the original ABT algorithm in (Yokoo et al. 1992) where assignments are only announced to relevant agents. A twist where values are only revealed by relation was sometimes also claimed to increase this pri- (Silaghi, Sam-Haroud, and Faltings 2000b; vacy Brito and Meseguer 2003). unclear whether cryptographic algorithms offer this pri- vacy or it as in them (Silaghi 2004a; no proposal is made Silaghi and Mitra 2004). simply irrelevant them, for It is is • Privacy of Algorithms The privacy of algorithm is typically defined as a pri- vacy with respect to what reasoning processes are performed by an agent, other than the constraints on possible messages enforced by the commonly agreed solver protocol. Such privacy of algorithms is de- scribed in (Silaghi and Faltings 2002a). In solutions based on cryptographic protocols the equivalent con- cept for enforcing privacy of algorithm is introduced in (Silaghi and Friedrich 2005; Silaghi 2005a), allow- ing for the definition of a stronger level of privacy: requested privacy. Definition 11 (requested privacy) Given secret inputs σ, the prior knowledge Γ of t colluders and a multi-party computation process Π with answer α that can be decomposed in a desired data α∗ and an algo- rithmic dependent unrequested data ¯α, we say that an algorithm A achieves requested t-privacy if the proba- bility distribution of the secrets that an attacker con- trolling any at most t participants can learn is condi- tionally independent on Π, A and ¯α given requested data α∗ and prior knowledge Γ. P (σ α, Γ, Π, A) = P (σ α∗, Γ) This concept of requested privacy is the strongest known relevant privacy concept and can be relaxed for computational purposes to: Definition 12 (non-uniform requested privacy) Given secret inputs σ, the prior knowledge Γ of t colluders and a multi-party computation process Π with answer α that can be decomposed in a desired data α∗ and an algorithmic dependent unrequested data ¯α, we say that an algorithm A achieves non-uniform requested t-privacy if for any secret ¯σ ∈ σ that is not deterministically revealed given requested data α and prior knowledge Γ, it is also not deterministically revealed given Π, A, and ¯α, to any attacker controlling any at most t participants. ∀¯σ ∈ σP (¯σ α∗, Γ) < 1 ⇒ P (¯σ α, Γ, Π, A) < 1 • Privacy of Constraint Topology in was shown It (Silaghi 2004a; Silaghi, Faltings, and Petcu 2006) that erasing topological information information by combining all constraints into a big n-ary function maximizes privacy. the protection of topological information can be a secret in itself, with application to protecting trade secrets, as in (Silaghi et al. 2001; Silaghi and Faltings 2002c). Further, • Agent Identity Privacy Hiding the real identity of the agents participating in a DCOP is a concern discussed in (L´eaut´e and Faltings 2013). This refers to the general concept of anonymity, where software agents cannot be linked to human owners. Steganographic DCOPs Besides the previous types of privacy, we raise the next one: • Privacy of Existence of Secrets In an observation we introduce here, note that for many of the studied DCOP frameworks, the agents that participate in the computation may need to suggest or admit that there exist secrets involved in the computation, without which the additional costs of the setup and execution would not be warranted. But admitting that secrets may be involved is some- times a significant privacy loss in itself, for example when agents want to claim openness for political rea- sons. This concept is introduced in this paper, with the Steganographic DCOP. Steganography is the area of cryptography concerned with hiding of the existence of secret information. While the oldest example of steganography dates from Histaeus' 499BC message tattooed under the hair of a slave's head, as described by Herodotus, the best known modern techniques range from invisible ink to hiding of data in poetry and images. Steganography is more appropriate then cryptogra- phy for the case of secrets with social impact, as is often the case with distributed constraint optimiza- tion (DCOP). With social impacts, an agent may lose status simply by stating that it has secrets, and trying to safeguard them. Politicians (and not only them) try to claim openness and lack of secrets. In our experience, the need to protect existence of secrets proved to be a key impediment to the large scale adoption of cryptographic DCOP solvers, since many potential users find it difficult to publicly admit a need, or to call for secret problem solving. In this work we identify the steganographic potential of traditional DCOP solvers where the technical dis- tribution of the problem can be claimed as the unique cause of avoiding decentralization and straightforward full revelation, allowing to hide privacy needs. The basic framework recommended for stegano- graphic DCOPs, rephrasing variable ownership require- ments as unary constraint, is: Definition 13 (StegDomDCOP) A Steganographic Domain Distributed Constraint Optimization Problem (StegDomDCOP) consists of n agents A = {a1, ..., an} and n variables V = {x1, x2, ...xn}. Each agent ai has a unary constraint on xi and its domain Di. Agents know weighed constraints on subsets of these variables, specifying costs and rewards induced by assignments to concerned variables. Each agent can also identify, for each cost it associates with a total assignment, a secret reward for not revealing it. The goal is to choose values for variables such that an objective function is minimized or maximized, while each agent's participation is rational, performing only acts with positive expected sum of costs and rewards. The objective function described is addition over costs, but can be any associative, commutative, monotonic ag- gregation operator defined over a totally ordered set of valuations, with minimum and maximum elements. The requirement of having the number of vari- ables equal the number of agents can be dropped, together with the requirement that each agent has a unary constraint on a distinct variable, ob- framework that maps eas- taining an equivalent ier to the problems in (Silaghi, Sam-Haroud, and Faltings 2000b): to certain problems, like Definition 14 (StegCosDCOP) A Steganographic Cost Distributed Constraint Optimization Problem (StegCosDCOP) consists of m agents A = {a1, ..., am} and n variables V = {x1, x2, ...xn}. Agents know weighed constraints on subsets of these variables, specifying costs and rewards induced by assignments to concerned variables. Each agent can also identify, for each cost it associates with a total assignment, a secret reward for not revealing it. The goal is to choose values for variables such that an objective function is minimized or maximized, while each agent's participation is rational, performing only acts with positive expected sum of costs and rewards. The objective function described is addition over costs, but can be any associative, commutative, monotonic ag- gregation operator defined over a totally ordered set of valuations, with minimum and maximum elements. in the of fact that the last The power theory expression two frameworks are guaranteed equivalent by dual CSP conver- sions (Bacchus and Van Beek 1998). However, mod- eling strategies and solvers may perform significantly better with one rather than the other approach, since the transformation between representations can be laborious. primal is The StegDCOP framework has the advantage that its agent problems can be masqueraded as instances of any common DCOP framework, and can be used in asynchronous search protocols to protect privacy without making other agents suspicious, as in (Silaghi and Faltings 2002a). The framework enables the support of privacy of con- straints and privacy of domains in a similar way to UD- COP and DCSPs. Frameworks Disambiguation Distributed Constraint Optimization Frameworks can be classified along the following independent dimen- sions: • X1 Boolean, Fuzzy, semiring/value system: Weighted, Probabilistic • X2 structure of the problem: Static, Dynamic Local (D L), Dynamic Topology (D T) • X3 whether domains are known only by some agents, by everyone with variable ownership (enabling exten- sional constraint representation), or with distribution of constraints. Combinations are possible. The corre- sponding distribution reasons are named: Domains, Variables, Costs, Domains Costs (D C), ... • X4 privacy of decision specification, for opening all assignments in the adopted solution to everyone, or only to specified agents: Open, Closed • X5 privacy management, where participants claim no privacy and do not secretly quantify privacy con- cerns, where participants claim privacy and secretly quantify its loss, where participants claim no pri- vacy but secretly quantify privacy concerns, and where secrecy is admitted and enforced with cryp- tography: Public, Quantified, Steganographic, Cryptographic1 • X6 optimization function with utilitarian summa- tion of constraints, egalitarian Leximin, or egalitarian Theil index2: Utilitarian, Leximin, Theil A nomenclature can be designed to easily differen- tiate between the different types of frameworks, or to enable studies of framework relations and hierarchy. To maximize compatibility with past practice and the logic of the framework design, the proposed name struc- ture is: X6X5X4DisX3X2X1COP (1) where each of the names can be in camel case on any unambiguous prefix of the above properties, to enable disambiguation and easy extensions. The new nomenclature (1) is long and shorter versions can be obtained by selecting as defaults: X6=Utilitarian, X5=Public, X4=Open, X3=Domain, X2=Static, X1=Weighted. Each default parameter can be skipped if all other parameters it separates from the base string "Dis" are at their default value. Further, the Boolean- COP can also be historically named as CSP. With these conventions, for example: 1See (Yokoo, Suzuki, and Hirayama 2002; Silaghi 2003; Grinshpoun et al. 2019; Silaghi and Mitra 2004; Grinshpoun and Tassa 2016) 2See (Matsui et al. 2018; Netzer and Meisels 2011) • DCOP. The common DCOP in Defini- tion 5 (Modi et al. 2002) becomes Utilitarian- Public-Open-Dis-Domains-Static-Weighted-COP or shorter UPODisDSWCOP With the use of defaults, the acronym becomes DisCOP. • UDCOP. UDCOPs (Savaux et al. 2017) with full camel case specifiers becomes Utilitarian- Quantified-Open-Dis-Domains-Static-Weighted- COP or shorter, UQODisDSWCOP. With the use of defaults, the acronym becomes QODisCOP. • The new StegDomDCOP framework in Defini- tion 13 becomes Utilitarian-Steganographic- Open-Dis-Domains-Static-Weighted-COP. With the use of defaults, the acronym becomes SODisCOP. • The new StegCosDCOP framework in Defini- tion 14 becomes Utilitarian-Steganographic- Open-Dis-Costs-Static-Weighted-COP. With the use of defaults, the acronym becomes SODisCCOP. • DisWCSP. The DisWCSP (Silaghi 2005a) becomes tion 7 Cryptographic-Closed-Dis-Costs-Static- Weighted-COP or UCCDisCSWCOP. With the use of defaults, the acronym becomes CCDisCCOP. Defini- in Utilitarian- • DCSP. With this proposal, the original DCSP is denoted in Definition 3 (Yokoo et al. 1992) Utilitarian-Public-Open-Dis-Domains-Static- Boolean-COP or UPODisDSBCOP. With the use of defaults, acronym becomes DisCSP, as that acronym was used in several previous works (Silaghi, Sam-haroud, and Faltings 2001). the • DisCSP With the use of defaults, 4 private constraints (Silaghi and Faltings 2005) SPs with tion Utilitarian-Public-Open-Dis-Domains Costs- Static-Boolean-COP or UPODisD CSBCOP and would be denoted shortly DisD CCSP. the DisC- from Defini- denoted is We note that the term "Open" that we pro- pose for specifying the publication of the final de- cisions, has been used in the past as synonym for dynamism (Silaghi and Faltings 2002b). The term "Utilitarian" DCOP has been used for specifying ex- plicit representation of Privacy, but also to express util- itarian maximization of social welfare (Walras 1896). However, the possible acronyms in the new nomencla- ture will not conflict with prior notations (that did not already have conflict, like DCSP). Conclusion We have identified a new type of privacy relevant to Distributed Constraint Optimization, namely privacy of existence of secrets, leading to the formal introduc- tion of the Steganographic DCOPs. The new framework allows for disambiguate between problems where pri- vacy is public but quantified and when privacy require- ments are hidden (steganography). We review existing frameworks offering privacy in constraint optimization to position the new type of privacy in the literature, and propose a new nomenclature system that, if adopted, can help disambiguate all privacy frameworks found in use. Further, the obtained classification can be used to identify gaps in the literature as well as valuable new frameworks to study. References [Bacchus and Van Beek 1998] Bacchus, and Van Beek, P. 1998. On the conversion between non-binary and binary constraint satisfaction prob- lems. In AAAI/IAAI, 310 -- 318. F., [Bistarelli et al. 1999] Bistarelli, S.; Montanari, U.; Rossi, F.; Schiex, T.; Verfaillie, G.; and Fargier, H. Semiring-based csps and valued csps: Frameworks, properties, and comparison. Constraints 4(3):199 -- 240. 1999. [Brito and Meseguer 2003] Brito, I., and Meseguer, P. 2003. Distributed forward checking. In International Conference on Principles and Practice of Constraint Programming, 801 -- 806. Springer. [Doshi et al. 2008] Doshi, P.; Matsui, T.; Silaghi, M.; Yokoo, M.; and Zanker, M. 2008. Distributed pri- vate constraint optimization. In Proceedings of the 2008 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology-Volume 02, 277 -- 281. IEEE Computer Society. [Faltings, L´eaut´e, and Petcu 2008] Faltings, B.; L´eaut´e, T.; and Petcu, A. 2008. Privacy guarantees through distributed constraint satisfaction. In Web Intelligence and Intelligent Agent Technology, 2008. WI-IAT'08. IEEE/WIC/ACM International Conference on, vol- ume 2, 350 -- 358. IEEE. [Franzin et al. 2002] Franzin, M. S.; Freuder, E.; Rossi, F.; and Wallace, R. 2002. Multi-agent meeting schedul- ing with preferences: efficiency, privacy loss, and solu- tion quality. In Proceedings of the AAAI Workshop on Preference in AI and CP. [Freuder, Minca, and Wallace 2001] Freuder, E. C.; Minca, M.; and Wallace, R. J. 2001. Privacy/efficiency tradeoffs scheduling by constraint-based agents. IJCAI DCR, 63 -- 72. in distributed meeting In Proc. [Greenstadt, Pearce, and Tambe 2006] Greenstadt, R.; Pearce, J. P.; and Tambe, M. 2006. Analysis of privacy loss in distributed constraint optimization. In AAAI, volume 6, 647 -- 653. [Grinshpoun and Tassa 2016] Grinshpoun, and Tassa, T. 2016. P-syncbb: A privacy preserving branch and bound dcop algorithm. Journal of Artificial Intelligence Research 57:621 -- 660. T., [Grinshpoun et al. 2019] Grinshpoun, T.; Tassa, T.; Levit, V.; and Zivan, R. 2019. Privacy preserving re- gion optimal algorithms for symmetric and asymmetric dcops. Artificial Intelligence 266:27 -- 50. [Kasif 1990] Kasif, S. 1990. On the parallel complex- ity of discrete relaxation in constraint satisfaction net- works. Artificial Intelligence 45(3):275 -- 286. [L´eaut´e and Faltings 2013] L´eaut´e, T., and Faltings, B. 2013. Protecting privacy through distributed computa- tion in multi-agent decision making. Journal of Artifi- cial Intelligence Research 47:649 -- 695. [Maheswaran et al. 2006] Maheswaran, R. T.; Pearce, J. P.; Bowring, E.; Varakantham, P.; and Tambe, M. 2006. Privacy loss in distributed constraint reasoning: A quantitative framework for analysis and its applica- tions. Autonomous Agents and Multi-Agent Systems 13(1):27 -- 60. [Matsui et al. 2018] Matsui, T.; Matsuo, H.; Silaghi, M.; Hirayama, K.; and Yokoo, M. 2018. Leximin asymmet- ric multiple objective distributed constraint optimiza- tion problem. Computational Intelligence 34(1):49 -- 84. [Modi et al. 2002] Modi, P. J.; Shen, W.-M.; Tambe, M.; and Yokoo, M. 2002. An asynchronous complete method for general distributed constraint optimization. In Proc of Autonomous Agents and Multi-Agent Sys- tems Workshop on Distributed Constraint Reasoning. [Netzer and Meisels 2011] Netzer, A., and Meisels, A. 2011. Social dcop-social choice in distributed con- straints optimization. In Intelligent Distributed Com- puting V. Springer. 35 -- 47. [Savaux et al. 2016] Savaux, J.; Vion, J.; Piechowiak, S.; Mandiau, R.; Matsui, T.; Hirayama, K.; Yokoo, M.; El- mane, S.; and Silaghi, M. 2016. Discsps with privacy recast as planning problems for self-interested agents. In Web Intelligence (WI), 2016 IEEE/WIC/ACM In- ternational Conference on, 359 -- 366. IEEE. [Savaux et al. 2017] Savaux, J.; Vion, J.; Piechowiak, S.; Mandiau, R.; Matsui, T.; Hirayama, K.; Yokoo, M.; El- mane, S.; and Silaghi, M. 2017. Utilitarian approach to privacy in distributed constraint optimization prob- lems. In The Thirtieth International Flairs Conference. [Schiex et al. 1995] Schiex, T.; Fargier, H.; Verfaillie, G.; et al. 1995. Valued constraint satisfaction prob- lems: Hard and easy problems. IJCAI 95:631 -- 639. [Silaghi and Faltings 2002a] Silaghi, M.-C., and Falt- ings, B. 2002a. A comparison of distributed constraint satisfaction techniques with respect to privacy. In AA- MAS DCR Workshop. [Silaghi and Faltings 2002b] Silaghi, M.-C., and Falt- ings, B. 2002b. Openness in asynchronous constraint satisfaction algorithms. In 3rd DCR-02 Workshop, 156 -- 166. [Silaghi and Faltings 2002c] Silaghi, M.-C., and Falt- ings, B. V. 2002c. Self reordering for security in gener- alized english auctions (gea). In Proceedings of the first international joint conference on Autonomous agents and multiagent systems: part 1, 459 -- 460. ACM. [Silaghi and Faltings 2005] Silaghi, M.-C., and Faltings, B. 2005. Asynchronous aggregation and consistency discsp solvers for generalized vickrey auctions, complete and stochastic secure techniques. In In IJCAI05 DCR Workshop. Citeseer. [Silaghi 2005b] Silaghi, M.-C. 2005b. Hiding absence of solution for a distributed constraint satisfaction prob- lem. In FLAIRS Conference, 854 -- 855. [Wallace and Silaghi 2004] Wallace, R., and Silaghi, M. C. 2004. Using privacy loss to guide decisions in distributed csp search. In FLAIRS04. [Walras 1896] Walras, L. 1896. ´El´ements d'´economie politique pure, ou, Th´eorie de la richesse sociale. F. Rouge. [Yokoo et al. 1992] Yokoo, M.; Ishida, T.; Durfee, E. H.; and Kuwabara, K. 1992. Distributed constraint sat- isfaction for formalizing distributed problem solving. In 1992 12th International Conference on Distributed Computing System, 614 -- 621. IEEE. [Yokoo et al. 1998] Yokoo, M.; Durfee, E. H.; Ishida, T.; and Kuwabara, K. 1998. The distributed con- straint satisfaction problem: Formalization and algo- rithms. IEEE Transactions on knowledge and data en- gineering 10(5):673 -- 685. [Yokoo, Suzuki, and Hirayama 2002] Yokoo, M.; Secure dis- Suzuki, K.; and Hirayama, K. tributed constraint satisfaction: Reaching agreement without revealing private information. In International Conference on Principles and Practice of Constraint Programming, 387 -- 401. Springer. 2002. in distributed constraint satisfaction. Artificial Intelli- gence 161(1-2):25 -- 53. 2005. [Silaghi and Friedrich 2005] Silaghi, M. C., and Friedrich, G. Secure stochastic multi-party computation for combinatorial problems and a privacy concept that explicitely factors out knowledge about the protocol. Cryptology ePrint Archive, Report 2005/154. https://eprint.iacr.org/2005/154. [Silaghi and Mitra 2004] Silaghi, M.-C., and Mitra, D. Distributed constraint satisfaction and op- 2004. In Intelli- timization with privacy enforcement. gent Agent Technology, 2004.(IAT 2004). Proceedings. IEEE/WIC/ACM International Conference on, 531 -- 535. IEEE. [Silaghi et al. 2001] Silaghi, M.-C.; Sam-Haroud, D.; Calisti, M.; and Faltings, B. 2001. Generalized english auctions by relaxation in dynamic distributed csps with private constraints. In Proceedings of the IJCAI-2001 Workshop on Distributed Constraint Reasoning, 45 -- 54. Citeseer. [Silaghi et al. 2005] Silaghi, M.-C.; Abhyankar, A.; Zanker, M.; and Bart´ak, R. 2005. Desk-mates (sta- ble matching) with privacy of preferences, and a new distributed csp framework. AAAI Press. [Silaghi, Faltings, and Petcu 2006] Silaghi, M.-C.; Falt- ings, B.; and Petcu, A. 2006. Secure combinatorial optimization simulating dfs tree-based variable elimi- nation. In ISAIM. [Silaghi, Sam-Haroud, and Faltings 2000a] Silaghi, M. C.; Sam-Haroud, D.; and Faltings, B. 2000a. Dis- tributed asynchronous search with private constraints. In Proceedings of the fourth international conference on Autonomous agents, 177 -- 178. ACM. [Silaghi, Sam-Haroud, and Faltings 2000b] Silaghi, M.- C.; Sam-Haroud, D.; and Faltings, B. 2000b. Asyn- chronous search with aggregations. In AAAI/IAAI, 917 -- 922. [Silaghi, Sam-haroud, and Faltings 2001] Silaghi, M.- C.; Sam-haroud, D.; and Faltings, B. 2001. Abt with asynchronous reordering. In Intelligent Agent Tech- nology: Research and Development. World Scientific. 54 -- 63. [Silaghi 2003] Silaghi, M.-C. 2003. Arithmetic circuit for the first solution of distributed csps with cryptographic multi-party computations. In IAT, 609 -- 613. [Silaghi 2004a] Silaghi, M.-C. 2004a. Meeting schedul- ing guaranteeing n/2-privacy and resistant to statistical analysis (applicable to any discsp). In Web Intelligence, 2004. WI 2004. Proceedings. IEEE/WIC/ACM Inter- national Conference on, 711 -- 715. IEEE. [Silaghi 2004b] Silaghi, M.-C. 2004b. Secure multi-party computation for selecting a solution according to a uni- form distribution over all solutions of a general combi- natorial problem. Cryptology ePrint Archive, Report 2004/333. https://eprint.iacr.org/2004/333. [Silaghi 2005a] Silaghi, M. C. 2005a. Using secure
1604.08226	1	1604	2016-04-27T20:22:27	Influence of Agents Heterogeneity in Cellular Model of Evacuation	[ "cs.MA" ]	The influence of agents heterogeneity on the microscopic characteristics of pedestrian flow is studied via an evacuation simulation tool based on the Floor-Field model. The heterogeneity is introduced in agents velocity, aggressiveness, and sensitivity to occupation. The simulation results are compared to data gathered during an original experiment. The comparison shows that the heterogeneity in aggressiveness and sensitivity occupation enables to reproduce some microscopic aspects. The heterogeneity in velocity seems to be redundant.	cs.MA	cs	Influence of Agents Heterogeneity in Cellular Model of Evacuation Pavel Hrab´aka,∗, Marek Buk´acekb aBrno University of Technology, Faculty of Civil Engineering, Vever´ı 331/95, 602 00 Brno, bCzech Technical University in Prague, Faculty of Nuclear Sciences and Physical Engineering, Trojanova 13, 120 00 Prague, Czech Republic Czech Republic Abstract The influence of agents heterogeneity on the microscopic characteristics of pedes- trian flow is studied via an evacuation simulation tool based on the Floor-Field model. The heterogeneity is introduced in agents velocity, aggressiveness, and sensitivity to occupation. The simulation results are compared to data gathered during an original experiment. The comparison shows that the heterogeneity in aggressiveness and sensitivity occupation enables to reproduce some microscopic aspects. The heterogeneity in velocity seems to be redundant. Keywords: Travel time, Floor-Field model, Aggressiveness, Bonds principle, Heterogeneity, Pedestrian dynamics 1. Introduction The presented study of the heterogeneity in CA models directly extends the contribution on aggressiveness presented in [1]. Detailed study of the het- erogeneity in more aspects of agents in Floor-Field model is based on empir- ical observations related to variety of experiments conducted by our research group [2, 3, 4]. We aim to investigate, whether some microscopic aspects of the pedestrian flow can be mimicked introducing heterogeneity to some parameters of Floor-Field model. The experiments mentioned above were designed to study the boundary induced phase transition, which has been analysed theoretically in [5] for Floor- Field model. The object of such study is a rather small room with one exit and one multiple entrance, which may be considered as one segment of a large network. During the experiment, an important aspect of pedestrian behaviour has been observed: participants have different ability to push through the crowd, ∗Corresponding author. Tel.: +420 224 358 567; fax: +420 234 358 643 Email addresses: [email protected] (Pavel Hrab´ak), [email protected] (Marek Buk´acek) Preprint submitted to Journal of Computational Science July 29, 2021 what leads to significant variance in the time spent by the pedestrian in the room. The presented cellular model is based on the Floor-Field Model [6, 7, 8] with adaptive time-span [9] and principle of bonds [10]. The adaptive time span en- ables to model heterogeneous stepping velocity of pedestrians; the principle of bonds helps to mimic collective behaviour of pedestrians in lines. For compre- hensive summary of Floor-Field model modifications capturing different aspects of pedestrian flow and evacuation dynamics we refer the reader to [11]. In this article we focus on the heterogeneity in three aspects of the pedes- trian flow: the desired velocity, aggressiveness (or ability to win conflicts), and sensitivity to occupancy (related to line formation). 2. Experiment Presented simulation study leans over the experiment "passing-through", which set-up is schematically depicted in in Figure 1. Detailed analyses of the experiment with respect to microscopic aspects of pedestrian flow can be found in [3, 4]. Selected results of the above mentioned studies are summarized in this section. Videos capturing some aspects of the experiment are available at http://gams.fjfi.cvut.cz/peds. Figure 1: Taken from [3]. Left: Setting of the experiment, a = 7.2 m, b = 4.4 m. Right: sketch of pedestrian's hat used for automatic image recognition. In the experiment participants were instructed to walk through the rect- angular room with one 60 cm wide exit and multiple entrance. The inflow of pedestrians α [ped/s] has been artificially controlled in order to investigate the boundary induced phase transition, which is not the subject of this article. Nevertheless, thanks to that the pedestrian behaviour could be observed under variety of conditions. The size of the crowd in front of the exit played the main role. During each run of the experiment, pedestrians were passing repeatedly through the room in order to keep stable inflow for sufficiently long time. Due to that there are 20 to 40 records for each participant. Moreover, the unique codes assigned to all participants enabled the study of individual properties of the pedestrians under variety conditions. 2 26252423222120control bitbinary codecentre of massrecognitionmarkUsually the maximal outflow/capacity is related to bottleneck width b. In [12] the dependence is suggested Jmax = 1.9 · b. Here we note that the measured ca- pacity of the bottleneck in the experiment was approximately 1.4 ped/s, which is higher than suggested. It was caused by high motivation of pedestrians to exit the room nad by the fact that the narrowing was fallowed by a slightly wider corridor. Despite that the motion of pedestrians within the following corridor was strictly one-lane without overtaking. 2.1. Travel-Time The key investigated quantity is the travel time T T denoting the time in- terval between the entrance at Tin and the egress at Tout of each pedestrian, i.e., T T = Tout − Tin. To capture the pedestrians behaviour under variety of conditions, the travel time is investigated with respect to the average number of pedestrians in the room Nmean defined as Nmean = 1 Tout − Tin(cid:90) Tout Tin N (t)dt , (1) where N (t) stands for the number of pedestrians in the room at time t. As expected the average T T increases with respect to the number of pedestrians in the room N , referred further as the occupancy. Surprisingly, the variance of the T T increases dramatically with occupancy as well. Figure 2: Taken from [4]. Scatter plot of the travel time T T with respect to the occupancy Nmean extracted from the experiment. Three participants are highlighted. Their travel time is approximated by the piecewise linear model (2). We can see that Ped. 2 has lower desired velocity in free regime but higher ability to push through the crowd in comparison to Ped. 4. Figure 2 shows the scatter plot of all pairs (Nmean, T T ) gathered over all runs of experiment and all participants. Records corresponding to three chosen pedestrians are highlighted. We can observe that the reaction of participants to the occupancy N significantly differs. The mean travel time in the free-flow regime (0 - 7 pedestrians) reflects the pedestrian desired velocity; the slope of the travel-time dependence on the occupation N in the congested regime (10 - 3 0510152025303540455001020304050N_mean [peds]TT [s] Ped. 17 dataPed. 17 modelPed. 4 dataPed. 4 modelPed. 2 dataPed. 2 model45 pedestrians) reflects the pedestrian ability to push through or walk around the crowd. This observation corresponds to the piece-wise linear model for each pedestrian S v0(i) T T = + 1{N >7}(N − 7) · slope(i) + noise (2) where S = 7.2 m, v0(i) is the free-flow velocity of the pedestrian i, slope(i) is the unique coefficient of the linear model for pedestrian i. The breakpoint N = 7 depends from the room geometry. The weighted mean of the R2 value of the model (2) is 0.688. 2.2. Heterogeneity and Travel-Time The high variance of the T T caused by the heterogeneous reaction to crowd is reflected in the distribution of the relative travel-time T TR, i.e., normalized T T with respect to the average of records corresponding to similar occupancy Nmean. Figure 3 present the histograms of T TR for free flow regime (0-7 pedestrians) and for congested regime (30-50 pedestrians). Figure 3: Taken from [4]. Realative travel-time in free-flow 0-7 ped (left) and congested regime 30-45 ped (right). There are two aspects explaining the high variance of the T T accompanied with extremely low and high values in congested regime: the aggressiveness and better path choice. We have observed that some of the pedestrians were pushing effectively through the crowd in order to get to the exit. This caused that some less aggressive pedestrians stayed trapped within the crowd or on its edge for significantly longer time. Furthermore, observing the path choice of "fast" and "slow" pedestrians we can make a conclusion that the lower travel-time was reached by walking around the crowd as well, see Figure 4. 3. Model Definition The model used for this study comes out from the Floor-Field cellular model with several modifications introduced in [10, 9] and studied in more detail in [13] extended by the aggressiveness element [1]. The playground of the model is represented by the rectangular two-dimensional lattice L ⊂ Z2 consisting of cells x = (x1, x2). Every cell may be either occupied by one agent or empty. 4 00.511.522.533.5020406080TT_R00.511.522.533.5050100150TT_RFigure 4: Taken from [4]. Trajectories of pedestrians in front of the exit grouped according to the related travel-time, high T T left and low T T right. Agents are moving along the lattice by hopping from their current cell x ∈ L to a neighbouring cell y ∈ N (x) ⊂ L, where the neighbourhood N (x) is Moore neighbourhood, i.e., N (x) = {y ∈ L; maxj=1,2 xj − yj ≤ 1} . 3.1. Choice of the New Target Cell Agents choose their target cells y from N (x) stochastically according to probabilistic distribution P (y x; state of N (x)), which reflects the "attrac- tiveness" of the cell y to the agent. Part of the "attractiveness" is expressed by means of the static field S, where S(y) =(cid:112)y12 + y22 is the distances of cell y ∈ L to the exit cell E = (0, 0) acting as the common target for all agents. As usual, P (y x) ∝ exp{−kSS(y)}, for y ∈ N (x), where kS ∈ [0, +∞) denotes the parameter of sensitivity to the field S. The probabilistic choice of the target cell is further influenced by the occu- pancy of neighbouring cells and by the diagonality of the motion. An occupied cell is considered to be less attractive, nevertheless, it is meaningful to allow the choice of an occupied cell while the principle of bonds is present (explanation of the principle of bonds follows below). Furthermore, the movement in diagonal direction is penalized in order to suppress the zig-zag motion in free flow regime and support the symmetry of the motion with respect to the lattice orientation. Technically this is implemented as follows. Let Ox(y) be the identifier of agents occupying the cell y from the point of view of the agent sitting in cell x, i.e. Ox(x) = 0 and for y (cid:54)= x Ox(y) = 1 if y is occupied and Ox(y) = 0 if y is empty. Then P (y x) ∝ (1 − kOOx(y)), where kO ∈ [0, 1] is again the parameter of sensitivity to the occupancy (kO = 1 means that occupied cell will never be chosen). Similarly can be treated the diagonal motion defining the diagonal movement identifier as Dx(y) = 1 if (x1 − y1) · (x2 − y2) (cid:54)= 0 and Dx(y) = 0 otherwise. Sensitivity parameter to the diagonal movement is denoted by kD ∈ [0, 1] (kD = 1 implies that diagonal direction is never chosen). The probabilistic choice of the new target cell can be than written in the final form P (y x) = exp(cid:8) − kSS(y)(cid:9)(cid:0)1 − kOOx(y)(cid:1)(cid:0)1 − kDDx(y)(cid:1) (cid:80)z∈N (x) exp(cid:8) − kSS(z)(cid:9)(cid:0)1 − kOOx(z)(cid:1)(cid:0)1 − kDDx(z)(cid:1) . (3) It is worth noting that the site x belongs to the neighbourhood N (x), therefore the equation (3) applies to P (x x) as well. 5 Figure 5: Choice of the target cell with respect to sensitivity to occupancy. Left: kO = 0, bond is created. Right: kO = 1, motion to second most attractive cell. 3.2. Updating Scheme The used updating scheme combines the advantages of fully-parallel up- date approach, which leads to necessary conflicts, and the asynchronous clocked scheme [14] enabling the agents to move at different rates. Each agent carries as his property the own period denoted as τ , which rep- resents the minimal time between two steps, leading to desired times of update to be t = kτ , k ∈ Z. The time-line is divided into isochronous intervals of the length h > 0. During each algorithm step k ∈ Z such agents are updated, whose desired time of the next actualization lies in the interval(cid:2)kh, (k + 1)h(cid:1). A wise choice of the interval length h in dependence on the distribution of τ enables to model heterogeneous velocity of agents while keeping sufficient number of conflicts. Here we note that we use the concept of adaptive time-span, i.e., the time of the desired actualization is recalculated after each update of the agent, since it can be influenced by the essence of the motion, e.g., diagonal motion leads to a time-penalization, since it is √2 times longer. For more detail see e.g. [9]. This is an advantage over the probabilistic approach introduced in [15]. 3.3. Principle of Bonds and Sensitivity to Occupation The principle of bonds is closely related to the possibility of choosing an occupied cell. An agent who chooses an occupied cell builds a bond to the agent sitting in there. This bond lasts until the motion of the blocking agent or until the next activation of the bonded agent. The idea is that the bonded agents attempt to enter their chosen cell immediately after it becomes empty. The parameter kO (sensitivity to occupation) influences the willingness of agents to bond. The effect of different occupancy values is illustrated in Figure 5. If kO = 0, agents choose the next target cell regardless whether it is occupied or empty, i.e, in case the most attractive cell (closer to the exit) is occupied, the agent creates a bond to the occupying agent assuming his motion within the step. This strategy supports the motion in lines and can lead to fluent flow through the bottleneck. On the other hand, the ignorance of occupancy leads to the increase of conflicts, which may block the motion in front of the 6 000111Figure5:Occupancy.3.2.UpdatingSchemeTheusedupdatingschemecombinestheadvantagesoffully-parallelup-dateapproach,whichleadstonecessaryconflicts,andtheasynchronousclockedscheme[16]enablingtheagentstomoveatdifferentrates.Eachagentcarriesashispropertytheownperioddenotedasτ,whichrep-130resentstheminimaltimebetweentwosteps,leadingtodesiredtimesofupdatetobet=kτ,k∈Z.Thetime-lineisdividedintoisochronousintervalsofthelengthh>0.Duringeachalgorithmstepk∈Zsuchagentsareupdated,whosedesiredtimeofthenextactualizationliesintheinterval(cid:2)kh,(k+1)h(cid:1).Awisechoiceoftheintervallengthhindependenceonthedistributionofτenablesto135modelheterogeneousvelocityofagentswhilekeepingsufficientnumberofcon-flicts.Herewenotethatweusetheconceptofadaptivetime-span,i.e.,thetimeofthedesiredactualizationisrecalculatedaftereachupdateoftheagent,sinceitcanbeinfluencedbytheessenceofthemotion,e.g.,diagonalmotionleadstoatime-penalization,sinceitis√2timeslonger.Formoredetailseee.g.[11].140Thisisanadvantageovertheprobabilisticapproachintroducedin[17].3.3.PrincipleofBondsandSensitivitytoOccupationTheprincipleofbondsiscloselyrelatedtothepossibilityofchoosinganoccupiedcell.Anagentwhochoosesanoccupiedcellbuildsabondtotheagentsittinginthechosencell.Thisbondlastsuntilthemotionoftheblockingagent145oruntilthenextactivationofthebondedagent.Theideaisthatthebondedagentsattempttoentertheirchosencellimmediatelyafteritbecomesempty.TheparameterkO(sensitivitytooccupation)influencesthewillingnessofagentstobond.TheeffectofdifferentoccupancyvaluesisillustratedinFigure5.IfkO=0,agentschoosethenexttargetcellregardlesswhetheritisoccupied150orempty,i.e,incasethemostattractivecell(closertotheexit)isoccupied,theagentcreatesabondtotheoccupyingagentassuminghismotionwithinthestep.Thisstrategysupportsthemotioninlinesandcanleadtofluentflowthroughthebottleneck.Ontheotherhand,theignoranceofoccupancyleadstotheincreaseofconflicts,whichmayblockthemotioninfrontofthe155exit.HereitisimportanttomentionthattheparameterkO<1influencesthecellattractivenessbeforethenormalization.Therefore,incaseofextremely6Figure 6: Conflict solution for γ1 < γ2. Left: More aggressive wins the conflict over two less aggressive. Right: The conflict of two more aggressive can resolve by the blocking of the movement. exit. Here it is important to mention that the parameter kO < 1 influences the cell attractiveness before the normalization. Therefore, in case of extremely attractive cell (e.g. the exit), the probability distribution of choosing the next target cells does not change dramatically. 3.4. Solution of Conflicts and Aggressiveness The in essence parallel updating scheme is inevitable accompanied by con- flicts, when more agents decide to enter the same cell. In such case, one of the agents is usually chosen at random to win the conflict. There are more ap- proaches, how the randomness is executed, e.g. uniformly, or proportionally to the hopping probabilities [6]. Important role in models of pedestrian evacuation play the unresolved conflicts, i.e., the aim to attempt the same cell leads to the blocking of the motion. This is captured by the friction parameter µ denoting the probability that none of the agent wins the conflict. An improvement is given by the friction function [16], which raises the friction according to the number of conflicting agents. For purposes of this article, we introduce the choice of the winning agent based on his ability to win conflicts represented by an additional parameter γ ∈ [0, 1], which is referred to as the aggressiveness. Similar heterogeneity in agents behaviour has been used in [17], where the "aggressiveness" has been represented by the willingness to overtake. The conflict is always won by agents with highest γ among conflicting agents. If there are two or more agents with the highest γ, the friction parameter µ plays a role. In this article we assume that the higher is the aggressiveness γ, the less should be the probability that none of the agents wins the conflict. Therefore, the conflict is not solved with probability µ(1 − γ) (none of the agents move). With complement probability 1− µ(1− γ) the conflict resolves to the motion of one of the agents. This agent is chosen randomly with equal probability from all agents involved in the conflict having the highest γ. The conflict solution among heterogeneous group of agents is depicted in Figure 6. 7 6P.Hrab´akM.Buk´acekIftherearetwoormoreagentswiththehighestγ,thefrictionparameterµplaysarole.Inthisarticleweassumethatthehigheristheaggressivenessγ,thelessshouldbetheprobabilitythatnoneoftheagentswinstheconflict.Therefore,theconflictisnotsolvedwithprobabilityµ(1−γ)(noneoftheagentsmove).Withcomplementprobability1−µ(1−γ)theconflictresolvestothemotionofoneoftheagentswithhighestγ.Thisagentischosenatrandom.Themechanismofthefrictioncanbeeasilymodified.AnexampleofconflictsolutionisdepictedinFigure3.γ2γ1γ1γ2γ1γ1γ1γ2γ2µ(1−γ2)γ1γ2γ21−µ(1−γ2)2γ1γ2γ21−µ(1−γ2)2γ1γ2γ2Fig.3.Conflictsolutionforγ1<γ2.Left:Moreaggressivewinstheconflictovertwolessaggressive.Right:Theconflictoftwomoreaggressivecanresolvebytheblockingofthemovement.4ImpactoftheAggressivenessElementTheeffectoftheaggressivenesshasbeenstudiedbymeansofthesimulation.ResultsstressedinthisarticlecomefromthesimulationswithparametersgivenbyTable1.Thevaluesofτandγaredistributedamongagentsuniformlyandindependentlyoneachother.Table1.Valuesofparametersusedforsimulation.kSkOkDhµτγ3.510.70.1s0.5{.25,.4}{0,1}Thesimulationset-uphasbeendesignedaccordingtotheexperiment,i.e.,theroomofthesize7.2m×4.4mhasbeenmodelledbytherectangularlat-tice18siteslongand11siteswide.Thesizeofonecellthereforecorrespondsto0.4m×0.4m.Theexitisplacedinthemiddleoftheshorterwall,theopenboundaryismodelledbyamultipleentranceontheoppositewall.Newagentsareenteringthelatticestochasticallywiththemeaninflowrateα[pedes-4. Simulation Setting Main goal of this article is to study the influence of heterogeneity in chosen parameters. In this article we focus on the heterogeneity in free-flow velocity (represented by own frequency), aggressiveness, and sensitivity to occupation. In order to have comparable results to the experiment, the parameters have been calibrated to give similar values of important macroscopic quantities as free flow velocity (1.57 m/s in experiment) and maximal outflow (1.4 ped/s in experiment). The used set of parameters is given in Table 1. Table 1: Parameter values and description Parameter Value ∆x [cm] kS kD µ h [s] τ [s] 40 3.5 0.7 0.9 0.2 0.2 {0.15, 0.4} 0.14 {0, 1} 0.9 {0.1, 0.95} γ kO Range Description Lattice constant Sensitivity to potential Penalization of diag. motion Friction parameter [0,∞) [0, 1] [0, 1] (0,∞) Length of one time step (0,∞) Homogeneous own period Heterogeneous own period Homogeneous aggressiveness [0, 1] Heterogeneous aggressiveness Homog. sensitivity to occupation Heter. sensitivity to occupation [0, 1] The free flow simulation (without interactions) is directly influenced by kS, kD and τ . The diagonal penalization kD together with time penalization of diagonal motion have been tested in previous research. The values of kS and τ have been chosen to agree with the mean and variance of the free-flow veloc- ity. Here we note that the pedestrians in the experiment walked relatively fast (1.57 m/s), which motivated us to set the algorithm step h = 0.2 s of real time to balance significant decrease of velocity in congested regime. The motion of agents in crowd is influenced by parameters µ, γ, and kO. These parameters have been calibrated by means of the maximal outflow from the exit in congested situation, i.e, the exit capacity (1.4 ped/s). The significant decrease of velocity in crowd is modelled by means of relatively high friction µ = .9. Here we note that such high friction is necessary to compensate the conflict solution mechanism related to aggressiveness and the motion in lines related to the sensitivity to occupation. Our goal is to illustrate the effects of heterogeneity in chosen parameters. Therefore, the values of outflow had not been calibrated directly to the value 1.4 m/s, but sufficiently close to it, see Figure 7. Such approach enables to fit the homogeneous and heterogeneous values of each parameter independently and therefore can be used in all considered scenarios -- 1: hom (homogeneous in all parameters), 2: tau (heterogeneous in velocity), 3: agr (heterogeneous in aggressiveness), 4: obs (heterogeneous in sensitivity to occupation), 5: agr,obs 8 Figure 7: The box-plots of outflow Jout from the congested room (N = 50) measured for 20 runs of the simulation experiment for each parameter set. (heterogeneous in aggressiveness and occupation independently), 6: agr+obs (heterogeneous in aggressiveness and occupation with dependence, i.e., more aggressive is more sensitive to occupation). The simulation has been performed for system with periodic boundaries, i.e., the egress of an agent causes the entrance of another one (contrary to [1], where open boundaries were used giving the same results). The properties of agents (τ, γ, kO) were drawn from uniform distribution on the groups of parameters. For each set of parameters the simulation was performed for the occupancy N ∈ {1, 3, 5, 7, 10, 12, 14, 17, 20, 30, 40, 45, 50, 75, 100}. The simulation was performed until 1000 agents walked through the exit and it was repeated 20 times. All quantities are averaged. 5. Simulation Results In Table 2 the measured values of free flow velocity v0 (velocity for agents under the occupancy N ≤ 4), maximal outflow Jout, and average travel-time T T N for all 6 settings are compared to the experimental values. From the values we can see that the average travel-time is slightly, but not significantly, overestimated by the model. The lower free-flow velocity in case 2: tau is caused by the heterogeneity in velocity. The used set of parameters did not allow to fit both, the free-flow velocity and the outflow in this case. The reason lies in the synchronous update with h = 0.2 s. As will be discussed below, the heterogeneity in velocity defined by the own frequency is not desirable for the presented model with respect to the performed experiment. Table 2: Average quantities measured for different parameter settings. v0 [m/s] Jout [ped/s] T T 45 [s] T T 100 [s] Exp. 1.57 1.40 24.31 -- 1:hom 2:tau 1.57 1.39 30.76 66.83 1.11 1.42 30.74 67.74 3:agr 1.57 1.37 30.72 67.84 4:obs 1.57 1.38 31.17 67.52 5:agr,obs 1.57 1.35 32.01 67.63 6:agr+obs 1.57 1.30 33.05 70.59 9 1: hom2: tau3: agr4: obs5: agr, obs6: agr+obs1.21.31.41.5Jout [ped/s]The main stress is given to the travel-time study and dependence on the average occupancy Nmean, which reflects the heterogeneity in the reaction to the crowd. The T T − Nmean plots for studied parameter sets are given in Figure 8. In the graphs the mean travel time is plotted according by groups with the same parameters, the overall mean and quantiles are present for completeness. From the graphs is evident that the model is able to mimic the piecewise linear dependence (2) of T T on Nmean, which is present in the experimental study. The break-point of the model is in agreement with the experimental observation N = 7. We can see quite good agreement with the experiment regarding the trend of the dependence of T T on the occupancy not only in average but also in the minimal and maximal measured values, see Figure 2. Due to to significantly lower volume of data from the experiment it is reasonable to compare the 0.1 and 0.9 quantiles. Here we note that the lower average T T in experiment (24.31 s) than produced by the model (approx. 30 s) is caused by the small volume of data related to N ≈ 45 in the experiment -- such crowded conditions were kept for relatively short time due to the small number of participants (75). We can see that the differences in the slope of the linear dependence can be observed in all heterogeneous scenarios. Nevertheless, for further studies we have neglected the heterogeneity in the own updating period τ , related to veloc- ity. Even the homogeneous setting of free-flow velocity can reproduce variances in the free-flow travel-time in sufficient manner due to the stochastic nature of Floor-Field model. Therefore, another heterogeneity in the parameter τ is redundant. On the other hand, the concept can be used in case of evident heterogeneity where the desired velocity significantly differs [18, 19]. The heterogeneity in aggressiveness parameter γ becomes evident in occu- pancy N ∈ (10, 50), the effect does not rise significantly for higher occupancy, see sub-figure 3. Complementary evinces itself the heterogeneity in sensitiv- ity to occupancy kO (sub-figure) 4, which becomes most evident for occupancy N > 50. Interesting results brings the combination of these two parameters. The scenario 5 shows that the heterogeneity in γ and kO combines both effects from 3 and 4 in superposition manner. The main differences in the slope of T T shows the scenario 6, where the aggressiveness and occupancy are dependent, i.e., the more aggressive agents are more sensitive to occupation. Here we note that the sensitivity to occupation should be interpreted as the willingness to join an existing queue in order to move along this queue. It is therefore reasonable to suppose that the aggressive behaviour is related to the willingness to overtake, i.e., kO → 1 For further comparison let us focus on the distribution of the relative travel time T TR. In Figure 9 we can see the histograms of T TR for scenarios 1, 3, and 6, i.e., with increasing heterogeneity. We can make a conclusion that with increasing heterogeneity increases the relative frequency of low values of T TR (first bin). Similarly, the modus of the distribution is closer to lower values for higher heterogeneity (bins 4, 3, 2). Looking at the histograms of T TR related to heterogeneous scenarios, we conclude that the final distribution can be considered as a mixture of two uni- 10 Figure 8: Average travel against mean occupancy plotted by groups. Black line represents the mean over all entries, gray lines correspond to 0.1, 0.5, and 0.9 quantils. 11 020406080100050100150TT [s]1: hom hommean0204060801000501001502: tau tau=.15tau=.40mean020406080100050100150TT [s]3: agr agr=0agr=1mean0204060801000501001504: obs obs=.10obs=.95mean020406080100050100150Nmean [ped]TT [s]5: agr, obs a=0, o=.1a=0, o=.95a=1, o=.1a=1, o=.95mean020406080100050100150Nmean [ped]6: agr+obs a=0, o=.1a=1, o=.95meanmodal distribution corresponding to two groups of agents, which seems to be the case of the experiment as well, see Figure 3. Figure 9: normalized histograms of relative travel-time T TR for scenarios 1: hom, 3: agr, and 6: agr+obs. Although some differences between scenarios 3 and 6 six are evident, they might be considered as marginal in comparison with the homogeneous case 1. Let us therefore focus on another aspect observed during the experiment -- In Figure 10 we can see the snapshots from the simulation the path choice. showing a representing situation of the simulation. In scenario 3, the more aggressive agents are more successful in pushing through the crowd, but they do not evince any preference in path-choice. In scenario 6, the aggressive agents prefer walking around the crowd and hopping to the exit from the left or right. As a consequence, less aggressive agents standing in lines remain often trapped few cell away from the exit, as often observed during the experiment. Figure 10: Snapshots from the simulation of scenarios 1: hom, 3: agr, and 6: agr+obs approximately 60 s after initiation. By red dots are marked more aggressive pedestrians. 6. Summary and Conclusions The article focuses on the heterogeneity in three parameters of the original cellular model based on Floor-Field model. These parameters are then consid- 12 0123400.010.020.030.040.050.06TTR1: hom all0123400.010.020.030.040.050.06TTR3: agr allagr=0agr=10123400.010.020.030.040.050.06TTR6: agr+obs allagr=0, obs=.1agr=1, obs=.9502468101214161820051015Time = 34.8, Occupancy = 50, A = 5002468101214161820051015Time = 68.4, Occupancy = 50, A = 5002468101214161820051015Time = 40.6, Occupancy = 50, A = 50ered as the properties carried by individual agents, specifically the heterogeneity in own period (velocity), aggressiveness, and sensitivity to occupancy is studied. It is important to mention that the introduced concept of aggressiveness and principle of bonds can be used in other rule-based cellular model of pedestrian dynamics. The introduction of heterogeneity into the model has been inspired by the results of conducted experiments. It has been show that pedestrians react dif- ferently to the crowd: some are able to push through the crowd effectively, some remain trapped in front of the exit. Moreover, this is not a property of individual trajectories, but this ability is closely related to the pedestrians. In this study we aim to show that the heterogeneity in the ability to win con- flicts is necessary to reproduce the above mentioned effects in cellular models. Such property can be useful for proper modelling of an evacuation of a complex structure, where the heterogeneity in the ability to win conflicts is expected. In such cases a group of people can remain trapped within the facility for un- justifiable long time, although the average flows and evacuation times fulfil the expectations. The introduced aspects of heterogeneity can be summarized as follows: 1. Velocity: The heterogeneity in velocity causes undesired bi-modal his- togram in the free-flow regime. The observed heterogeneity in pedestrian sample can be sufficiently modelled by the stochastic nature of the de- cisions, the variance in the travel time is then related to the number of deviations from the direct path. Nevertheless, the concept can be used in dramatically heterogeneous scenario where the average speed of one group of pedestrians is two times higher than the average velocity of other group. 2. Aggressiveness: This conflict-solution method, where the conflict is won by the agent with higher value of aggressiveness, seems to be very effective in the reproduction of high variance of the T T in the congested regime. In combination with the heterogeneity of the velocity we are able to sim- ulate a situation when slower agent is more aggressive than a fast one. It is important to note that the term aggressiveness may be a little bit misleading, since it corresponds to the ability to win conflicts, which may be given by some rules of preference as well. 3. Sensitivity to Occupancy: This aspect influences mainly the space usage by the agent in given conditions. The lower the sensitivity is, the higher is the average T T for the agent, since he waits in lines and can be overtaken and trapped in front of the exit. This parameter plays very important role in combination with the parameter of aggressiveness. Acknowledgements This work was supported by the Czech Science Foundation grant GA15- 15049S and by the Czech Technical University grant SGS15/214/OHK4/3T/14. 13 References References [1] P. Hrab´ak, M. Buk´acek, Conflict solution according to "aggressiveness" of agents in floor-field-based model, in: R. Wyrzykowski (Ed.), PPAM 2015, Part II, Vol. 9574 of Lecture Notes in Computer Science, Springer International Publishing Switzerland, 2016, pp. 1 -- 10. [2] M. Buk´acek, P. Hrab´ak, M. Krb´alek, Experimental analysis of two- dimensional pedestrian flow in front of the bottleneck, in: M. Chraibi, M. Boltes, A. Schadschneider, A. Seyfried (Eds.), Traffic and Granular Flow '13, Springer International Publishing, 2015, pp. 93 -- 101. [3] M. Buk´acek, P. Hrab´ak, M. Krb´alek, Experimental study of phase transi- tion in pedestrian flow, in: W. Daamen, D. C. Duives, S. P. Hoogendoorn (Eds.), Pedestrian and Evacuation Dynamics 2014, Vol. 2 of Transportation Research Procedia, Elsevier Science B.V., 2014, pp. 105 -- 113. [4] M. Buk´acek, P. Hrab´ak, M. Krb´alek, Individual microscopic results of bot- tleneck experiments, in: Traffic and Granular Flow '15, Springer Interna- tional Publishing, to be published, arXiv:1603.02019 [physics.soc-ph]. [5] T. Ezaki, D. Yanagisawa, K. Nishinari, Analysis on a single segment of evacuation network, Journal of Cellular Automata 8 (5-6) (2013) 347 -- 359. [6] C. Burstedde, K. Klauck, A. Schadschneider, J. Zittartz, Simulation of pedestrian dynamics using a two-dimensional cellular automaton, Physica A: Statistical Mechanics and its Applications 295 (3-4) (2001) 507 -- 525. [7] A. Kirchner, A. Schadschneider, Simulation of evacuation processes using a bionics-inspired cellular automaton model for pedestrian dynamics, Physica A: Statistical Mechanics and its Applications 312 (12) (2002) 260 -- 276. [8] T. Kretz, Pedestrian traffic, simulation and experiments, Ph.D. thesis, Uni- versitat Duisburg-Essen, Germany (2007). [9] M. Buk´acek, P. Hrab´ak, M. Krb´alek, Cellular model of pedestrian dynam- ics with adaptive time span, in: R. Wyrzykowski (Ed.), PPAM 2015, Part II, Vol. 8385 of Lecture Notes in Computer Science, Springer Berlin Hei- delberg, 2014, pp. 669 -- 678. [10] P. Hrab´ak, M. Buk´acek, M. Krb´alek, Cellular model of room evacuation based on occupancy and movement prediction: Comparison with experi- mental study, Journal of Cellular Automata 8 (5-6) (2013) 383 -- 393. [11] A. Schadschneider, D. Chowdhury, K. Nishinari, Stochastic Transport in Complex Systems: From Molecules to Vehicles, Elsevier Science B. V., Amsterdam, 2010. 14 [12] A. Seyfried, O. Passon, B. Steffen, M. Boltes, T. Rupprecht, W. Klingsch, New insights into pedestrian flow through bottlenecks, Transportation Sci- ence 43 (3) (2009) 395 -- 406. [13] M. Buk´acek, P. Hrab´ak, Case study of phase transition in cellular models of pedestrian flow, in: J. Was, G. Sirakoulis, S. Bandini (Eds.), Cellular Automata, Vol. 8751 of Lecture Notes in Computer Science, Springer In- ternational Publishing AG, 2014, pp. 508 -- 517. [14] D. Cornforth, D. G. Green, D. Newth, Ordered asynchronous processes in multi-agent systems, Physica D 204 (1-2) (2005) 70 -- 82. [15] S. Bandini, L. Crociani, G. Vizzari, Heterogeneous pedestrian walking speed in discrete simulation models, in: M. Chraibi, M. Boltes, A. Schad- schneider, A. Seyfried (Eds.), Traffic and Granular Flow '13, Springer In- ternational Publishing, 2015, pp. 273 -- 279. [16] D. Yanagisawa, A. Kimura, A. Tomoeda, R. Nishi, Y. Suma, K. Ohtsuka, K. Nishinari, Introduction of frictional and turning function for pedestrian outflow with an obstacle, Phys. Rev. E 80 (2009) 036110. [17] X. Ji, X. Zhou, B. Ran, A cell-based study on pedestrian acceleration and overtaking in a transfer station corridor, Physica A: Statistical Mechanics and its Applications 392 (8) (2013) 1828 -- 1839. [18] P. Kleczek, J. Was, Simmulation of pedestrian behavior in shopping mall, in: J. Was, G. Sirakoulis, S. Bandini (Eds.), Cellular Automata, Vol. 8751 of Lecture Notes in Computer Science, Springer International Publishing AG, 2014, pp. 650 -- 659. [19] E. Spartalis, I. Georgoudas, G. Sirakulis, Ca crowd modeling for a retire- ment house evacuation with guidance, in: J. Was, G. Sirakoulis, S. Bandini (Eds.), Cellular Automata, Vol. 8751 of Lecture Notes in Computer Science, Springer International Publishing AG, 2014, pp. 481 -- 491. 15
1505.02595	1	1505	2015-05-11T13:07:43	Multi-Agent Distributed Coordination Control: Developments and Directions	[ "cs.MA", "eess.SY" ]	In this paper, the recent developments on distributed coordination control, especially the consensus and formation control, are summarized with the graph theory playing a central role, in order to present a cohesive overview of the multi-agent distributed coordination control, together with brief reviews of some closely related issues including rendezvous/alignment, swarming/flocking and containment control.In terms of the consensus problem, the recent results on consensus for the agents with different dynamics from first-order, second-order to high-order linear and nonlinear dynamics, under different communication conditions, such as cases with/without switching communication topology and varying time-delays, are reviewed, in which the algebraic graph theory is very useful in the protocol designs, stability proofs and converging analysis. In terms of the formation control problem, after reviewing the results of the algebraic graph theory employed in the formation control, we mainly pay attention to the developments of the rigid and persistent graphs. With the notions of rigidity and persistence, the formation transformation, splitting and reconstruction can be completed, and consequently the range-based formation control laws are designed with the least required information in order to maintain a formation rigid/persistent. Afterwards, the recent results on rendezvous/alignment, swarming/flocking and containment control, which are very closely related to consensus and formation control, are briefly introduced, in order to present an integrated view of the graph theory used in the coordination control problem. Finally, towards the practical applications, some directions possibly deserving investigation in coordination control are raised as well.	cs.MA	cs	MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 1 Multi-Agent Distributed Coordination Control: Developments and Directions Xiangke Wang1∗, Xun Li1, Yirui Cong12, Zhiwen zeng1 and Zhiqiang Zheng12 1College of Mechantronic Engineering and Automation, National University of Defense Technology, 410073, 2 Research School of Engineering, the Australian National University, Canberra ACT 2601, Australia Changsha, China SUMMARY In this paper, the recent developments on distributed coordination control, especially the consensus and formation control, are summarized with the graph theory playing a central role, in order to present a cohesive overview of the multi-agent distributed coordination control, together with brief reviews of some closely related issues including rendezvous/alignment, swarming/flocking and containment control. In terms of the consensus problem, the recent results on consensus for the agents with different dynamics from first- order, second-order to high-order linear and nonlinear dynamics, under different communication conditions, such as cases with/without switching communication topology and varying time-delays, are reviewed, in which the algebraic graph theory is very useful in the protocol designs, stability proofs and converging analysis. In terms of the formation control problem, after reviewing the results of the algebraic graph theory employed in the formation control, we mainly pay attention to the developments of the rigid and persistent graphs. With the notions of rigidity and persistence, the formation transformation, splitting and reconstruction can be completed, and consequently the range-based formation control laws are designed with the least required information in order to maintain a formation rigid/persistent. Afterwards, the recent results on rendezvous/alignment, swarming/flocking and containment control, which are very closely related to consensus and formation control, are briefly introduced, in order to present an integrated view of the graph theory used in the coordination control problem. Finally, towards the practical applications, some directions possibly deserving investigation in coordination control are raised as well. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc INTERNATIONAL JOURNAL OF ROBUST AND NONLINEAR CONTROL Int. J. Robust. Nonlinear Control 2013; 00:2–28 Received . . . Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/rnc KEY WORDS: Multi-agent system; Coordination control; Graph theory; Consensus; Formation control 1. INTRODUCTION Researchers have long noticed and carried on detailed analysis on many coordinated behaviors, for example, the forage for food or defense against predators of insects, birds and fishes in nature, and self-organization or self-excitation of particles in physics [1–5]. These drove researchers to consider seriously why the creature and particles take initiative to coordinate, and motivated the studies and applications on multi-agent coordination. Though it appears to be more complicated than single-agent systems, there are indeed many significant advantages in the coordinations of MAS (multi-agent system) over the single-agent system, for example [6–11]: • distributed sensors and actuators, as well as inherent parallelism; • larger redundancy, higher robustness and great fault tolerance; • performing tasks that single-agent system cannot do; • performing usually the tasks that can be finished by single costly agent with lower cost and higher excellent performance. No doubt that there are more advantages as well. In addition, with the improvements of the embedded computing and communicating technologies, and the developments of distributed and decentralized methodologies, the distributed coordinations of MASs have become easy to materialize. With these inspirations, the problem of distributed coordination control of a network of mobile autonomous agents was of interest in control and robotics in the past decade. ∗Correspondence to: College of Mechantronic Engineering and Automation, National University of Defense Technology, 410073, Changsha, China. E-mail: [email protected]. Contract/grant sponsor: the Research Projection of National University of Defense Technology; contract/grant number: JC13-03-02 Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Prepared using rncauth.cls [Version: 2010/03/27 v2.00] MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 3 Generally speaking, distributed coordination control uses local interactions between agents to achieve collective behaviors of multiple agents, and therefore to accomplish global tasks. It has a broad range of potential applications, such as in military, aerospace, industry, and entertainment. Figure 1 shows some typical applications of multi-agent distributed coordinations. In the military field, multiple mobile robot systems can adopt a proper geometric pattern to perform military tasks for taking the place of human soldiers, such as reconnaissance, searching, mine clearance, and patrol under adverse/hazardous circumstances. Taking the reconnaissance mission as an example, a single robot has limited ability to gain environmental information, however, if multiple robots keep proper formation to cooperatively apperceive the surrounding, they are likely to rapidly and accurately obtain the environmental information. In the aerospace field, satellite formation is the leading technique in the space application in 21th century, which opens up a brand-new direction for the application of satellites, especially for mini-satellites. Satellite formation cannot only greatly reduce the cost and enhance the reliability and survivability, but also broaden and override the function of individual satellites and achieve the tasks that multiple single spacecrafts cannot finish. In the industrial field, multiple mobile robots can deal with some dull, dirty and dangerous work in formation. For example, when multiple robots cooperatively carry large scale goods in a poisonous environment, their positions and orientations are strictly restricted in order to meet the requirements of load balance. In the entertainment field, for example, multiple dancing robots or soccer robots, in order to keep neat or meet tactical needs, must keep some special patterns, and may dynamically switch the patterns for avoiding obstacles. Distributed coordination control of MASs has attracted tremendous attention with various aspects including consensus, formation, rendezvous, alignment, synchronization, swarming, flocking, containment control, gossip, cooperative searching and reconnaissance, cooperative location and mapping, to name a few. A number of methods have been proposed for achieving distributed coordination control of MAsS, and conventional methods include the leader-follower method [12– 16], behavior-based method [17–20], virtual structure method [21–24], graph-based method [6, 7, 25–27], etc. But now, these methods are gradually mixed together, and it is hard to clearly Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 4 X. WANG, ET AL. Figure 1. Some typical applications of multi-agent distributed coordinations distinguish from each other. In particular, the graph-based method has become the dominator, because MASs can be modeled by a graph naturally, and almost all the aspects of coordination control can be then studied by utilizing the graph theory. Therefore, this paper mainly focuses on the recent developments of consensus, formation control, and some closely related topics such as rendezvous/alignment, swarming/flocking and containment control, with the graph theory paying central roles, in order to present a cohesive overview of the multi-agent distributed coordination control. Also, some topics which might be interesting for future investigation are discussed. 2. RECENT DEVELOPMENTS A MAS is in general distinguished with larger scale (often decided by the numbers of agents) and decentralized perception, communication and control structures, and forms inter-connected network between agents consequently. Thus, it can be naturally modeled by a graph with vertices being used to describe agents and the edges being used to represent topological relationships between agents such as perception, communication and control links. Both directed and undirected graphs can be used to describe the MAS. For example, when the ith agent is required to keep the predetermined distance to the jth agent, while the jth agent is not required to keep the distance to the ith agent, the edge ~ij is directed from vertex i to j, and at this time a directed graph is used to model the Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 5 MAS. Conversely, when the interconnection between agents ith and jth is bi-directional, when the ith agent could perceive the jth agent if and only if the jth agent could perceive the ith agent for example, the edge ij is undirected, and at this time the graph is also undirected. After modelled by a graph, as noted above, the mature graph theory can be borrowed to study the coordination control of MASs. In recent years, many monographs about the research results of the graph theory on multi-agent coordination control problems, e.g. [8–11], are published. Therefore, in the sequel, we will mainly pay attention on the survey of recent developments on some typical multi-agent coordination problems, including consensus, formation control, rendezvous/alignment, swarming/flocking and containment control, with graph theory playing a central role. The relationships among the investigated coordination control problems are summarized as Fig. 2. (cid:38)(cid:82)(cid:81)(cid:86)(cid:72)(cid:81)(cid:86)(cid:88)(cid:86) (cid:54)(cid:87)(cid:68)(cid:87)(cid:88)(cid:86)(cid:3) (cid:68)(cid:85)(cid:72)(cid:3) (cid:83)(cid:82)(cid:86)(cid:76)(cid:87)(cid:76)(cid:82)(cid:81)(cid:86) (cid:36)(cid:83)(cid:83)(cid:79)(cid:76)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:86) (cid:36) (cid:71) (cid:71) (cid:76) (cid:81) (cid:74) (cid:3) (cid:83) (cid:68) (cid:74) (cid:72) (cid:87) (cid:87) (cid:82) (cid:80) (cid:72) (cid:85) (cid:72) (cid:87) (cid:81) (cid:85) (cid:76) (cid:70) (cid:3) (cid:53)(cid:72)(cid:81)(cid:71)(cid:72)(cid:93)(cid:89)(cid:82)(cid:88)(cid:86)(cid:15) (cid:36)(cid:79)(cid:76)(cid:74)(cid:81)(cid:80)(cid:72)(cid:81)(cid:87)(cid:15) (cid:54)(cid:90)(cid:68)(cid:85)(cid:80)(cid:76)(cid:81)(cid:74)(cid:15) (cid:41)(cid:79)(cid:82)(cid:70)(cid:78)(cid:76)(cid:81)(cid:74)(cid:15) (cid:38)(cid:82)(cid:81)(cid:87)(cid:68)(cid:76)(cid:81)(cid:80)(cid:72)(cid:81)(cid:87)(cid:3) (cid:38)(cid:82)(cid:81)(cid:87)(cid:85)(cid:82)(cid:79) (cid:41)(cid:82)(cid:85)(cid:80)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3) (cid:74)(cid:72)(cid:81)(cid:72)(cid:85)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3) (cid:68)(cid:81)(cid:71) (cid:80)(cid:68)(cid:76)(cid:81)(cid:87)(cid:68)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74) (cid:41)(cid:82)(cid:85)(cid:80)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3) (cid:38)(cid:82)(cid:81)(cid:87)(cid:85)(cid:82)(cid:79) (cid:42)(cid:85)(cid:68)(cid:83)(cid:75)(cid:3)(cid:87)(cid:75)(cid:72)(cid:82)(cid:85)(cid:92) (cid:38)(cid:82)(cid:82)(cid:85)(cid:71)(cid:76)(cid:81)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:38)(cid:82)(cid:81)(cid:87)(cid:85)(cid:82)(cid:79) Figure 2. Relationships among consensus, formation control and some closely related topics, i.e., rendezvous, alignment, swarming, flocking and containment control. 2.1. Consensus In the networks of agents, the consensus means to reach an agreement regarding a certain quantity of interest that depends on the states of all agents. A consensus protocol is an interaction rule that specifies the information exchange between an agent and all of its neighbors in the network [28,29]. In general, there are two key elements considered in consensus problems, that is, the dynamics of agents and the communications among agents, as illustrated in Fig. 3. To be more precise, two parts are normally taken into account in the agents’ dynamics, which are the dynamics of the agent Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 6 X. WANG, ET AL. itself and the protocol used among agents. The former is the inherent model of the agents, such as the first/second/high-order models, linear/nonlinear models, and continuous/discrete models; the latter is used to modify the agent inherent model in order to make the MAS achieve a consensus. The communications among agents can also be divided into communicating structures such as the shapes and changing of the communication topology, and data transmissions including time delays and transmission failure. Dynamics Dynamics Agent ( Inherent Models ) Protocals (Modified Models) Communications Communications Structures Data Transmissions Figure 3. Key elements considered in consensus problems On the other hand, the consensus problems can also be divided into two categories by considering the predominance of agents. If an agent has priority so that the other agents must follow its trajectory, then such an agent is called the “leader”; accordingly, if the MAS has at least one leader, then the consensus is leader-follower consensus; otherwise, it is leaderless consensus. Even though it seems there is an obvious difference between leader-follower and leaderless consensus, they have the same foundations, because the topology and consensus protocol can make a common agent act as a leader. Afterwards, we will mainly focus on the developments of leaderless consensus and then point out some important results about leader-follower consensus. The beginning of researches on consensus is [30], in which Vicsek et al. propose a simple discrete- time model of particles all moving in the plane at the same speed but with different headings. Afterwards, the algebraic graph theory, which studies the relationships between the structure of graphs and different matrix representations of graphs, is widely employed in the consensus [31]. The most important concepts of the algebraic graph theory used in MASs are the adjacent matrix and the Laplacian matrix. For a graph G = (V, E) consisting of a nonempty vertex set V = {1, . . . , N } and an edge set E ⊆ V × V, the adjacency matrix A(G) = [aij ] is an N × N matrix given by Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 7 aij = 1, aij > 0 if and only if (i, j) ∈ E, and aij = 0, otherwise. The degree of vertex i is given by di = Pj aij, and the Laplacian matrix of G is Ln(G) = diag(d1, . . . , dN ) − A(G). The pioneering work to deal with consensus by virtue of the Laplacian matrix is literature [18], in which a theoretical explanation for the observed behavior in [30] is provided. After that a large number of literatures use the algebraic graph theory to study the consensus problem, for example, literature [29, 32–40], to name a few. The most basic case of consensus is that the considered agents are governed by linear dynamics, especially by single-integrator dynamics. The theoretical framework for reaching a consensus for networked agents with single-integrator dynamics under fixed/switching communicating topology was introduced by Olfati-Saber and Murray in literature [32], building on their earlier work of [25]. The subsequent studies mainly followed up on the way of [32]. For single-integrator dynamics, the commonly used continuous-time consensus protocol is [29, 32–34, 39] xi(t) = − n X j=1 aij (t)(cid:0)xj (t) − xi(t)(cid:1), i = 1, . . . , n; (1) for first-order discrete systems, the commonly used discrete-time consensus protocol is [18, 29, 34, 40] xi[k + 1] = n X j=1 aij[k]xj[k], i = 1, . . . , n, (2) in which xi(t) and xi[k] represent the states of the ith agent, aij(t) and aij[k] are the (i, j) entry of the adjacent matrix of the associated communication graph of MASs at time t or k(t). Note that protocols (1) and (2) only use the relative state information of the adjacent agent, so it is local and distributed. Then, by using the tools of the algebraic graph theory, different conclusions for different cases are made. Under the condition of the fixed topology, protocols (1) and (2) can guarantee a consensus if the communication graph is connected for an undirected graph or has a rooted spanning tree for a directed graph. By contrast, when the topology is switching, protocols (1) and (2) result in a consensus if there exists an infinite sequence of contiguous, uniformly bounded time intervals, with the property that across each interval, the union of the communication graphs is connected for an undirected case or has a rooted directed spanning tree for a directed case. Furthermore, by Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 8 X. WANG, ET AL. using the maximum and minimum eigenvalues of the Laplacian matrix, the converging rate of the consensus protocol is analyzed [41]. Time delays associated with both message transmission and processing, for instance caused by the communication congestion, are inevitable and also taken into account in consensus problems. Let δij denote the time delay communicated from the jth agent to the ith agent, protocol (1) is modified as xi(t) = − n X j=1 aij (t)(cid:0)xj (t − δij ) − xi(t − δij )(cid:1). (3) Reference [32] investigated the simplest case where δij = δ and the communication topology was fixed, undirected, and connected, and concluded that the consensus was achieved if and only if 0 ≤ δ < π/(2λmax(L)), where L denoted the Laplacian of the communication graph. Along the way developed by [32], different time delays and communication topologies are considered in the consensus problem. If the time delay affects only the state being transmitted, protocol (3) is modified as xi(t) = − n X j=1 aij(t)(cid:0)xj(t − δij) − xi(t)(cid:1), (4) and the consensus for switching directed topologies remains valid for an arbitrary time delay when δij = δ [42]. The constant or time-varying, uniformly or non-uniformly distributed time delays are considered in [43], and sufficient conditions for the existence of average consensus under bounded communication delays are given correspondingly. The consensus problem with noise- perturbed under fixed and switching topologies as well as time-varying communication delays is investigated in [44]. It is shown that the consensus is reached for arbitrarily large constant, time-varying, or distributed delays if consensus is reached without delays [45]. For the discrete consensus protocol (2), conclusions similar to that of the continuous case can be obtained. It is shown in [46] that if a consensus is reached under a time-invariant undirected communication topology, then the presence of communication delays does not affect the consensus. In addition, sufficient conditions for the consensus under dynamically changing communication topologies and bounded time-varying communication delays are shown in reference [47]. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 9 The results of the first-order system have been extended to the consensus of the second-order and high-order linear dynamical MASs, and similar consensus protocols and graph conditions are obtained. Consensus for multiple agents governed by the second-order linear dynamics has been considered by the similar framework of using the algebra graph theory [48–51]. For example, in the representative work [48], a linear distributed consensus protocol for the second-order MAS is designed with aid of the Laplacian matrix without requiring velocity information of neighbors; a variable structure control law is used to design the consensus protocol by taking the second- order system as two cascade fist-order systems in [51]. In terms of high-order consensus problems, the first discussion is completed by W. Ren in [52], where the second-order MAS is generalized to the lth-order integrator MAS. Afterwards, the consensus protocol is modified in [53] and χ- consensus is investigated under undirected communication topology. It should be noted that both the aforementioned literatures have no self-feedback information in the consensus protocols. With adding the self-feedback items, reference [54] investigates the constant-value consensus problem for high-order MASs under fixed and switching directed topology. Based on those meaningful explorations, the robust analysis and the converging results with time delays are further provided, for example, the work in [55–57]. The consensus governed by more complex high-order linear dynamics, such as the SISO (single input single output) and the MIMO (multiple inputs multiple outputs) linear dynamics are also investigated. At the beginning, researcher did not notice the relationship between the consensus in MAS and the controllability in individual agents; therefore, many studies are done with the assumption that the isolated agent is controllable, e.g. the work in [58]. It is proven that this technical assumption is not necessary in [59], excluding the necessity of a sufficient connection of the graph topology is required for consensus stability for a system with unstable agent dynamics. Afterwards, the studies on communication delays and robustness are conducted, e.g. [60, 61]. And further, a necessary and sufficient condition for high-order consensus with unknown communication delays is given for the existence of solution to heterogeneous multi- agent systems in [62]. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 10 X. WANG, ET AL. Considering that most physical systems are inherently nonlinear in nature, the consensus where the agents are governed by nonlinear dynamics has also aroused the attention of some researchers recently. It can be shown, for example, in [63, 64] where the continuously differentiable nonlinear dynamics are considered; in [65–67] where the nonlinear dynamics satisfy the global Lipschitz condition under a directed communicating graph, respectively, without and with a leader agent; and in [68] where an adaptive control method is introduced to study the synchronization of uncertain nonlinear networked systems. 2.2. Formation Control The formation control is that a team comprised of multiple agents keeps a predetermined geometric pattern and adapts to the environmental constraints (e.g. obstacle avoidance) at the same time during the movement towards a specific goal. In general, the main control problems in formation are summarized as follows [69, 70]: • Formation generation: how to design a formation pattern for MASs and then achieve it? • Formation maintaining: how to maintain the formation pattern for MASs during the movement? • Formation transformation: how to transit the formation pattern, which means transiting one formation pattern into another? • Obstacle avoidance with formation: how to change a motion plan or a formation pattern for MASs in order to avoid obstacles during movement? • Self-adaptation: how to automatically change the formation in order to best adapt to the dynamical unknown environment? The graph theory is a powerful tool to study formation control involving all the above five aspects. Firstly, graphs are naturally used to generate and maintain a formation, and consequently to achieve transformation between different formation patterns. In early 2001, directed graphs were used to describe the topologies and patterns of formation, and then study the formation evolution problem [12]. Saber et al adopted the combination graph theory to obtain a unique Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 11 determined formation pattern through designing the weights of edges [71]. Desai et al in [13] further adopted directed acyclic graphs to represent the control graph between agents, and designed the control strategy, afterwards, through enumerating all possible control graphs to realize transformation between any two formation patterns. It is worthy pointing out that by adding a pre-specified geometric pattern, the formation generation and maintaining can also be achieved in the corresponding consensus problem, which is usually known as consensus-based formation control, e.g., in [72,73]. Consequentially, the algebra graph theory is widely employed in formation control too. As aforementioned, the neighbors of vertex i in the graph represented MAS is exactly the collection of agents that have a topological relationship such as perception with the ith agent. Therefore, by using the properties of the Laplacian matrix, local, distributed and scalable formation controllers can be designed, and their stabilities can also be verified by virtue of the eigenvalue of the Laplacian matrix, for a typical example, the work in the classic literature [26]. And then, Fax and Murray set up the relation between the formation controller and the topological structure of the communicating network with the Laplacian matrix, and proved that if the local controller was stable, then the stability of formation with linear dynamics depended on the stability of the information flow [25]; Lin et. al proved that if and only if there is a global accessible vertex in the perception graph, the formation is stable by using tools of the algebraic graph theory [74]. In addition, the formation stability concepts are defined with the aid of the graph theory. Tanner et. al defined the leader-to-formation stability notion based on the graph and input-to-state stability concept, which includes both transient and steady-state errors [75]. In contrast, the pairwise asymptotic stability is defined based on directed graphs by involving a single pair of neighboring agents in [76], which implies that any two agents can asymptotically achieve and maintain a desired relative attitude and position though the pair of agents which may not be neighbors and do not interact with each other directly. Another important tool of the graph theory employed in multi-agent formation control is the rigid graph theory. An undirected graph is rigid if for almost every structure, the only possible continuous moves are those which preserve every inter-agent distance; further, an undirected graph Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 12 X. WANG, ET AL. is called minimally rigid if it is rigid and if there exists no rigid graphs with the same number of vertices and a smaller number of edges (see Fig. 4 for more illustrations of rigid and minimally rigid graphs). The main tools of rigid graphs are Henneberg sequence and Laman’s theorems [77]. Figure 4. (a) rigid graph (b) minimally rigid graph (c) non-rigid graph The former was raised by Henneberg to construct two-dimensional minimally rigid graphs, and the latter were raised by Laman in 1970 to verify if a two-dimensional graph was rigid [78]. With the concepts of rigid graphs, some primitive operators are defined to deal with the operations such as splitting and restructuring of rigid formation [71, 79]; also, a formation control law is proposed by only using the distances between agents in order to ensure rigidity or generic rigidity of formation [25]. Note that the concept of rigidity is mainly for the undirected graph, which means the interrelations between vertices are bi-directional. However, in practical applications, the MAS is often modeled by a directed graph for reducing the cost of communication and perception. Hence on the basis of rigidity, a group led by Prof. Brian D.O. Anderson at the Australian National University developed a “directed rigidity” concept, named persistence. A directed graph is persistent if and only if its underlying graph is rigid, and itself is constraint consistent [27], as illustrated in Fig. 5 †. Similar to a minimally rigid one, a graph is minimally persistent if it is persistent and if no single edge can be removed without losing persistence. The main difference between rigidity and persistence is that rigidity assumes all the constraints are satisfied, as if they were enforced by an external agency or through some mechanical properties, while persistence considers each † In R2, the graph represented in (a) is not persistent. For almost all uncoordinated displacements of 2, 3 and 4 (even if they satisfy their constraints), 4 is indeed unable to satisfy its three constraints. This problem cannot happen for the graph represented in (b), which is persistent however. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 13 constraint to be the responsibility of a single agent. Based on the concept of persistence, many (cid:20) (cid:21) (cid:20) (cid:21) (cid:22) (cid:23) (cid:22) (cid:23) Figure 5. (a) nonpersistent graph in R2 (b) persistent graph in R2 works have been investigated by the group led by Prof. Brian D.O. Anderson, such as developing the persistent/minimum persistent notions from the rigid/minimum rigid concepts [27]; fixing the transformation, splitting and reconstruction of formation in a two-dimensional space through defining three primitive graph operators by virtue of Henneberg sequence [80, 81]; designing a formation control law to keep the persistence for formation in two-dimensional space [82]; and further proposing distributed formation control laws in order to keep the minimum persistence for two kinds (leader-remote-follower and coleader) of formation with small perturbations [83]. It is worthy pointing out that the persistence notions have been applied to practical multiple robots systems. For example, B. Smith et al designed an automatic generation algorithm of persistent formations for multiple robots under distance and communication constraints [84], and a method for accomplishing such persistent formations was presented and further demonstrated with a prototype network of robots in a NASA project for researches in Antarctica [85]. 2.3. Some closely related issues 2.3.1. Rendezvous/Alignment The rendezvous problem refers to designing a distributed local control strategy to make multiple agents reach the same specified position at the same time. To some extent, rendezvous can be taken as the application of consensus in actual systems, such as robots and spacecrafts [86]. This problem was firstly proposed in literature [87], in which a distributed simple memoryless point rendezvous algorithm was proposed with proven convergence. And then, literature [88] and [89] extended this algorithm to a “stop-and-go” strategy under synchronous Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 14 X. WANG, ET AL. and asynchronous situations, respectively, and analyzed the convergence of the strategies. Similar to other multi-agent coordinations, the graph theory also played key roles in the research of the rendezvous problem. Cortes et. al presented a robust rendezvous algorithm in an arbitrarily dimensional space with the aid of the proximity graph [90]; and a connectivity-preserving protocol consisting of a set of distributed control rules was proposed in the case of the link in communication graph failure in literature [91]. In addition, in view of different kinds of practical systems, the rendezvous problem has received extensive researches. For multiple omni-directional mobile robots equipped with line-of-sight limited-range sensors moving in a connected, nonconvex, unknown environment, literature [92] presented a perimeter minimizing rendezvous algorithm by using the notions of connectivity-preserving constraint sets and proximity graphs. For multiple nonholonomic unicycles, a discontinuous and time-invariant rendezvous algorithm was designed with tools from the nonsmooth Lyapunov theory and the graph theory [93]. The discontinuous rendezvous policies were further investigated in literature [94], in which the new proposed sufficient conditions for characterizing the control policies were less restrictive than those presented in the above-mentioned literature. For a multiple agent system moving like a Dubins car, a simple quantized control law was designed with three values to make agents achieve rendezvous given a connected initial assignment with minimalism in sensing and control [95]. The attitude alignment problem, sometimes called attitude consensus or attitude synchronization problem, also received extensive attention in multi-agent fields, especially multiple spacecrafts [96], multiple satellites [97] and multiple marine robots [98]. Similar to the rendezvous problem, the attitude alignment is required to design a distributed control strategy to make the attitude of multiple agents tend to be consistent simultaneously. Lawton et al [99] adopted the behavior-based method to design two kinds of control strategies that made a group of aircrafts achieve attitude alignment; and further, this work was extended to more general scenes which did not need bi- directional communication [96]. For the attitude alignment with limited communication and no reference signal, Sarlette et al proposed attitude alignment strategies based on artificial potential methods [100]. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 15 2.3.2. Swarming/Flocking The swarming is often inspired by biological and physical systems, which refers to the prevalent collective behavior and the self-organization phenomenon, such as ant colonies, bee colonies, flocks of birds and schools of fishes [2–4]. And recently it has emerged in MAS with focuses on physical embodiment and realistic interactions among the individuals themselves and also between the individuals and the environment [101]. Generally speaking, swarming is characterized by the following: 1) flexibility: adaptability to the environment; 2) robustness: anti-jamming to the internal and external disturbances; 3) dispersion: dynamical behavior based on individuals; 4) self organization: obvious overall system properties in evolution, namely emerging [102]. As a special case of swarming, the flocking refers to the phenomena that the MAS (usually with the second-order integrator dynamics) presents certain coordinated behaviors by decentralized control with the aid of local perception and reaction between individuals. The classical flocking model is proposed by Reynolds in [103], in which the animation of flocking behaviors, such as bird flock and fish school, are realized through three rules, i.e. collision avoidance, velocity matching and flock centering. These three rules can make the distance between agents converge to an expected value, the speed of agents tend to be consistent, and agents cannot collide with each other. After then, for the linear systems with the second-order integrator dynamics, flocking algorithms are usually designed by local artificial potential fields integrated with the graph theory [104–106]. In [104], a theoretical framework based on the algebraic graph theory and the spatially induced graphs is provided for the design and analysis of scalable flocking algorithms under a connected topology. Further, this work was extended in [105] from two directions: one is the case where only a fraction of agents are informed and the other is that where the velocity of the virtual leader is varying. In [106], the stability properties of a group of agents governed by decentralized, nearest-neighbor interaction rules are analyzed. The flocking problem without considering collision between multi- agents has also got extensive attentions in recent years, e.g. in [107,108]. While little work has been done on flocking with nonlinear dynamics, except that Dong proposed a backsteping based control Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 16 X. WANG, ET AL. law design method for the flocking of multiple nonholonomic wheeled mobile robots with directed topology in [109]. 2.3.3. Containment control The containment control can be taken as an expansion of consensus or flocking with pinning control, which refer to designing a distributed control law to drive the states of the followers to the convex hull spanned by the states of multiple leaders. In particular, when the multiple leaders are moving, some literature call it as set tracking problem [110]. In early 2005, the simplest containment control, which took two agents as leaders and drove the rest converge to the line decided by the two leaders, was discussed in [26]. For the containment control with the first-order dynamics under fixed undirected graphs, a simple “stop-and-go” strategy is proposed and analyzed in literature [111]. Further, the containment control with single- or double- integrator dynamics was studied in [111–115], of which, literature [114] proposed two containment control algorithms via only position measurements where the leaders were neighbors of only a subset of the followers, and literature [115] examined the containment control for a second-order multi-agent system with random switching topologies. Moreover, the containment control strategy for multiple Lagrangian system in directed topological structure was proposed in [116]; and the containment control of multiple rigid-body systems with uncertainty was studied in [117]. For the more general nonlinear dynamics, G. Shi et al [110] investigated the distributed set tracking problem with unmeasurable velocities under switching directed topologies, and provided the necessary and sufficient conditions for set input-to-state stability and set integral input-to-state stability. Remark 1 The studies on coordination control are still ongoing, both in depth and width. On the one hand, the studied agents are found with more and more complicated dynamics and limited perceptions, for example, from high-order dynamics [62] to complex environmental constraints [118], again to communication bandwidth limitation [119–121], and quantitative communication [122]. On the other hand, more and more attention has been paid to various practical multi-robot systems, Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 17 for example, multiple micro-satellites [123], multiple spacecrafts [124, 125], multiple marine robots [126], multiple wheeled mobile robots [127, 128], and multiple Lagrangian systems [129]. 3. DIRECTIONS The theoretical work on multi-agent coordination control has made significant achievements in the past decade, which have however rare practical applications. In fact, the behaviors of actual robots are much more complicated than those considered in the existing work. Therefore, how to promote the theoretical work for serving in practice in order to shrink the gap between theory and practice is still an open problem which should be considered seriously. In the authors’ opinion, towards this direction, there are several topics deserving further investigation in the study of multi-agent coordination control, which listed as follows: 1) Coordination control with strongly nonlinear dynamics The existing coordination control often considers the agents governed by linear dynamics, especially by the single- or double-integrator dynamics; however, most physical systems are inherently nonlinear in nature. At present, coordination control with nonlinear dynamics has received seldom consideration, and few related work assume the nonlinear dynamics to be continuously differentiable [63, 64] or globally Lipschitz [65, 66]. In contrast, the dynamics of robots, vehicles or UAVs in practice are generally found with strong nonlinearities, which are hard to be described simply by using single/double-integrator or continuously differentiable/global Lipschitz continuous function. So it is necessary and beneficial to study multi-agent coordination control in the presence of the nonlinearities by means of the nonlinear control theory. Literature [130,131] carried on some preliminary research on this way via the input-to-state stability concept, but more problems, such as the finite time convergence, converging rate analysis, with topological switching and time-delay, are still worthy of further exploration. 2) Coordination control in three-dimensional space The existing work about multi-agent coordinations usually considers agents in a plane; while the coordinations in the more general three-dimensional space, composed of three rotational and Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 18 X. WANG, ET AL. three translational degrees of freedoms, are not extensively investigated. In essence, any general rigid motion (including rotation and translation) is described in three-dimensional space and its subspaces, and the coordination control in a plane is just a special case in three-dimensional space. Moreover, the applications of multi-agent coordination control are not confined to be in the plane. For instance, in the applications of multiple spacecraft formation flights [123], multiple marine robots cooperative exploration [126], flight array [132], and multiple flying vehicles cooperative handling objects [124] or playing tennis [125], the coordination controls are required to be considered in the three-dimensional space with attitude and position control simultaneously. Such problems are worthwhile to be solved, in the authors’ opinion, for coordination control in three- dimensional space include two aspects: a) the graph theory in three-dimensional space. As noted above, the graph theory is an important tool used in multi-agent coordination control, however, the existing results on the graph theory cannot be directly extended to the three-dimensional space. For example, Laman’s theorem and Henneberg sequence in three-dimensional space are still open problems [82]. Improving and expanding the existing foundations of the graph theory in order to make them valid in three-dimensional space is still a problem that has not been effectively solved. b) designing the coordination control law in a three-dimensional space. In recent years, different mathematical tools have been used to study the multi-agent coordination control in the three- dimensional space, such as the decoupling method (namely carrying on the attitude control and the position control independently) [123, 125], the Lie-group abstraction [133], the homogeneous transformation matrix [134, 135], and the dual quaternion approach [76], etc. Whatever, the above methods are still not systematic, and it is necessary to combine with the graph theory to further design coordination control laws under different topology structures, with different dynamic models and adaptation to external disturbances. 3) Coordination control with practical constraints In practice, there are many constraints for the coordinations of MASs, such as non-ideal communications and perceptions, and heterogeneous dynamics. Challenges from pure theories to Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 19 practical applications should be attached with more extensive concern, in order to shrink the gap between theories and applications. Communications are very important for multi-agent coordinations. Communications for actual systems are usually not ideal.For example, time-delays, unstable signals, and limited bandwidth are inevitable in the actual systems. Therefore, multi-agent coordination control with non-ideal communication is a challenging problem, and many work can be further considered, such as the design of fault tolerance topological structure which is robustness against the loss of communicating nodes or links, coordination algorithms with limited communicating bandwidth or quantitative communication, or even without communications, and underlying communication protocols with reliable robustness. In addition, it is usually the ideal perceptions that is considered by the existing MASs, which means agents can obtain all the required information in real-time. However in practice, sensors equipped by agents certainly have some perceptional limitations. For example, the commonly used cameras have conical perceptional fields, and the laser ranging sensor (typically the Hokuyo’s series products) can only perceive the information in a sector. Meanwhile, the perceived information is usually accompanied by noise and time-delay, and further more information, such as other agents’ relative velocities and accelerations, is generally difficult to obtain precisely. Therefore, the coordination control with perceptional constraints is also a problem that requires to be fixed, such as coordination control with perceptional directions, coordination control without velocity measurement, and coordination control with time-delay and noisy information. Thirdly, in the existing work on coordination control, the agents are usually considered to be isomorphic, which means all agents are found with the same dynamics. However, in many practical applications, for example, in the cooperative reconnaissance carried out by manned/unmanned aerial vehicles, and the cluster spacecraft system, the dynamics between agents are quite different. So, the coordination control for heterogeneous MASs is also a direction worth to pay attentions to in the future. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 20 X. WANG, ET AL. Finally, most of the existing theoretical results are only verified by simulations, rather than by actual systems, partly due to the high cost and various restrictions of the experiments. Therefore, to verify and apply the theoretical results to actual multi-robot systems is a most pressing issue too. 4) Combination with other collective behaviours For the last ten years, the work on coordination control in fact has led the research of the distributed networked system and provided some necessary supports to different types of collective behaviors, for e.g. the wireless sensor network. Therefore, how to expand and fuse the existing results of coordination control to other collective behaviours such as wireless sensor network is also a notable direction. 4. CONCLUSION This paper reviews the recent developments on multi-agent coordination control with focus on the consensus, formation control, and some closely related topics including rendezvous/alignment, swarming/flocking and containment control, with the graph theory playing a central role, in order to present a cohesive overview of the multi-agent distributed coordination control, and further provides some directions possibly deserving to be investigated in coordination control. The work on coordination control involves extremely extensive contents, and the authors believe that in the coming decade, more problems related to coordination control will be solved along with the developments of other related disciplines and technologies, and will also get more practical applications in more fields. REFERENCES 1. Hubbard S, Babak B, Sigurdsson ST, Magnusson KG. A model of the formation of fish schools and migrations of fish. Ecological Modelling 2004; 174(4):359–374. 2. Toner J, Tu Y. Hydrodynamics and phases of flocks. Annals of Physics 2005; 318(1):170–244. 3. Janson S, Middendorf M, Beekman M. Honey bee swarms: How do scouts guide a swarm of uninformed bees? Animal Behaviour 2005; 70(2):349–358. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 21 4. Couzin ID, Krause J, Franks NR, Levin SA. Effective leadership and decision-making in animal groups on themove. Nature 2005; 433(3):513–516. 5. Levine H, Rappel WJ. Self organization in systems of self-propelled particles. Physical Review E 2001; 63(1):208– 211. 6. Ren W, Beard RW, Atkins EM. Information consensus in multivehicle cooperative control. IEEE Control System Magazine 2007; 27(2):71–82. 7. Anderson BDO, Yu C, Fidan B, Hendrickx JM. Rigid graph control architecture for autonomous formation. IEEE Control System Magazine 2008; 28(6):48–63. 8. Lin Z, Francis B, Maggiore M. Distributed Control and Analysis of Coupled Cell Syatems. VDM Verlag: Germany, 2008. 9. Ren W, Beard R. Distributed Consensus in Multi-vehicle Cooperative Control. Spring: New York, 2008. 10. Bullo F, Cortes J, Martinez S. Distributed Control of Robotic Networks: A Mathematical Approach to Motion Coordination Algorithms. Princeton University Press: Princeton and Oxford, 2009. 11. Mesbahi M, Egerstedt M. Graph Theoretic Methods in Multiagent Networks. Princeton University Press: Princeton and Oxford, 2010. 12. Desai JP, Ostrowski JP, Kumar V. Modeling and control of formations of nonholonomic mobile robots. IEEE Transactions on Robotics and Automation 2001; 17(6):905–908. 13. Desai JP. A graph theoretic approach for modeling mobile robot team formations. Journal of Robotic Systems 2002; 19(11):511–525. 14. Consolini L, Morbidi F, Prattichizzo D, Tosques M. Leader-follower formation control of nonholonomic mobile robots with input constraints. Automatica 2008; 44(5):1343–1349. 15. Dimarogonasa DV, Tsiotrasb P, Kyriakopoulosc KJ. Leader-follower cooperative attitude control of multiple rigid bodies. Systems & Control Letters 2009; 58(6):429–435. 16. Chen J, Sun D, Yang J, Chen H. Leader-follower formation control of multiple non-holonomic mobile robots incorporating a receding-horizon scheme. International Journal of Robotics Research 2010; 29(6):727–747. 17. Balch T, Arkin RC. Behavior-based formation control for multirobot teams. IEEE Translations on Robotics and Automation 1998; 14(6):926–939. 18. Jadbabaie A, Lin J, Morse AS. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Transactions on Automatic Control 2003; 48(6):988–1001. 19. Antonelli G, Arrichiello F, Chiaverini S. Experiments of formation control with multirobot systems using the null-space-based behavioral control. IEEE Transactions on Control Systems Technology 2009; 17(5):1173–1182. 20. Arrichiello F, Chiaverini S, Indiveri G, Pedone P. The null-space based behavioral control for mobile robots with velocity actuator saturations. International Journal of Robotics Research 2010; 29(10):1317–1337. 21. Lewis MA, Tan KH. High precision formation control of mobile robots using virtual structures. Autonomous Robots 1997; 4(4):387–403. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 22 X. WANG, ET AL. 22. Ren W, Beard RW. Decentralized scheme for spacecraft formation flying via the virtual structure approach. Journal of Guidance, Control, and Dynamics 2004; 27(1):73–82. 23. Li NHM, Liu HHT. Formation UAV flight control using virtual structure and motion synchronization. Proceedings of the 2008 American Control Conference, Washington, USA, 2008; 1782–1787. 24. van den Broek THA, van de Wouw N, Nijmeijer H. Formation control of unicycle mobile robots: a virtual structure approach. Proceedings of the Joint 48th IEEE Conference on Decision and Control and 28th Chinese Control Conference, Shanghai, China, 2009; 8328–8333. 25. Fax JA, Murray RM. Information flow and cooperative control of vehicle formations. IEEE Transactions on Automatic Control 2004; 49(9):1465–1476. 26. Lin Z, Francis B, Maggiore M. Necessary and sufficient graphical conditions for formation control of unicycles. IEEE Transactions on Automatic Control 2005; 50(1):121–127. 27. Yu C, Hendrickx JM, Fidan B, Anderson BDO, Blondel VD. Three and higher dimensional autonomous formations: rigidity, persistence and structure persistence. Automatica 2007; 43(3):387–402. 28. Olfati-saber R, Fax JA, Murray RM. Consensus and cooperation in networked multi-agent systems. Proceedings of the IEEE 2007; 95(1):215–233. 29. Ren W, Beard W, Atkins EM. Information consensus in multivehicle cooperative control. IEEE Control Systems Magazine 2007; 27(2):71–82. 30. Vicsek T, Czirook A, Ben-Jacob E, Cohen I, Shochet O. Novel type of phase transition in a system of self-deriven particles. Physical Review Letter 1995; 75(6):1226–1229. 31. Godsil C, Royle G. Algebraic Graph Theory. Springer-Verlag: New York, 2004. 32. Olfati-Saber R, Murray RM. Consensus problems in networks of agents with switching topology and time-delays. IEEE Transactions on Automatic Control 2004; 49(9):1520–1553. 33. Lin Z, Broucke M, Francis B. Local control strategies for groups of mobile autonomous agents. IEEE Transactions on Automatic Control 2004; 49(4):622–629. 34. Ren W, Beard RW. Consensus seeking in multiagent systems under dynamically changing interaction topologies. IEEE Transactions on Automatic Control 2005; 50(5):655–661. 35. Zelazo D, Mesbahi M. Edge agreement: Graph-theoretic performance bounds and passivity analysis. IEEE Transactions on Automatic Control 2011; 56(3):544–555. 36. Lin Z, Francis B, Maggiore M. State agreement for continuous-time coupled nonlinear system. SIAM Journal on Control and Optimization 2007; 46(1):288–307. 37. Cao M, Morse AS, Anderson BDO. Reaching a consensus in a dynamically changing environment: A graphical approach. SIAM Journal on Control and Optimization 2008; 47(2):575–600. 38. Rahmani A, Ji M, Mesbahi M, Egerstedt M. Controllability of multi-agent systems from a graph-theoretic perspective. SIAM Journal on Control and Optimization 2009; 48(1):162–186. 39. Su Y, Huang J. Stability of a class of linear switching systems with applications to two consensus problems. IEEE Transactions on Automatic Control 2012; 57(6):1420–1430. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 23 40. Moreau L. Stability of multiagent systems with time-dependent communication links. IEEE Transactions on Automatic Control 2005; 50(2):169–182. 41. Xiao L, Boyd S. Fast linear iterations for distributed averaging. Systems & Control Letters 2004; 53(1):65–78. 42. Moreau L. Stability of continuous-time distributed consensus algorithms. Proceedings of 43th IEEE Conference on Decision and Control, Paradise Island, Bahamas, USA, 2004; 3998C4003. 43. Blimana PA, Ferrari-Trecateb G. Average consensus problems in networks of agents with delayed communications. Automatica 2008; 44(3):1985–1995. 44. Sun Y, Zhao D, Ruan J. Consensus in noisy environments with switching topology and time-varying delays. Physica A: Statistical Mechanics and its Applications 2010; 389(19):4149–4161. 45. Munz U, Papachristodoulou A, Allgower F. Consensus in multi-agent systems with coupling delays and switching topology. IEEE Transactions on Automatic Control 2011; 56(12):2976–2982. 46. Tanner HG, Christodoulakis DK. State synchronization in local-interaction networks is robust with respect to time delays. Proceedings of the 44th IEEE Conference on Decision and Control, and the European Control Conference, Seville, Spain, 2005; 4945–4950. 47. Xiao F, Wang L. State consensus for multi-agent systems with switching topologies and time-varying delays. International Journal of Control 2006; 79(10):1277–1284. 48. Xie G, Wang L. Consensus control for a class of networks of dynamic agents. International Journal of Robust and Nonlinear Control 2007; 17(10-11):941–959. 49. Su H, Wang X, Lin Z. Synchronization of coupled harmonic oscillators in a dynamic proximity network. Automatica 2009; 45(10):2286–2291. 50. Qin J, Zheng WX, Gao H. Consensus of multiple second-order vehicles with a time-varying reference signal under directed topology. Automatica 2011; 47(9):1983–1991. 51. Cao Y, Ren W. Distributed coordinated tracking with reduced interaction via a variable structure approach. IEEE Transactions on Automatic Control 2012; 57(1):33–48. 52. Ren W, Moore K, Chen YQ, Ieee. High-order consensus algorithms in cooperative vehicle systems. Proceedings of the 2006 IEEE International Conference on Networking, Sensing and Control, 2006; 457–462. 53. Jiang FC, Xie GM, Long W, Chen XJ. The chi-consensus problem of high-order multi-agent systems with fixed and switching topologies. Asian Journal of Control 2008; 10(2):246–253. 54. Jiang FC, Wang L. Consensus seeking of high-order dynamic multi-agent systems with fixed and switching topologies. International Journal of Control 2010; 83(2):404–420. 55. Liu Y, Jia YM. Consensus problem of high-order multi-agent systems with external disturbances: An H-infinity analysis approach. International Journal of Robust and Nonlinear Control 2010; 20(14):1579–1593. 56. Mo L, Jia Y. H-infinity consensus control of a class of high-order multi-agent systems. IET Control Theory and Applications 2011; 5(1):247–253. 57. Lin P, Li Z, Jia Y, Sun M. High-order multi-agent consensus with dynamically changing topologies and time- delays. IET Control Theory and Applications 2011; 5(8):976–981. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 24 X. WANG, ET AL. 58. Seo JH, Shim H, Back J. Consensus of high-order linear systems using dynamic output feedback compensator: Low gain approach. Automatica 2009; 45(11):2659–2664. 59. Xi J, Cai N, Zhong Y. Consensus problems for high-order linear time-invariant swarm systems. Physica A- Statistical Mechanics and Its Applications 2010; 389(24):5619–5627. 60. Liu Y, Jia YM. H-infinity consensus control for multi-agent systems with linear coupling dynamics and communication delays. International Journal of Systems Science 2012; 43(1):50–62. 61. Zhang Y, Tian YP. Allowable delay bound for consensus of linear multi-agent systems with communication delay. International Journal of Systems Science 2013; . 62. Tian YP, Zhang Y. High-order consensus of heterogeneous multi-agent systems with unknown communication delays. Automatica 2012; 48(6):1205–1212. 63. Zhao Y, Li B, Qin J, Gao H, Karimi HR. H∞ consensus and synchronization of nonlinear systems based on a novel fuzzy model. IEEE Transactions on Cybernetics 2013; in Press. 64. Li H, Liao X, Huang T. Second-order locally dynamical consensus of multiagent systems with arbitrarily fast switching directed tpologies. IEEE Transactions on Cybernetics 2013; in Press. 65. Yu W, Chen G, Cao M, Kurths J. Second-order consensus for multiagent systems with directed topologies and nonlinear dynamics. IEEE Trans. Syst. Man Cybern. Part B 2010; 40(3):881–891. 66. Song Q, Cao J, Yu W. Second-order leader-following consensus of nonlinear multi-agent systems via pinning control. Systems & Control Letters 2010; 59(9):553–562. 67. Song Q, Liu F, Cao J, Yu W. M-matrix strategies for pinning-controlled leader-following consensus in multiagent systems with nonlinear dynamics. IEEE Transactions on Cybernetics 2013; in Press. 68. Das A, Lewis FL. Cooperative adaptive control for synchronization of sencond-order systems with unknown nonlinearities. International Journal of Robust and Nonlinear Control 2011; 21(13):1509–1524. 69. Lemay M, Michaud F, Letourneau D, marc Valin J. Autonomous initialization of robot formations. Proceedings of the IEEE International Conference on Robotics and Automation, New Orleans, LA, 2005; 3018–3023. 70. Ren D, Lu G. Thinking in formation control. Control and Decision 2005; 20(6):601–606. 71. Olfati-Saber R, Murray RM. Graph rigidity and distributed formation stabilization of multi-vehicle systems. Proceedings of the IEEE International Conference on Decision and Control, Las Vegas, Nevada, USA, 2002; 2965–2971. 72. Ren W. Consensus based formation control strategies for multi-vehicle systems. Proceedings of the 2006 American Control Conference, Minneapolis, Minnesota, USA, 2006; 4237–4242. 73. Wu Z, Guan Z, Wu X, Li T. Consensus based formation control and trajectory tracing of multi-agent robot systems. Journal of Intelligent Robot System 2007; 48(3):397–410. 74. Lin Z, Broucke M, Francis B. Local control strategies for groups of mobile autonomous agents. IEEE Transactions on Automatic Control 2004; 49(4):622–629. 75. Tanner HG, Pappas GJ, Kumar V. Leader-to-formation stability. IEEE Transactions on Robotics and Automation 2004; 20(3):443–455. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 25 76. Wang X, Yu C, Lin Z. A dual quaternion solution to attitude and position control for rigid-body coordination. IEEE Transactions on Robotics 2012; 28(5):1162–1170. 77. Graver J, Servatius B, Servatius H. Combinatorial Rigidity (Graduate Studies in Mathematics). American Mathematical Society: New York, USA, 1994. 78. Laman G. On graphs and rigidity of plane skeletal structures. Journal of Engineering Mathematical 1970; 4(4):331–340. 79. Eren T, Anderson B, AMorse, Whiteley W, PBelhumeur. Operations on formations of autonomous agents. Communications in Information and Systems 2004; 3(4):223–258. 80. Hendrickx JM, Fidan B, Yu C, Anderson BDO, Blondel VD. Formationa reorganization by primitive operations on directed graphs. IEEE Transactions on Automation 2008; 53(4):968–979. 81. Hendrickx JM, Anderson BDO, Delvenne JC, Blondel VD. Directed graphs for the analysis of rigidity and persistence in autonomous agent system. International Journal of Robust and Nonlinear Control 2007; 17(10 - 11):960–981. 82. Yu C, Anderson BDO, Dasgupta S, Fidan B. Control of minimally persistent formations in the plane. SIAM Journal of Control Optimal 2009; 48(1):206–233. 83. Summers TH, Yu C, Dasgupta S, Anderson BDO. Control of minimally persistent leader-remote-follower and coleader formations in the plane. IEEE Transactions on Automatic Control 2011; 56(12):2778–2792. 84. Smith BS, Egrstedt MB, Howard A. Automatic generation of persistent formations for multi-agent networks under range constrains. Mobile Networks and Applications 2009; 14(3):322–335. 85. Smith BS, Wang J, Egrstedt MB. Persistent formation control of multi-robot networks. Proceeding of the 47th IEEE Conference on Decision and Control, Cancun, Mexico, 2008; 471–476. 86. Kranakis E, Krizanc D, Rajsbaum S. Mobile agent rendezvous: A survey. Structural Information and Communication Complexity, vol. 4056/2006. Springer Berlin / Heidelberg: New York, 2006; 1–9. 87. Ando H, Oasa Y, Suzuki I, Yamashita M. Distributed memoryless point convergence algorithm for mobile robots with limited visibility. IEEE Transactions on Robotics and Automation 1999; 15(5):818–828. 88. Lin J, Morse AS, Anderson BDO. The multi-agent rendezvous problem. part 1: The synchronous case. SIAM Journal of Control Optimal 2007; 46(6):2096–2119. 89. Lin J, Morse AS, Anderson BDO. The multi-agent rendezvous problem. part 2: The asynchronous case. SIAM Journal of Control Optimal 2007; 46(6):2120–2147. 90. Cortes J, Martinez S, Bullo F. Robust rendezvous for mobile autonomous agents via proximity graphs in arbitrary dimensions. IEEE Transactions on Automatic Control 2006; 51(8):1289–1298. 91. Xiao F, Wang L, Chen T. Connectivity preservation for multi-agent rendezvous with link failure. Automatica 2012; 48(1):25–35. 92. Ganguli A, Cortes J, Bullo F. Multirobot rendezvous with visibility sensors in nonconvex environments. IEEE Transctions on Robotics 2009; 25(2):340–352. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 26 X. WANG, ET AL. 93. Dimarogonas DV, Kyriakopoulos KJ. On the rendezvous problem for multiple nonholonomic agents. IEEE Transactions on Automatic Control 2007; 52(5):916–922. 94. Conte G, Pennesi P. The rendezvous problem with discontinuous control policies. IEEE Transactions on Automatic Control 2010; 55(1):279–283. 95. Yu J, LaValle SM, Liberzon D. Rendezvous without coordinates. IEEE Transactions on Automatic Control 2012; 57(2):421–434. 96. Ren W. Formation keeping and attitude alignment for multiple spacecraft through local interactions. Journal of Guidance, Control and Dynamics 2007; 30(2):633–638. 97. Bondhus AK, Pettersen KY, Gravdahl JT. Leader/follower synchronization of satellite attitude without angular velocity measurements. Proceedings of the 44th IEEE Conference on Decision and Control, and the European Control Conference 2005, Seville, Spain, 2005; 7270–7277. 98. Smith TR, Hanssmann H, Leonard NE. Orientation control of multiple underwater vehicle with symmetry- breaking potentials. Proceedings of the 40th IEEE Conference on Decision and Control, Orlando, Florida, USA, 2001; 4598–4603. 99. Lawton JR, Beard RW. Synchronized multiple spacecraft rotations. Automatica 2002; 38(8):1359–1364. 100. Sarlette A, Sepulchre R, Leonard NE. Autonomous rigid body attitude synchronization. Automatica 2009; 45(2):572–577. 101. Barca JC, Sekercioglus YA. Swarm robotics reviewed. Robotica 2013; 31(3):349 – 359. 102. Cheng D, Chen H. From swarm to social behavior control. Science & Technology Review 2004; (8):4–7. 103. Reynolds CW. Flocks, herds, and schools: A distributed behavioral model. Computer Graphics 1987; 21(4):25–34. 104. Olfati-Saber R. Flocking for multi-agent dynamic systems: Algorithms and theory. IEEE Transations on Automatic Control 2006; 51(3):401–420. 105. Su H, Wang X, Lin Z. Flocking of multi-agents with a virtual leader. IEEE Transactions on Automatic Control 2009; 54(2):293–307. 106. Tanner HG, Jadbabaie A, Pappas GJ. Flocking in fixed and switching networks. IEEE Transations on Automatic Control 2007; 52(5):863–868. 107. Lee D, Spong MW. Stable flocking of multiple inertial agents on balanced graphs. IEEE Transations on Automatic Control 2007; 52(8):1469–1475. 108. Moshtagh N, Jadbabaie A. Distributed geodesic control laws for flocking of nonholonomic agents. IEEE Transations on Automatic Control 2007; 52(4):681–686. 109. Dong W. Flocking of multiple mobile robots based on backstepping. IEEE Transations on System, Man and Cybernetics: Part B 2011; 41(2):414–424. 110. Shi G, Hong Y, Johansson KH. Connectivity and set tracking of multi-agent systems guided by multiple moving leaders. IEEE Transactions on Automatic Control 2012; 57(3):663–676. 111. Ji M, Ferrari-Trecate G, Egerstedt M, Buffa A. Containment control in mobile networks. IEEE Transactions on Automatic Control 2008; 53(5):1972–1975. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc MULTI-AGENT DISTRIBUTED COORDINATION CONTROL: DEVELOPMENTS AND DIRECTIONS 27 112. Cao Y, Stuart D, Ren W, Meng Z. Distributed containment control for multiple autonomous vehicles with doubleintegrator dynamics: Algorithms and experiments. IEEE Transactions on Control Systems Technology 2011; 19(4):929–938. 113. Cao Y, Ren W. Containment control with multiple stationary or dynamic leaders under a directed interaction graph. Proceedings of the IEEE Conference on Decision Control, Shanghai, China, 2009; 3014–3019. 114. Li J, Ren W, Xu S. Distributed containment control with multiple dynamic leaders for double-integrator dynamics using only position measurements. IEEE Transactions on Automatic Control 2012; 57(6):1553–1559. 115. Lou Y, Hong Y. Target containment control of multi-agent systems with random switching interconnection topologies. Automatica 2012; 48(5):879–885. 116. Mei J, Ren W, Ma G. Distributed containment control for lagrangian networks with parametric uncertainties under a directed graph. Automatica 2012; 48(4):653–659. 117. Meng Z, Ren W, You Z. Distributed finite-time attitude containment control for multiple rigid bodies. Automatica 2010; 46(12):2092–2099. 118. Ganguli A, Cortses J, Bullo F. Multirobot rendezvous with visibility sensors in nonconvex environments. IEEE Transactions on Robotics 2009; 25(2):340–352. 119. Li T, Xie L. Distributed consensus over digital networks with limited bandwidth and time-varying topologies. Automatica 2011; 47(9):2006–2015. 120. Li T, Xie L, Zhang J. Distributed consensus with limited communication data rate. IEEE Transactions on Automatic Control 2011; 56(2):279–292. 121. Lavaei J, Murray RM. Quantized consensus by means of gossip algorithm. IEEE Transactions on Automatic Control 2012; 57(1):19–32. 122. Kashyap A, Basar T, Srikant R. Quantized consensus. Automatica 2007; 43(7):1192–1203. 123. Schlanbusch R, Nicklasson PJ. Synchronization of target tracking cascadeded leader-follower spacecraft formation. Advances in Spacecraft Technologies, Hall J (ed.). 2011; 563–584. 124. Michael N, Fink J, Kumar V. Cooperative manipulation and transportation with aerial robots. Auton Robot 2011; 30(1):73–86. 125. Muller M, Lupashin S, D’Andrea R. Quadrocopter ball juggling. Proceedings of the 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, San Francisco, CA, USA, 2011; 5113–5120. 126. Girdhar Y, Xu A, Dey BB, Meghjani M, Shkurti F, Rekleitis I, Dudek G. Mare: Marine autonomous robotic explorer. Proceedings of the 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, San Francisco, CA, USA, 2011; 5048–5053. 127. Dong W. Tracking control of multiple-wheeled mobile robots with limited information of a desired trajectory. IEEE Transactions on Robotics 2012; 28(1):262–268. 128. Consolini L, Morbidi F, Prattichizzo D, Tosques M. On a class of hierarchical formations of unicycles and their internal. IEEE Transactions on Automatic Control 2012; 57(4):845–859. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc 28 X. WANG, ET AL. 129. Mastellone S, Mejia JS, Stipanovi DM, Spong MW. Formation control and coordinated tracking via asymptotic decoupling for lagrangian multi-agent systems. Automatica 2011; 47(12):2355–2363. 130. Shi G, Hong Y. Set-tracking of multi-agent systems with variable topologies guided by moving multiple leaders. Proceedings of the 49th IEEE Conference on Decision and Control, Atlanta, GA, 2010; 2245–2250. 131. Wang X, Liu T, Qin J. Second-order consensus with unknown dynamics via cyclic-small-gain method. IET Control Theory and Application 2012; 6(18):2748–2756. 132. Oung R, D’Andrea R. The distribute flight array. Mechatronics 2011; 21(5):908–917. 133. Nathan M, Vijay K. Planning and control of ensembles of robots with nonholonomic constraints. The International Journal of Robotics Research 2009; 28(8):962–975. 134. Igarashi Y, Hatanaka T, Fujita M, Spong MW. Passivity-based attitude synchronization in SE(3). IEEE Transactions on Control Systems Technology 2009; 17(5):1119–1134. 135. Hatanaka T, Igarashi Y, Fujita M, Spong MW. Passivity-based pose synchronization in three dimensions. IEEE Transactions on Automatic Control 2012; 57(2):360–375. Copyright c(cid:13) 2013 John Wiley & Sons, Ltd. Int. J. Robust. Nonlinear Control (2013) Prepared using rncauth.cls DOI: 10.1002/rnc
1802.08020	4	1802	2018-11-20T12:55:09	On Rational Delegations in Liquid Democracy	[ "cs.MA", "cs.GT" ]	Liquid democracy is a proxy voting method where proxies are delegable. We propose and study a game-theoretic model of liquid democracy to address the following question: when is it rational for a voter to delegate her vote? We study the existence of pure-strategy Nash equilibria in this model, and how group accuracy is affected by them. We complement these theoretical results by means of agent-based simulations to study the effects of delegations on group's accuracy on variously structured social networks.	cs.MA	cs	On Rational Delegations in Liquid Democracy ∗ Daan Bloembergen1, Davide Grossi2, and Martin Lackner3 1Centrum Wiskunde & Informatica, Amsterdam, The Netherlands 2Bernoulli Institute, University of Groningen, The Netherlands 3TU Wien, Vienna, Austria [email protected], [email protected], [email protected] November 21, 2018 Abstract Liquid democracy is a proxy voting method where proxies are delegable. We propose and study a game-theoretic model of liquid democracy to address the following question: when is it rational for a voter to delegate her vote? We study the existence of pure-strategy Nash equilibria in this model, and how group accuracy is affected by them. We complement these theoretical results by means of agent-based simulations to study the effects of delegations on group's accuracy on variously structured social networks. 1 Introduction Liquid democracy (Blum & Zuber, 2016) is an influential proposal in recent debates on democratic reforms in both Europe and the US. Several grassroots campaigns, as well as local parties, experimented with this novel type of decision making procedure. Examples include the German Piratenpartei1 and the EU Horizon 2020 project WeGovNow (Boella et al., 2018), which have incorporated the LiquidFeedback2 platform in their decision making, as well as grass-roots organizations such as the Democracy Earth Foundation3. Liquid democracy is a form of proxy voting (Miller, 1969; Tullock, 1992; Alger, 2006; Green-Armytage, 2015; Cohensius et al., 2017) where, in contrast to classical proxy voting, proxies are delegable (or transitive, or transferable). Suppose we are voting on a binary issue, then each voter can either cast her vote directly, or she can delegate her vote to a proxy, who can again either vote directly or, in turn, delegate to yet another proxy, and ∗This paper (without Appendix) appears in the proceedings of AAAI'19. We are indebted to the anonymous reviewers of IJCAI/ECAI'18 and AAAI'19 for many helpful comments on earlier versions of this paper. We are also grateful to the participants of the LAMSADE seminar at Paris Dauphine University, and the THEMA seminar at University Cergy-Pontoise where this work was presented, for many helpful comments and suggestions. Daan Bloembergen has received funding in the framework of the joint programming initiative ERA-Net Smart Energy Systems' focus initiative Smart Grids Plus, with support from the European Union's Horizon 2020 research and innovation programme under grant agreement No 646039. Davide Grossi was partially supported by EPSRC under grant EP/M015815/1. Martin Lackner was supported by the European Research Council (ERC) under grant number 639945 (ACCORD) and by the Austrian Science Foundation FWF, grant P25518 and Y698. 1https://www.piratenpartei.de/ 2https://liquidfeedback.org/ 3https://www.democracy.earth/ 1 so forth. Ultimately, the voters that decided not to delegate cast their ballots, which now carry the weight given by the number of voters who entrusted them as proxy, directly or indirectly. Contribution The starting point of our analysis is an often cited feature of liquid democ- racy: transitive delegations reduce the level of duplicated effort required by direct voting, by freeing voters from the need to invest effort in order to vote accurately. The focus of the paper is the decision-making problem that voters, who are interested in casting an accurate vote, face between voting directly, and thereby incurring a cost in terms of effort invested to learn about the issue at hand, or delegating to another voter in their network, thereby avoiding costs. We define a game-theoretic model, called delegation game, to represent this type of interaction. We establish pure strategy Nash equilibrium existence results for classes of delegation games, and study the quality of equilibria in terms of the average accuracy they enable for the population of voters, both analytically and through simulations. Proofs of the two main results (Theorems 1 and 2) are presented in full, while we provide proofs of the simpler secondary results as an Appendix only. By means of simulations we also study the effects of different network structures on delegation games in terms of: performance against direct voting, average accuracy and the probability of a correct majority vote, the number and quality of voters acting as ultimate proxies (so-called gurus) and, finally, the presence of delegation cycles. To the best of our knowledge, this is the first paper providing a comprehensive study of liquid democracy from a game-theoretic angle. Related Work Although the idea of delegable proxy was already sketched by Dodg- son (1884), only a few very recent papers have studied aspects of liquid democracy in the (computational) social choice theory (Brandt et al., 2016) literature. Kling et al. (2015) provide an analysis of election data from the main platform implementing a liquid democracy voting system (Liquid Feedback) for the German Piratenpartei. They focus on network theoretic properties emerging from the structure of delegations -- with partic- ular attention to the number of highly influential gurus or 'super-voters'. Inspired by their experimental analysis, Golz et al. (2018) propose and analyze a variant of the liquid democracy scheme able to restrict reliance on super-voters. Skowron et al. (2017) study an aspect of the Liquid Feedback platform concerning the order in which proposals are ranked and by which they are brought to the attention of the community. Boldi et al. (2011) investigate applications of variants of the liquid democracy voting method (called viscous democracy) to recommender systems. Brill (2018) presents some research direc- tions in the context of liquid democracy. A general, more philosophical discussion of liquid democracy is provided by Blum & Zuber (2016). More directly related to our investigations is the work by Christoff & Grossi (2017) and, especially, by Kahng et al. (2018). The first paper studies liquid democracy as an aggregator -- a function mapping profiles of binary opinions to a collective opinion -- in the judgment aggregation and binary voting tradition (Grossi & Pigozzi, 2014; Endriss, 2016). The focus of that paper are the unintended effects that transferable proxies may have due to delegation cycles, and due to the failure of rationality constraints normally satisfied by direct voting. The second paper addresses one of the most cited selling arguments for liquid democ- racy: delegable proxies guarantee that better informed agents can exercise more weight on group decisions, thereby increasing their quality. Specifically, Kahng et al. (2018) study the level of accuracy that can be guaranteed by liquid democracy (based on vote delega- 2 tion with weighted majority) vs. direct voting by majority. Their key result consists in showing that no 'local' procedure to select proxies can guarantee that liquid democracy is, at the same time, never less accurate (in large enough graphs) and sometimes strictly more accurate than direct voting. In contrast to their work, we assume that agents incur costs (effort) when voting directly, and on that basis we develop a game-theoretic model. Also, we assume agents aim at tracking their own type rather than an external ground truth, although we do assume such a restriction in our simulations to better highlight how the two models are related and to obtain insights applicable to both. 2 Preliminaries 2.1 Types, Individual Accuracy and Proximity We are concerned with a finite set of agents (or voters, or players) N = {1, . . . , n} having to decide whether x = 1 or x = 0. For each agent one of these two outcomes is better, but the agent is not necessarily aware which one. We refer to this hidden optimal outcome as the type of agent i and denote it by τi ∈ {0, 1}. Agents want to communicate their type truthfully to the voting mechanism, but they know it only imperfectly. This is captured by the accuracy qi of an agent i: qi determines the likelihood that, if i votes directly, she votes according to her type τi. We assume that an agent's accuracy is always ≥ 0.5, i.e., at least as good as a coin toss. We distinguish two settings depending on whether the agents' types are deterministic or probabilistic. A deterministic type profile T = (cid:104)τ1, . . . , τn(cid:105) simply collects each agent's type. In probabilistic type profiles types are independent random variables drawn according to a distribution P. Given a probabilistic type profile, the likelihood that any two agents i, j ∈ N are of the same type is called the proximity pi,j where pi,j = P(τ (i) = τ (j)) = P(τ (i) = 1) · P(τ (j) = 1) + (1 − P(τ (i) = 1)) · (1 − P(τ (j) = 1)). In the probabilistic setting we assume agents know such value although, importantly, they do not know P. In a deterministic type profile, we have pi,j = 1 if τi = τj and pi,j = 0 otherwise. Following standard equilibrium theory, our theoretical results assume agents act as if they have access to the accuracy of each agent. More realistically, in our simulations we assume agents have access to such information only with respect to neighbors on an underlying interaction structure. 2.2 Interaction Structure and Delegations Agents are nodes in a network (directed graph) represented by a relation R ⊆ N 2. For i ∈ N the neighborhood of i in (cid:104)N, R(cid:105) is denoted R(i), i.e., the agents that are directly connected to i. Agents have the choice of either voting themselves, thereby relying solely on their own accuracy, or delegating to an agent in their neighborhood. A delegation profile is a vector d = (cid:104)d1, . . . , dn(cid:105) ∈ N n. Given a delegation profile d we denote by di the proxy selected by i in d. Clearly a delegation profile can be viewed as a functional graph on N or, equivalently, as a map in d : N → N where d(i) = di. When the iterated application of d from i reaches a fixed point we denote such fixed point as d∗ i and call it i's guru (in d). In the following, we write N∗ to denote the set of voters whose delegation does not lay on a path ending on a cycle, i.e., the set of voters i for which d∗ i exists. We write d(cid:48) = (d−i, j) as a short form for d(cid:48) = (cid:104)d1, . . . , di−1, j, di+1, . . . , dn(cid:105). As agents may only be able to observe and interact with their direct network neighbors, structural properties of the interaction network may play a role in the model dynamics. In 3 our simulations we will focus on undirected graphs (that is, R will be assumed to be sym- metric, as social ties are normally mutual) consisting of one single connected component (that is, N 2 is included in the reflexive transitive closure of R). Under these assumptions, we consider four typical network structures that are well represented in the literature on social network analysis (cf. Jackson 2008): 1) the random network, in which each pair of nodes has a given probability of being connected (Erdos & R´enyi, 1959); 2) the regular network, in which all nodes have the same degree; 3) the small world network, which features a small average path length and high clustering (Watts & Strogatz, 1998); and 4) the scale free network, which exhibits a power law degree distribution (Barab´asi & Albert, 1999).4 3 A Model of Rational Delegations 3.1 Individual Accuracy under Delegable Proxy Each agent i has to choose between two options: either to vote herself with accuracy qi or to delegate, thereby inheriting the accuracy of another voter (unless i is involved in a delegation cycle). These choices are recorded in the delegation profile d and can be used to compute the individual accuracy for each agent i ∈ N∗ as follows: if i ∈ N∗ if i /∈ N∗ i ) · (1 − pi,d∗ i ) i · pi,d∗ 0.5 i + (1 − qd∗ (cid:26)qd∗ ∗ i (d) = q (1) In Equation (1) i's accuracy equals the likelihood that i's guru has the same type and votes accurately plus the likelihood that i's guru has the opposite type and fails to vote accurately. Note that if i votes directly, i.e., di = i, then q∗ i (d) = qi. Observe that if i's delegation leads to a cycle (i /∈ N∗), i's accuracy is set to 0.5. The rationale for this assumption is the following. If an agent delegates into a cycle, even though she knows her own accuracy and she actively engages with the voting mechanism by expressing a delegation, she fails to pass information about her type to the mechanism. No information is therefore available to decide about her type. It may be worth observing that, by Equation (1), in a deterministic type profile we i ); have that pi,j ∈ {0, 1} and therefore i's accuracy reduces to: qd∗ 1 − qd∗(i) if i ∈ N∗ and τ (i) (cid:54)= τ (d∗ i ); and 0.5 if i /∈ N∗. i if i ∈ N∗ and τ (i) = τ (d∗ Before introducing our game theoretic analysis, we make the following observation. Agents have at their disposal an intuitive strategy to improve their accuracy: simply delegate to a more accurate neighbor. We say that a delegation profile d is positive if for j (d) > qj. Furthermore, we say that a delegation from i to a all j ∈ N either dj = j or q∗ neighbor j is locally positive if qj · pi,j + (1 − qj) · (1 − pi,j) > qi. Proposition 1. Let d be a positive delegation profile. Further, let s, t ∈ N , ds = s, and d(cid:48) = (d−s, t), i.e., agent s votes directly in d and delegates to t in d(cid:48). If the delegation from s to t is locally positive, then d(cid:48) is positive (proof in Appendix A). However, locally positive delegations do not necessarily correspond to optimal del- egations. This can be easily seen in an example where agent i is not a neighbor of a very competent agent j, but would have to delegate via an intermediate agent k (who delegates to j). If this intermediate agent k has a lower accuracy than i, then the dele- gation from i to k would not be locally positive, even though it is an optimal choice. So 4Although random and regular graphs are not generally applicable to real-world settings, they serve as a useful baseline to illustrate the effects of network structure on delegations. 4 utility-maximization may require information which is inherently non-local (accuracy of 'far' agents). 3.2 Delegation Games We assume that each agent i has to invest an effort ei to manifest her accuracy qi. If she delegates, she does not have to spend effort. Agents aim therefore at maximizing the trade-off between the accuracy they can achieve (either by voting directly or through proxy) and the effort they spend. Under this assumption, the binary decision set-up with delegable proxy we outlined above can be used to define a natural game -- called delegation game -- G = (cid:104)N, P, R, Σi, ui(cid:105), with i ∈ N , where N is the set of agents, P is the (possibly degenerate) distribution from which the types of the agents in N are drawn, R the underlying network as defined above, Σi ∈ N is the set of strategies of agent i (voting, or choosing a specific proxy), and (cid:40) q∗ i (d) qi − ei ui(d) = if di (cid:54)= i if di = i (2) is agent i's utility function. The utility i extracts from a delegation profile equals the accuracy she inherits through proxy or, if she votes, her accuracy minus the effort spent.5 In delegation games we assume that qi − ei ≥ 0.5 for all i ∈ N . This is because if qi − ei < 0.5, then i would prefer a random effortless choice over taking a decision with effort. A few comments about the setup of Equation (2) are in order. First of all, as stated earlier, we assume agents to be truthful. They do not aim at maximizing the chance their type wins the vote, but rather to relay their type to the mechanism as accurately as possible.6 Secondly, notice that the utility an agent extracts from a delegation profile may equal the accuracy of a random coin toss when the agent's delegation ends up into a delegation cycle (cf. Equation (1)). If this happens the agent fails to relay information about her type, even though she acted in order to do so. This justifies the fact that 0.5 is also the lowest payoff attainable in a delegation game. The following classes of delegation games will be used in the paper: games with de- terministic profiles, i.e., where P is degenerate and all players are assigned a crisp type from {0, 1}; homogeneous games, where all players have the same (deterministic) type;7 and effortless voting games, where for each i ∈ N we have ei = 0. As an example, a homogeneous game in matrix form is given in Table 1, where N = {1, 2}, R = N 2 and the distribution yields the deterministic type profile T = (cid:104)1, 1(cid:105). Interestingly, if we assume that qi − ei > 0.5 with i ∈ {1, 2}, and that8 q−i > qi − ei (i.e., the opponent's accuracy is higher than the player's individual accuracy minus her effort), then the game shares the ordinal preference structure of the class of anti-coordination games: players need to avoid coordination on the same strategy (one should vote and the other delegate), with two coordination outcomes (both players voting or both delegating) of which the second (the delegation cycle) is worst for both players. Notice that, were the 5No utility is accrued for gaining voting power in our model. 6Notice however that our modeling of agents' utility remains applicable in this form even if agents are not truthful but the underlying voting rule makes truthfulness a dominant strategy -- such as majority in the binary voting setting used here. normally assumed by jury theorems (Grofman et al., 1983). 7This is the type of interaction studied, albeit not game-theoretically, by Kahng et al. (2018) and 8We use here the usual notation −i to denote i's opponent. 5 vote delegate (to 2) vote q1−e1, q2−e2 q2, q2 − e2 delegate (to 1) q1 − e1, q1 0.5, 0.5 Table 1: A two-player delegation game. The row player is agent 1 and the column player is agent 2. underlying network not complete (i.e., R ⊂ N 2), the matrix would be shrunk by removing the rows and columns corresponding to the delegation options no longer available. The introduction of effort has significant consequences on the delegation behavior of voters, and we will study it in depth in the coming sections. It is worth noting immediately that the assumptions of Proposition 1 no longer apply, since agents may prefer to make delegations that are not locally positive due to the decreased utility of voting directly. 3.3 Existence of Equilibria in Delegation Games In this section we study the existence of pure strategy Nash Equilibria (NE) in two classes of delegation games. NE describe how ideally rational voters would resolve the effort/accuracy trade-off. Of course, such voters have common knowledge of the delega- tion game structure -- including, therefore, common knowledge of the accuracies of 'distant' agents in the underlying network. Our simulations will later lift some of such epistemic assumptions built into NE. Deterministic Types deterministic type profiles. In the following we provide a NE existence result for games with Theorem 1. Delegation games with deterministic type profiles always have a (pure strat- egy) NE. Proof. First of all, observe that since the profile is deterministic, for each pair of agents i and j, pi,j ∈ {0, 1}. The proof is by construction. First, we partition the set of agents N into N1 = {i ∈ N τ (i) = 1} and N0 = {i ∈ N τ (i) = 0}. We consider these two sets separately; without loss of generality let us consider N1. Further we consider the network R1 = {(i, j) ∈ N1 × N1 : (i, j) ∈ R}. Since (N1, R1) can be seen as a directed graph, we can partition it into Strongly Connected Components (SCCs). If we shrink each SCC into a single vertex, we obtain the condensation of this graph; note that such a graph is a directed acyclic graph (DAG). We construct a delegation profile d by traversing this DAG bottom up, i.e., starting with leaf SCCs. Let S ⊆ N1 be a set of agents corresponding to a leaf SCC in the condensation DAG. We choose an agent i in S that has maximum qi − ei. Everyone in S (including i) delegates to i. Now let S ⊆ N1 be a set of agents corresponding to an inner node SCC in the condensation DAG and assume that we have already defined the delegation for all SCCs that can be reached from S. As before, we identify an agent i ∈ S with maximum qi − ei. Further, let T ⊆ N1 \ S be the set of all voters j that can be reached from S in (N1, R1), and for which q∗ j > qi−ei. We distinguish two cases. (i) If T (cid:54)= ∅, then we choose an agent k ∈ T with q∗ j and all agents in S directly or indirectly delegate to k. (ii) If T = ∅, all agents in S delegate to i. This concludes our construction (as for N0 the analogous construction applies); let d be the corresponding delegation profile. It remains to verify that this is indeed a NE: Let i be some agent in an SCC S, and, without loss of generality, let i ∈ N1. Observe that since we have a deterministic k = maxj∈T q∗ 6 j (d) > q∗ i ∈ S but di (cid:54)= i, and (3) d∗ profile, if agent i changes her delegation to j, then i's utility changes to q∗ j (d) if i ∈ N1 and 1 − q∗ j (d) if i ∈ N0. First, note that for all agents k ∈ N , q∗ k(d) ≥ qk − ek ≥ 0.5. Hence, we can immediately exclude that for agent i delegating to an agent in j ∈ N0 is (strictly) beneficial, as it would yield an accuracy of at most 1 − q∗ j ≤ 0.5. Towards a contradiction assume there is a beneficial deviation to an agent j ∈ N1, i.e., there is an agent j ∈ R(i) ∩ N1 with q∗ i (d). Let us now consider the three cases: (1) di = i, (2) d∗ i /∈ S. In case (1), everyone in S delegates to i. Hence, if j ∈ S, a cycle would occur yielding a utility of 0.5, which is not sufficient for a beneficial deviation. If a delegation to j /∈ S is possible but was not chosen, then by construction q∗ j ≤ qi − ei and hence this deviation is not beneficial. We conclude that in case (1) such an agent j cannot exist. In case (2), everyone in S delegates to d∗ i . Hence, if j ∈ S, then d∗ j = d∗ i , a contradiction. If j /∈ S, the same reasoning as before applies and hence also here we obtain a contradiction. In case (3), by construction, d∗ i /∈ S had been chosen to maximise accuracy, hence j ∈ S. Since for all k ∈ S, d∗ i , only a deviation to i itself can be beneficial, i.e., j = i. However, since d∗ d∗(i) > qi − ei, no beneficial deviation is possible. i was chosen because q∗ k = d∗ It follows that also homogeneous games always have NE. Effortless Voting Effortless voting (ei = 0 for all i ∈ N ) is applicable whenever effort is spent in advance of the decision and further accuracy improvements are not possible. Theorem 2. Delegation games with effortless voting always have a (pure strategy) NE. Proof. We prove this statement by showing that the following procedure obtains a NE: We start with a strategy profile in which all players vote directly, i.e., player i's strategy is i. Then, we iteratively allow players to choose their individual best response strategy to the current strategy profile. Players act sequentially in arbitrary order. If there are no more players that can improve their utility by changing their strategy, we have found a NE. We prove convergence of this procedure by showing that a best response that increases the player's utility never decreases the utility of other players. We proceed by induction. Assume that all previous best responses have not reduced any players' utility (IH). Assume player i now chooses a best response that increases her utility. Let d be the delegation profile; further, let d∗ i = s. By assumption, i's utility started with qi−ei = qi and has not decreased since, i.e., ui(d) ≥ qi. Since i's best response strictly increases i's utility, it cannot be a delegation to herself. So let a delegation to j (cid:54)= i be i's best response and consider profile d(cid:48) = (d−i, j). Further, let d∗ j = t, i.e., i now delegates to j and by transitivity to t, i.e., d(cid:48)∗ j = t. Let k be some player other than i. We define the delegation path of k as the sequence (d(k), d(d(k)), d(d(d(k))), . . . ). If k's delegation path does not contain i, then k's utility remains unchanged, i.e., uk(d(cid:48)) ≥ uk(d). If k's delegation path contains i, then k now delegates by transitivity to t, i.e., we have d∗ k = s and d(cid:48)∗ k = t. By Equation (2), we have i = d(cid:48)∗ uk(d) = qs · pk,s + (1 − qs) · (1 − pk,s) ) = qt · pk,t + (1 − qt) · (1 − pk,t). uk(d (cid:48) and (3) (4) We have to show that k's utility does not decrease, i.e., uk(d(cid:48)) ≥ uk(d), under the as- sumption that i chooses a best response, i.e., ui(d(cid:48)) > ui(d), with: ui(d) = qs · pi,s + (1 − qs) · (1 − pi,s) ) = qt · pi,t + (1 − qt) · (1 − pi,t). ui(d (cid:48) and (5) (6) 7 In the following we will often use the fact that, for a, b ∈ [0, 1], if ab + (1− a)(1− b) ≥ 0.5, then either a, b ∈ [0, 0.5] or a, b ∈ [0.5, 1]. By IH, since accuracies are always at least 0.5, it holds that ui(d) ≥ qi ≥ 0.5 and by Equation (5) we have qs · pi,s + (1− qs)· (1− pi,s) ≥ 0.5 and hence pi,s ≥ 0.5. Analogously, Equation (3) implies that pk,s ≥ 0.5. Furthermore, we use the fact that pk,i = pk,sps,i + (1 − pk,s)(1 − ps,i) + (−2(2xk − 1) · (2xi − 1) · (xs − 1)xs) (7) where xj = P(τ (j) = 1) for j ∈ {k, i, s}. Observe that, by the definition of utility in Equation (2), the assumptions made on d and d(cid:48), and the fact that for a, b ∈ [0, 1], if ab + (1 − a)(1 − b) ≥ 0.5, then either a, b ∈ [0, 0.5] or a, b ∈ [0.5, 1]. So we have that either xj ≥ 0.5 for j ∈ {k, i, s}, or xj ≤ 0.5 for j ∈ {k, i, s}. We work on the first case. The other case is symmetric. Let also γk,s,i = −2(2xk − 1) · (2xi − 1) · (xs − 1)xs. From the above it follows that 0.5 ≥ γk,i,s ≥ 0. Furthermore, given that pi,s = ps,i ≥ 0.5, we can also conclude that pk,i ≥ 0.5. Now by substituting pk,s = pk,ipi,s + (1 − pk,i)(1 − pi,s) + (−2(2xk − 1) · (2xs − 1) · (xi − 1)xi (cid:124) (cid:123)(cid:122) γk,i,s (cid:125) ) (8) in Equation (3), we obtain (cid:122) (cid:122) (cid:125)(cid:124) ui(d) (cid:125)(cid:124) ui(d(cid:48)) (cid:123) (cid:123) uk(d) = (2pk,i − 1)( 2qspi,s − qs − pi,s + 1) + 1 − pk,i + γk,i,s(2qs − 1). Similarly, using the appropriate instantiation of Equation (7) for xj with j ∈ {k, i, t}, by substituting pk,i · pi,t + (1 − pk,i)(1 − pi,t) + γk,i,t for pk,t in Equation (4) we obtain (cid:48) uk(d ) = (2pk,i − 1) · ( 2qtpi,t − qt − pi,t + 1) + 1 − pk,i + γk,i,t(2qt − 1). (9) Now observe that, since pk,i ≥ 0.5 we have that (2pk,i − 1) ≥ 0. It remains to compare γk,i,s(2qs − 1) with γk,i,t(2qt − 1), showing the latter is greater than the former. Observe that both expressions have a positive sign. We use the fact that ab + (1 − a)(1 − b) < cd + (1 − c)(1 − d) implies cd > ab under the assumption that a, b, c, d ∈ [0.5, 1]. On the basis of this, and given that ui(d(cid:48)) > ui(d), we obtain that qs · pi,s < qt · pi,t and therefore that qs · xs < qt · xt, from which we can conclude that (cid:122) (cid:0) <(cid:0) (cid:124) (cid:125)(cid:124) γk,i,s (cid:123)(cid:122) γk,i,t −2(2xk − 1) · (2xs − 1) · (xi − 1)xi −2(2xk − 1) · (2xt − 1) · (xi − 1)xi · (2qs − 1) · (2qt − 1). (cid:123) (cid:1) (cid:1) (cid:125) It follows that the assumption ui(d(cid:48)) > ui(d) (player i chose a best response that increased her utility) together with Equations (8) and (9) implies that uk(d(cid:48)) > uk(d) (and a fortiori that uk(d(cid:48)) ≥ uk(d)). We have therefore shown that if some player chooses a best response, the utility of other players does not decrease. This completes the proof. Discussion The existence of NE in general delegation games remains an interesting open problem. It should be noted that the proof strategies of both Theorems 1 and 2 do not work in the general case. Without a clear dichotomy of type it is not possible to assign delegations for all agents in an SCC (as we do in the proof of Theorem 1). And the key 8 property upon which the proof of Theorem 2 hinges (that a best response of an agent does not decrease the utility of other agents) fails in the general case due to the presence of non- zero effort. Finally, it should also be observed that Theorem 2 (as well as Proposition 1) essentially depend on the assumption that types are independent random variables. If this is not the case (e.g., because voters' preferences are correlated), delegation chains can become undesirable. Example 1. Consider the following example with agents 1, 2 and 3. The probability distribution P is defined as P(τ (1) = 1 ∧ τ (2) = 1 ∧ τ (3) = 0) = 0.45, P(τ (1) = 0 ∧ τ (2) = 1 ∧ τ (3) = 1) = 0.45, and P(τ (1) = 1 ∧ τ (2) = 1 ∧ τ (3) = 1) = 0.1. Consequently, p1,2 = 0.55, p2,3 = 0.55, and p1,3 = 0.1. Let us assume that the agents' accuracies are q1 = 0.5001, q2 = 0.51, and q3 = 0.61. A delegation from agent 1 to 2 is locally positive as q2 · p1,2 + (1− q2)· (1− p1,2) = 0.501 > q1. Furthermore, a delegation from 2 to 3 is locally positive as q3 · p2,3 + (1 − q3) · (1 − p2,3) = 0.511 > q2. However, the resulting delegation from 1 to 3 is not positive since q3 · p1,3 + (1 − q3) · (1 − p1,3) = 0.412. 3.4 Quality of Equilibria in Delegation Games In delegation games players are motivated to maximize the tradeoff between the accuracy they acquire and the effort they spend for it. A natural measure for the quality of a delegation profile is, therefore, how accurate or informed a random voter becomes as a result of the delegations in the profile, that is, the average accuracy (i.e., ¯q∗(d) = i (d)) players enjoy in that profile. One can also consider the utilitarian social i∈N ui(d) of a delegation profile d. This relates to average accuracy (cid:80) welfare SW(d) =(cid:80) i∈N q∗ 1 n as follows: ∗ ¯q (d) = id(i)=i ei . SW(d) +(cid:80) n It can immediately be noticed that equilibria do not necessarily maximize average accuracy. On the contrary, in the following example NE yields an average accuracy of close to 0.5, whereas an average accuracy of almost 1 is achievable. Example 2. Consider an n-player delegation game where all players have the same type and (i, j) ∈ R for all j and all i > 1, i.e., player 1 is a sink in R and cannot delegate to anyone, but all others can delegate to everyone. Further, we have e1 = 0 and ei = 0.5 − for i ≥ 2. The respective accuracies are q1 = 0.5 + 2 and qi = 1. If player i ≥ 2 does not delegate, her utility is 0.5 + . Hence, it is always more desirable to delegate to player 1 (which yields a utility of 0.5 + 2 for i). Consider now the profiles in which all players delegate to player 1 (either directly or transitively). Player 1 can only vote directly (with utility 0.5 + 2). All such profiles are NE with average accuracy 0.5 + 2. If, however, some player j ≥ 2 chose to vote herself, all players (except 1) would delegate to j thereby obtaining an average accuracy of 1 − 0.5−2 , which converges to 1 for n → ∞. This is not a NE, as j could increase her utility by delegating to 1. n The findings of the example can be made more explicit by considering a variant of the price of anarchy for delegation games, based on the above notion of average accuracy. That is, for a given delegation game G, the price of anarchy (PoA) of G is given by PoA(G) = maxd∈N n ¯q∗(d) mind∈NE(G) ¯q∗(d) , where NE(G) denotes the set of pure-strategy NE of G. 9 Fact 1. PoA is bounded below by 1 and above by 2 (see Apendix A). An informative performance metrics for liquid democracy is the difference between the group accuracy after delegations versus the group accuracy achieved by direct voting. This measure, called gain, was introduced and studied by Kahng et al. (2018). Here we adapt it to our setting as follows: G(G) =(cid:0)mind∈NE(G) ¯q∗(d)(cid:1) − ¯q where ¯q = ¯q∗((cid:104)1, . . . , n(cid:105)). That is, the gain in the delegation game G is the average accuracy of the worst NE minus the average accuracy of the profile in which no voter delegates. It turns out that the full performance range is possible: Fact 2. G is bounded below by −0.5 and above by 0.5 (see Apendix A). The above bounds for PoA and gain provide only a very partial picture of the perfor- mance of liquid democracy when modeled as a delegation game. The next section provides a more fine-grained perspective on the effects of delegations. 4 Simulations We simulate the delegation game described above in a variety of settings. We restrict ourselves to homogeneous games. This allows us to relate our results to those of Kahng et al. (2018). Our experiments serve to both verify and extend the theoretical results of the previous section. In particular we simulate the best response dynamics employed in the proof of Theorem 2 and show that these dynamics converge even in the setting with effort, which we could not establish analytically. In addition, we investigate the dynamics of a one-shot game scenario, in which all agents need to select their proxy simultaneously at once. Setup We generate graphs of size N = 250 of each of the four topologies random, regular, small world, and scale free, for different average degrees, while ensuring that the graph is connected. Agents' individual accuracy and effort are initialized randomly with qi ∈ [0.5, 1] and qi − ei ≥ 0.5. We average results over 2500 simulations for each setting (25 randomly generated graphs × 100 initializations). Agents correctly observe their own accuracy and effort, and the accuracy of their neighbors. The game is homogeneous, so proximities are 1. Each agent i selects from her neighborhood R(i) (which includes i herself) the agent j that maximizes her expected utility following Equation (2). We compare two scenarios. The iterated best response scenario follows the procedure of the proof of Theorem 2, in which agents sequentially update their proxy to best-respond to the current profile d using knowledge of their neighbors' accuracy q∗ i (d). In the one-shot game scenario all agents choose their proxy only once, do so simultaneously, and based only on their neighbors' accuracy. The latter setup captures more closely the epistemic limitations that agents face in liquid democracy. 4.1 Iterated Best Response Dynamics These experiments complement our existence theorems. They offer insights into the effects of delegations on average voter's accuracy in equilibrium, and on the effects of different network structures on how such equilibria are achieved. We initialize qi ∼ N (0.75, 0.05) and first investigate the case in which ei = 0 for all i (effortless voting). Across all combinations of network types and average degrees ranging from 4 to 24, we find that the best response dynamics converges, as predicted by 10 Table 2: Total number of best response updates by individual agents and corresponding full passes over the network required for convergence. Reported are the mean (std.dev.) over all network types. Note: not all agents update their delegation at each full pass, but any single update requires an additional pass to check whether the best response still holds. Degree BR updates (effortless) Full passes (effortless) BR updates (with effort) Full passes (with effort) 4 298.1 (18.2) 3.6 (0.5) 294.7 (18.4) 3.6 (0.5) 8 261.7 (11.1) 3.0 (0.1) 259.4 (10.6) 3.0 (0.3) 12 254.2 (6.9) 2.9 (0.2) 252.9 (6.6) 2.8 (0.6) 16 251.6 (4.5) 2.8 (0.4) 250.9 (4.8) 2.6 (0.8) 20 250.6 (3.3) 2.7 (0.5) 250.2 (4.9) 2.4 (0.9) 24 250.0 (2.6) 2.5 (0.5) 249.9 (4.3) 2.4 (1.0) Table 3: Comparing the maximum accuracy across all agents and the mean accuracy under delegation d for different network degrees, averaged across network types. The differences are statistically significant (paired t-test, p = 0.05). Degree maxj qj ¯q∗(d) 4 0.8908 0.8906 8 0.8908 0.8903 12 0.8904 0.8897 16 0.8909 0.8899 20 0.8904 0.8890 24 0.8910 0.8893 Theorem 2, and does so optimally with d∗ i = arg maxj qj for all i. We observe minimal differences between network types, but see a clear inverse relation between average degree and the number of iterations required to converge (Table 2, top). Intuitively, more densely connected networks facilitate agents in identifying their optimal proxies (further details are provided in Appendix B). We accumulate results across all network types and compare the effortless setting to the case in which effort is taken into account. When we include effort ei ∼ N (0.025, 0.01), we still observe convergence in all cases and, interestingly, the number of iterations required does not change significantly (Table 2, bottom). Although the process no longer results in an optimal equilibrium, each case still yields a single guru j with qj ≈ maxk qk (less than 1% error) for all k ∈ N . In this scenario, the inclusion of effort means that a best response update of agent i no longer guarantees non-decreasing accuracy and utility for all other agents, which was a key property in the proof of Theorem 2. This effect becomes stronger as the average network degree increases, and as a result higher degree networks allow a greater discrepancy between the maximal average accuracy achievable and the average accuracy obtained at stabilization (Table 3). In lower degree graphs (e.g. degree 4) we further observe differences in convergence speed between the four different network types which coincide with differences between the average path lengths in those graphs: a shorter average distance between nodes yields a lower number of best response updates. This is intuitive, as larger distances between nodes mean longer delegation chains, but we have not yet conducted statistical tests to verify this hypothesis. 11 Figure 1: Top: Without effort. Bottom: With effort ei ∼ N (0.025, 0.01). Left: mean accuracy under liquid democracy, ¯q∗(d). The solid (dashed) line shows the mean (std. dev.) of the initial accuracy q; the dotted line shows maxi qi. Right: probability of a correct majority vote under d. 4.2 One-Shot Delegation Games Here we study one-shot interactions in a delegation game: all agents select their proxy (possibly themselves) simultaneously among their neighbors; no further response is possi- ble. This contrasts the previous scenario in which agents could iteratively improve their choice based on the choices of others. While Kahng et al. (2018) study a probabilistic model, we instead assume that agents deterministically select as proxy the agent j ∈ R(i) that maximizes their utility, as above. We compare ¯q and ¯q∗ (the average network accuracy without and with delegation, respectively), as well as the probability of a correct major- ity vote under both direct democracy PD and liquid democracy PL where gurus carry as weight the number of agents for whom they act as proxy. The difference PL − PD is again similar to the notion of gain (Kahng et al., 2018). In line with Condorcet's jury theorem (see for instance Grofman et al. 1983) PD → 1 as N → ∞, and indeed for N = 250 we obtain PD ≈ 1. First we again look at the effortless setting. Figure 1 (top) shows both metrics for the four different network types and for different average degrees. We observe that while ¯q∗(d) increases as the network degree increases (and in fact is always higher than ¯q without 12 tancebetweennodesyieldsalowernumberofbestresponseupdates.9Thisisintuitive,aslargerdistancesbetweennodesmeanlongerdelegationchains,butwehavenotyetcon-ductedstatisticalteststoverifythishypothesis.One-ShotDelegationGamesHerewestudyone-shotinteractioninadelegationgame:allagentsselecttheirproxy(possiblythemselves)simulta-neouslyamongtheirneighbors;nofurtherresponseispos-sible.Thiscontraststhepreviousscenarioinwhichagentscoulditerativelyimprovetheirchoicebasedonthechoicesofothers.WhileKahng,Mackenzie,andProcaccia(2018)studyaprobabilisticmodel,weinsteadassumethatagentsdeterministicallyselectasproxytheagentj2R(i)thatmaximizestheirutility,asabove.Wecompare¯qand¯q⇤(theaveragenetworkaccuracywithoutandwithdelegation,re-spectively),aswellastheprobabilityofacorrectmajorityvoteunderbothdirectdemocracyPDandliquiddemocracyPLwhereguruscarryasweightthenumberofagentsforwhomtheyactasproxy.ThedifferencePLPDistheno-tionofgain(Kahng,Mackenzie,andProcaccia2018).InlinewithCondorcet'sjurytheorem(seeforinstanceGrof-man,Owen,andFeld1983)PD!1asN!1,andin-deedforN=250weobtainPD⇡1.Firstweagainlookattheeffortlesssetting.Figure1(top)showsbothmetricsforthefourdifferentnetworktypesandfordifferentaveragedegrees.Weobservethatwhile¯q⇤(d)increasesasthenetworkdegreeincreases(andinfactisal-wayshigherthan¯qwithoutdelegation),theprobabilityofacorrectmajorityoutcome,PL,simultaneouslydecreases.ThisconfirmstheanalysisofKahng,Mackenzie,andPro-caccia(2018).Wealsoobservethatthenumberofgurusdecreasesexponentiallyasthedegreeincreases(Figure2,left).Simplyput,givingallvotingweighttoasmallgroupofgurusincreasesthechanceofanincorrectmajorityvote,assumingthatgurushavealessthanperfectaccuracy.Whenweincludeeffort(Figure1,bottom),therebymov-ingawayfromthemodelofKahng,Mackenzie,andProcac-cia(2018),weobserveadrasticdecreaseinaveragenetworkaccuracycombinedwithalowerprobabilityofacorrectma-jorityoutcomeunderliquiddemocracy,withbothdecreas-ingasnetworkdegreeincreases.Themainreasonistheex-istenceofdelegationcyclesinthiscase.Thiscontraststhebestresponsesettingabovewhereagentscoulditerativelyreconsidertheirchoiceofproxyandthusavoidcycles.Now,evenwithrelativelyloweffort(mean0.025),uptohalfofallagentsarestuckinacycle(andtherebyfailtopassinforma-tionabouttheirtype)whendegreeincreases.ThisconfirmsresultsontheprobabilityofdelegationcyclesfromChristoffandGrossi(2017)andstressestheimportanceofcyclereso-lutioninconcreteimplementationsofliquiddemocracysuchasLiquidFeedback.Finally,Figure1highlightsdifferencesbetweenthefournetworktypes.Scalefreenetworksyieldalowerprobabilityofacorrectmajorityoutcomeacrossalldegrees,aswellasalargernumberofguruswithaloweraverageaccuracyand9Moredetailedresultssupportingthisfindingarepresentedinthesupplementarymaterial,AppendixB.4812162024Degree0.60.650.70.750.80.850.90.95Mean network accuracy4812162024Degree0.750.80.850.90.9511.05Prob. correct majorityRandomRegularSmall WorldScale Free4812162024Degree0.60.650.70.750.80.850.90.95Mean network accuracy4812162024Degree0.750.80.850.90.9511.05Prob. correct majorityRandomRegularSmall WorldScale Free4812162024Degree0.60.650.70.750.80.850.90.95Mean network accuracy4812162024Degree0.750.80.850.90.9511.05Prob. correct majorityRandomRegularSmall WorldScale FreeFigure1:Top:Withouteffort.Bottom:Witheffortei⇠N(0.025,0.01).Left:averagenetworkaccuracyunderliq-uiddemocracy.Thesolid(dashed)lineshowsthemean(std.dev.)ofq;thedottedlineshowsmaxiqi.Right:probabilityofacorrectmajorityvoteunderliquiddemocracy.4812162024Degree00.050.10.150.20.250.3Percentage of gurus4812162024Degree11.21.41.61.82Mean distance to guruRandomRegularSmall WorldScale FreeFigure2:Percentageofgurunodesunderd(left)andmeandistancebetween(non-guru)nodesandtheirgurus(right).longerdelegationchains(Figure2,right).Intuitively,thisindicatesthatone-shotinteractioninscalefreenetworksismorelikelytoendupinalocaloptimum.Incontrast,smallworldnetworkshaveshortaveragedistancesandthusagentsaremorelikelytobeclosetotheiroptimalguru.ConclusionsandFutureWorkThepaperintroduceddelegationgamesasafirstgame-theoreticmodelofliquiddemocracy.Bothourtheoreticalandexperimentalresultsshowedthatvotingeffortisakeyingredientforunderstandinghowdelegationsformandwhattheireffectsare.Ourempiricalfindingsprovidedfurtherin-sightsintotheinfluenceofinteractionnetworksonthequal-ityofcollectivedecisionsinliquiddemocracy.Thepaperopensupseveraldirectionsofresearch.Agen-eralNEexistencetheoremisthemainopenquestion.Theframeworkcanthenbegeneralizedalongnaturallines,e.g.:Figure 2: Percentage of guru nodes under d (left) and mean distance between (non-guru) nodes and their gurus (right). delegation), the probability of a correct majority outcome, PL, simultaneously decreases. This confirms the analysis of Kahng et al. (2018). We also observe that the number of gurus decreases exponentially as the degree increases (Figure 2, left). Simply put, giving all voting weight to a small group of gurus increases the chance of an incorrect majority vote, assuming that gurus have a less than perfect accuracy. When we include effort (Figure 1, bottom), thereby moving away from the model of Kahng et al. (2018), we observe a drastic decrease in average network accuracy combined with a lower probability of a correct majority outcome under liquid democracy, with both decreasing as network degree increases. The main reason is the existence of delegation cycles in this case. This contrasts the best response setting above where agents could iteratively reconsider their choice of proxy and thus avoid cycles. Now, even with relatively low effort (mean 0.025), up to half of all agents are stuck in a cycle (and thereby fail to pass information about their type) when the degree increases. This confirms results on the probability of delegation cycles from Christoff & Grossi (2017) and stresses the importance of cycle resolution in concrete implementations of liquid democracy. Figure 1 further highlights differences between the four network types. Scale free networks yield a lower probability of a correct majority outcome across all degrees, as well as a larger number of gurus with a lower average accuracy and longer delegation chains (Figure 2, right). Intuitively, this indicates that one-shot interactions in scale free networks are more likely to end up in a local optimum. In contrast, small world networks have short average distances and thus agents are more likely to be close to their optimal guru. Finally, our experiments highlight a key feature of liquid democracy: the trade-off between a reduction in (total) effort against a loss in average voting accuracy. 5 Conclusions and Future Work The paper introduced delegation games as a first game-theoretic model of liquid democ- racy. Both our theoretical and experimental results showed that voting effort is a key ingredient for understanding how delegations form and what their effects are. Our em- pirical findings provided further insights into the influence of interaction networks on the quality of collective decisions in liquid democracy. The paper opens up several directions of research. A general NE existence theorem 13 4812162024Degree00.050.10.150.20.250.3Percentage of gurus4812162024Degree11.21.41.61.82Mean distance to guruRandomRegularSmall WorldScale Freeis the main open question. Our model can then be generalized in many directions, e.g.: by making agents' utility dependent on voting outcomes; by dropping the independence assumption on agents' types; or by assuming the voting mechanism has better than 0.5 accuracy in identifying the types of agents involved in cycles. References Alger, D. (2006). Voting by proxy. Public Choice, 126 (1-2), 1 -- 26. Barab´asi, A.-L. & Albert, R. (1999). Emergence of scaling in random networks. science, 286 (5439), 509 -- 512. Blum, C. & Zuber, C. I. (2016). Liquid democracy: Potentials, problems, and perspectives. Journal of Political Philosophy, 24 (2), 162 -- 182. Boella, G., Francis, L., Grassi, E., Kistner, A., Nitsche, A., Noskov, A., Sanasi, L., Savoca, A., Schifanella, C., & Tsampoulatidis, I. (2018). WeGovNow: A map based platform to engage the local civic society. In WWW '18 Companion: The 2018 Web Conference Companion, (pp. 1215 -- 1219)., Lyon, France. International World Wide Web Conferences Steering Committee. Boldi, P., Bonchi, F., Castillo, C., & Vigna, S. (2011). Viscous democracy for social networks. Communications of the ACM, 54 (6), 129 -- 137. Brandt, F., Conitzer, V., Endriss, U., Lang, J., & Procaccia, A. D. (Eds.). (2016). Hand- book of Computational Social Choice. Cambridge University Press. Brill, M. (2018). Interactive democracy. In Proceedings of the 17th International Confer- ence on Autonomous Agents and MultiAgent Systems, (pp. 1183 -- 1187). International Foundation for Autonomous Agents and Multiagent Systems. Christoff, Z. & Grossi, D. (2017). Binary voting with delegable proxy: An analysis of liquid democracy. In Proceedings of TARK'17, volume 251, (pp. 134 -- 150). EPTCS. Cohensius, G., Mannor, S., Meir, R., Meirom, E., & Orda, A. (2017). Proxy voting for In Proceedings of the 16th Conference on Autonomous Agents and better outcomes. MultiAgent Systems, (pp. 858 -- 866). International Foundation for Autonomous Agents and Multiagent Systems. Dodgson, C. L. (1884). The Principles of Parliamentary Representation. Harrison and Sons. Endriss, U. (2016). Judgment aggregation. In F. Brandt, V. Conitzer, U. Endriss, J. Lang, & A. D. Procaccia (Eds.), Handbook of Computational Social Choice. Cambridge Uni- versity Press. Erdos, P. & R´enyi, A. (1959). On random graphs, I. Publicationes Mathematicae (Debre- cen), 6, 290 -- 297. Golz, P., Kahng, A., Mackenzie, S., & Procaccia, A. D. (2018). The fluid mechanics of liquid democracy. In WADE'18, arXiv:1808.01906. Green-Armytage, J. (2015). Direct voting and proxy voting. Constitutional Political Economy, 26 (2), 190 -- 220. 14 Grofman, B., Owen, G., & Feld, S. (1983). Thirteen theorems in search of truth. Theory and Decision, 15, 261 -- 278. Grossi, D. & Pigozzi, G. (2014). Judgment Aggregation: A Primer. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers. Jackson, M. O. (2008). Social and Economic Networks. Princeton University Press. Kahng, A., Mackenzie, S., & Procaccia, A. (2018). Liquid democracy: An algorithmic perspective. In Proc. 32nd AAAI Conference on Artificial Intelligence (AAAI'18). Kling, C. C., Kunegis, J., Hartmann, H., Strohmaier, M., & Staab, S. (2015). Voting behaviour and power in online democracy: A study of liquidfeedback in germany's pirate party. In Ninth International AAAI Conference on Web and Social Media. Miller, J. C. (1969). A program for direct and proxy voting in the legislative process. Public choice, 7 (1), 107 -- 113. Skowron, P., Lackner, M., Brill, M., Peters, D., & Elkind, E. (2017). Proportional rank- ings. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI 2017, Melbourne, Australia, August 19-25, 2017, (pp. 409 -- 415). Tullock, G. (1992). Computerizing politics. Mathematical and Computer Modelling, 16 (8- 9), 59 -- 65. Watts, D. J. & Strogatz, S. H. (1998). Collective dynamics of 'small-world' networks. Nature, 393 (6684), 440. 15 A Proofs of Proposition 1, Fact 1, and Fact 2 Here we provide the proofs op Proposition 1, Fact 1, and Fact 2. Proof of Proposition 1. First we show that q∗ s (d) = qs. If dt = t, the statement holds since the delegation from s to t is locally positive. Otherwise, let dt = r. Since d is positive, we know that s (d(cid:48)) > q∗ q ∗ t (d) = qr · pt,r + (1 − qr) · (1 − pt,r) > qt. Now let xi = P(τ (i) = 1) for i ∈ {s, t, r}. Since types are independent random variables, it holds that pi,j = xixj + (1 − xi)(1 − xj). Using this fact, we can verify that ps,r = ps,tpt,r + (1 − ps,t)(1 − pt,r) + (−2(2xs − 1) · (2xr − 1) · (xt − 1)xt). We now want to prove that −2(2xs − 1) · (2xr − 1) · (xt − 1)xt ≥ 0. (10) (11) First note that a locally positive delegation from i to j implies that pi,j ≥ 0.5 (since qi, qj ≥ 0.5); hence ps,t ≥ 0.5 and pt,r ≥ 0.5. Second, observe that for any a, b ∈ [0, 1], if ab + (1 − a)(1 − b) ≥ 0.5, then either a, b ∈ [0, 0.5] or a, b ∈ [0.5, 1]. Now, since ps,t = xsxt + (1 − xs)(1 − xt) ≥ 0.5 and pt,r = xtxr + (1 − xt)(1 − xr) ≥ 0.5, we conclude that either xs, xt, xr ∈ [0, 0.5] or xs, xt, xr ∈ [0.5, 1]. In both cases Equation (11) holds. From Equation (11) and Equation (12) it follows that: ps,r ≥ ps,tpt,r + (1 − ps,t)(1 − pt,r) = 2ps,tpt,r − ps,t − pt,r + 1. (12) Thus we obtain: (cid:48) ∗ s (d q ) = qrps,r + (1 − qr)(1 − ps,r) = 2qrps,r − qr − ps,r + 1 (2ps,tpt,r − pt,r − ps,t + 1 ≥ ps,r by Equation (12) ≥ 2qr(cid:124)(cid:123)(cid:122)(cid:125)≥1 (cid:124) (cid:123)(cid:122) (cid:125) ) − qr − (2ps,tpt,r − pt,r − ps,t + 1) + 1 = 4qrps,tpt,r − 2qrpt,r − 2qrps,t + 2qr − qr − 2ps,tpt,r + pt,r + ps,t − 1 + 1 = 4qrps,tpt,r − 2ps,tpt,r − 2qrps,t + 2ps,t − 2qrpt,r + qr + pt,r − 1 + 1 − ps,t = (2ps,t − 1 )(2qrpt,r − qr − pt,r + 1 ) + 1 − ps,t (cid:125) (cid:124) (cid:123)(cid:122) >0 (cid:125) > qt since d is positive (cid:124) (cid:123)(cid:122) (cid:123)(cid:122) (cid:125) > (2ps,t − 1)qt + 1 − ps,t = 2qtps,t − qt − ps,t + 1 (cid:124) > qs as d(cid:48) s = t is locally positive > qs. We conclude that q∗ holds that q∗ fact that the delegations from z to s and from s to r are positive. Hence d(cid:48) is positive. s (d(cid:48)) > qs. It remains to show that for an agent z with dz = s, it still z (d(cid:48)) > qz. This can be shown by the same argument as above, now using the 16 Proof of Fact 1. For the upper bound of 2, Example 2 shows that maximal average accu- racy can be made arbitrarily close to 1, while the average accuracy of a worse NE can be made arbitrarily close to 0.5. For the lower bound of 1 it suffices to consider any delegation game were R = ∅ (that is, no delegation is possible). Proof of Fact 2. For the lower bound we can use again Example 2, where the average accuracy of the worse NE can be made arbitrarily close to 0.5, and the average accuracy of direct voting (which given the structure of the example is maximized by direct voting) can be made arbitrarily close to 1. For the upper bound, consider an effortless homogeneous game where q1 = 1, and qi = 0.5 + for any i (cid:54)= 1 in N . Assume furthermore that R is such that any agent can delegate only to 1. In such a game, the average accuracy of the unique NE is 1, while the average accuracy of direct voting can be made arbitrarily close to 0.5. We thereby obtain the desired bound of 0.5. B Additional Simulation Results Detailed results for the "Simulations" subsection "Iterated Best Response Dynamics". Figure 3 shows convergence speed results in the effortless setting for different network types. The larger differences between the required number of best response updates for lower degree graphs of different types (e.g. degree 4) coincide with differences between the mean distance between nodes in those graphs: a shorter average distance yields a lower number of best response updates. The latter is shown in Table 4. Figure 3: Total number of best response updates of the individual agents (left) and corresponding number of full iterations over the network (right) for different network types and degrees in the effortless setting. Table 4: Mean distance between nodes and required number of best response updates for degree-4 networks of different types, sorted from high to low. Type Mean distance BR updates Small World 4.97 317.88 Regular 4.39 308.98 Random 4.03 287.88 Scale Free 3.41 277.55 17 4812162024Degree240250260270280290300310320Number of BR updatesRandomRegularSmall WorldScale Free4812162024Degree2.42.62.833.23.43.63.84Number of iterations
0902.4218	1	0902	2009-02-24T19:10:41	On graph theoretic results underlying the analysis of consensus in multi-agent systems	[ "cs.MA", "cs.DM", "math.CO", "math.OC" ]	This note corrects a pretty serious mistake and some inaccuracies in "Consensus and cooperation in networked multi-agent systems" by R. Olfati-Saber, J.A. Fax, and R.M. Murray, published in Vol. 95 of the Proceedings of the IEEE (2007, No. 1, P. 215-233). It also mentions several stronger results applicable to the class of problems under consideration and addresses the issue of priority whose interpretation in the above-mentioned paper is not exact.	cs.MA	cs	On graph theoretic results underlying the analysis of consensus in multi-agent systems = ,...,1 n i - tx )( & i (1) Pavel Chebotarev1 Institute of Control Sciences of the Russian Academy of Sciences 65 Profsoyuznaya Street, Moscow 117997, Russia Key words: consensus algorithms, cooperative control, flocking, graph Laplacians, networked multi-agent systems The objective of this note is to give several comments regarding the paper [1] published in the Proceedings of the IEEE. As stated in the Introduction of [1], “ Graph Laplacians and their spectral properties […] are important graph-related matrices that play a crucial role in convergence analysis of consensus and alignment algorithms.” In particular, the stability properties of the distributed consensus algorithms ( ) = (cid:229) txta tx ,)( )( )( i ij j iNj for networked multi-agent systems are completely determined by the location of the Laplacian eigenvalues of the network. The convergence analysis of such systems is based on the following lemma [1, p. 221]: Lemma 2: (spectral localization) Let G be a strongly connected digraph on n nodes. Then rank( L) = n − 1 and all nontrivial eigenvalues of L have positive real parts. Furthermore, suppose G has c ‡ 1 strongly connected components, then rank(L) = n − c. Here, L = [lij] is the Laplacian matrix of G, i.e., L = D – A, where A is the adjacency matrix of G, and D is the diagonal matrix of vertex out-degrees. Four comments need to be made concerning this lemma. First, the last statement of the lemma is not correct. Indeed, recall that the strongly connected components (SCC) of a digraph G are its maximal strongly connected subgraphs. For instance, if G is a converging tree , i.e., G is a directed tree with root r such that every vertex of G can be linked to r via a directed path, and n > 1, then G has c = n strongly connected components, but rank(L) = n − 1 > n − c = 0. The statement under consideration becomes valid if one replaces strongly connected components with weakly connected components (WCC) and additionally requires that these WCC’s are strong. A weakly connected component of G is a maximal subgraph of G whose vertices are mutually reachable by violating the edge directions. A more general correct statement results by substituting, in the same place, sink SCC’s, where a sink strongly connected component is an SCC having no edges directed outwards. This result was proved in [2] as well as some other Laplacian related results applicable to the cooperative control. Second, the proof of the rank property (the first statement of Lemma 2) is attributed in [1] to [3]. Let me note that a stronger fact was proved earlier in [2]. More specifically, Proposition 11 of [2] states that rank(L) = n – d, where d is the so-called in-forest dimension of G, i.e., the minimum possible number of converging trees in a spanning converging forest of G. It was also shown (Proposition 6) that the in-forest dimension of G is equal to the number of its sink SCC’s and that the forest dimension of a strongly L) = n − 1, connected digraph is one (Proposition 7). Consequently, for a strongly connected digraph, rank( which coincides with the first statement of Lemma 2. In addition, according to Proposition 8, “the forest dimension of a digraph is no less than its number of weak components 2 and does not exceed the number of its strong components and the number of its unilateral components.” 1 E-mail: [email protected]; [email protected] . 2 A weak component = a weakly connected component; a strong component = a strongly connected component. - , , + = kx ( i kx )( i + + )1 = i = ,...,1 n where (3) (2) Third, Remark 1 given after the proof of Lemma 2 says 3: “Lemma 2 holds under a weaker condition of existence of a directed spanning tree for G. […] This type of condition on existence of directed spanning trees have appeared in [4]–[6].” Here, by Lemma 2 the authors conceivably mean the conclusion that rank(L) = n − 1. Let us observe that the existence of a directed spanning tree for G implies that d = 1, so this statement follows from Proposition 11 of [2]. Fourth, the statement of Lemma 2 that “all nontrivial eigenvalues of L have positive real parts” holds true in the general case, and not only for strongly connected digraphs or digraphs that contain directed spanning trees. This was shown in [7, Proposition 9]. In Section II.C of [1] a discrete-time counterpart of the consensus algorithm (1) is considered: n ( ) (cid:229) kxa kx ,)( )( e i ij j j 1 = is the step size. In the matrix form, (2) is represented as follows: kx kxP ( )( )1 0>e IP L is referred to in [1] as the Perron matrix with parameter e of G. where e-= IP L e-= were studied in [2] and [7]; in particular, (i) of Lemma 3 in [1] coincides with The matrices Proposition 12 of [2]. The asymptotic behavior of the process (3) is determined by the properties of the sequence P, P2, P3,…. If the stochastic matrix P is primitive, i.e., it has only one eigenvalue with modulus 1, P = vw lim k T , where v and w are the right and left eigenvectors of P then, as stated in Lemma 4 of [1], k ¥fi 1=wvT . In the case of a corresponding to the eigenvalue 1, respectively, with a normalization that provides general nonnegative Perron matrix P, the sequence P, P2, P3,… need not have a limit, so the long-run m (cid:229) = ¥P always exists and, by the P m P lim k 1 - ¥ = transition matrix is considered. The matrix m ¥fi k 1 normalized matrix J of maximal in- Markov chain tree theorem proved in [8], [9], it coincides with the forests of G. J is the eigenprojector of L; by Proposition 11 of [2] rank ( J = d, where d is the in-forest ) J are the eigenvectors of L corresponding to the eigenvalue 0; dimension of G. The columns of consequently, they determine the consensus trajectories of the process (1) and the flocking trajectories [10]. The elements of J were characterized in Theorems '2 and 3 of [2]. An algebraic method for calculating J was presented in [7]. As has been shown above, [2] and [7] contained a number of results on the Laplacians of directed graphs which were useful for the cooperative control of multi-agent systems. A number of additional results were presented in [11] and [12]. Some of them are surveyed in [13]. In January 2001 Alex Fax, one of the authors of [1], sent me a message, where he asked about the eigenstructure of digraph Laplacians and requested to send copies of related papers. During the subsequent correspondence, later in 2001, I sent him [2] and [7]. Recently, I was pleased to familiarize myself with [1] and to learn that our early results proved to be useful in the analysis of consensus and cooperation algorithms of decentralized control. However, I was surprised that, instead of references to [2] and [7], this article contained references to papers published several years later. 3 The bibliographic references are redirected here to the list of references of this note. REFERENCES [1] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proc. IEEE. vol. 95, pp. 215–233, Jan. 2007. [2] R. P. Agaev and P. Yu. Chebotarev, “The matrix of maximum out forests of a digraph and its applications,” Automation and Remote Control, vol. 61, pp. 1424-1450, Sept. 2000. [3] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Trans. Autom. Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004. [4] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile autonomous agents using nearest neighbor rules,” IEEE Trans. Autom. Control, vol. 48, pp. 988–1001, June 2003. [5] L. Moreau, “Stability of multiagent systems with time-dependent communication links,” IEEE Trans. Autom. Control, vol. 50, pp. 169–182, 2005. [6] W. Ren and R. W. Beard, “Consensus seeking in multiagent systems under dynamically changing interaction topologies,” IEEE Trans. Autom. Control, vol. 50, pp. 655–661, 2005. [7] R. P. Agaev and P. Yu. Chebotarev, “Spanning forests of a digraph and their applications,” Automation and Remote Control, vol. 62, pp. 443–466, March 2001. [8] A. D. Wentzell and M. I. Freidlin, “On small random perturbations of dynamical systems,” Russian Mathematical Surveys, vol. 25, no. 1, pp. 1–55, 1970. [9] T. Leighton and R. L. Rivest, “The Markov chain tree theorem,” Computer Science Technical Report MIT/LCS/TM–249, Laboratory of Computer Science, MIT, Cambridge, Mass., 1983. [10] J. J. P. Veerman, G. Lafferriere, J. S. Caughman, and A . Williams, “Flocks and formations,” J. Statistical Physics, vol. 121, no. 5–6, pp. 901–936, 2005. [11] P. Chebotarev and R. Agaev, “Forest matrices around the Laplacian matrix,” Linear Algebra and Its Applications, vol. 356, pp. 253–274, 2002. [12] R. Agaev and P. Chebotarev, “On the spectra of nonsymmetric Laplacian matrices,” Linear Algebra and Its Applications, vol. 399, pp. 157–168, 2005. [13] P. Yu. Chebotarev and R. P. Agaev, “Coordination in multiagent systems and Laplacian spectra of digraphs,” Automation and Remote Control, vol. 70, no. 3, March 2009. In press.
1801.00743	1	1801	2018-01-02T17:45:28	Um Sistema Multiagente no Combate ao Braqueamento de Capitais	[ "cs.MA", "cs.AI" ]	Money laundering is a crime that makes it possible to finance other crimes, for this reason, it is important for criminal organizations and their combat is prioritized by nations around the world. The anti-money laundering process has not evolved as expected because it has prioritized only the signaling of suspicious transactions. The constant increasing in the volume of transactions has overloaded the indispensable human work of final evaluation of the suspicions. This article presents a multiagent system that aims to go beyond the capture of suspicious transactions, seeking to assist the human expert in the analysis of suspicions. The agents created use data mining techniques to create transactional behavioral profiles; apply rules generated in learning process in conjunction with specific rules based on legal aspects and profiles created to capture suspicious transactions; and analyze these suspicious transactions indicating to the human expert those that require more detailed analysis.	cs.MA	cs	Um Sistema Multiagente no Combate ao Branqueamento de Capitais Claudio Alexandre, João Balsa [email protected], [email protected]. Faculdade de Ciências da Universidade de Lisboa, BioISI-MAS, Campo Grande, 1749-016, Lisboa, Portugal DOI: 10.17013/risti.25.1-17 Resumo: Branqueamento de capitais é um crime que possibilita o financiamento de outros crimes, por isso ele é importante para as organizações criminosas e seu combate é motivo de mobilização das nações do mundo inteiro. O processo de anti- branqueamento de capitais não evoluiu como esperado pois tem priorizado a sinalização de transações suspeitas. O crescente aumento no volume de transações tem sobrecarregado o indispensável trabalho humano de avaliação final das sinalizações. Este artigo apresenta um sistema multiagente que objetiva ir além da captura de transações suspeitas, buscando auxiliar o especialista humano na análise das suspeições. Os agentes criados utilizam técnicas de data mining para criação de perfis de comportamento transacional; aplicam as regras obtidas no aprendizado em conjunto com regras especificas baseadas em aspectos legais e nos perfis criados para captura de transações suspeitas; e analisam estas transações sinalizadas indicando ao especialista humano aquelas que necessitam de análise mais detalhada. Palavras-chave: Sistemas multiagente; agentes inteligentes; data mining; anti- branqueamento de capitais. A Multi-Agent System in the Combat Against Money Laundering Abstract: Money laundering is a crime that makes it possible to finance other crimes, for this reason, it is important for criminal organizations and their combat is prioritized by nations around the world. The anti-money laundering process has not evolved as expected because it has prioritized only the signaling of suspicious transactions. The constant increasing in the volume of transactions has overloaded the indispensable human work of final evaluation of the suspicions. This article presents a multiagent system that aims to go beyond the capture of suspicious transactions, seeking to assist the human expert in the analysis of suspicions. The agents created use data mining techniques to create transactional behavioral profiles; apply rules generated in learning process in conjunction with specific rules based on legal aspects and profiles created to capture suspicious transactions; and analyze these suspicious transactions indicating to the human expert those that require more detailed analysis. Keywords: multi-agent systems; intelligent agents; data mining; anti-money laundering. RISTI, N.º 25, 12/2017 1 1. Introdução Ações de prevenção e combate ao crime de branqueamento de capitais (BC) são priorizadas por quase todos os governos do mundo, no mínimo, no mesmo nível das grandes questões globais (Madinger, 2012). BC é tipicamente um crime que consiste em tornar lícita a origem ilícita de um determinado ganho financeiro. A estimativa global anual de BC é em torno de 2% a 5% do Produto Interno Bruto Global (UNODC, 2014). Sobre fraude financeira, somente a América do Norte estima que sua perda passe de $3,1 mil milhões em 2017 para $4,8 mil milhões em 2018 (Cser, 2017). Este volume financeiro perdido anualmente já é motivo suficiente para o assunto ser tratado com prioridade, no entanto, outro fator leva os governos a priorizarem o combate a este crime: a sua comprovada ligação com outras práticas criminosas como narcotráfico, fraude, corrupção, sequestro, terrorismo, contrabando de armas, entre outros (Schott, 2006). As instituições financeiras, em sua maioria, já utilizam processos semi-automatizados para sinalização de transações suspeitas de BC, baseados em informações de cadastro, médias, desvios padrões e regras fixas pré-estabelecidas, geralmente oriundas de observações empíricas ou da experiência humana dos Analistas de anti-branqueamento de capitais (ABC). Porém, o crescente aumento no volume das transações realizadas, aliado à frequente publicação de novas regulamentações, nacionais e internacionais, acabam por provocar ineficiência neste processo de sinalização. Visando tornar o processo de ABC mais ágil e eficiente, este artigo apresenta a estrutura de um sistema multiagente para suporte à tomada de decisão neste contexto. Os agentes inteligentes utilizam técnicas de data mining para a criação de perfis de comportamento transacionais históricos, analisam e sinalizam transações suspeitas e auxiliam o Analista de ABC na tomada de decisão sobre as sinalizações. O modelo BDI (Belief, Desire, Intention) incorporado pela metodologia de desenvolvimento adotada (Prometheus), permite utilizar os perfis comportamentais encontrados, bem como possibilita a implementação de regras especificas em função do risco envolvido. Estas características, dentre outros benefícios, aumentam a eficiência do processo, enfrentando o aumento no volume de transações com a redução gradativa da intervenção humana. Este artigo segue a seguinte organização: indicação dos trabalhos relacionados com propostas de mineração de dados e arquiteturas de sistemas que buscam o ABC; apresentação da estratégia baseada no comportamento transacional dos atores para combate a este tipo de fraude; descrição da abordagem conservadora adotada no tratamento do risco envolvido; detalhamento do sistema e dos agentes definidos; apresentação dos resultados obtidos e as conclusões. 2. Trabalhos Relacionados A capacidade de adaptação do modus operandi dos fraudadores e a falta de informação sistematizada que associe as transações suspeitas com a comprovação do crime, são obstáculos para um avanço mais rápido da automatização do processo de prevenção e combate ao crime de branqueamento de capitais (BC). 2 RISTI, N.º 25, 12/2017 O primeiro sistema amplamente divulgado na área do anti-branqueamento de capitais (ABC) foi o FinCEN Artificial Intelligence System (FAIS) (Senator et al., 1995), desenvolvido e utilizado pelo Financial Crimes Enforcement Network (FinCEN), órgão do Departamento do Tesouro dos Estados Unidos. Em seguida, a Subcomissão Permanente de Investigações do Comitê do Senado Americano para Assuntos Governamentais solicitou ao Office of Technology Assessement (OTA)1 avaliação sobre a utilização de técnicas de pesquisa baseadas em inteligência artificial (IA), visando monitorizar o tráfego e transferências bancárias com o propósito de reconhecer transações suspeitas. Em seu relatório de 1995, o OTA concluiu que "o conceito original na sua formulação mais simples – o monitoramento do tráfego de transferência bancária, de forma contínua e em tempo real, usando técnicas de inteligência artificial – não é viável". Contudo, ponderou que existiam alternativas na tecnologia da informação a serem utilizadas para apoiar e reforçar a aplicação da lei contra o BC. Sugeriu, então, a utilização de técnicas tais como: knowledge acquisition, machine learning, clustering, knowledge sharing e data transformation (OTA, 1995). Em 1998, por orientação do relatório do United States General Accounting Office (GAO), a gestão decidiu pela não implementação de novos produtos utilizando técnicas de IA. Curiosamente, desde então, os estudos têm direcionado propostas para utilização das técnicas recomendadas buscando a identificação de anomalias ou situações suspeitas, conforme bem demonstrado em Chandola et al. (2009) e Sabau (2012). Contudo, conforme será comentado na seção 2.2, são poucas as propostas de sistemas baseados em agentes inteligentes para suportar todo o processo de ABC. 2.1. Mineração de Dados As técnicas de data-mining, machine learning e clustering têm sido fortemente utilizadas na tentativa de identificar casos suspeitos de BC. Em Zhang et al. (2003) um conjunto de dados são discretizados, mapeados para espaço dimensional Euclidiano, projetados numa linha de tempo discretizada, formando um histograma. Os clusters são criados, utilizando o algoritmo K-means, com base nos segmentos do histograma. Análises de correlação locais e globais são então aplicadas para detectar padrões suspeitos. É uma boa abordagem para a análise de comportamentos individuais e/ou de grupo baseado em picos anormais no histograma. No entanto, na análise de grande quantidade de clientes e transações durante longo período de tempo, a detecção de casos suspeitos pode ser dificultada, pois podem existir poucos ou nenhum pico no histograma. Em Kingdon (2004), é proposta a utilização de técnicas de IA e de uma máquina de vetores de suporte (Support Vector Machine-SVM) numa perspectiva diferente daquela até então utilizada, ou seja, montar o perfil histórico dos clientes e buscar utilizações fora do padrão ao invés de focar em comportamento suspeito baseado apenas em perfil cadastral. Independente da tecnologia utilizada, esta proposta fortalece o conceito da política Know Your Customer (KYC) e torna-se ainda mais útil se aplicada de forma incremental e constante. 1 http://ota.fas.org/ - escritório do Congresso Americano que funcionou de 1972 a 1995. RISTI, N.º 25, 12/2017 3 Em Tang & Yin (2005) os autores propõem outra extensão da SVM para analisar as transações dos clientes e detectar comportamento fora do padrão. É apresentada uma combinação de um kernel com melhoramentos do Radial Basis Function (RBF) Scholkopf et al. (2001) com definição de distâncias distintas e algoritmos SVM supervisionados e não- supervisionados. Uma vantagem desta abordagem é ela conseguir lidar com conjuntos de dados heterogêneos, porém, a avaliação de desempenho foi feita apenas com dados de simulação. Combinar a técnica de clustering com Multilayer Perceptron (MLP) foi a proposta de Le- Khac et al. (2009). O algoritmo K-means, é utilizado para formação dos clusters e esta técnica baseia-se em duas características principais (fundo de investimento e investidor) que, em seguida, são usados como entrada do processo de formação de um MLP. Os resultados apresentados mostram que a sua abordagem é eficiente. No entanto, o número de características, o número de padrões de treinamento utilizados foi pequeno e isso pode afetar a precisão. Outros métodos estatísticos foram propostos, como em Liu & Zhang (2010) que utiliza scan statistics, onde as transações realizadas num período de tempo são selecionadas aleatoriamente e agrupamentos incomuns são buscados. Adequado para trabalhos de auditoria em função da aleatoriedade adotada, no entanto, para um processo cotidiano de ABC nenhuma transação pode ser desprezada. Em Le-Khac & Kechadi (2010) os autores apresentam um estudo de caso em que aplicam uma solução para geração de uma base de conhecimento, combinando técnicas de mineração de dados, clustering utilizando K-means, redes neurais e algoritmos genéticos para detectar padrões de BC. Analisando somente por este documento, a solução pode apresentar um baixo custo benefício, considerando a alta complexidade da implementação proposta. Larik & Haider (2011) focam seu trabalho nas informações de débito e crédito realizadas pelos clientes de numa instituição financeira, visando identificar transações suspeitas. Propõem um novo algoritmo e um índice para avaliar e classificar as transações. Uma boa ideia, porém, limitada pela utilização somente das informações de débito e crédito. 2.2. Arquiteturas de Sistemas Em Gao et al. (2006), uma arquitetura de sistema é definida utilizando um conjunto de agentes especializados, tais como: agentes de coleta de dados (sistemas internos e informações externas); agentes de monitoramento para acompanhamento do perfil cadastral do cliente e das transações realizadas; e um agente que emite relatórios e alertas sobre possíveis operações de BC. Apesar da boa proposta de arquitetura, o problema crucial do volume de análises submetidas ao analista humano não é enfrentando, aliás, pode até ser agravado com a automatização da fase de sinalização de transações suspeitas. Outra abordagem baseada em agentes foi apresentada em Xuan & Pengzhu (2007). A arquitetura proposta é semelhante à descrita em Gao et al. (2006), porém, com evolução dos agentes incluindo as características de negociação, diagnóstico e autoaprendizagem. As 4 RISTI, N.º 25, 12/2017 características dos agentes é o ponto forte desta proposta, contudo, também encerra sua contribuição nas sinalizações para o analista humano. Em Xu & Gao (2010) os autores propõem uma alteração na arquitetura apresentada anteriormente, incorporando as fases e técnicas de um modelo de tomada de decisão, aplicado em tempo real. A arquitetura do sistema proposto anteriormente foi melhorada, mas em termos de processo não houve avanços, persistindo a falta de apoio aos analistas humanos. Rajput et al. (2014) utilizam ontologias e regras na criação de um sistema especialista para detecção de transações suspeitas de BC. Uma boa abordagem para a criação das regras a serem aplicadas, testada com um volume significativo de dados, contudo, a quantidade de 2% de transações suspeitas detectadas numa base real torna-se inviável num processo rotineiro que necessita da validação do especialista humano. 3. Estratégia de Combate a Fraudes A política Know Your Customer (KYC) definida pelo Comitê de Basiléia2 é um guia de melhores práticas no combate a fraudes, pois detalha os procedimentos a serem seguidos. Baseado nestes documentos e no benchmark realizado pela Hewlett-Packard (Laxman, 2014) foi criado um fluxo genérico de combate a fraudes ou burla no setor financeiro (Figura 1), quando realizadas em processos total ou parcialmente automatizados. Figura 1 – Fluxo Genérico de Combate a Fraudes O branqueamento de capitais (BC) é considerado um crime adjacente. Em Canas (2004) é explicado que "para haver Branqueamento teria de haver um crime anterior que proporcionasse ilicitamente ao seu autor proventos que posteriormente ele, ou outrem, 2 http://www.bis.org/publ/bcbs85.pdf e http://www.bis.org/publ/bcbs04a.htm RISTI, N.º 25, 12/2017 5 pretendessem camuflar", na sua essência é possível afirmar que ele é uma instância do fluxo genérico de fraude. Dessa forma, a atividade de anti-branqueamento de capitais (ABC) pode ser vista como sendo o fluxo genérico mostrado na Figura 1, aplicado ao setor financeiro. Um típico fluxo de ABC praticado pelas instituições financeiras foi mostrado e detalhado em Alexandre & Balsa (2015). Porém, mesmo sem a utilização de técnicas como mineração de processo (Norambuena & Zepeda, 2017), é fácil observar que o fluxo de ABC sendo tratado como uma instância do fluxo genérico de combate a fraudes é ineficiente, pois ele falha na identificação e sinalização das situações suspeitas. Casos não identificados impedem análise detalhada e a consequente geração de novas normas e recomendações, surgindo, dessa forma, o risco de uma falha sistêmica no processo geral de ABC. A Figura 2 mostra uma nova versão do fluxo geral de combate a fraudes, cujo objetivo é mitigar o risco identificado. A criação de perfis dos atores participantes da atividade, baseados no histórico completo das operações realizadas; a substituição dos atuais parâmetros fixos por regras de produção, embasadas nestes perfis e nas normas e recomendações existentes; a eliminação da utilização de somente operações mais recentes; e o uso de inteligência em alguns pontos do processo eleva qualitativamente o nível na captura de operações suspeitas e, principalmente, da tomada de decisão pelo especialista (Pinto et al., 2014). Figura 2 – Novo Fluxo Genérico de Combate a Fraudes 3.1. Abordagem baseada no comportamento transacional Uma base com dados reais refletindo o comportamento transacional de clientes, possui uma significativa quantidade de atributos necessários para o controle e gestão do negócio envolvido, sendo que nem todos são relevantes para uma busca de transações suspeitas. A seleção dos atributos relevantes e a possível geração de novos atributos, reflete a importância da fase de pré-processamento, que será descrita na seção seguinte. 6 RISTI, N.º 25, 12/2017 Os atributos selecionados ou gerados, precisam refletir as principais características do universo de dados onde a anomalia é buscada, no caso, devem refletir a base de transações e permitir identificar comportamentos suspeitos (Paula et al., 2016). Neste trabalho é proposta a criação de vários atributos, que agregam quantidades e segmentam características, com o objetivo de estabelecer um perfil do comportamento transacional, dentro de um ciclo temporal, para cada ator do processo. O ciclo temporal tem relação direta com a natureza do negócio envolvido, apresentando a duração máxima possível (trimestral, semestral, anual, etc.). 3.2. Pré-processamento dos dados e formação dos perfis Neste trabalho são utilizados dados reais de um banco brasileiro, referente a dois anos de transações do produto contas correntes. As contas dos 5,2 milhões de correntistas deste banco recebem, em média, 85 milhões de transações anuais. A análise da base de dados mostrou que menos de 10% dos clientes são constituídos por pessoas jurídicas (empresas comerciais, industrias, governos, etc.), no entanto, são responsáveis por mais de 90% dos valores totais envolvidos. Desta forma, visando uma melhor caracterização dos clientes, foi realizada a divisão da base de dados: uma somente com clientes singulares e outra com clientes tipo pessoa jurídica. Todo o procedimento descrito foi executado separadamente para cada base e está representado na Figura 3. Entrada: Base Total de Transações-BTT 1 Transações Relevantes-TR  análise da BTT 2 Perfis de Clientes-PC  seleção/geração de atributos da TR 3 Número de Clusters-k  Número de Atributos de PC – 1 (Passo 1) (Passo 2) (Passo 3) 4 enquanto k > 1 faça 5 Conjunto de Clusters-Cl  Algoritmo de Classificação ( algc(PC , k) ) 6 Vetor Erro1-E1(k)  Calcula Erro na Classificação ( algr1(Cl) ) 7 Vetor Erro2-E2(k)  Calcula Erro na Classificação ( algr2(Cl) ) k  k - 1 8 9 fim enquanto 10 Índice do Menor Erro-idx  Encontra o Menor (E1) 11 Classificação Final-CF  algc(PC, idx) 12 Conjunto de Regras1-R1  algr1(CF) 13 Índice do Menor Erro-idx  Encontra o Menor (E2) 14 Conjunto de Regras2-R2  algr2(CF) 15 retorna CF, R1, R2 Figura 3 – Algoritmo do Processo de Aprendizagem (Passo 4) (Passo 5) RISTI, N.º 25, 12/2017 7 De cada base foram excluídas as transações cujas características impedem a prática de branqueamento de capitais (tarifas, comissões, juros, impostos, etc.) (Figura 3 - Passo 1). A soma final das bases resultou em 35 milhões de transações relevantes. Um ano de transações foi utilizado para a geração do perfil de comportamento transacional, ou seja, para cada cliente presente na base de transações foram agrupadas informações tais como: idade da conta; quantidade de movimentos gerados e de serviços utilizados; percentual de transferência de recursos para outros bancos e para contas do próprio banco; além da quantidade de movimentos agregados em 6 faixas de valores. O 12º atributo, denominado percentual de débito, representa, de forma ponderada no período analisado, o tempo que o recurso permaneceu na conta do cliente (Alexandre & Balsa, 2016) (Figura 3 - Passo 2). 3.3. Aprendizagem e geração de regras A tabela de perfis dos clientes ativos no ano analisado ficou com 2,4 milhões de linhas, cada linha representando univocamente a tripla cliente, agência e conta. Esta tabela foi utilizada num processo de aprendizado indutivo não-supervisionado com clustering, para formação de grupos de clientes com características semelhantes e mutuamente exclusivos. O algoritmo K-means (algc) foi utilizado para classificação e os algoritmos PART e J48 (algr1 e algr2) para geração das regras de produção, executados 11 vezes (número de atributos menos 1) (Figura 3 - Passo 3). O modelo de cluster utilizado mostrou-se adequado ao permitir a interpretação dos agrupamentos sem a utilização de sofisticados esquemas de visualização (Castillo-Rojas et al., 2017). Dois conjuntos de regras, um de cada algoritmo utilizado, com a menor quantidade de instâncias incorretamente classificadas foram selecionados (Figura 3 - Passo 4) e conjuntamente com os clusters que originaram essas regras representam o resultado final do processo (Figura 3 - Passo 5). 4. Abordagem Conservadora com Relação a Risco A análise dos clusters gerados, para os dois segmentos de clientes, permitiu a identificação de características tais como: grande movimentação de valores elevados, com transferência integral para outras instituições financeiras (risco 3); ou movimentação de valores próximos do limite de comunicação aos órgãos reguladores (risco 2). Desta análise resultou a classificação mostrada na Tabela 1. Com esta classificação é possível definir uma melhor estratégia, oferecendo tratamento diferenciado aos grupos de clientes, conforme seu nível de risco. Apesar do ótimo nível de acerto obtido na avaliação das regras geradas, em torno de 99% para ambos os segmentos de clientes, um por cento de erro representa mais de 26 mil transações e não podem ser desprezadas. A matriz de confusão gerada pelos algoritmos identificou as regras que por serem aplicáveis a dois ou mais grupos de clientes compõem o 1% de erro mencionado. Ou seja, regras classificam clientes como pertencentes a mais de um perfil. A decisão foi então reclassificar os perfis que não representam risco ou têm baixo risco (perfis 1, 2 e 3), conforme mostra a 8 RISTI, N.º 25, 12/2017 Tabela 1. Dessa forma, a transação que pertencer a um desses 3 grupos será reclassificada, somente para efeito de análise, para uma regra nos perfis de risco mais elevado que satisfaça a condição. Por exemplo, na base de dados utilizada para data mining, 33 regras classificam os clientes pessoas singulares como pertencentes aos perfis de risco 2 e 3, correspondendo a 1,85% do total. Contudo, estas regras também classificam 0,06% de clientes originalmente pertencentes ao perfil padrão. A reclassificação consiste em, durante o processo de busca por transação suspeita, considerar esses clientes padrão como pertencentes aos grupos de risco, sem modificar a classificação original. 4.1. Criação de regras normativas e baseadas no comportamento A classificação dos perfis também permite a criação de regras especificas, quer sejam baseadas nos normativos vigentes, quer sejam inspiradas no comportamento transacional. Conforme já mencionado, este trabalho utilizou um ano de informações para geração dos perfis, obtendo totais mensais e permitindo selecionar valores máximos dentro do ano para cada atributo relevante. Foi estabelecido que a busca por transações suspeitas retroagirá sempre um mês a partir da data solicitada para análise, desta forma será sempre utilizado o comportamento de 1 mês de transações para efeito de comparação com os perfis. Para os perfis de risco, as regras utilizarão sempre cálculos em torno do maior valor mensal encontrado para cada atributo e, para os demais perfis, será utilizado como limite o valor total anual, conforme consta da Tabela 1. Tabela 1 – Classificação dos perfis gerados Perfil Pessoa Singular Outro Tipo de Cliente Reclassifica Perfil Valor Limite 1. Baixa Utilização Cluster4 2. Cliente Padrão Cluster2 Cluster4 Cluster1 3. Risco 1 (baixo) Cluster1 - 4. Risco 2 (médio) Cluster5 5. Risco 3 (alto) Cluster3 Cluster3 Cluster2 Sim Sim Sim Não Não Total anual Total anual Máximo mensal Máximo mensal Máximo mensal O banco de regras criado deve ser o mais estável possível, posto que se transformará em crenças a serem utilizadas pelos agentes, possibilitando tomada de decisão consistentes e coerentes. No entanto, é possível que a instituição financeira utilizadora do sistema deseje, numa determinada análise, ser mais rigorosa quanto aos valores limites, o que levaria, inevitavelmente, a modificação nas regras. Para contornar esta situação, foi parametrizado um percentual intitulado Margem Adicional de Risco (MAR), cuja atribuição ficará a critério da instituição utilizadora e que aplicado sobre os valores limites reduzem-no pelo percentual indicado. Nos testes realizados para este artigo, referido percentual foi mantido em zero. RISTI, N.º 25, 12/2017 9 A Figura 4 mostra que quando a MAR é maior que zero (o que está indicado em setembro, e no Total) os perfis de risco 2 e 3 têm seu limite reduzido pelo percentual indicado. O mesmo ocorre com os demais perfis, sendo que estes em relação ao valor total anual. O perfil de risco 1, por ser perfil de alerta, não é afetado pela MAR. Quando a MAR é igual a zero os perfis de risco utilizam o mesmo valor limite e os demais perfis utilizam o valor total anual, sem reduções. Figura 4 – Aplicação da Margem Adicional de Risco 5. Modelo dos Agentes A maioria dos sistemas existentes atualmente para suportar o processo de anti- branqueamento de capitais (ABC), concentra-se na fase do processo referente a captura de transações suspeitas, transferindo para o Analista de ABC a tarefa de comprovação da suspeição. Além disso, boa parte dessas soluções são fortemente baseadas em parametrizações que, basicamente, aplicam os normativos vigentes. O desempenho desses sistemas está refletido no resultado da pesquisa realizada em 2014 com 317 profissionais especialistas em ABC e técnicos de conformidade, de instituições financeiras de 48 países, mostrando que apenas 58% acreditam que os sistemas existentes 10 RISTI, N.º 25, 12/2017 em suas organizações são capazes de monitorar adequadamente as transações ocorridas nas diversas linhas de negócios3. Esta situação motiva este trabalho na proposição de uma solução que integra técnicas de data mining a um sistema multiagente com o propósito de capturar transações suspeitas de branqueamento de capitais e auxiliar o Analista de ABC na tomada de decisão. Na modelagem do sistema foi utilizada a metodologia Prometheus (Padgham & Winikoff, 2004). A implementação está sendo realizada no framework JaCaMo (Boissier et al., 2013) que utiliza a linguagem Jason (Bordini, Hübner & Wooldridge, 2007). A Figura 5 mostra, de forma simplificada, um dos diagramas da Prometheus Design Tool4(PDT), com a arquitetura final do sistema. O diagrama ressalta os agentes, as bases de dados externas (DB), as bases de conhecimento (KB) e a interação dos agentes com o ambiente e os Analistas de ABC. O grupo de agentes responsáveis pela captura de transação suspeita (CTS) que são especializados por produto da linha de negócios de uma instituição financeiro (contas correntes, câmbio, fundos de investimento, empréstimos, etc.). Esta especialização oferece duas grandes vantagens: a primeira é que cada agente pode ser aprimorado com as especificidades do produto e dos perfis dos clientes, que podem mudar em cada produto; a segunda diz respeito aos princípios de manutenibilidade e de escalabilidade, ou seja, um produto vigente pode ser descontinuado (morte de um CTS) e um novo produto pode ser criado (novo CTS) sem que isso interfira no funcionamento dos demais agentes. Na análise das transações os CTS utilizam as regras de produção vigentes, geradas com base nos perfis dos clientes, nos normativos sobre ABC e nas normas internas da instituição financeira. Atuam em duas modalidades: busca por transação ou busca por cliente. Na busca por transação a base histórica de transações é analisada integralmente, dentro do período informado, enquanto na busca por cliente, somente as transações relacionadas ao cliente informado são analisadas. Um agente é responsável pela gerencia da captura de transações (GCT), ele pode receber uma solicitação externa de análise e comandar para execução pelo CTS especialista ou comandá-la autonomamente. Ao receber a informação de um CTS de que uma transação suspeita foi identificada, comanda uma análise na modalidade busca por cliente para os demais CTSs responsáveis por produtos que constem do perfil do referido cliente. Somente o GCT tem conhecimento de quantos e quais são os CTSs existentes. Após todos os CTSs acionados terem enviado mensagens de resposta ao comando de análise, o GCT comunica ao agente auxilia processo decisório (APD) a existência de transações suspeitas. O agente APD é responsável por aprofundar a análise e decidir sobre a suspeição da transação, confirmando-a ou não. A utilização de conhecimento específico sobre o produto e de uma matriz de decisão baseada no histórico de decisões tomadas, permitem ao agente decidir sobre a suspeição ou, quando não é possível chegar a uma decisão, sinalizar para o 3https://home.kpmg.com/xx/en/home/insights/2014/01/global-anti-money-laundering- survey.html 4 https://sites.google.com/site/rmitagents/software/prometheusPDT RISTI, N.º 25, 12/2017 11 Analista de ABC sobre a existência de um caso complexo. As decisões tomadas pelo agente, bem como aquelas informadas pelo Analista, são guardadas no histórico de decisões e são utilizadas no aprendizado para evolução da matriz de decisão. Este agente também sugere alterações na matriz de decisão e pela atualização desta base de conhecimento, após as sugestões serem validadas. Figura 5 – Arquitetura do Sistema Proposto O agente evolui base de perfis (EBP) atua na análise do histórico de transações para geração de perfis de clientes e posterior comparação com a base de perfis existente. Este processo pode ser acionado por uma solicitação do usuário ou de forma autônoma pelo agente. Os possíveis novos perfis surgidos são sugeridos para o Analista de ABC, para análise. O agente EBP atualiza a base de perfis com os perfis que forem validados. As evoluções das bases de conhecimento e o aprendizado previstos no sistema visam, primordialmente, mitigar o risco da ocorrência de falso positivo e/ou falso negativo, existente em sistema baseados unicamente em um conjunto de regras e padrões de comportamento (Gao et al., 2006) e (Le-Khac & Kechadi, 2010). 6. Análise dos Resultados Os dados reais utilizados neste trabalho referem-se a 2 anos, com 30,5 milhões e 35,2 milhões de transações relevantes, respectivamente. Os perfis de comportamento transacionais dos clientes foram gerados a partir dos dados do primeiro ano, constituindo a base de referência. A busca por transações suspeitas foi realizada em um mês de 12 RISTI, N.º 25, 12/2017 GCTCTSAPDEBPtransações relevantes do segundo ano, resultando em 2,6 milhões de transações. Sobre estas transações foram gerados 516.942 perfis de comportamento transacionais dos clientes no período. O processo de busca por transações suspeitas, implementado até o momento, pode ser dividido em 3 fases: a reclassificação dos perfis ou ajuste da matriz de confusão; a captura das transações suspeitas; e a análise das transações capturadas. A Figura 6 mostra um resumo com informações do ambiente e a quantidade de transações suspeitas. Figura 6 – Resultado Fases 1 e 2 O sistema utiliza como crenças um conjunto de 129 regras, bem como a classificação dos clientes nos perfis indicados. Essas crenças são reavaliadas uma única vez, ocasião em que 418 perfis foram reclassificados, saindo do perfil Padrão para perfis de risco, somente nesta análise. No final 182 perfis são indicados como suspeitos e suas transações precisam ser analisadas. O resultado da fase 3 do processo é mostrado na Figura 7. A distribuição dos suspeitos por perfil, com seu respectivo percentual, alerta para a concentração de 52% dos suspeitos no perfil de risco 3 (Alto Risco). Com o objetivo de auxiliar o especialista na análise das RISTI, N.º 25, 12/2017 13 transações, os suspeitos são agrupados pela regra acionada na sua captura (mais de uma regra pode ser acionada para o mesmo perfil). Figura 7 – Resultado Fase 3 Em função do sigilo legal envolvido nas informações, a Figura 7 mostra apenas dois perfis com parte dos detalhes envolvidos, devidamente descaracterizados. No entanto, é possível observar o texto de algumas regras que ressaltam a relevância das suspeitas, como por exemplo a regra legal BCXX2016003, originada no normativo do Banco Central Brasileiro que sinaliza uma queda acentuada na movimentação de um perfil que, normalmente, tem alta movimentação. Os normativos dos bancos centrais dos países, geralmente, recomendam investigar esta mudança de comportamento pois pode representar a parada de um período no qual ocorreu o branqueamento de capital. O percentual de 0,0409% de suspeitos, para um mês de transações, mostra-se exequível para a análise humana. O sistema atualmente em utilização na instituição financeira fornecedora dos dados é fortemente baseado em valores limites pré-fixados e informações cadastrais, tipo renda/faturamento e atividade econômica. Assim sendo, as transações suspeitas agora indicadas incluem apenas algumas das suspeitas efetivamente identificadas e comunicadas 14 RISTI, N.º 25, 12/2017 aos órgãos de controle. As suspeitas baseadas no comportamento transacional dos clientes estão em análise visando atestar seu grau de precisão. 7. Conclusões É crescente a preocupação da indústria financeira e dos governos com o crime de branqueamento de capitais, quer seja pelos recursos perdidos, quer seja pelas consequências deste crime que, normalmente, financiam outros crimes. Os sistemas atualmente em utilização pelas instituições financeiras são fortemente baseados em informações cadastrais, a proposta aqui apresentada analisa o comportamento transacional dos clientes, estabelece um perfil e utiliza-o como balizador para comportamento futuro. Os resultados obtidos mostram a viabilidade de utilização sistemática e estabelece nova frente de combate a este crime. A verificação real das suspeitas sinalizadas estão em andamento e o próximo passo é fazer com que o agente de auxílio ao processo decisório aprenda com as decisões tomadas pela Analista de Anti-Branqueamento de Capitais e passe a sinalizar somente os casos inéditos. Referências Alexandre, C., & Balsa, J. (2015). A Multiagent Based Approach to Money Laundering Detection and Prevention. In Proceedings of the ICAAI. 1, pp. 230-235. Lisbon: SciTePress. doi:10.5220/0005281102300235 Alexandre, C., & Balsa, J. (2016). Integrating Client Profiling in an Anti-money Laundering in Information Systems and Multi-agent Based System. In New Advances Technologies (pp. 931-941). doi:10.1007/978-3-319-31232-3_88 Boissier, O., Bordini, R. H., Hübner, J. F., Ricci, A., & Santi, A. (2013, June). Multi-agent Oriented Programming with JaCaMo. Sci. Comput. Program., 78, 747-761. doi:10.1016/j.scico.2011.10.004 Bordini, R. H., Hübner, J. F., & Wooldridge, M. (2007). Programming Multi-Agent Systems in AgentSpeak Using Jason (Wiley Series in Agent Technology). John Wiley & Sons. Canas, V. (2004). O Crime de Branqueamento: Regime de Prevenção e de Repressão. (L. Almedina, Ed.) Coimbra-PT. Castillo-Rojas, Wilson, Medina-Quispe, Fernando, & Vega-Damke, Juan. (2017). Esquema de Visualización para Modelos de Clústeres en Minería de Datos. RISTI, (21), 67-80. https://dx.doi.org/10.17013/risti.21.67-80. Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly Detection: A Survey. ACM Comput. Surv., 41, 15:1--15:58. doi:10.1145/1541880.1541882 Cser, A., Kount Inc. (Producer). (2017). The Truth About Machine Learming in Fraud from Retrieved March [Video Prevention http://info.kount.com/webinar/machine-learning. webinar]. 23, 2017, RISTI, N.º 25, 12/2017 15 Gao, S., Xu, D., Wang, H., & Wang, Y. (2006). Intelligent Anti-Money Laundering System. Service Operations and Logistics, and Informatics, 2006. IEEE International Conference on, (pp. 851-856). doi:10.1109/SOLI.2006.328967 Kingdon, J. (2004). AI fights money laundering. Intelligent Systems, IEEE, 19, 87-89. doi:10.1109/MIS.2004.1 Larik, A. S., & Haider, S. (2011). Clustering based anomalous transaction reporting. Procedia Computer Science, 3, 606-610. doi: 10.1016/j.procs.2010.12.101 Laxman, S. (2014). The Fight Against Fraud. Retrieved March, 2017, from https://iaonline.theiia.org/the-fight-against-fraud Le-Khac, N., Markos, S., & Kechadi, M. T. (2009). Towards a New Data Mining-Based Approach for Anti-Money Laundering in an International Investment Bank. ICDF2C (pp. 77-84). doi:10.1007/978-3-642-11534-9_8 Le-Khac, N., & Kechadi, M. T. (2010). Application of Data Mining for Anti-money Laundering Detection: A Case Study. In Proceedings of the 2010 IEEE ICDMW (pp. 577-584). doi:10.1109/ICDMW.2010.66 Liu, X., & Zhang, P. (2010). A Scan Statistics Based Suspicious Transactions Detection Model for Anti-money Laundering (AML) in Financial Institutions. Mediacom 2010 (pp. 210-213). doi:10.1109/MEDIACOM.2010.37 Madinger, J. (2012). Money Laundering: A Guide for Criminal Investigators (Third ed.). Taylor & Francis Group. Norambuena, Brian Keith, & Zepeda, Vianca Vega. (2017). Minería de procesos de software: una revisión de experiencias de aplicación. RISTI, (21), 51-66. https://dx.doi.org/10.17013/risti.21.51-66. OTA. (1995). Information Technologies for the Control of Money Laundering (2 ed., Vol. from 2). Washington, DC: Congress of the U.S. Retrieved March, 2017, https://books.google.pt/books?id=TPdkrX2qkEAC Padgham, L., & Winikoff, M. (2004). Developing Intelligent Agent Systems: A Practical Guide. New York, NY, USA: John Wiley & Sons Inc. Paula, E., Ladeira, M., Carvalho, R., & Marzagão, T. (2016). Deep Learning Anomaly Detection as Support Fraud Investigation in Brazilian Exports and Anti-Money Laundering. 15th IEEE ICMLA (pp. 954-960). doi:10.1109/ICMLA.2016.0172 Pinto, T., Vale, Z., Sousa, T., Praça, I., Santos, G., & Morais, H. (2014). Adaptive learning in agents behaviour: A framework for electricity markets simulation. Integrated Computer-Aided Engineering, 21, 399-415. doi:10.3233/ICA-140477 Rajput, Q., Khan, N. S., Larik, A. S., & Haider, S. (2014). Ontology Based Expert-System for Suspicious Transactions Detection. Computer and Information Science, 7, 103-114. Retrieved March, 2017, from doi:10.5539/cis.v7n1p103 16 RISTI, N.º 25, 12/2017 Sabau, A. S. (2012). Survey of Clustering based Financial Fraud Detection Research. (I. EconomicÃ¡, Ed.) Informatica Economica, 16, 110-122. Scholkopf, B., Platt, J. C., Shawe-Taylor, J. C., Smola, A. J., & Williamson, R. C. (2001). Estimating the Support of a High-Dimensional Distribution. Neural Computing, 13, 1443-1471. doi:10.1162/089976601750264965 Schott, P. A. (2006). Reference Guide to Anti-Money Laundering and Combating the Financing of Terrorism: Second Edition and Supplement on Special Recommendation IX. Washington, DC: The World Bank and The IMF. Senator, T. E., Goldberg, H. G., Wooton, J., . . . Wong, R. W. (1995). The Financial Crimes Enforcement Network AI System (FAIS) Identifying Potential Money Laundering from Reports of Large Cash Transactions. AI Magazine, 16, 21-39. Retrieved March, 2017, from doi:10.1609/aimag.v16i4.1169 Tang, J., & Yin, J. (2005). Developing an intelligent data discriminating system of anti- money laundering based on SVM. Proceedings of 2005 International Conference on, 6, pp. 3453-3457. doi:10.1109/ICMLC.2005.1527539 UNODC. (2014). United Nations Office on Drugs and Crime - Annual Report 2014. Retrieved March, 2017, from https://www.unodc.org/documents/AnnualRepo rt2014/Annual_Report_2014_WEB.pdf Xu, D., & Gao, S. (2010). Real-Time Exception Management Decision Model (RTEMDM): Applications in Intelligent Agent-Assisted Decision Support in Logistics and Anti- Money Laundering Domains. 43rd HICSS, (pp. 1-10). doi:10.1109/HICSS.2010.314 Xuan, L., & Pengzhu, Z. (2007). An Agent Based Anti-Money Laundering System (pp. 5472-5475). for Financial Supervision. WiCom 2007 Architecture doi:10.1109/WICOM.2007.1340 Zhang, Z., Salerno, J. J., & Yu, P. S. (2003). Applying Data Mining in Investigating Money Laundering Crimes. In Proceedings of the Ninth ACM SIGKDD (pp. 747-752). New York, NY, USA: ACM. doi:10.1145/956750.956851 RISTI, N.º 25, 12/2017 17
1907.00327	1	1907	2019-06-30T06:12:48	Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning	[ "cs.MA", "cs.AI", "cs.CV" ]	There are many AI tasks involving multiple interacting agents where agents should learn to cooperate and collaborate to effectively perform the task. Here we develop and evaluate various multi-agent protocols to train agents to collaborate with teammates in grid soccer. We train and evaluate our multi-agent methods against a team operating with a smart hand-coded policy. As a baseline, we train agents concurrently and independently, with no communication. Our collaborative protocols were parameter sharing, coordinated learning with communication, and counterfactual policy gradients. Against the hand-coded team, the team trained with parameter sharing and the team trained with coordinated learning performed the best, scoring on 89.5% and 94.5% of episodes respectively when playing against the hand-coded team. Against the parameter sharing team, with adversarial training the coordinated learning team scored on 75% of the episodes, indicating it is the most adaptable of our methods. The insights gained from our work can be applied to other domains where multi-agent collaboration could be beneficial.	cs.MA	cs	Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning Niranjan Balachandar, Justin Dieter, Govardana Sachithanandam Ramachandran *all authors contributed equally 9 1 0 2 n u J 0 3 ] A M . s c [ 1 v 7 2 3 0 0 . 7 0 9 1 : v i X r a Abstract There are many AI tasks involving multiple inter- acting agents where agents should learn to coop- erate and collaborate to effectively perform the task. Here we develop and evaluate various multi- agent protocols to train agents to collaborate with teammates in grid soccer. We train and evaluate our multi-agent methods against a team operating with a smart hand-coded policy. As a baseline, we train agents concurrently and independently, with no communication. Our collaborative proto- cols were parameter sharing, coordinated learn- ing with communication, and counterfactual pol- icy gradients. Against the hand-coded team, the team trained with parameter sharing and the team trained with coordinated learning performed the best, scoring on 89.5% and 94.5% of episodes respectively when playing against the hand-coded team. Against the parameter sharing team, with adversarial training the coordinated learning team scored on 75% of the episodes, indicating it is the most adaptable of our methods. The insights gained from our work can be applied to other do- mains where multi-agent collaboration could be beneficial. 1. Introduction There are numerous tasks where collaboration is beneficial: driving, multi-player games, etc. In the future, many of the cars on the road will be autonomous (Heineke et al., 2017), so it would be useful for these cars to learn to collaborate to avoid accidents and improve efficiency. In multi-player games, it would be beneficial for agents to operate under a coordinated protocol that leverages unified strategies with specialization of responsibilities. A good example of such a game is soccer, where players have a common team strategy while performing individual roles. While learning to collaborate for these tasks seems intuitive to humans, for a given task, it is unclear how to most ef- fectively balance AI collaboration with specialization. We might consider treating the entire multi-agent system as a single agent and train this agent with a centralized controller. However, for a centralized controller the state and action spaces rise exponentially with the number of sub-agents, so this would not be logistically practical or computationally tractable when the state-space is large or there are many agents. Thus, decentralized methods that are capable of allowing collaboration and communication between agents are de- sirable for these situations. Ideally, the policies learned by the agents via decentralized methods will be adaptable to variations in the policies of collaborating agents, general- izable, and capable of achieving similar performance to a that of a centralized controller. Here we explore and eval- uate various multi-agent reinforcement learning methods in application to grid soccer, a multi-player soccer game (Iravanian), our objective is to train our multi-agent rein- forcement learning teams to defeat a team operating with an intelligent handed-coded strategy. 2. Background/Related Work A number of methods for multi-agent control have been developed in literature. 2.1. Centralized This approach involves a centralized controller that maps states of all the agents to a joint action (an action for each agent) (Gupta et al., 2017). This is a Multi-agent Partially observable Markov decision process (MPOMDP) policy. Advantages: This scheme allows tighter coordination be- tween agents and the dynamics are stationary. Disadvantages: The main disadvantage of the centralized approach is the exponential increase in the state and action spaces with increase in number of agents. Manipulations can be performed to reduce the joint action space size to nA (Moradi, 2016) for a discrete action space. For example, we could reduce the action space by factoring action probability i P (ai) where ai are the individual actions of agents, which reduces the action space from An to nA. as P ((cid:126)a) =(cid:81) Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning However, when there are many collaborating agents, this method may still be impractical. Thus for our paper, we will focus on decentralized methods for multi-agent control. 3. Approach 3.1. Grid Soccer Simulator 2.2. Concurrent Learning In concurrent learning, each agent is independent and learns its own unique policy (Gupta et al., 2017). In this scheme each agent has an independent observation and indepen- dently executes actions based on its observation. Here the reward is shared between all the agents. Agents should learn their individual roles while being evaluated with team performance, so this method should effectively decompose coordination in a distributed way. Advantages: The agents learn heterogeneous policies and the agents take specific roles, for example, in the soccer simulator, an agent could learn to be defender, another can learn to be the striker. Another advantage is the there is no communication between agents, which may be useful in a system where there are bandwidth constraints. Disadvantages: Since the agents do not share their experi- ence with each other, the training of unique policies does not scale with a large number of agents. This adds to the sample complexity of the scheme. Another disadvantage is that since all agents are learning policies independently, from each agent's point of view the dynamics are not stationary. 2.3. State sharing In this scheme, neighboring agents share their observation and action space with each other (Chernova & Veloso, 2008). Here the neighboring agents' actions and observation are stacked as the state space of each agent. This reduces the problem to distributed constraint optimization (DCOP). 2.4. Counterfactual Policy Gradients Counterfactual policy gradients is an actor critic method that leverages the utility of a centralized value approximation network as with parameter sharing but also gives a method for credit assignment to ensure that agents are properly given credit for individual actions rather than the actions of the group and allows for specialization of individual agents (Foerster et al., 2017) . The agents are represented by decentralized policy networks. The counterfactual algorithm uses TD learning to update the centralized critic and gradients of the policy network to perform updates. The policy gradients use a specialized advantage function that analyzes the contributions of each agents individual actions to the overall received reward. Grid Soccer Simulator is a discrete grid simulation of multi- player soccer (Iravanian). We built a python API to interact with this grid soccer simulator. The two teams have 3 play- ers each that play on a grid of size 10 × 18, as shown in Figure 1. Goals are indicated by the bolded lines on the right and left sides of the grids, and at any given timestep, one of the 6 players has the ball. In order to score, a player must physically move to the goal. It is possible (but obviously not advisable) for a player to score in its own goal, thus rewarding the opposing team. At each timestep, players can hold (stay still), move to one of the 8 adjacent grid positions, or pass to a teammate. choosing to move to an invalid position causes a player to remain at the same position. After each goal, a player from the team that did not score starts with the ball. Players can steal (cause a turnover of the ball) by running into a player of the opposing team who has the ball and is not holding or standing in between teammates of the opposing team that are passing to each other. The grid soccer game comes with a team that operates with an intelligent hand-coded strategy that is based on real- world soccer strategies. For example, when the opposing team has the ball and is close to the hand-coded team's goal, the players on the hand-coded team will go towards the hand-coded team's goal to play defense. 3.2. States We intend to leverage the ability of Convolutional Neu- ral Networks to analyze images, so at each timestep, we represent the state of each agent as a 4-channel image S ∈ RH×W×4 where H and W are the height and width of the grid respectively as illustrated in Figure 2. Each pixel of S is a boolean indicating the presence of a certain object. The first channel indicates the agent location: (cid:40) (cid:40) Si,j,1 = 1 0 location of agent is (i, j) otherwise . The second channel indicates the teammates' locations: Si,j,2 = 1 0 location of any teammate is (i, j) otherwise . (cid:40) The third channel indicates the opposing team members' locations: Si,j,3 = 1 0 location of any opposing team member is (i, j) otherwise . Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning Figure 1. Sample rendering of grid soccer game state. Two of our AI teams are playing against each other. Action Index 0 1 2 3 4 5 6 7 8 8 + k for k = 1 to n − 1 Action stay at (i,j) move to (i+1,j) move to (i,j-1) move to (i-1,j) move to (i,j+1) move to (i+1,j+1) move to (i+1,j-1) move to (i-1,j-1) move to (i-1,j+1) pass to teammate k Table 1. possible actions for each agent if current agent location is (i, j) and each team has n players Timestep Result agent own-goal team own-goal agent scored goal team scored goal opponent scored goal opponent own goal agent turns ball over team turns ball over agent steals ball team steals ball agent illegal movement agent successful pass agent hold agent legal movement Reward -100 -75 50 50 -50 10 -10 -10 10 10 -3 -1 -1 -2 Table 2. Rewards for each possible timestep result Figure 2. illustration of state representation as boolean 4-channel images The fourth channel indicates the ball's location: 1 0 location of ball is (i, j) otherwise . (cid:40) Si,j,4 = 3.3. Actions There are a total of n + 8 actions, where n is the number of players per team (n = 3 for our experiments). The possible actions for an agent currently located at (i, j) are listed in Table 1. 3.4. Rewards We constructed a handcrafted reward function to penalize getting scored on, turnovers, and stalling to incentive steal- ing the ball and scoring quickly. At each timestep, all agents in the team receive a reward, which depends on the result of the timestep. The possible results of a timestep and the corresponding rewards are summarized in Table 2. Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning 3.5. Deep Q-Learning Because the state space is very large (around (2n)HW +1 = 6181 for n = 3 players per team, tabular Q-learning is impractical, so we approximate the Q-value with a Convo- lutional Neural Network (CNN). Our CNN takes as input an agent state (H × W × 4) image and outputs a vector of size A of Q-values for each action for the input state. Our CNN has 3 convolutional layers followed by 2 fully connected layers, and the architecture is summarized in Fig- ure 3. Each layer except the final fully connected layer has ReLU activation. For stability, we include a target network as included (Mnih et al., 2015). We update the target net- work weights update after every iteration in which a goal is scored. We optimize the following objective function for gradient descent: Q(s(cid:48), a(cid:48), θ−) − Q(s, a, θ) (cid:17)2 (cid:16) L(θ) = r + γ max a(cid:48) where θ are the network weights and θ− are the target net- work weights, and Q is the feed-forward output of the CNN. During training, we perform an epsilon-greedy exploration with decayed over time and gradient descent with Adam (Kingma & Ba, 2014). 3.6. Concurrent Learning Our baseline approach is concurrent learning with no com- munication. Here each agent learns its own set Q-network weights. Since this approach has no explicit communica- tion or collaboration protocols, we expect this approach to perform the poorest. 3.7. Parameter Sharing In the parameter sharing scheme, the agents are homoge- neous. This allows model to be trained with experiences of all agents, which makes the algorithm data efficient. Here the learning is centralized while the control is decentral- ized by sharing learned parameters of the model between the agents. The agents still have different behavior since each agent receive different observations. This means a single policy is learned during train time and each agent executes the policy during inference with their independent observation. Advantage: The experience of all agents are used by the learner, hence making them sample efficient. Since the model is learned by a centralized learner and shared across all agents, they are computationally efficient. Computational resources such as GPU can be allocated to the centralized learner and each agent should use only light computational resources since inference is not computationally intensive. Disadvantage: It requires a centralized leaner and all of the agent's observation, reward and action needs to be transmit- ted to the leaner. This might be a limitation for an appli- cation that has limited bandwidth. The model lacks tighter coordination between agents as there is no possibility of communication. Here the model learns co-ordination im- plicitly by learning protocols an individual agent should follow based solely on its state. Here we use Q-Learning to update our value function ap- proximations. State, action, reward and the next sate from agents are transmitted to the centralized learner. For each time step of each agent, the centralized learner computes the Q-learning gradient to update the Q network. All agents share the same Q-network weights. 3.8. Coordinated Learning with Communication The main limitation of parameter sharing is that they lack tight co-ordination between agents. We address this by defin- ing a communication mechanism between agents. Many approaches in the past that use communication for tighter co-ordination, formulate the problem as Distributed Con- straint optimization (DCOP), which involves usage of the whole system or at least a smaller subsystem of agents. This uses significant communication bandwidth and causes la- tency in action selection for each agent. This becomes more intractable as the number of agents in the system increases. We formulate a new form of communication that uses a very small amount of bandwidth where the action selection is decentralized and does not involve computation for a group of agents. We achieved this by defining a joint action with communication a∗: a∗ = (ai, ag) here ai is the action taken by an individual agent. ag is one of a discrete number of communications that is broadcasted to other agents. The communications from neighboring agents are appended to an agents own observation xit, to form the state space sit = xit, ajt∀j, j (cid:54)= i In our DQN, we allow agents to receive this communication by having the following revised agent state representation. Instead of the original 4 channels, there will now be 3+Ag channels. The agent location channel, opponents location channel, and ball location channel remain. But now instead of a teammates locations channel, there will be Ag chan- nels, where channel ag will have a 1 at pixel (i, j) if there is a teammate at (i, j) that broadcasted communication action ag at the most recent timestep. Now our Q-network will output to joint action space instead of action space. Our implementation of coordinated learning also includes param- eter sharing (each agent has the same Q-network weights) for greater sample efficiency. One of the limitations of this approach is that the action space increases to ai × ag. Hence discretion is required Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning Figure 3. Q-value approximation network architecture in choosing the number of available group actions. In our experiments on grid soccer, we achieve this by defining a joint action: a∗ = arg max at Q(st, at) Another limitation is that since each agent is learning to communicate the group actions and since the neighboring agents' group actions are used as part of this agent's state space, the dynamics of the model are not stationary. To mitigate this issue, we use minibatches and a replay buffer for the model updates. For each time step for each agent, we capture experience et = (st, at, rt, st+1) and we update a replay buffer R = e1, e2, ...en. At each epoch, we sample a minibatch from the replay buffer R. Use of minibatch and replay buffer improves stability. Though coordinated learning with communication performs best, it requires a lot of epochs to train. One of the main reasons for this is because, each agents receives the same reward. We address this by using appropriate credit as- signment. The approach we explored is difference reward (Colby et al., 2015), which is defined as Di(z) = G(z) − G(zi + ci) Here G(z) is the global reward of unified state z of the system, zi is the system state without agent i, and ci is the counterfactual term. Since it is non-trivial to estimate G(zi + ci), we formulated a simple strategy for credit assignment. We assign a reward for each agent based on the following: Ri = R.Qi(st, at)( 1 N Qi(sit, ait)) N(cid:88) i=1 Here N is number of agents in the system, R is global re- ward signal, and Qi(sit, ait) is the state action value for agent i at timestep t. The intuition behind this is that if the agent directly contributed to the reward, the correspond- ing Qi would be greater than for an agent which did not contribute to overall performance of the system. 3.9. Counterfactual Policy Gradients The implementation of the counterfactual actor critic policy gradient method works as follows: 3.9.1. CENTRALIZED CRITIC The centralized critic is a function approximator Q : S × A → R where A is the joint action space of all agents. The usual representation of value approximation with a neural network that outputs a vector of size A will not work as A will grow exponentially with the number of agents and will become infeasible to represent as the output of a neural network. Instead, the centralized critic can output a vector of size A to provide the Q values of a single agent where A is the action space of a single agent. As an input, the network takes in the other agents actions as well as a one-hot vector representing which agent Q-values are being produced for. This network structure is demonstrated in Figure 4. The details of the actual model implemented are as follows. The states were represented as a 3 dimensional tensor, with the first two dimensions the dimensions of the grid followed by one channel each for every player on the counterfactual team, a single channel for every player on the opposing team, and a channel for the location of the ball. This was fed into a convolutional layer with 32 output channels, a 3×3 kernel, and a stride of 1. This was then fed into another convolutional layer with 64 output channels, a 4 × 4 kernel, and a stride of 2. This is then flattened and a one hot rep- resentation of u−a as well as a one hot representation of a Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning 3.9.3. ACTOR TRAINING AND COUNTERFACTUAL ADVANTAGES The unique representation of the centralized critic allows for a very useful advantage function to be computed. Aa(s, u) = Q(s, u) −(cid:88) πa(u(cid:48)aτ a)Q(s, (u−a, u(cid:48)a)) u(cid:48)a Figure 4. Structure of the centralized critic for the counterfactual −a algorithm. ut denotes the joint action at timestep t and u t denotes the joint action excluding the action of agent a (Foerster et al., 2017) This counterfactual advantage provides a metric of how much better/worse an agents actions will be compared to if it had chosen a different action with the actions of other agents held fixed. This provides a method of credit assign- ment to ensure that agents are only rewarded for how their actions contribute to the success of the team verses getting rewarded/penalized for results they did not contribute to. This advantage is then fed into the typical policy gradient update rule with advantages: (cid:88) ∆θ = α ∇θlogπa(uaτ a)Aa(s, u) is appended before being processed by two fully connected layers with ReLu activation and respective sizes of 128 and 64. Finally a linear layer is added at the end to output a vec- tor of size A. The network is trained using the SARSA(λ) algorithm which gives a combination of SARSA style re- turns and Monte Carlo returns (Foerster et al., 2017). We used λ = 0.8 to be empirically effective but the model could potentially benefit from more fine tuning of this parameter. 3.9.2. POLICY NETWORK To represent the stochastic policies of the actors, a policy network with a softmax output of dimension A is used. To allow the agents to share what they have learned parameters are shared but a one-hot representation of which agent is using the network is passed as an input to enable special- ization of agents. The exact structure of the neural network is as follows. The state is represented as a 3 dimensional tensor with a channel for the current agent's locations, a channel for the ball's location, a channel for the location of teammates, and a channel for the location of opposing players. This is fed into a convolutional layer with 32 out- put channels, a 3 × 3 kernel, and a stride of 1. This was then fed into another convolutional layer with 64 output channels, a 4 × 4 kernel, and a stride of 2. This is then flat- tened and a one hot representation of a is appended before being processed by two fully connected layers with ReLu activation and respective sizes of 128 and 64. A softmax layer is added at the end to give an output dimension of A. Additionally a decaying randomness factor of is added such that π(as) = (1 − )π(as) + to ensure that the agent is exploring sufficiently. is decayed from 0.5 to 0.05 over 300000 timesteps. a where θ is the parameters of the policy network. α is the learning rate, empirically chosen to be 0.001 for our ex- periments. Also, all the theoretical guarantees of conver- gence with normal actor critic methods hold for this actor critic method with counterfactual advantages (Foerster et al., 2017). 4. Experiment Results The current code for all our simulators, APIs, and learn- ing algorithms is included in https://github.com/ jdietz31/CS234-MultiagentProject. We eval- uate the performance of our algorithms by measuring the "goal ratio", which is the proportion of goals scored by the algorithm's team of the last 200 goals. So the worst possible goal ratio an algorithm can achieve is 0, and the best possi- ble goal ratio an algorithm can achieve is 1. We train each of our methods for hundreds of thousands of iterations. For each of our models, we use a learning rate of 0.001, initial epsilon of 0.5, final epsilon of 0.05, and perform linear ep- silon decay over the course of training. We vary the decay rate to account for differing sample efficiencies. For coordi- nated learning, we set the minibatch size to 1, 000 and replay buffer size to 50, 000. Hyperparameters for Counterfactual are specified in the methods section. We then trained our two best models- parameter sharing vs coordinated learning against each other to further de- lineate performance. The results of this adversarial train- ing are shown in Figure 7. Videos of the parameter shar- ing team playing the coordinated learning team are in- cluded here: https://drive.google.com/open? id=1jntEENSbnW_oyoFZJuCobBHybO_jdpUy. The results in Figure 5 and Figure 6 shows the training per- formance and final performance of all of our different multi- Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning Figure 5. Performance of multiagent models throughout training in terms of ratio of goals scored against the hand-coded team. Figure 6. Performance of multiagent models in the last 1000 iterations of training. Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning Figure 7. Goal ratio of the coordinated learning team when training adversarially against parameter sharing after both are pretrained against the hand-coded team. agent models in terms of the ratio of goals scored against the hand-coded agent. The Concurrent model achieves de- cent performance, but we observe that one of the agents tends to learn offense and defense, while the other two learn very little, so there is almost no teamwork. Counterfactual learns quickly but converges to a local optima where it only plays offense and no defense. Parameter Sharing learns to be almost perfect but is slightly unstable in its performance throughout training. The coordinated model learns the slow- est but is stable and converges to near perfect performance. Also, when allowed to train adversarially against other mod- els after being pretrained against the hand-coded team, coor- dinated learning performs the best as demonstrated in Figure 7. 5. Conclusion Our results show that the model with communication is able to perform the best against both the hand coded agent and against our other reinforcement learning agents. It is a simple enough model that it is able to learn very effectively and its communication model allows it to learn methods of cooperating and team strategy that simpler models cannot compete with. The counterfactual policy gradient method is the most versatile and complex model and with sufficient hyper-parameter tuning it should be able to surpass other models. The complexity of the model makes it promising to be able to learn difficult tasks but also makes it very unstable and not very robust with respect to performing well with a variety of hyper-parameters. Future work can expand on making this counterfactual method more robust and less unstable. Also, our experiments only worked with 3 agents and real world multiagent systems can be composed of many more agents. Future work should investigate into how these models can perform in more complicated scenarios with many more agents. References Chernova, Sonia and Veloso, Manuela. Teaching multi- robot coordination using demonstration of communica- tion and state sharing. In Proceedings of the 7th Inter- national Joint Conference on Autonomous Agents and Multiagent Systems - Volume 3, AAMAS '08, pp. 1183 -- 1186, Richland, SC, 2008. International Foundation for Autonomous Agents and Multiagent Systems. ISBN 978-0-9817381-2-3. URL http://dl.acm.org/ citation.cfm?id=1402821.1402826. Colby, Mitchell K., Curran, William, and Tumer, Ka- gan. Approximating difference evaluations with lo- In Proceedings of the 2015 Interna- cal information. tional Conference on Autonomous Agents and Multia- gent Systems, AAMAS '15, pp. 1659 -- 1660, Richland, SC, 2015. International Foundation for Autonomous Agents and Multiagent Systems. ISBN 978-1-4503-3413- 6. URL http://dl.acm.org/citation.cfm? id=2772879.2773372. Foerster, Jakob N., Farquhar, Gregory, Afouras, Triantafyl- los, Nardelli, Nantas, and Whiteson, Shimon. Counterfac- tual multi-agent policy gradients. CoRR, abs/1705.08926, 2017. Gupta, Jayesh K., Egorov, Maxim, and Kochenderfer, Mykel J. Cooperative multi-agent control using deep Collaboration of AI Agents via Cooperative Multi-Agent Deep Reinforcement Learning reinforcement learning. Workshop, 2017. doi: 10.1007/978-3-319-71682-4 5. In Adaptive Learning Agents Heineke, Kersten, Kampshoff, Philipp, Mkrtchyan, Self-driving car tech- the road? Armen, and Shao, Emily. nology: When will https://www.mckinsey.com/industries/ automotive-and-assembly/our-insights/ self-driving-car-technology-when- will-the-robots-hit-the-road, May 2017. the robots hit Sina. Iravanian, agents. GridSoccerSimulator. and https://github.com/sinairv/ Grid-soccer simulator Kingma, Diederik P. and Ba, Jimmy. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David, et al. Human-level control through deep reinforcement learn- ing. Nature, 518(7540):529 -- 533, February 2015. ISSN 00280836. URL http://dx.doi.org/10.1038/ nature14236. Moradi, Milad. A centralized reinforcement learning method for multi-agent job scheduling in grid. CoRR, abs/1609.03157, 2016. URL http://arxiv.org/ abs/1609.03157.
1807.08531	1	1807	2018-07-23T11:10:53	Multisensor Management Algorithm for Airborne Sensors Using Frank-Wolfe Method	[ "cs.MA" ]	This study proposes an airborne multisensor management algorithm for target tracking, taking each of multiple unmanned aircraft as a sensor. The purpose of the algorithm is to determine the configuration of the sensor deployment and to guide the mobile sensors to track moving targets in an optimal way. The cost function as a performance metric is defined as a combination of the D-optimality criterion of the Fisher information matrix. The convexity of the cost function is proved and the optimal solution for deployment and guidance problem is derived by the Frank-Wolfe method, also known as the conditional gradient descent method. An intuitive optimal approach to deal with the problem is to direct the sensor to the optimal position obtained by solving a nonlinear optimization problem. On the other hand, the proposed method takes the conditional gradient of the cost function as the command to the deployed sensors, so that the sensors are guaranteed to be in the feasible points and they achieve the current best performance. Simulation results demonstrate that the proposed algorithm provides better performance than directing each sensor to its optimal position.	cs.MA	cs	Multisensor Management Algorithm for Airborne Sensors Using Frank-Wolfe Method Youngjoo Kim* and Hyochoong Bang Department of Aerospace Engineering, Korea Advanced Institute of Science and Technology *Corresponding author: [email protected] Abstract: This study proposes an airborne multisensor management algorithm for target tracking, taking each of multiple unmanned aircraft as a sensor. The purpose of the algorithm is to determine the configuration of the sensor deployment and to guide the mobile sensors to track moving targets in an optimal way. The cost function as a performance metric is defined as a combination of the D- optimality criterion of the Fisher information matrix. The convexity of the cost function is proved and the optimal solution for deployment and guidance problem is derived by the Frank-Wolfe method, also known as the conditional gradient descent method. An intuitive optimal approach to deal with the problem is to direct the sensor to the optimal position obtained by solving a nonlinear optimization problem. On the other hand, the proposed method takes the conditional gradient of the cost function as the command to the deployed sensors, so that the sensors are guaranteed to be in the feasible points and they achieve the current best performance. Simulation results demonstrate that the proposed algorithm provides better performance than directing each sensor to its optimal position. Keywords: multisensor management; Fisher information matrix; Frank-Wolfe method; conditional gradient descent; convex optimization 1. Introduction Multisensor system has emerged as an important research subject in military as well as civilian applications. Whereas a single sensor generally can obtain limited information, multiple sensors are able to provide sufficient information obtained from different viewpoints and focuses. Given the advancement in sensor technologies, the amount of data the sensors can process has increased. This motivates researches on automatic management of multiple sensors to improve overall perception performance. Multisensor management is formally described as a system or a process that attempts to manage a set of sensors in a dynamic, uncertain environment, to improve the performance of data fusion [1]. Details of purpose, role, architecture, and problem of the multisensor management are well described in [1-3]. Some multisensor management algorithms have been developed within the domain of antisubmarine warfare [4, 5]. In particular, a general methodology for the management of multisensor systems can be found in [4]. Expanding the single-target tracking scheme, a problem of managing an array of sensors in order to track multiple targets was considered in [6, 7]. These works provide convincing answers to this application area in regard to efficient and autonomous deployment of sonobuoy resources in submarine tracking. Important steps of the multisensor management technique are to quantify and control the accuracy of the target state estimation. The Fisher information matrix (FIM) provides the means of achieving this aim in an efficient way [8, 9]. The inverse of the FIM denotes a lower bound on the performance of an estimator. The purpose of this study is to develop a multisensor management algorithm applicable to airborne warfare, considering each unmanned aircraft as a sensor system. Our previous work [10] has attempted to achieve the goal, but it is limited to the management of stationary sensors. One major difference of the unmanned aircraft from the sonobuoys is that the former are mobile sensors. In order to obtain better performance, the sensors can move to appropriate positions to chase the moving targets. The fact calls our attention to the problem in the sensor management as to determining the appropriate positions of the deployed sensors. We have addressed path planning of 2 of 18 a mobile sensor where the location of sensor deployment is different of that of an operational area [11]. Another recent work [12] has addressed control of multiple mobile sensors using a potential field obtained by gradient of a goal function, but it is limited to maneuvers in 2D space without consideration of sensor coverage. An airborne sensor needs to be in the feasible region where the measurement is available. By using FIM to quantify the performance, in this paper, we propose a multisensor management algorithm for determining 1) when a new sensor is required, 2) where to deploy the new sensor, and 3) how to guide the deployed sensors. The first two decisions, of when and where to deploy a new sensor, have been addressed in [10]. In this study, on the other hand, we imposed constraints to the optimization problems. That is, we took sensor coverage in account for making decisions of deployment location and guidance of the sensors. Our contribution lies in the method to provide the guidance command to the deployed sensors. We formulated the problem of finding the best sensor location as a constrained convex optimization problem and provided the negative conditional gradient of the cost function as the guidance command, which is the solution of the subproblem of Frank-Wolfe optimization technique. Solving the subproblem every time step significantly reduced computation time while the sensors approached the optimal location, maintaining the targets in the sensors' coverage. The proposed multisensor management algorithm was verified by numerical simulations with stationary and moving targets. The remainder of the paper is structured as follows. Section 2 presents the problem formulation for which the multisensor management algorithm is designed. Section 3 addresses how to define the performance metric and its convexity is proved therein. In Section 4, the multisensor management algorithm is given. The algorithm consists of determining the configuration of the sensor deployment and generating command to the deployed sensors. Section 5, some examples are given to demonstrate the effectiveness of the proposed algorithm. Summary and conclusions are presented in Section 6. 2. Problem Formulation In this section, we present the problem formulation for the management algorithm for an airborne multisensor system to track multiple ground targets. It is assumed that each ground target is moving independently, and the number of the targets is 𝐿. The number of the unmanned aircraft is 𝑁𝑚𝑎𝑥 and that of the deployed is 𝑁. Suppose that each unmanned aircraft is equipped with a sensor system, which provides angle and range measurements. For example, one can use a radar system, a LIDAR system, or a camera system with terrain elevation data for the framework to be described in this paper. The purpose of the multisensor management is to obtain the best estimate of the target position. Each sensor system in this paper is assumed able to detect the false measurement and distinguish multiple measurements from the multiple targets. The modelling methodology for the false alarms and measurement association, generally considered in radar systems, can be referred to [4, 6, 7]. A major characteristic of the airborne multisensor system is that the sensors are mobile. In this study, we restrict the unmanned aircraft as a rotary-wing type that is able to hover and move to every direction unlike fixed-wing aircraft. This helps us to focus on the multisensor management, not the path planning. A fusion center as a sensor manager, which can be a ground control station or an unmanned aircraft, collects sensory data and determines when, where, and for which target group the unmanned aircraft should be deployed. At the same time, each mobile sensor is directed to chase the moving targets after its deployment by continuously calculating appropriate directions to track the targets to achieve better performance. Within the problem domain described above, there are three inter-related problems to be addressed. The first problem is the time for the new sensors to be deployed, given the current estimate of the target state. The target tracking algorithm is supposed to satisfy a criterion as the lower bound of the state estimate performance. The time will be determined as the time when the accuracy of the state estimate is lower than a prespecified criterion. The second problem is to determine the configuration of the next deployment; the target group to be tracked and the location of the newly deployed sensor. In other words, given a certain time for 3 of 18 the next deployment, and the predicted target state at that time, the algorithm determines which targets should be tracked by the new sensor and where the sensor should be positioned in an optimal way. An optimization problem for maximizing information of targets the new sensor gathers should be solved to determine the configuration of the next deployment. The third problem is the guidance of each sensor to track the moving targets. That is, the direction in which each sensor should move is supposed to be determined by solving the optimization problem. 2.1. Target Model In this paper, the target system is described as a stochastic vector state space model. Each target is assumed to be moving independently with a constant velocity. The system model for each target 𝑙 = 1,2, … , 𝐿 can be described as a near-constant-velocity model [13] by 𝑿𝑘+1 𝑙 = 𝐹𝑘 𝑙𝑿𝑘 𝑙 + 𝐺𝑘 𝑙 𝝂𝑘 𝑙 (1) The target state at time k is defined as 𝑿𝑘 𝑙 = [𝑥𝑘, 𝑦𝑘, 𝑧𝑘, 𝑥𝑘, 𝑦𝑘, 𝑧𝑘]𝑇 = [𝒑𝑘, 𝒗𝑘]𝑇. The process noise 𝑙 obeys a zero-mean Gaussian distribution whose covariance is expressed as 𝝂𝑘 𝑙 = [ 𝑄𝑘 𝑙)2 (𝜎𝑥 0 0 2 0 𝑙 ) (𝜎𝑦 0 0 0 𝑙)2 (𝜎𝑧 ] (2) where [𝜎𝑥 matrices 𝐹𝑘 𝑙 𝜎𝑧 𝑙] 𝑙 𝜎𝑦 𝑙 and 𝐺𝑘 𝑇 denotes the standard deviation of the noise and ∆𝑡 is the sampling interval. The 𝑙 are constructed as 𝑙 = [ 𝐹𝑘 𝐼3×3 03×3 𝐼3×3∆𝑡 𝐼3×3 ] 𝑙 = [ 𝐺𝑘 ∆𝑡2 𝐼3×3 2 𝐼3×3∆𝑡 ] (3) and 𝐼𝑚×𝑛 is an m by n identity matrix. The covariance of 𝐺𝑘 𝑙 𝝂𝑘 𝑙 is 𝐺𝑘 𝑙 𝑄𝑘 𝑙 (𝐺𝑘 𝑙 ) 𝑇 . 2.2. Observation Model Consider the observation model regarding the measurement from the target 𝑙 by the sensor 𝑛 = 1,2, … , 𝑁 given by 𝑛,𝑙 = 𝒉𝑘 𝒁𝑘 𝑛,𝑙(𝑿𝑘 𝑙 ) + 𝝎𝑘 𝑛,𝑙 (4) 𝑛 be a set of targets assigned to the sensor n at time 𝑘. Each sensor is supposed to track the 𝑛 and provides measurement to the fusion center. Invalid measurements and Let 𝕋𝑘 target 𝑙 ∈ 𝕋𝑘 measurements from other targets are discarded. Each sensor provides angle measurements in azimuth and elevation, and a range measurement, 𝑛,𝑙 = [𝐴𝑧, 𝐸𝑙, 𝑅]𝑇. Each measurement from the sensor n tracking the target l is expressed as as 𝒁𝑘 𝑥𝑙−𝑥𝑛 𝑦𝑙−𝑦𝑛) tan−1 ( √(𝑥𝑙 − 𝑥𝑛)2 + (𝑦𝑙 − 𝑦𝑛)2 + (𝑧𝑙 − 𝑧𝑛)2] [ 2 +(𝑦𝑙−𝑦𝑛) 2 √(𝑥𝑙−𝑥𝑛) tan−1 ( 𝑧𝑙−𝑧𝑛 ) + 𝝎𝑘 𝑛,𝑙 (5) 𝑛,𝑙 = [ 𝒁𝑘 𝐴𝑧 𝐸𝑙 𝑅 ] + 𝝎𝑘 𝑛,𝑙 = 4 of 18 𝑛,𝑙 is the zero-mean Gaussian measurement noise whose covariance matrix is 𝑅𝑘 𝑛,𝑙 . The where 𝝎𝑘 corresponding observation matrix 𝐻𝑘 𝑛,𝑙 is obtained by 𝑛,𝑙 = 𝐻𝑘 𝜕𝒉𝑘 𝑛,𝑙(𝑿𝑘 𝑙 ) 𝑙 𝜕𝑿𝑘 (6) Note that 𝐻𝑘 𝑛,𝑙 = 𝜕𝒉𝑘 𝑛,𝑙 𝜕𝑿𝑘 𝑙 ⁄ has zero elements in the last three columns as 𝑛,𝑙 = [ 𝐻𝑘 𝜕𝒉𝑘 𝑛,𝑙(𝑿𝑘 𝑙 ) 𝑙 𝜕𝒑𝑘 𝜕𝒉𝑘 𝑛,𝑙(𝑿𝑘 𝑙 ) 𝑙 𝜕𝒗𝑘 ] = [ 𝜕𝒉𝑘 𝑛,𝑙(𝑿𝑘 𝑙 ) 𝑙 𝜕𝒑𝑘 03×3] (7) Letting 𝑥 = 𝑥𝑙 − 𝑥𝑛, 𝑦 = 𝑦𝑙 − 𝑦𝑛, 𝑧 = 𝑧𝑙 − 𝑧𝑛, the observation matrix can be expressed as 𝑦 𝑛,𝑙 = 𝐻𝑘 −𝑥𝑧 𝑥2+𝑦2 (𝑥2+𝑦2+𝑧2)√𝑥2+𝑦2 [ √𝑥2+𝑦2+𝑧2 𝑥 −𝑥 𝑥2+𝑦2 −𝑦𝑧 0 1 (𝑥2+𝑦2+𝑧2)√𝑥2+𝑦2 √𝑥2+𝑦2 𝑦 𝑧 √𝑥2+𝑦2+𝑧2 √𝑥2+𝑦2+𝑧2 03×3 (8) ] The measurement is available only when the target 𝑙 is in the coverage of the sensor. For simplifying the problem, assume each sensor always looks downward. The sensor position is restricted as tan−1 ( 𝑥𝑙−𝑥𝑛 𝑧𝑙−𝑧𝑛) < tan−1 ( 𝑧𝑛 > 𝐻𝑚𝑖𝑛 𝜃𝑥 2 𝑦𝑙−𝑦𝑛 𝑧𝑙−𝑧𝑛) < 𝜃𝑦 2 (9) where 𝜃𝑥 and 𝜃𝑦 denote angle of views of the sensor in x and y direction, respectively. 𝐻𝑚𝑖𝑛 denotes the minimum altitude of the sensor. By rearranging (9), a set of linear constraints for the sensor position is obtained as ±(𝑥𝑙 − 𝑥𝑛) − (𝑧𝑙 − 𝑧𝑛) tan 𝜃𝑥 2 ±(𝑦𝑙 − 𝑦𝑛) − (𝑧𝑙 − 𝑧𝑛) tan 𝑛 𝑧𝑛 > 𝐻𝑚𝑖𝑛 for all 𝑙 ∈ 𝕋𝑘 < 0, 𝜃𝑦 2 < 0, (10) The above equations will be used as the constraints of the optimization problems in section 4. 2.3. Sequential Extended Kalman Filter In this paper, we adopted the sequential extended Kalman filter (EKF) to estimate the target state since every measurement and state transition are independent and the observation matrix can be expressed as the first-order derivative of the observation model. Brief introduction to the sequential EKF is given here and the details can be referred to the literature [14]. Its time update step is the same as that of typical EKF. The measurement update step of the sequential EKF is as follows. + 𝐾𝑘+1, 𝑖 (𝒁𝑘+1, 𝑖 − 𝒉𝑖(𝑋𝑘+1, 𝑖−1 + )) + 𝑿 𝑘+1, 𝑖 + = 𝑿 𝑘+1, 𝑖−1 𝑇 + 𝐾𝑘+1, 𝑖 = 𝑃𝑘+1, 𝑖−1 𝑃𝑘+1, 𝑖 + 𝐻𝑘+1, 𝑖 + = 𝑃𝑘+1, 𝑖−1 (𝐻𝑘+1, 𝑖𝑃𝑘+1, 𝑖−1 + 𝑇 𝐻𝑘+1, 𝑖 + 𝑅𝑘+1, 𝑖) + −1 − 𝐾𝑘+1, 𝑖𝐻𝑘+1, 𝑖𝑃𝑘+1, 𝑖−1 (11) + where 𝑿 𝑘+1, 0 available measurement at the time step. − and 𝑿 𝑘+1 + = 𝑿𝑘+1, 𝐼 = 𝑿 𝑘+1 + , and so does for 𝑃 . In the above, 𝑖 = 1,2, … , 𝐼 denotes 3. Performance Metric 3.1. Fisher Information Matrix 5 of 18 After the target state is estimated, the fusion center calculates the FIM for the multisensor system. Let 𝑿𝑘 be an unknown, random state vector to be estimated. Its unbiased estimate is defined as 𝑿 𝑘(𝒁𝑘) when the measurement 𝒁𝑘 is obtained. Then, the posterior Cramer-Rao lower bound (PCRLB) is defined to be the inverse of the FIM 𝐽𝑘. It provides a lower bound of the error covariance matrix of 𝑿 𝑘(𝒁𝑘) as 𝐶𝑘 ≜ 𝔼 {[𝑿 𝑘(𝒁𝑘) − 𝑿𝑘][𝑿 𝑘(𝒁𝑘) − 𝑿𝑘] 𝑇 } ≥ 𝐽𝑘 −1 (12) where 𝔼 denotes expectation over 𝑿𝑘 and 𝒁𝑘 . The inequality in the above equation means that 𝐶𝑘 − 𝐽𝑘 −1 is a positive semi-definite matrix. The posterior FIM 𝐽𝑘+1 can be evaluated by the following formula given in [8, 9] as 𝐽𝑘+1 = (𝑄𝑘 + 𝐹𝑘𝐽𝑘 −1𝐹𝑘 𝑇)−1 + 𝐽𝑍,𝑘+1 (13) The first term of the FIM 𝐽𝑘+1 contains the prior information of the target states at time 𝑘 + 1. The initial FIM is defined as 𝐽0 = (𝑃0)−1. It is assumed that the sensors can identify each target so that the measurement-to-target associations are known. If the association is unknown, its effect to the FIM can be evaluated by an information reduction matrix [6]. Note that the matrix 𝐽𝑍,𝑘+1 in (13) gives the measurement contribution to the FIM. For calculating the measurement contribution to the FIM, consider the observation model regarding the measurement of the target 𝑙 by the 𝑛th sensor given by (5) revisited as 𝑛,𝑙 = 𝒉𝑘 𝒁𝑘 𝑛,𝑙(𝑿𝑘 𝑙 ) + 𝝎𝑘 𝑛,𝑙 (14) 𝑛,𝑙 is a nonlinear function relating the target state and the measurement. 𝝎𝑘 where 𝒉𝑘 Gaussian measurement noise vector whose covariance is 𝑅𝑘 simplifying the measurement contribution given in [4], 𝐽𝑍,𝑘+1 𝑛,𝑙 𝑛,𝑙 is a zero-mean 𝑛,𝑙. By expanding the discussion about can be expressed as 𝑛,𝑙 = (𝐻𝑘+1 𝑛,𝑙 ) 𝐽𝑍,𝑘+1 𝑇 −1 𝑛,𝑙 ) (𝑅𝑘+1 (𝐻𝑘+1 𝑛,𝑙 ) (15) Since the sensors have independent measurement processes, the measurement contribution 𝐽𝑍,𝑘+1 can be written as 𝑙 𝑙 𝐽𝑍,𝑘+1 𝑛 = ∑ 𝑛,𝑙 𝑁 𝐽𝑍,𝑘+1 𝑛=1 𝑙 ≠ 0 and 𝑤𝑘+1 for 𝑙 ∈ 𝕋𝑘+1 (16) 𝑙 denotes the measurement contribution to the FIM of tracking the target 𝑙. We adopted where 𝐽𝑍,𝑘+1 a weight vector 𝒘 to contain tracking priority on each target. For example, 𝒘𝑘 = [2,1,1,0] if an operator attempts to track the first three targets with higher priority on the first target. If a sensor is tracking the fourth target, the information of the fourth target will be ignored. It is important to make the sensor management algorithm predictive because it needs sufficient time to plan and make the deployment. Therefore, if we perform the optimization at time 𝑘, it is necessary to predict the FIM in the subsequent sampling times 𝑘 + 1, … for determining the configuration and the time of the next deployment. Given the FIM 𝐽𝑘 at time 𝑘, we can determine the predictive FIM 𝐽𝑘+𝑟 at each time 𝑘 + 𝑟, 𝑟 = 1,2, … by using the Riccati-like recursion, (13). 3.2. Performance Score 6 of 18 In this section, two kinds of scores are introduced. One is a measure of the amount of information of a certain target the fusion sensor gathers, called 'information score'. The other, called 'performance score', is a measure of the performance of a certain sensor to track the targets assigned to the sensor. Note that the optimization problem in this paper is formulated as a minimization problem, so that the lower score implies better performance. 3.2.1. Information of Each Target Given the FIM, denoted as 𝐽𝑘+𝑟 𝑙 , for tracking the target l at time 𝑘 + 𝑟, the information score for the target position can be defined by using the D-optimality criterion as 𝑙 where Ψ𝑘+𝑟 as 𝑙 = − 𝑙𝑛 𝑑𝑒𝑡(Ψ𝑘+𝑟 𝑐𝑘+𝑟 𝑙 ) (17) is a FIM for target position estimation taken as the first 3 × 3 elements in the FIM 𝐽𝑘+𝑟 𝑙 𝑙 = [ 𝐽𝑘+𝑟 𝑙 Ψ𝑘+𝑟 𝐽21 𝐽12 𝐽22 ] (18) and 𝐽12, 𝐽21, 𝐽22 are 3 × 3 matrices. The D-optimality criterion is based on the determinant of the FIM, which corresponds to the volume of the uncertainty ellipsoid. In addition to this, common criteria for optimal observation can be found in various literature [15, 16]. The individual information score 𝑐𝑘+𝑟 is used in determining the time for the new deployment. 𝑙 3.2.2. Performance of Each Sensor In target state estimation, a general goal is to minimize the target position estimation error. In tracking multiple targets, we are interested in maximizing overall performance of each sensor for tracking targets assigned to it. The performance score will be used in determining both configurations of the new deployment and the guidance command to track the moving targets in an optimal way. The cost function describing the tracking performance is defined based on the measurement contribution to the FIM 𝐽𝑍,𝑘+1 and the observation matrix is calculated as 𝐻𝑘+1 where the state for the FIM is taken as the sensor state 𝑿𝑘+1 𝑛,𝑙 = 𝜕𝒉𝑘+1 𝑛,𝑙 𝜕𝑿𝑘+1 𝑛,𝑙 ⁄ . 𝑛 𝑛 The tracking performance score of the 𝑛th sensor at time 𝑘 + 1 is defined as 𝑛 𝑠𝑍,𝑘+1 where = ∑ 𝑐𝑍,𝑘+1 𝐿 𝑙=1 𝑛,𝑙 𝑙 𝑤𝑘+1 for 𝑙 ∈ 𝕋𝑘+1 𝑛 (19) 𝑛,𝑙 = − ln det(Ψ𝑍,𝑘+1 𝑐𝑍,𝑘+1 𝑛,𝑙 ) (20) 𝑛,𝑙 and Ψ𝑍,𝑘+1 denotes the measurement contribution to the FIM of the senor n estimating the position of the target l at time 𝑘 + 1. Note that, because the measurements have no observability in velocity, 𝐻𝑘+1 has zero columns as (7) and the measurement contribution to the FIM has a form of 𝑛,𝑙 = 𝜕𝒉𝑘+1 𝑛,𝑙 𝜕𝑿𝑘+1 ⁄ 𝑛 𝑛,𝑙 = (𝐻𝑘+1 𝑛,𝑙 ) 𝐽𝑍,𝑘+1 𝑇 −1 𝑛,𝑙 ) (𝑅𝑘+1 𝑛,𝑙 ) = [ (𝐻𝑘+1 𝑛,𝑙 Ψ𝑍,𝑘+1 03×3 03×3 03×3 ] (21) The performance score of the sensor 𝑛, (19), is a function of the sensor position and formulated as a weighted sum of the performance scores of the sensor 𝑛 to track the target 𝑙, denoted as 𝑐𝑍,𝑘+1 . The information of the target whose weight is 0 is discarded. Only the measurement contribution term of (13) is used in defining the cost function because the information affected by the system model is independent of the sensor configuration. Note that as the determinant of the FIM becomes larger, 𝑛,𝑙 7 of 18 the performance scores, (17) and (19), become lower. Thus, the global minimum of (19) is the optimal position of the sensor. 3.2.3. Convexity of the Performance Score Before solving the optimization problem of target tracking, the convexity of the cost function has to be supported by theory. This subsection is dedicated to prove that the performance score of target tracking, 𝑠𝑍,𝑘+1 , is convex. The definitions and theorems introduced in the following paragraphs are adopted from the literature [17-20] and some notations are adjusted according to the conventions of this paper. 𝑛 Definition 1 (Convex set [18]). Let 𝕊 be a subset of ℝ𝑛. We say that 𝕊 is a convex set if 𝛼𝒙 + (1 − 𝛼)𝒚 ∈ 𝕊, 𝑓𝑜𝑟 𝑎𝑙𝑙 𝒙, 𝒚 ∈ 𝕊 𝑎𝑛𝑑 𝛼 ∈ [0, 1]. Definition 2 (Convex function). A function 𝑓 ∶ 𝕊 ⟼ ℝ is called convex over a convex set 𝕊 if 𝑓(𝛼𝒙 + (1 − 𝛼)𝒚) ≤ 𝛼𝑓(𝒙) + (1 − 𝛼)𝑓(𝒚), 𝑓𝑜𝑟 𝑎𝑙𝑙 𝒙, 𝒚 ∈ 𝕊 𝑎𝑛𝑑 𝛼 ∈ [0, 1]. defining the performance of target tracking has its domain, 𝕊, as a set of 𝑛 ∈ 𝕊 = ℝ3. The set 𝕊 is a convex set, and we need to prove the 𝑛 The cost function 𝑠𝑍,𝑘+1 positions of the sensor n, i.e., 𝒑𝑘+1 function 𝑠𝑍,𝑘+1 𝑛 ∶ 𝕊 ⟼ ℝ is convex over 𝕊. 𝑛,𝑙 Firstly, let us deal with the function 𝑐𝑍,𝑘+1 ∶ 𝕊 ⟼ ℝ defined in (20). It has been proved that the D-optimality criterion is convex in general as described in the following theorem. Theorem 1 (Convexity of the D-optimality criterion [20]). The function Λ ∶ Ψ ⟼ − ln det(Ψ) is convex if Ψ is positive semi-definite and strictly convex if Ψ is positive definite. Therefore, it indicates that the function 𝑐𝑍,𝑘+1 𝑛 tracking the target 𝑙 , denoted as Ψ𝑍,𝑘+1 definiteness is introduced below. 𝑛,𝑙 𝑛,𝑙 is convex if the FIM of the observation by the sensor , is positive semi-definite. The definition of positive Definition 3 (Positive definite [17]). A matrix Ψ is positive definite if 𝒂𝑇Ψ𝒂 > 0, and positive semi-definite if 𝒂𝑇Ψ𝒂 ≥ 0 for all 𝒂 ≠ 0. Define 𝑛,𝑙 = [ 𝐻𝑘+1 𝜕𝒉𝑘+1 𝑛 𝑛,𝑙 (𝒑𝑘+1 𝑛 ) 𝜕𝒑𝑘+1 0] = [ 𝑇 𝒉1 ⋮ 𝑇 𝒉𝑚 01×3 ⋮ ] (22) 01×3 where 𝒉𝑖 are 3 × 1 vectors where 𝑖 = 1, … , 𝑚 and 𝑚 is the number of measurements of the sensor 𝑛. Consequently, the measurement contribution to the target position can be expressed as 𝑛,𝑙 = ∑ 𝜎𝑖 Ψ𝑍,𝑘+1 𝑚 𝑖=1 −2𝒉𝑖 𝑇𝒉𝑖 (23) Choose any 3 × 1 vector 𝒂 ∈ ℝ3, 𝒂 ≠ 𝟎, and multiply both sides of Ψ𝑍,𝑘+1 𝑛,𝑙 by 𝒂 as 𝒂𝑇Ψ𝑍,𝑘+1 𝑛,𝑙 𝒂 = ∑ 𝜎𝑖 𝑚 𝑖=1 −2(𝒂𝑇𝒉𝑖)2 ≥ 0 (24) Therefore, Ψ𝑍,𝑘+1 𝑛,𝑙 is positive semi-definite, and we can conclude 𝑐𝑍,𝑘+1 𝑛,𝑙 is convex. By expending the above discussion, it can be shown that the Fisher information matrix is positive semi-definite in general. If Ψ𝑍,𝑘+1 exist a 𝒂 ∈ ℝ3, 𝒂 ≠ 𝟎 such that 𝒂𝑇𝒉𝑖 = 0. Then, 𝜎𝑖 𝑛,𝑙 is full-rank, besides, null ((Ψ𝑍,𝑘+1 𝑛,𝑙 𝑇 ) −2(𝒂𝑇𝒉𝑖)2 > 0 and Ψ𝑍,𝑘+1 ) = ∅. Thus, there does not 𝑛,𝑙 is positive definite. 8 of 18 Lemma 1 (Operation that preserves convexity [19]). If 𝑓1, … , 𝑓𝑚 ∶ 𝕊 ⟼ ℝ are convex, then the following function is convex as well. 𝑋 ⟼ 𝛼1𝑓1(𝑋) + ⋯ + 𝛼𝑚𝑓𝑚(𝑋) for 𝛼1, … , 𝛼𝑚 ≥ 0 𝑛,𝑙 The performance score of the sensor 𝑛, 𝑠𝑍,𝑘+1 shown in (19). Therefore, 𝑠𝑍,𝑘+1 𝑛 𝑠𝑍,𝑘+1 is also a global minimum. 𝑛 𝑛 as is convex as well. It is concluded that there is a local minimum of , is calculated by a convex combination of 𝑐𝑍,𝑘+1 4. Multisensor Management Algorithm The proposed algorithm is described as a block diagram in Fig. 1. A cycle of the loop is executed every time a new measurement is arrived. The first and second steps, estimating target states and calculating the individual information score, are described in section 2 and 3 in detail. In this section, how to determine the configuration of the new sensor and how to guide the deployed sensors will be addressed. Given estimated target states, if any target exists where the information score of the target 𝑙 > 𝑆1, the fusion center prepare a new sensor deployment is larger than a prespecified value, i.e., 𝑐𝑘+𝑟 as follows. Figure 1. Block diagram of the proposed multisensor management algorithm. 4.1. Determining the Configuration of New Deployment The configuration of the deployment means the location of the new sensor and the set of targets to be tracked by the sensor. For L targets, there are possible 2𝐿 − 1 target groups. The number of the candidates can be reduced by selecting the candidate groups which contains the target 𝑙− whose 𝑙− information has been considered insufficient, i.e., 𝑐𝑘+𝑟 > 𝑆1. Also, the target groups whose size is larger than a prespecified value 𝑉 are discarded. The maximum number of targets, 𝑉, reflects the limitation in processing power of sensor systems. Therefore, for each candidate of the target group 𝑁+1,𝑖 and the 𝑁+1,𝑖 ∈ {𝕋𝑘+𝑟 𝕋𝑘+𝑟 corresponding sensor position are determined by solving an optimization problem. The candidate 𝑁+1,𝑖 is identified as the target group to be tracked if the best score of the configuration target group 𝕋𝑘+𝑟 has the largest value among all the candidates. The algorithm for determining the new configuration is as follows. 𝑁+1,𝑖 ≤ 𝑉} where 𝑖 = 1, … , 2𝐿 − 1, the best score 𝑠𝑍,𝑘+𝑟 𝑁+1,𝑖 𝑙− ∈ 𝕋𝑘+𝑟 𝑁+1,𝑖 and 𝕋𝑘+𝑟 9 of 18 Algorithm 1 (Configuration of new deployment). Get a set of candidate target groups: 𝑁+1,𝑖 and 𝕋𝑘+𝑟 𝑁+1,𝑖 𝑙− ∈ 𝕋𝑘+𝑟 𝑁+1,𝑖 ≤ 𝑉} {𝕋𝑘+𝑟 𝑁+1,𝑖, Set the best score 𝑠∗ = ∞ For each 𝕋𝑘+𝑟 Solve Problem 1: ∗ 𝑁+1,𝑖) Get the best score (𝑠𝑍,𝑘+𝑟 , 𝑁+1,𝑖) and the sensor position (𝒑𝑘+𝑟 } , 𝒑∗ = arg 𝑠∗ ∗ ∗ 𝑁+1,𝑖) 𝑠∗ = min {𝑠∗, (𝑠𝑍,𝑘+𝑟 Deploy a new sensor at 𝒑∗ 𝑁 = 𝑁 + 1 END ∗ 𝑁+1,𝑖) = arg(𝑠𝑍,𝑘+𝑟 The convex optimization problem is defined as follows. Problem 1 (Best sensor position). minimize 𝑠𝑍,𝑘+𝑟 subject to 𝒑𝑘+𝑟 𝑁+1,𝑖) 𝑁+1,𝑖(𝒑𝑘+𝑟 𝑁+1,𝑖 ∈ 𝑃𝑘+𝑟 𝑁+1,𝑖 𝑁+1,𝑖 is defined as a set of sensor positions which satisfy the constraints specified in (10). 𝑃𝑘+𝑟 We solve Problem 1 using the conditional gradient method [18], also known as Frank-Wolfe method. The formulation for solving the optimization problem regarding a cost function 𝑠 which is a function of a position 𝒑 is given as follows. In order to generate a feasible direction 𝒑𝑗 − 𝒑𝑗 that satisfies the descent condition ∇𝑠(𝒑𝑗) ∙ (𝒑𝑗 − 𝒑𝑗) < 0, we use the method 𝒑𝑗+1 = 𝒑𝑗 + 𝛼 𝑗(𝒑𝑗 − 𝒑𝑗) (25) which is to solve the optimization problem minimize ∇𝑠(𝒑𝑗) ∙ (𝒑 − 𝒑𝑗) subject to 𝒑 ∈ 𝑃 (26) where 𝛼 𝑗 ∈ (0,1] is the stepsize which can be determined by some rules or set as a constant. Obtain 𝒑𝑗 as the solution, that is, 𝒑𝑗 = arg min 𝒑∈𝑃 ∇𝑠(𝒑𝑗) ∙ (𝒑 − 𝒑𝑗) (27) 𝑁+1,𝑖 at the 𝑗th iteration of the For Problem 1, 𝒑𝑗 in the above equations denotes sensor position 𝒑𝑘+𝑟 optimization, and 𝑠(𝒑) denotes the cost function of tracking performance. 𝒑 is a point of 𝑃 that lies furthest along the negative gradient direction −∇𝑠(𝒑) . Because the subproblem, (26), is to minimize the linear approximation of the problem and 𝑃 is specified by linear constraints, it is a linear program, which is relatively easy to solve. A corollary to Jacobi's formula [21] states that for any invertible matrix Ψ, 𝜕 𝜕𝑡 ln det Ψ(𝑡) = tr (Ψ−1 𝜕 𝜕𝑡 Ψ) (28) where tr(∙) denotes the trace of a matrix. Thus, the gradient of 𝑐𝑛,𝑙 = − ln det Ψ𝑛,𝑙 can be expressed as ∇𝑐𝑛,𝑙 = [ 𝜕𝑐𝑛,𝑙 𝜕𝒑𝑛 ] = − [tr (Ψ−1 𝜕 𝜕𝑥 Ψ) tr (Ψ−1 𝜕 𝜕𝑦 Ψ) tr (Ψ−1 𝜕 𝜕𝑧 Ψ)] (29) 10 of 18 The gradient of the tracking performance ∇𝑠𝑛 is then obtained as a linear combination of ∇𝑐𝑛,𝑙 as ∇𝑠𝑛 = ∑ 𝑤𝑙 𝐿 𝑙=1 ∇𝑐𝑛,𝑙 for 𝑙 ∈ 𝕋𝑛 (30) following the definition of 𝑠𝑛, given in (19). 4.2. Guidance of the Deployed Sensors Since the positions of the targets are changing, the performance of tracking the targets is changing too. An optimal position for tracking a group of targets at a time step will no longer be the best position in next time step. Thus, the fusion center needs to determine the next position of the sensors to chase the moving targets every time the target state is estimated. It is essentially an optimization problem similar to the discussion of the previous subsection. An intuitive way to generate commands to the sensors is to solve the optimization problem and take the solution as the position command as follows. Algorithm 2.1 (Guidance of sensors). For each deployed sensor 𝑛, Solve the following convex optimization problem: 𝑛 𝑛 Take the solution (𝒑𝑘+1 minimize 𝑠𝑍,𝑘+1 subject to 𝒑𝑘+1 (𝒑𝑘+1 ) 𝑛 𝑛 ∈ 𝑃𝑘+1 )∗ as the position command to the sensor 𝑛 END The algorithm given above seems straightforward. It can be solved by the conditional gradient method addressed in (25-27). However, one can imagine that it will take time for the unmanned aircraft to reach the input position. When the targets are not stationary, the optimal position for the sensor is not stationary as well. Thus, the optimal position should be calculated at every time when new measurements are available. Moreover, executing Algorithm 2.1, i.e., solving N nonlinear optimization problems, requires high computational power. An alternative to Algorithm 2.1 is proposed as follows. Algorithm 2.2 (Guidance of sensors - proposed). For each deployed sensor n, Solve the following linear optimization problem: 𝑛 minimize ∇𝑠𝑍,𝑘+1 subject to 𝒑𝑛 ∈ 𝑃𝑛 𝑛 (𝒑𝑘+1 ) ∙ (𝒑𝑛 − 𝒑𝑘+1 𝑛 ) Take the solution 𝒑𝑛 as the position command to the sensor END Note that the linear optimization problem in Algorithm 2.2 is the subproblem of the optimization problem in Algorithm 2.1, revisited as minimize 𝑠𝑍,𝑘+1 subject to 𝒑𝑘+1 (𝒑𝑘+1 ) 𝑛 ∈ 𝑃𝑘+1 𝑛 𝑛 𝑛 (31) Each iteration of the conditional gradient method is executed at every time step for Algorithm 2.2. , is supposed to move toward the furthest feasible point 𝒑𝑛. The sensor, which is currently at 𝒑𝑘+1 The conditional gradient method is still valid even if the sensor couldn't reach the point 𝒑𝑛 at the 𝑛 + next time step. Moving in the feasible direction, each sensor will be positioned at 𝒑𝑘+2 𝛼𝑘+1(𝒑𝑛 − 𝒑𝑘+1 ) where 𝛼𝑘+1 ∈ (0,1]. This way, the sensors always can be at feasible points which are the current best positions, requiring less computation because the algorithm is to solve a set of 𝑛 = 𝒑𝑘+1 𝑛 𝑛 11 of 18 linear programming. Moving along the shortest path to the optimal position, by Algorithm 2.1, would take the minimum time to arrive. However, the points on the shortest path are not guaranteed feasible, so that the sensor may fail to track some targets when moving on the way to the optimal position. The graphical illustration of the guidance algorithms is given in Fig. 2. Given a configuration of the target tracking of a sensor, the performance score can be plotted as a function of the sensor position. For simplicity, the performance score is presented as a function of x-y coordinates. At the bottom of the graph is the contour of the performance score. The solid line and dashed line denote the paths of the sensor guided by Algorithm 2.2 and Algorithm 2.1, respectively. As the dashed arrow shows, Algorithm 2.1 directs the sensor to its optimal position and the sensor is taking the shortest path to its destination. However, sensing on some points on the way may provide the higher (worse) performance and there is a possibility that the sensor is on the infeasible point, at which some of the targets are out of the sensor's coverage. On the other hand, Algorithm 2.2 directs the sensor in the feasible direction at each time step, and finally to the optimal sensor position. Figure 2. Graphical illustration of the guidance algorithms. 5. Simulation We simulated the problem of tracking ground targets by sensors equipped on multiple unmanned aircraft. There are 5 unmanned aircraft and 10 targets, i.e., 𝑁𝑚𝑎𝑥 = 5 and 𝐿 = 10. Fig. 3 shows the targets at their initial positions. The squares denote the targets and their numbers are presented at the upper right side of the squares. The squares drawn with a thinner line denote the targets that the operator attempts not to track. The maximum number of targets a sensor can track is set 𝑉 = 4 and the operator attempts not to track the targets 3, 4, 5, i.e., 𝒘 = [1, 1, 0, 0, 0, 1, 1, 1, 1, 1]. The sensors are not allowed to be below 500 𝑚, i.e., 𝐻𝑚𝑖𝑛 = 500 𝑚. It is assumed that the new deployment is available 5 seconds after the last deployment. The time for the new deployment is determined as the time when the information score is larger than 𝑆1 = 4. The initial information score 𝑙 = is 𝑐0 𝑙 = diag([1, 1, 1]) diag([1, 1, 1, 0.25, 0.25, 0.25]) and therefore the position part of the FIM becomes Ψ0 𝑙)−1. The information score will increase as the time passes if no sensor is tracking the where 𝐽0 𝑙] = target until the score meets 𝑆1 . The standard deviation of accelerometer noise is [𝜎𝑥 [0.1, 0.1, 0.1] 𝑚/𝑠2 and the measurement is assumed to be corrupted by the noise with the standard deviation of [𝜎𝐴𝑧 𝑙 = 0 for each target 𝑙 = 1, … , 𝐿 since the initial covariance matrix for target state is set as 𝑃0 𝑛] = [0.007 𝑟𝑎𝑑, 0.007 𝑟𝑎𝑑, 0.05 𝑚]. 𝑛 , 𝜎𝐸𝑙 𝑛 , 𝜎𝑅 𝑙, 𝜎𝑦 𝑙 , 𝜎𝑧 𝑙 = (𝑃0 12 of 18 Figure 3. The initial configuration of the targets. 5.1. Stationary Targets, no Guidance Firstly, let's see how the management algorithm works in the cases with stationary targets. In the first case, the guidance algorithm (Algorithm 2.2) will not be executed after the configuration of the new sensor is determined by Algorithm 1. Thus, the deployed sensors will remain stationary at their optimal position. Fig. 4 shows the time history of information about each target. Note that lower value of the information score means better tracking performance. The dashed lines denote the threshold set by the operator. If the information score meets the threshold, the fusion center deploys the new sensor. The fusion sensor predicted that the information of the target 1 will be insufficient at 𝑡 = 7.85 𝑠, and consequently the first sensor was deployed at 𝒑𝑛=1 = [−149, 358, 1105]𝑇 𝑚 to track the targets {1, 7, 8, 9} from that time on. It can be observed that the information score of the targets being tracked has decreased after the deployment. That is, the Algorithm 1 determined that the configuration provides the best target tracking performance score. The next deployment was occurred at 𝑡 = 12.85 𝑠. As a result, the final target groups 𝕋𝑓 for all sensors can be expressed as an 𝑁 × 𝐿 matrix as 𝕋𝑓 = [ 1 0 0 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 1 ] (32) where 1 in nth row and lth column in 𝕋𝑓 means that the target l is being tracked by the sensor n. The final configuration is shown in Fig. 5. 13 of 18 Figure 4. The time history of the information score of each target when the targets are stationary and the sensors are deployed at their optimal positions. Figure 5. The final locations of the targets and the deployed sensors for case 5.1. 5.2. Stationary Targets, with Guidance As the second case with the stationary targets, the sensors are deployed at the places that are not optimal for the tracking performance and they are supposed to track the targets assigned as 𝕋0 = [ 1 0 0 0 0 1 1 1 0 0 0 1 0 0 0 0 1 1 1 0 0 1 0 0 0 0 0 0 1 1 ] (33) 14 of 18 Throughout the simulation, the sensor movement is assumed as the first-order lag model whose speed is limited by 3 m/s for each axis. By executing the guidance algorithm (Algorithm 2.2), we observe the behavior of the sensors as Fig. 6. The sensors were directed to the optimal position and the performance score has the lower value as each sensor moves toward the optimal position as depicted in Fig. 7. Figure 6. The time history of the sensor positions drawn with their optimal positions when each sensor is deployed at a random position and guided by Algorithm 2.2. Figure 7. The time history of the performance score of each sensor guided by Algorithm 2.2 to track the stationary targets. 5.3. General Case 15 of 18 In this case, each target is moving independently with the speed of 0 -- 3 m/s. Initially, three sensors are assigned to track the targets as 𝕋0 = [ 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 1 1 0 ] (34) The fusion center executes Algorithm 1 (Configuration of new deployment) and Algorithm 2.2 (Guidance of sensors) as depicted in Fig. 1. The resultant information score of the targets are plotted against time as Fig. 8. It is observed that, at 𝑡 = 7.85 𝑠 , the information of the target 10 was considered insufficient and the targets {8, 9, 10} started being tracked by the newly deployed sensor 4. After the deployment, there was no target whose information score is larger than the threshold 𝑆1. The time history of performance score of each sensor is plotted in Fig. 9, compared with those in Algorithm 2.1 and with no guidance command. In the case with no guidance, the sensors are being stationary at their initial position. We observe that the performance scores of the sensors are increasing as the time passes when no guidance algorithm is executed while the sensors guided by Algorithm 2.1 or Algorithm 2.2 provide better (lower) performance scores. Note that Algorithm 2.2 gives better performance than Algorithm 2.1 which solves the optimization problem for the performance score of a sensor and takes the optimal solution as the guidance command. When the targets are moving, the optimal position to track the targets is changing and the sensors may not make it to the optimal position at next time steps. Instead, the sensors guided by Algorithm 2.2 move in the furthest feasible direction and achieve the current best performance score. This tendency is same for the most cases in the simulations except the cases with stationary targets. When the targets are stationary, Algorithm 2.1 shows the better performance since Algorithm 2.1 guides each sensors directly to the (fixed) optimal position and the sensor takes the shortest path to its destination. Furthermore, Algorithm 2.1 requires much more computation time than Algorithm 2.2. It is obvious since Algorithm 2.1 is supposed to iteratively solve its subproblem (31) which is same to Algorithm 2.2 itself. Thus, the required computation time is proportional to the number of iterations in Algorithm 2.1. In the simulation, the average computation time of executing Algorithm 2.1 was 74 times longer than that of executing Algorithm 2.2 that implies that the average number of iterations in Algorithm 2.1 was about 74. 16 of 18 Figure 8. The time history of the information score of each target when the targets are moving and the sensors are deployed at their optimal positions. Figure 9. The performance scores of the sensors resulted in being guided by Algorithm 2.2 (solid line), Algorithm 2.1 (dashed line) are plotted against time, with the results with no guidance command (dotted line). 6. Conclusion 17 of 18 This paper proposed a framework for airborne multisensor management for target tracking. The performance metric was newly defined and its convexity was proved. The convex optimization was used in determining the configuration of the new sensor deployment as well as in the guidance of the deployed sensors to improve their target tracking performance. In providing optimal position to each sensor, an intuitive way that directs the sensor to its optimal position by solving the nonlinear optimization problem did not perform well as the targets are moving. Instead, we proposed an algorithm that provides the negative gradient of the cost function by solving the subproblem of the nonlinear optimization problem. The simulation result showed that the proposed algorithm obtained better performance score requiring much less computation time. In conclusion, the proposed framework addressed in this paper is a powerful tool that allows the efficient management of the unmanned aircraft as sensors. References [1] N. Xiong and P. Svensson, "Multi-sensor management for information fusion: issues and approaches," Information Fusion, vol. 3, no. 2, pp. 163-186, 6// 2002. [2] G. W. Ng and K. H. Ng, "Sensor management -- what, why and how," Information Fusion, vol. 1, no. 2, pp. 67-75, 2000. [3] K. A. Tarabanis, P. K. Allen, and R. Y. Tsai, "A survey of sensor planning in computer vision," IEEE transactions on Robotics and Automation, vol. 11, no. 1, pp. 86-104, 1995. [4] M. Hernandez, T. Kirubarajan, and Y. Bar-Shalom, "Multisensor resource deployment using posterior Cramér-Rao bounds," IEEE Transactions on Aerospace and Electronic Systems, vol. 40, no. 2, pp. 399- 416, 2004. [5] D. E. Penny, "Sensor management in an ASW data fusion system," in Sensor Fusion: Architectures, Algorithms, and Applications III, 1999, vol. 3719, pp. 418-430: International Society for Optics and Photonics. [6] R. Tharmarasa, T. Kirubarajan, M. L. Hernandez, and A. Sinha, "PCRLB-based multisensor array management for multitarget tracking," IEEE Transactions on Aerospace and Electronic Systems, vol. 43, no. 2, pp. 539-555, 2007. [7] R. Tharmarasa, T. Kirubarajan, A. Sinha, and T. Lang, "Decentralized Sensor Selection for Large-Scale Multisensor-Multitarget Tracking," IEEE Transactions on Aerospace and Electronic Systems, vol. 47, no. 2, pp. 1307-1324, 2011. [8] P. Tichavsky, C. H. Muravchik, and A. Nehorai, "Posterior Cramér-Rao bounds for discrete-time nonlinear filtering," IEEE Transactions on signal processing, vol. 46, no. 5, pp. 1386-1396, 1998. [9] B. Ristic and M. S. Arulampalam, "Tracking a manoeuvring target using angle-only measurements: algorithms and performance," Signal processing, vol. 83, no. 6, pp. 1223-1238, 2003. [10] Y. Kim and H. Bang, "Airborne multisensor management for multitarget tracking," in Unmanned Aircraft Systems (ICUAS), 2015 International Conference on, 2015, pp. 751-756: IEEE. [11] Y. Kim, W. Jung, and H. Bang, "Real-time path planning to dispatch a mobile sensor into an operational area," Information Fusion, vol. 45, pp. 27-37, 2019. [12] J. Hong, Y. Kim, and H. Bang, "Cooperative circular pattern target tracking using navigation function," Aerospace Science and Technology, vol. 76, pp. 105-111, 2018. [13] X. R. Li and V. P. Jilkov, "Survey of maneuvering target tracking. Part I. Dynamic models," IEEE Transactions on aerospace and electronic systems, vol. 39, no. 4, pp. 1333-1364, 2003. 18 of 18 [14] D. Willner, C. Chang, and K. Dunn, "Kalman filter algorithms for a multi-sensor system," in Decision and Control including the 15th Symposium on Adaptive Processes, 1976 IEEE Conference on, 1976, vol. 15, pp. 570-574: IEEE. [15] M. Patan, Optimal observation strategies for parameter estimation of distributed systems. University Press Zielona Góra, 2004. [16] S. S. Ponda, R. M. Kolacinski, and E. Frazzoli, "Trajectory optimization for target localization using small unmanned aerial vehicles," Master's thesis, Massachusetts Institute of Technology, Cambridge, MA, 2009. [17] D. Ucinski, Optimal measurement methods for distributed parameter system identification. CRC Press, 2004. [18] D. P. Bertsekas, Nonlinear programming. Athena scientific Belmont, 1999. [19] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. [20] A. Pázman, Foundations of optimum experimental design. Springer, 1986. [21] J. R. Magnus and H. Neudecker, "Matrix differential calculus with applications in statistics and econometrics," 1995.
1705.06130	1	1705	2017-05-13T14:16:56	Stability and Performance of Coalitions of Prosumers Through Diversification in the Smart Grid	[ "cs.MA", "cs.SI", "eess.SY" ]	Achieving a successful energetic transition through a smarter and greener electricity grid is a major goal for the 21st century. It is assumed that such smart grids will be characterized by bidirectional electricity flows coupled with the use of small renewable generators and a proper efficient information system. All these bricks might enable end users to take part in the grid stability by injecting power, or by shaping their consumption against financial compensation. In this paper, we propose an algorithm that forms coalitions of agents, called prosumers, that both produce and consume. It is designed to be used by aggregators that aim at selling the aggregated surplus of production of the prosumers they control. We rely on real weather data sampled across stations of a given territory in order to simulate realistic production and consumption patterns for each prosumer. This approach enables us to capture geographical correlations among the agents while preserving the diversity due to different behaviors. As aggregators are bound to the grid operator by a contract, they seek to maximize their offer while minimizing their risk. The proposed graph based algorithm takes the underlying correlation structure of the agents into account and outputs coalitions with both high productivity and low variability. We show then that the resulting diversified coalitions are able to generate higher benefits on a constrained energy market, and are more resilient to random failures of the agents.	cs.MA	cs	Stability and Performance of Coalitions of Prosumers Through Diversification in the Smart Grid Nicolas Gensollen, Vincent Gauthier, Monique Becker and Michel Marot CNRS UMR 5157 SAMOVAR, Telecom SudParis/Institut Mines Telecom Email: {nicolas.gensollen, vincent.gauthier, michel.marot, monique.becker}@telecom-sudparis.eu 7 1 0 2 y a M 3 1 ] A M . s c [ 1 v 0 3 1 6 0 . 5 0 7 1 : v i X r a Abstract-Achieving a successful energetic transition through a smarter and greener electricity grid is a major goal for the 21st century. It is assumed that such smart grids will be characterized by bidirectional electricity flows coupled with the use of small renewable generators and a proper efficient information system. All these bricks might enable end users to take part in the grid stability by injecting power, or by shaping their consumption against financial compensation. In this paper, we propose an algorithm that forms coalitions of agents, called prosumers, that both produce and consume. It is designed to be used by aggregators that aim at selling the aggregated surplus of production of the prosumers they control. We rely on real weather data sampled across stations of a given territory in order to simulate realistic production and consumption patterns for each prosumer. This approach enables us to capture geographical correlations among the agents while preserving the diversity due to different behaviors. As aggregators are bound to the grid operator by a contract, they seek to maximize their offer while minimizing their risk. The proposed graph based algorithm takes the underlying correlation structure of the agents into account and outputs coalitions with both high productivity and low variability. We show then that the resulting diversified coalitions are able to generate higher benefits on a constrained energy market, and are more resilient to random failures of the agents. I. INTRODUCTION Designing stable power systems is a classical engineering challenge since blackouts can have catastrophic consequences. One obvious condition for stability revolves around sustaining at any time a power generation that meets the demand. If one is larger than the other, the system deviates from its synchronous state. If nothing is done to return to the synchro- nized equilibrium, this can lead to catastrophic cascades of failures [1] [2]. With traditional power plants based on fossil energies, the production can be scheduled in order to sustain the predicted consumption. Deviations from this schedule can then be supported almost on real time by using fast response power plants, interconnections with border countries, as well as regulator entities. Even if most individual consumers may not realize it, the electricity prices are not constant rates and evolve with the production/consumption conditions on the grid. It is often assumed that the granularity of these prices is meant to increase in the future as a way to pass on the production conditions on the end users [3]. The pricing of electricity and the necessity for the grid operator to have different emergency reserves lead to an economy setting for electricity, where operations and various kind of contracts are decided on a market [4]. Such a market environment clearly necessitates communication in order to monitor and obtain information from the grid as well as exchange between participants. These kind of settings are already being used at the transport level of current power grid with large renewable plants in the production portfolio. Using electricity markets appears thus as a way to manage the reliability of the whole system, especially when the use of renewables is important. The smart grid vision however goes further and revolu- tionizes this top-down centralized architecture by assuming that bidirectional power flows are allowed. This would change completely the nature of the traditional grid since production could be located down to the very end of the distribution networks. Nodes that were simply pure loads yesterday could behave tomorrow as generators or loads [5]. On the other hand, the user of renewables in the production is continuously growing, and is expected to become majority in a near future. These plants completely rely on the presence and the intensity of their respective resources (wind, sun, tiles...). Balancing production and demand in such a scenario appears much more challenging and it seems inevitable that the grid should modernize its infrastructures to sustain this transition [4]. The key to drive safely such a system is assumed to be information, and more precisely, the capacity to measure, communicate, and analyze data on real time. in cases when it In this paper, we focus on these "nodes" in the distribution network that can produce and consume electricity. More precisely, we consider a set of agents that own both renewable distributed energy resources (DER) and electrical loads. The production of an agent can be, of course, used to meet its own demand, but is over-producing, we consider that it has the possibility to inject, against financial compensation, its extra-production in the grid. Such an agent is known as a "prosumer" [6] (and will be called model accordingly in this paper). A key point of this work is to merge the interests of the grid operator with the prosumers in a single utility function. While the latter intend to maximize their benefits with higher production contracts, the operator primary concerns are related to the quality and stability of the injections. The coalitions that we wish to form should thus be both stable and productive. This compromise can be difficult to obtain in real situations, especially when renewable generation and complex spatiotemporal correlation happen. More precisely, in our model, any entity that wishes to sell its production on some energy market has to estimate and announce a production capacity for an upcoming period of time. If a contract is bought, the entity commits to injecting exactly, at any time of the contracted period, the announced amount of power, under financial penalties if it fails. Since pro- sumers use exclusively renewables, contracts come naturally with some amount of risk due to the intermittent nature of most renewables. Using storage as buffers in order to reduce risks is a popular approach. Actually, a whole branch of the smart grid literature is even studying the possibility to use electric vehicles as moving storage capacities for stabilizing the grid. Without storage and proper control, the over-producing state of the agents, and thus the power they inject in the grid, might be rather unstable. This is clearly unacceptable for the grid operator that cannot ensure system stability if it has to deal with numerous small unpredictable entities. Forming coalitions/aggregations is an envisaged solution to both the number of entities soliciting the operator, and their high variability. Indeed, it has been argued that using a multi- level aggregators architecture for the control could lead to better performance on the communication side [7], since an ag- gregator can be considered as a single node by the level above it. On the other hand, it is well-known that diversification of the assets is a way of minimizing risks when constructing a portfolio. One thus expect a more stable and predictable energy production for a coalition than for a single agent. Nevertheless, all coalitions are obviously not equally stable or productive, which means that special attention should be paid to the aggregation step. Recall that prosumers have both a consumption and a production component, each depending on location and time, meaning that there are complex underlying correlations between the agents. This is a central topic of the present paper : given N prosumers, what coalitions should be formed so that the compromise between expected production and variability is optimized ? We will see that the variability in the aggregated produc- tions can be quantified to a certain extent by the correlation among the agents forming the coalitions. Understanding the correlation relationships among the agents can thus give an indication about what coalitions to form and how much they should sell. More precisely, we use a graph representation of the correlation structure to gain insight about the expected production to risk ratios of different coalitions. We build a framework in which the system operator specifies both the minimum production acceptable to enter the market and restrain the amount of risk he is willing to take. We then propose a graph heuristic that uses decorrelated cliques in order to form diversified productive and stable coalitions, and we compare the results with other formation strategies (see section VII). Rather than maximizing the profit of each agent in the grid (prosumers and grid operator), our algorithm tries to form the most productive coalitions given a maximum amount of risk acceptable. In other words, it tries to maximize coalitions prof- its without considering individual retributions to single agents (a task often studied through game theory). Furthermore, the pricing strategies for both the prosumer and the grid is beyond the scope of this paper. Because agents are susceptible to fail for diverse reasons, the propensity of the system to undertake these failures is critical [8]. We therefore investigate in this paper the resilience of the coalitions when prosumers fail. Despite the fact that losing agents is usually detrimental to the coalitions, we will see that the coalitions formed with our algorithm tend to be less impacted by random failures. The paper is organized as follows, section II gives a brief overview of the related literature, section III clarifies how we generated realistic prosumer production traces based on weather data. In section IV, we define most of the notations and explain why correlation between prosumers is a quantity of interest for our objective. Based on the conclusions of sec- tion IV, we develop in section V a utility function quantifying how much power a coalition can announce on the market given an accepted risk level. We then develop, in section VI a greedy optimization algorithm that uses decorrelated cliques as inputs and improves their utility over a correlation-constrained environment. Finally, section VII provides some results both on performance of the method and resilience of the coalitions formed. II. RELATED WORK As explained in the section above, allowing entities using mainly renewables to inject power in the grid is a difficult challenge. Indeed, on the contrary to fossil plants whose production can be scheduled in advance to meet the expected consumption, renewables by definition only produce when the resource is present. Unfortunately, these moments do not necessarily coincide with the consumption peak hours [9]. One possible solution consists in shifting some of the loads towards high productivity periods. Demand side management techniques implemented in the end users' smart meters, can mitigate the consumption peaks and wast less production [9] [10]. Since end users tend to seek a maximum utility at a minimum cost, dynamic pricing is believed to be a good way to give incentives to them. By carefully scheduling the prices, it is assumed that the load curve can be shaped to some extend. Dynamic pricing is an interesting and useful tool, but it has also some limitations. It is likely that dynamic pricing will serve as a shaping mass tool while finer techniques will be needed locally. A popular approach in this direction consists in deploying storage devices in the network and using them as electricity buffers. Basically these storages would be charged when there is a surplus of production, and discharged when the consumption exceeds the production. Although quite simple, this idea causes numerous challenges when it comes to its real implementation. As long as the system considered remains small, a centralized control of these buffers could be envisioned. However, for real large systems it is likely that more sophisticated decentralized algorithms will be necessary. In [11], the authors introduce a distributed energy management system with a high use of renewables such that power is scheduled in a distributed fashion. In [12] the optimal storage 2 capacity problem is addressed. There is indeed an interesting trade-off between the costs of the equipments and the expected availability of power. The authors develop a framework that enables them to exhibit a Pareto front of efficient solutions. These are only a few possibilities for balancing production and demand in the smart grid. Most of the time, these techniques will be coupled with predictions of the upcoming load curves and weather conditions. Combining all these technologies enables aggregators to quantify their expected production and the inherent risk that comes with it. The optimization of expected returns to risk is a traditional goal in finance, and a wide literature exists on this topic. It is well-known for instance, that the more risk one is willing to take, the higher his potential gains. On the contrary, when investing exclusively on low risks assets, one should expect relatively small gains. This trade-off is formalized in the Markowitz' portfolio theory [13]. More precisely, given a set of assets for which we have some historic data of returns, the objective is to find a linear combination of these assets (the so-called portfolio) which maximizes the expected value while minimizing the variance of the portfolio's return. Markowitz's answer is a set of efficient portfolios that all optimize in some sense this trade-off. If one is able to put a number on his risk acceptance or on the target expected return, the corresponding efficient portfolio is a priori the best option. One of the most controversial assumptions in the portfolio theory is that returns are jointly normally distributed (or, at least, that the returns distribution is jointly elliptical). Some economist have pointed out the fact that this assumption might not capture well the reality of financial markets [14]. Nevertheless, one of the key point in the Markowitz theory is to consider explicitly the correlation between the assets since they impact directly the variances of the portfolios. Since the work of [15], an interesting approach consists in computing a distance metric based on the correlation coefficients in order to organize the series in a correlation graph. Nodes represent the series considered while the edges are weighted by the metric. Because the metric can be computed for all pairs, these graphs are complete and of little use as is. Historically, the approach used by [15] was to compute a minimum spanning tree as to obtain a hierarchical clustering of the series. Later on, it was pointed out that, by definition, a spanning tree could not capture the underlying clustering structure hidden in the correlation graph. In this paper, we use another classical filtering technique called -graph [16]. It consists in selecting a threshold , and filtering out edges with smaller weights. As we will see further in this paper, this approach has the advantage of preserving clusters of correlated series. III. GENERATING REALISTIC PROSUMER PATTERNS An essential component of the smart grid is the smart meter which makes the interface between the end user and the rest of the system. Smart meters coupled with sensors measure quantities of interest (like instantaneous consumption), receive informations from the grid (electricity prices for instance), and take actions accordingly (demand side management programs). Smart meters are currently and gradually deployed, and will 3 probably provide interesting datasets to work on. Unfortu- nately, at the time this paper was written, production and consumption data for prosumers over a large region were not yet available to our knowledge. Some interesting experiments are nonetheless being conducted and data are progressively made public [17]. In this paper, we use weather quantities like wind speed or solar radiance as alternative data for generating realistic production and consumption series. Fortunately, these kinds of data are easier to find, and since the development of small personal weather stations, their geographical granularity keeps increasing. Since these quantities depend both on time and location, we discretize time into slots and space into zones in the following. A zone is simply a portion of the considered region of study for which we sampled data. Therefore, if prosumers i and j are positioned on the same zone, they are exposed to the same weather. Adding some intra-zone noise can easily be done though not considered in this paper. More formally, we denote by Pi(t) the instantaneous extra- production of agent i at time t : Pi(t) = P P i (t) − P D i (t) (1) i (t) its consumption at Where P P time t and P D i (t) represents the total production of agent i at time t. In other words, Pi(t) represents the instantaneous surplus of power that agent i is willing to sell at time t. As explained above, since large datasets containing this quantity over time are not yet available, we simulated these traces by considering separately i and P D P P i For a prosumer i, it is possible to write both quantities as a sum over the distributed energy resources (DERi) and loads (loadi) of i : . (cid:88) (cid:88) k∈DERi k∈loadi P P i (t) = P D i (t) = Pk(t) Pk(t) (2) (3) For simplicity, in this paper we only consider wind-turbines (WT) and photovoltaic panels (PV) as possible DERs for the agents (DERi = W Ti ∪ P Vi): P D i (t) = Pk(t) + Pk(t) (4) (cid:88) k∈W Ti (cid:88) k∈P Vi (cid:88) k∈P Vi We denote by νi(t) and Ψi(t) the wind speed (in m.s−1) and the solar radiance ( in W.m−2) at the location of agent i and at time t, so that : P P i (t) = FW T (νi(t)) + FP V (Ψi(t)) (5) (cid:88) k∈W Ti Where FW T (resp. FP V ) is the power curve for the wind- turbines (resp. photovoltaic panels). We made here the implicit assumption that all wind-turbines (resp. photovoltaic panels) have the same power curve. The model can be easily extended to multiple power curves accounting for different types of generators. More details about power curves and their approx- imations can be found in the appendix B and in [18]. The predictable. Let PS(t) =(cid:80) of coalition S at time t. i∈S Pi(t) be the extra-production S S S Suppose now that coalition S has to suggest a production value P CRCT to enter the market. This means that, during the time S is on the market, it will have to inject in the grid exactly P CRCT at any time t and will be rewarded proportionally to this amount, with penalties if it deviates. Obviously, the actual extra-production will not be constant at this value and will oscillate due to intermittences in the production and consumption. If S always produces more than , it will never have to pay penalties, but it is losing P CRCT some gains since it could have announced a higher contract value. If the production oscillates around P CRCT , by using batteries or demand side management techniques (see section II), S could be able to maintain its production to the contract value at any time. Nevertheless, if the oscillations are too important compared to the available storage capacity, S will probably break the contract and pay penalties. We can see that there is a return over risk trade-off here, meaning that coalitions should find the right balance between announcing too low and losing some potential gains, and claiming too high and paying penalties. S Let us illustrate the rest of the notations and concepts with a simple example. We consider only two agents i and j such that the distribution of their extra-production can be approximated by normal distributions : Pi ∼ N (µi, σi) and Pj ∼ N (µj, σj). This is only for explanation purposes as it is of course rather unrealistic in real situations where the distributions are skewed. Using simple statistics, we can write the distribution of the coalition S = {i, j} as P{i,j} ∼ N (µij, σij), where : (cid:40) µij = µi + µj (cid:113) σij = σ2 i + σ2 j + ρijσiσj (6) S S ρij being the Pearson's correlation coefficient between Pi and Pj. If the coalition {i, j} proposes a contract value P CRCT , all instants when {i, j} will produce less than P CRCT is critical. Indeed, in this kind of situations, {i, j} will either have to discharge batteries to keep up with its contract, or pay penalties to the grid. The probability that {i, j} is under- producing compared to the contract : P r[Pi,j ≤ P CRCT ] is thus an important indicator of the coalition's quality. A well- known result for normal distributions is that the cumulative distribution function can be written as : P r[Pij ≤ P CRCT (cid:33)(cid:35) 1 + erf (cid:32) (cid:34) (7) ] = S 1 2 where erf is the error function : erf (x) = dt. The contract a given coalition is willing to take depends on its capacity to compensate for under-producing (using batteries, backup generators...), and its risk acceptance. Se- lecting the right contract value appears thus as an interesting problem on its own that we plan to investigate in future works. In order to keep the present paper in a reasonable length, we simplify the contract value selection problem by giving some responsibilities to a third party named the grid operator. The role of the grid operator is to constrain the S − µij P CRCT √ σij 2 2√ π (cid:82) x 0 e−t2 Figure 1: Process diagram process for generating the Pi series is pictured in the first block of the process diagram (see figure 1). Note that a prosumer i is defined by his zone Zi as well as the sets DERi and loadi. That is, a prosumer can be configured to represent anything from a single wind-turbine for instance (DERi = {W T0} and loadi = ∅) to a pure load (DERi = ∅ and loadi = {L0}) through more complex combinations. In practice, we use random configurations for the agents. In the rest of the paper, we use french weather data [19] starting in January 2006 and ending in December 2012, with a sampling frequency of three hours, and generate N timeseries of extra-production over this date range. IV. NOTATIONS This section provides most of the notations and introduces important concepts for the rest of the paper. As explained in section III, we consider a set A = {a1, a2, ..., aN} of N prosumers configured randomly, and for each agent, we simulate its extra-production Pi(t), ∀i ∈ A from 2006 to 2012. Based on these historical values, our objective is now to form groups of prosumers (the so-called coalitions) so that the global power production resulting from the superposition of individual's extra-productions be both sufficiently high and 4 Time Serie AnalysisSimulationCoalition Formation1. Weather data resa-mpling2. Removal of season-al effect1. Space Dicretization2. Parametrization3. Production/con-sumption simulation4. Production sam-1. Generate correla-tion graph2. Clique percolation3. forming non over-lapping coalition123(cid:26) P r[Pij ≤ P CRCT S ≥ P M IN P CRCT S ] ≤ φ (9) On figure 2, P M IN is fixed to 2 units for illustration purpose. For φ = 0.1, only blue triangles and cyan circles coalitions are valid while red diamonds coalition is not. The Gaussian assumption of this small example is conve- nient as it allows us to write P CRCT (cid:63) analytically. Never- theless, such assumption is rather unrealistic in practice. In the following, we keep the same framework but release this Gaussian assumption unless the contrary is specified (see eq. 14). This assumption will indeed be convenient for computing some parameter estimates. S V. UTILITY FUNCTION In this section, we use the notions of contract values and valid coalitions developed in section IV in order to design a proper utility function. The contract basically indicates the rate at which a coalition has to inject power in the grid. It seems then natural that coalitions are remunerated proportionally to their contract values gain(S) ∝ P CRCT (cid:63) . More precisely, if λ is the unitary price rate for electricity, a coalition S injecting P CRCT (cid:63) in the grid during a period [t0, tk] earns : S S (cid:90) tk (cid:90) tk gain(S) = λP CRCT (cid:63) dt = P CRCT (cid:63) S S λdt (10) S t0 t0 is supposed to be a constant rate over the (since P CRCT (cid:63) contracted period). Using gain(S) directly as a utility function suffers a major drawback. It is indeed not a concave function of the coalitions' sizes, meaning that coalitions can grow as large as the number of agents allows it, without any counterbalance effects. Such a model, that virtually allows infinitely large coalitions and contract values, is in practice not realistic. There are indeed costs (communication costs for instance) that increase with the coalitions sizes. We take this observation into account by rescaling the utility of a coalition S by its size in term of number of agents (S): U(S) = S P CRCT (cid:63) P M AX , if S is valid, Sα 0, if S is not valid (11) where parameter α controls to what extent the size of a coalition impacts its utility, and P M AX is a normalizing factor. P M AX can be seen as the maximum production which can be injected in the grid. Based on U, the marginal contribution of an agent i can be expressed as δS(i) = U(S + {i}) − U(S). A coalition S has thus an interest in adding an additional agent i if this marginal contribution is positive :  1 (cid:18)S + 1 (cid:19)α S (12) δS(i) ≥ 0 ⇔ P CRCT (cid:63) S+{i} ≥ P CRCT (cid:63) S If α is set to zero, agents are added as long as they increase the contract value of the coalition. If α is greater 5 S S Figure 2: P CRCT (cid:63) depending on reliability parameter φ for Gaussian distributions (see equation 8). Blue curve with tri- angles stands for a coalition S with an expected production of 5 units and a standard deviation of 0.5. Under a grid policy of φ = 0.1, it is able to announce a contract value = 4.36. The same coalition in term of expected of P CRCT (cid:63) production (µ = 5), but with a higher variance (σ = 1.5, cyan curve with circles) can only afford a smaller contract = 3.07. The red curve with diamonds value of P CRCT (cid:63) stands for a coalition with a higher expected production (µ = 7), but with a very high unpredictability (σ = 5). For low values of φ, this coalition is thus heavily penalized and can only afford a contract of 0.59 units. Under grid policy (φ = 0.1, P M IN = 2), this last coalition is thus not allowed to enter the market (red dot below the horizontal dashed line). S market entry to coalitions able to propose both sufficiently high and sufficiently credible contract values. More formally, let φ ∈ [0, 1] be the reliability threshold fixed by the grid operator as a maximum value for the probability of under-producing. The highest contract value that a coalition can propose is thus ] = φ. In the Gaussian P CRCT (cid:63) example, it implies that coalition {i, j} is announcing : such that P r[Pij ≤ P CRCT (cid:63) S S P CRCT (cid:63) S = µij − √ 2σijerf−1(1 − 2φ) (8) S This is the best contract value that the coalition S can afford giving the stability policy φ of the grid operator. Figure 2 shows how P CRCT (cid:63) evolves according to the reliability parameter φ. For illustration, the range of φ values is shown from 0 to 1, but in practice, only small values of φ really make sense : φ = 1 for instance means that coalitions can announce absolutely anything since the probability of producing less than any contract value is necessarily less than one by trivial definition of a probability. As visible on figure 2, coalitions with high expected productions but presenting a high unpredictability are penalized and can only afford small contracts. In order not to overload the market with unrealistically small coalitions, the grid operator also specifies a lower bound P M IN on the contract values. We thus characterized a valid coalition as one satisfying the two conditions : 0.00.20.40.60.81.0φ505101520PCRCTSPCRCTS=4.36PCRCTS=3.07PCRCTS=0.59φGRID=0.1PMIN=2PCRCTS=f(φ) for gaussian assumptionµS=5,σS=0.5µS=5,σS=1.5µS=7,σS=5VI. COALITION FORMATION Section IV explained how contract values for the coalitions are computed, and in section V we related this quantity to the utility and gains of a coalition. Since the computation time of this quantity is not negligible, we derive in this section the heuristic we used to form the coalitions structure. A. Representing the correlation structure As seen in section IV, the variance of the aggregated pro- duction impacts directly the contract values, and depends on the covariances between the agents productions. We argue here that, by having some representation of the correlation structure between the agents, the search landscape for high utility coalitions could be reduced, such that good coalitions are more likely to be found quickly. Usually, this correlation structure is formalized with a covariance matrix or a correlation matrix that contains all the correlation coefficients between the agents : M = (ρij)∀i,j∈A2. By using a metric to map this matrix in a weighted adjacency matrix (see section II), it is possible to obtain a graph representation of the correlation relationships between the agents. In the following, we use two opposite distance metrics for this mapping : (cid:26) d1 ij = 1 − ρ2 ij, d2 ij = ρ2 ij = 1 − d1 ij (15) ij in G2. Clearly, d1 (resp. d2) maps two correlated series as close points (resp. distant) while two uncorrelated series are distant (resp. close). These metrics enable us to compute a corre- lation graph G1 = (A, E1) and a "de-correlation" graph G2 = (A, E2). For any i and j, the weight of the edge eij is d1 ij in G1 and d2 In both cases, we want to keep only the edges which weights are located in the lower tail of the distance distributions. In other words, we want to compute the -graphs of G1 and G2 such that only meaningful edges remain. Selecting the right filter is thus an important point since it affects the landscape search for the coalition formation. Unfortunately, there seems to be no clear consensus in the literature on how to select such a threshold. We will see later in this section that cliques in G2 are potential seeds for the coalitions. Since we want to generate NCOAL coalitions, we need at least NCOAL cliques of a given size to start. Besides, since we consider coalitions as disjoint, the starting cliques should be non overlapping. We thus select our optimal threshold for G2 as : (16) 2) ≥ NCOAL} (cid:63) = min∈[0,1] { s.t. Θk(G 2 is the de-correlation graph G2 filtered by , and where G Θk(G) is the set of non overlapping cliques of size k in a given graph G. In other words we select (cid:63) as the smallest threshold possible such that the filtered de-correlation graph contains at least NCOAL non overlapping cliques of size k. The existence of (cid:63) as defined in equation 16 is not guaranteed. The users has indeed to provide consistent values of NCOAL or k compared to the size of the agent population A. Figure 3: Gaussian mean approximation. Subplot a shows how the parameter α of the utility function should be chosen in function of the mean desired size of the coalitions (see equation 14). Subplot b displays the corresponding utility functions for different values of α. Blue curve with diamonds favors very small coalitions of 2 agents while the green one with triangles favors 5 agents coalitions. Finally, the red curve with squares has an optimal size of 15 agents. than zero, additional agents have to increase the contract value by some factor. The utility function with α is not necessarily convenient, here we relate α to the mean sizes of the coalitions ¯N. otherwise U(S) tends to form coalitions of size approximately in the order of ¯N, then : (cid:21) (cid:20) ∂U ∂S = 0 S= ¯N (13) In order to get an estimator for α, we solve equation 13 in a Gaussian case (as in section IV). Furthermore, since considering all the possible interactions between agents is analytically intractable, we use here a mean approximation. Any quantity x that varies over the agent set is thus simplified in its mean value ¯x. Solving equation 13 for α in these conditions leads to: ¯µ(cid:112) ¯N (¯ρ ¯N − ¯ρ + 1) + 1.4¯σerf−1(2φ − 1)(¯ρ ¯N − ¯ρ + 1) 0.7¯σ(¯ρ − 1)erf−1(2φ − 1) α(cid:63) ¯N = (14) Figure 3 shows how α(cid:63) and the utility function evolves according to the mean size of the coalitions ¯N. These curves are only valid in the simplified Gaussian example considered they will provide some guidance when here. Nonetheless, using real data. As can be pointed out, the purpose of U is not to study coalitions stability against player defection which could be done through game theory, nor to redistribute the coalition's utility in terms of individual payoffs. But we aim to design U as a measure of how good a given coalition is according to our criteria. In other terms does a given coalitions has a good production to risk ratio. 6 51015202530¯N0.000.010.020.030.040.050.06α¯Na.α¯N=f(¯N)051015202530¯N0.600.620.640.660.680.70Uαb.Uα=f(¯N)Uα2Uα5Uα15Uα15appealing. Nevertheless, the quality of the results seems to decrease as the sizes of the cliques increase. Indeed, the larger the desired cliques, the more dense G(cid:63) 2 becomes (see equation 16). There is a point where cliques results more from noisy edges than true de-correlation, which decreases the quality of the results. Directly mapping cliques to coalitions by this de-correlation oriented approach is thus not sufficient. It is indeed possible that adding agents to these cliques has the combined effect of increasing the expected production while decreasing its stability. The question revolves around measuring the benefits of this production surplus compared to the disadvantage of having coalition with high volatility. This can be quantified by the marginal benefit in equation 12. C. Algorithm The algorithm takes inputs from : • The agents : historical series of available productions Pi, • The grid operator : market entrance policy (P M IN , φ), • The "user" : Number of desired coalitions NCOAL and size of starting cliques k. The first steps consists in computing the de-correlation threshold (cid:63). Cliques graph G2 as well as the optimal of size k in G(cid:63) are considered as coalition seeds. The 2 next step is a local greedy improvement over the landscape represented by G(cid:63) 2 . Cliques add alternatively the node i(cid:63) in their neighborhood that yields the best marginal benefit M AXi∈N (clique)δclique(i) where N (clique) is the neighbor- hood of a given clique. This addition occurs only if i(cid:63) is not already involved in another coalition, and if δclique(i(cid:63)) ≥ 0, meaning that utilities are increasing. The algorithm stops when all nodes are distributed in a coalition or when the global utility stops increasing. See the details in algorithm 1 in the appendix. VII. RESULTS The algorithm presented in the previous section is supposed to generate a given number of coalitions that have good utilities. As it comprises mainly of a greedy optimization based on local improvements, there is no guarantee that the algorithm finds the global optimum. Since there is, to our knowledge, no state of the art algorithm that aggregates uncorrelated agents in an optimum way (see section II for related problems), we compare the results with : • Random sampling of coalitions : Coalitions are formed randomly without any other constraint that the desired size. This enables us to have an idea about the distribu- tions of utility values for coalitions of a given size. • Random sampling of coalition structures : Coalition structures are sampled randomly by shuffling and random divisions of the agents. Algorithm 2 uses such a sampling and returns the highest utility coalition structure sampled. This algorithm will be refered to as "random" in the following. • Correlated : This is the complete opposite of our al- gorithm. It uses the correlation graph G1 and performs a community detection. The resulting coalitions have Figure 4: Histograms of utility values for coalitions of size 3. Red bars stands for cliques in the decorrelation graph, and blue bars represents all the other possible triplets. As visible, cliques tend to exhibit higher utilities than randomly selected coalitions. B. Cliques In [16] the structural roles of weak and strong links on finan- cial correlation graphs is investigated. The author shows that strong links, accounting for strong correlation relationships, are responsible for the clustering, while weak links provide the connectivity between clusters. Indeed, if we consider three items, say a, b, and c such that a and b are strongly correlated and b and c are also strongly correlated, then it is likely that a and c are also strongly correlated. It can be easily shown using the cosine addition formula1, that if ρab > x and ρbc > x with x > 0, then ρac > 2x2 − 1). Correlation graphs capture this weak transitivity notion through clusters of correlated series. Nevertheless de-correlation seems like a more complex concept than correlation in the sense that there is not even a partial notion of transitivity when it comes to it. Therefore, the clustering coefficients of G 1 is much higher than the one of G 2. This can be seen as another formulation of [16] on the structural roles of weak and strong links on financial corre- lation graphs. Strong links, accounting for strong correlation relationships, are responsible for the clustering, while weak links provide the connectivity between clusters. Searching for 2 and hoping that this strategy will provide a nice clusters in G coalition structure of internally uncorrelated coalitions seems thus pointless. Consider now a clique in G 2, which is a complete subgraph 2. This is indeed a structure of interest for our purpose. of G Since there is a link for every pairs of nodes, we know, by construction, that a clique has a mean correlation and a maximum correlation less than . Figure 4 shows the distributions of the utility values for cliques of size 3 (triangles) in G(cid:63) 2 and for all the other possible triplets of agents. It is clearly visible that cliques tend to exhibit higher utilities because of their de-correlation prop- erty. Choosing cliques in G(cid:63) 2 as coalitions seems therefore 1 cos(a + b) = cos(a)cos(b) − sin(a)sin(b) 7 0.100.150.200.250.300.35Utility0246810121416FrequencyUtilities of 3 agents coalitions3-cliquesother tripletsFigure 5: Utility of random coalitions depending on their size. Blue dots show real mean utility values and the thick red curve its smoothed version by applying a Savitzky-Golay filter. On this plot the α parameter of the utility function was selected according to equation 14 in order to favor 40 agents coalitions. Figure 6: Evolution of the global utility U(CS) (blue diamond curve, left axis) and the number of agents involved in the coalitions (red circle curve, right axis) during the greedy optimization of algorithm 1 thus very high internal correlations. We thus expect this algorithm to perform very bad compared to the others. See algorithm 3. Before running the algorithms, we need to calibrate the utility function by choosing the value of the α parameter. Recall that the purpose of this parameter is to take into account some constraints on the coalition's sizes if needed. In this paper, neither the communication network nor the electrical grid are explicitly considered. Thus, we do not have any technical constraints on coalition sizes even if we designed the utility such that these could be taken into account. We select the desired size as being (cid:98)N/NCOAL(cid:99) (where (cid:98).(cid:99) means floor). Figure 5 shows how the mean utility of a coalition evolves with its size when the optimum size is set to 40 agents. Using equation 14 to estimate α based on the mean quantities and Gaussian approximations seems to give acceptable results for the utility function behavior. Figure 6 displays the evolution of the global utility and the number of involved agents during the course of the greedy algorithm 1. The transition from an invalid to a valid coalitions is clearly visible on the blue diamond curve and occurs between iteration 10 and 15. After this transition, coalition's utilities improve slowly up to a maximum point. Figure 7 shows the coalitions formed with the considered algorithms in the contract value / volatility space. The color map in the background indicates regions where we expect high utilities (red) and the ones where we expect very poor utility values (blue). The bottom right corner, with high contract values and low volatilities, is therefore the region where we wish to form our coalitions. A single coalition is represented by a marker and the color and shape of a marker indicates by which algorithm the coalition has been formed. Besides, the sizes of the coalitions are indicated on the markers, and the marker size is also proportional to the coalition size. We can see that the utility function results in approximatively balanced coalitions. Small yellow markers indicates the gravity 8 centers of their respective coalition structures. The coalitions of correlated agents (green squares) are clearly of poor quality according to our criteria since they can only afford small production contracts, and with a very high volatility. On figure 7, the decorrelated coalitions (blue dots) are closer to the bottom right corner indicating a much better quality in term of productivity over volatility ratio. The black dotted line indicates the mean values for the random coalitions sampling technique. Each small dot stands for the mean position of all sampled coalitions of this given size. Variances are not indicated for readability, but are usually quite large since this sampling only takes the size as a constraint. We can see that as coalitions get larger, they tend to increase on average their contract values, but at the price of a higher volatility. On figure 7, the results of the random coalition structure sampling are shown with the red ellipses that represent the distribution of the gravity centers of the sampled structures. Since the center of the ellipses stands for the mean and each ellipse adds one standard deviation, more than 99% of the sampled gravity centers are within the largest ellipse. The small yellow dot below the ellipses indicates the gravity center of our solution. It is thus visible that our greedy graph based algorithm is able to find a quite good coalition structure in terms of volatility and contract values. A key point for the coalitions, besides stability and pro- ductivity, is their resilience. The resilience of a system can be roughly described as its ability to perform its tasks when subject to failures of its components. Therefore, the notion of resilience we will use in the following can be seen as the ability of the coalition structures to inject stable power in the grid when node failures occur. According to our model, the grid operator specified two thresholds (P M IN and φ) such that the power injected by every coalition is constrained : ]. As long as a coalition can P CRCT propose a contract value higher than P M IN , it is valid and allowed to enter the energy market. We define the resilience of a coalition S as the probability that S produces more than ∈ [P M IN , P CRCT (cid:63) S S 20406080100Number of agents in the coalition0.2400.2420.2440.2460.2480.2500.252UtilityUtility of a coalition depending on its sizereal meansmoothed mean05101520253035Number of iterations0.00.10.20.30.40.5U(CS)020406080100120140160180Number of agents in CSFigure 7: Coalitions formed in the (contract value, volatility) space. The color map indicates qualities of portions of the plane. The closer to red the better (high contract values with small volatility). On the opposite, blue areas show poor quality (small contract values with high volatility). Blue dots stand for the decorrelated coalitions that we formed while green squares show correlated coalitions. The smaller yellow markers stand for the gravity centers of the coalition structures. The black dotted line shows how contract values and volatility evolve when the size of the coalitions increases (a subset of the points are labeled by the size of the coalition they represent). Each point is the average over 105 unconstrained draws of a random coalition. As we can also draw random coalition structures, we show the distribution of their gravity centers by the red ellipses (center is the mean, and each ellipse corresponds to one standard deviation). For instance in our algorithm the utility function favors balanced coalitions. We also loosely constrained the sizes of the coalitions in the utility function. the P M IN threshold : RS = P r[PS >= P M IN ] = 1 − P r[PS < P M IN ] And we extend this measure to the coalition structures : (17) (cid:0)1 − P r[PS < P M IN ](cid:1) (18) (cid:89) S∈CS RCS = We consider that prosumers fail randomly, and we denote by ψ ∈ [0, 1] the fraction of agents that failed. Figure 8 exhibits how the resilience of the coalition structures evolves according to ψ. On the top subplot, P M IN was voluntarily selected relatively low such that the resiliences of the three structures fit on the same figure. When the P M IN requirement increases, the differences between the algorithms also increase as visible on the bottom subplot of figure 8. The decorrelated coalitions seem to achieve a more resilient production on the market in the sense that they are able to sustain a higher fraction of node failures. VIII. CONCLUSION In this paper we studied how aggregations of prosumers could be authorized to sell their surplus of production to the 9 Figure 8: Resilience of the coalition structures when nodes fail randomly (see equation 18) for P M IN = 10M W (top subplot) and P M IN = 80M W (bottom subplot) grid operator. By relying on the past values of the agents, we constrained the market entry to both sufficiently productive and stable coalitions. The power that a coalition is able to propose on the market is therefore related to production and stability. As the correlations between the prosumers that form these coalitions impact directly their volatilities, we seek uncorrelated aggregations of agents. We used a graph representation of the correlation relationships between the agents as a reduced landscape for the coalition formation. A greedy algorithm that starts with cliques of the "de-correlation" graph of the agents and makes local improvements offers a good compromise between speed and quality of the results. We compare these results with random samplings, and an opposite strategy that clusters correlated agents together. We showed that the coalitions resulting from our algorithm are able to provide more power to the grid with a lower volatility. Because they tend to have globally a better production over volatility ratios, these coalitions will tend to use less storage and waste less energy than more unstable coalitions. We plan to study these benefits for the control of the aggregations in future works. Because in real situations, agents are prone to failure, resilience is also an important criterion for the quality of the aggregations. We therefore studied how the coalitions are able to remain on the market when their agents fail randomly one by one. We showed that, in this situation, the coalitions resulting from our algorithm better withstand losses of agents. REFERENCES [1] C. D. Brummitt, R. M. D'Souza, and E. a. Leicht, "Suppressing cas- cades of load in interdependent networks." Proceedings of the National Academy of Sciences of the United States of America, vol. 109, no. 12, pp. E680–9, Mar. 2012. [2] J.-W. Wang and L.-L. Rong, "Cascade-based attack vulnerability on the US power grid," Safety Science, vol. 47, no. 10, pp. 1332–1336, Dec. 2009. [3] T. Jiang, Y. Cao, L. Yu, and Z. Wang, "Load shaping strategy based on energy storage and dynamic pricing in smart grid," Smart Grid, IEEE Transactions on, vol. 5, no. 6, pp. 2868–2876, Nov 2014. [4] F. Van Hulle, "Integrating Wind," 2009. 0.60.81.01.21.4PCRCTS1e81234567Volatility1e73031282728261520253035342726342722Coalitions in the (PCRCTS,σPS) spacedecorrelatedcorrelated0.00.20.40.60.81.0ψ0.00.20.40.60.81.0RCSRCS=f(ψ)decorrelatedrandomcorrelated0.00.10.20.30.40.5ψ0.00.20.40.60.81.0RCSdecorrelatedrandomcorrelated[5] S. D. Ramchurn, P. Vytelingum, A. Rogers, and N. R. Jennings, "Putting the 'smarts' into the smart grid: A grand challenge for artificial intelligence," Commun. ACM, vol. 55, no. 4, pp. 86–97, Apr. 2012. [6] A. J. D. Rathnayaka et al., "Prosumer management in socio-technical smart grid," in Proceedings of the CUBE International Information Technology Conference, 2012, pp. 483–489. [7] E. Negeri and N. Baken, "Architecting the Smart Grid As a Holarchy," Proceedings of the 1st International Conference on Smart Grids and Green IT Systems, pp. 73–78, 2012. [8] S. Pahwa, C. Scoglio, S. Das, and N. Schulz, "Load-shedding strate- gies for preventing cascading failures in power grid," Electric Power Components and Systems, vol. 41, no. 9, pp. 879–895, 2013. [9] M. Milligan and B. Kirby, "Utilizing Load Response for Wind and Solar Integration and Power System Reliability," in Proc. of WindPower, 2010. [10] T. Logenthiran, D. Srinivasan, and T. Z. Shun, "Demand side manage- ment in smart grid using heuristic optimization," IEEE Transactions on Smart Grid, vol. 3, no. 3, pp. 1244–1252, 2012. [11] Y. Zhang, N. Gatsis, and G. B. Giannakis, "Robust energy management for microgrids with high-penetration renewables," IEEE Transactions on Sustainable Energy, vol. 4, no. 4, pp. 944–953, 2013. [12] M. Shadmand and R. Balog, "Multi-objective optimization and design of photovoltaic-wind hybrid system for community smart dc microgrid," Smart Grid, IEEE Transactions on, vol. 5, no. 5, pp. 2635–2643, Sept 2014. [13] H. M. Markowitz, Portfolio Selection: Efficient Diversification of Invest- ments. Yale University Press, 1959. [14] R. Chicheportiche and J.-P. Bouchaud, "The joint distribution of stock returns is not elliptical," International Journal of Theoretical and Ap- plied Finance, vol. 15, no. 03, p. 1250019, 2012. [15] R. Mantegna, "Hierarchical structure in financial markets," The Euro- pean Physical Journal B, vol. 11, no. 1, pp. 193–197, 1999. [16] A. Garas et al., "The structural role of weak and strong links in a financial market network," The European Physical Journal B, vol. 63, no. 2, pp. 265–271, 2008. [17] "Irish Social Science Data Archive." [Online]. Available: http: //www.ucd.ie/issda/data/commissionforenergyregulationcer/ [18] M. Lydia, S. S. Kumar, a. I. Selvakumar, and G. E. Prem Kumar, "A comprehensive review on wind turbine power curve modeling techniques," Renewable and Sustainable Energy Reviews, vol. 30, pp. 452–460, 2014. http://www.ncdc.noaa.gov/ [19] "Infoclimat." [Online]. Available: http://www.infoclimat.fr [20] "National Centers for Environmental Information." [Online]. Available: [21] C. Piedallu and J.-C. GÃ c(cid:13)gout, "Multiscale computation of solar radiation for predictive vegetation modelling," Annals of Forest Science, vol. 64, no. 8, pp. 899–909, 2007. [22] --, "Efficient assessment of topographic solar radiation to improve plant distribution models," Agricultural and Forest Meteorology, vol. 148, no. 11, pp. 1696 – 1706, 2008. APPENDIX A. Algorithms B. Net production series Data were collected from [19] (similar data can be found at [20] for the United States). The variables used in the simulation are : • Average wind speed (in m.s−1) • Nebulosity (integer in [0, 8]) • Temperature (in degree Celsius) 1) Wind power curve: Power curves are functions that, for a given type of generator, map some input quantity to the output power produced. For wind-turbines and solar arrays these functions are well studied and approximations have been proposed [18] [21], [22]. For the wind turbines, the power curve can be specified by 4 values : • Cut-in-speed : The wind speed at which the turbine first starts to rotate and generates power. • Rated-output-power : The maximum power that the tur- bine can generate. 10 Data: Pi series, Grid policy (P M IN , φ), Desired number of coalitions NCOAL, size of starting cliques k Result: CS = {S1, ..., SNCOAL} Compute G(cid:63) Find the NCOAL cliques in G(cid:63) 2 ; while U(CS) is improving do 2 ; for each clique do Find i(cid:63) ; if δclique(i(cid:63)) ≥ 0 then clique ← clique ∪ {i(cid:63)} ; end if ∃j ∈ clique, s.t δclique(j) < 0 then clique ← clique − {j} ; end end end Algorithm 1: Local greedy optimization algorithm Data: Agent set A, Desired number of coalitions NCOAL, Maximum number of iterations Loopmax Result: CS = {S1, ..., SNCOAL} N loop ← 0 ; CS(cid:63) ← ∅; while N loop < Loopmax do CS ← SelectRandomCS(); if U(CS) > U(CS(cid:63)) then CS(cid:63) ← CS; end N loop ← N loop + 1; end return CS(cid:63) Algorithm 2: Random algorithm Data: Pi series, Desired number of coalitions NCOAL, search step size β << 1 Result: CS = {S1, ..., SNCOAL} ← 1 ; CS ← ∅; while CS < NCOAL do 1 ; Compute G CS ← computeClusters(G 1); if CS = NCOAL then return CS; end else end end ← − β; Algorithm 3: Correlated algorithm • Rated-output-speed : The wind speed at which the turbine attains its rated output power. • Cut-out-speed : The speed at which the turbine is turned off as not to damage the rotor. The most interesting part is the increase of output power when the wind speed is in the cut-in-speed rated-output-speed range. Even if sometimes a simple linear model is used, the increase has been shown to be non linear and some more complex exponential fit can be found in the literature [18]. 2) Solar power curve: The input quantity desired for our power curve model for solar arrays is a radiance in W.m−2, which can be difficult to find in weather station available data. As we mainly collected nebulosity series, we used the Helios model described in [21], [22]. This model enabled us to compute perfect (clear blue sky situation) solar radiances at some specific locations on earth and at given timestamps. As nebulosity is a measure of the sky cloudiness, we can use the nebulosity series as degradation factors on the clear blue sky model (see [21], [22] for more details) :  Ψreal(t) = Ψperf ect(t)η(t) (cid:19)3.4 (cid:18) N (t) η(t) = 1 − 0.75 8 i where F heat (τ (t), t) is the power curve that maps the tem- perature to a heating consumption, and F elec (t) computes the consumption of agent i (other than heating) at a given hour of the day. In the simulation, all agents have a desired inside temperature Ti, supposed to be a constant for simplification. By using thermodynamic laws F heat (τ (t), t) can be approxi- mated by : i i F heat i (τ (t), t) = [Ti − τ (t)] Bi Ri (22) where Bi is the surface of thermal exchanges for agent i and Ri is their thermal resistance. We denote by Ωi the maximum consumption possible for agent i, which is basically the sum of all its appliances powers. We also denote by ωi(t) = {ωi(t0), ..., ωi(t24)} the vector of the average fraction of Ωi used for each hour. We can therefore write : F elec i (t) = Ωi(ωi(t) + ) (23) (19) where is a noise term. The vector ωi(t) enables us to easily differentiate agent consumption behaviors. Business or residential areas for instance can be easily distinguished with this kind of model. where Ψperf ect(t) and Ψreal(t) are respectively the clear blue sky and real radiances at time t, η(t) is the degradation factor at time t, and N (t) is the nebulosity index at time t. Once we have input data in the forms of radiances, we compute the production of a solar array with the following simplified power curve : FP V (Ψreal(t)) = SP V Ψreal(t)eP V (20) where SP V is the surface of the array, and eP V is its efficiency. The very simple form of this power curve is due to some simplifications in order not to overload the model. For instance, it does not take into account angles and orienta- tions degradations. These could be incorporated if needed by changing the power curve in the simulations. 3) Consumption: Modeling electric consumption has al- ready been widely tackled in the literature. Models can be basically divided into two main categories : Top-down and bottom-up approaches. Top-down techniques take aggregated consumption data as inputs and try to estimate individual consumption patterns while bottom-up methods use a fine modeling of users consumptions as to obtain realistic aggre- gated consumption curves. In this paper, we used a bottom- up model since the end user, or relatively small aggregations of end users, are in our interest. The main objective was to capture both daily patterns and seasonal variations of the consumptions. We assumed an additive model where the consumption of an agent is the sum of a seasonal heating term that depends on the outside temperature and an electronic consumption term that only depends on the hour of the day. By denoting τ (t) the outside temperature at timestamp t, we can express the consumption P D i (t) of agent i at time t : i (t) = F heat P D i (τ (t), t) + F elec i (t) (21) 11
1802.00435	2	1802	2019-08-20T03:19:24	Evolutionary model discovery of causal factors behind the socio-agricultural behavior of the ancestral Pueblo	[ "cs.MA" ]	Agent-based modeling of artificial societies offers a platform to test human-interpretable, causal explanations of human behavior that generate society-scale phenomena. However, parameter calibration is insufficient to conduct an adequate data-driven exploration of the importance of causal factors that constitute agent rules, resulting in models with limited causal accuracy and robustness. We introduce evolutionary model discovery, a framework that combines genetic programming and random forest regression to evaluate the importance of a set of causal factors hypothesized to affect the individual's decision-making process. We investigated the farm plot seeking behavior of the ancestral Pueblo of the Long House Valley simulated in the Artificial Anasazi model our proposed framework. We evaluated the importance of causal factors not considered in the original model that we hypothesized to have affected the decision-making process. Contrary to the original model, where closeness was the sole factor driving farm plot selection, selection of higher quality land and desire for social presence are shown to be more important. In fact, model performance is improved when agents select farm plots further away from their failed farm plot. Farm selection strategies designed using these insights into the socio-agricultural behavior of the ancestral Pueblo significantly improved the model's accuracy and robustness.	cs.MA	cs	Evolutionary model discovery of causal factors behind the socio-agricultural behavior of the ancestral Pueblo Chathika Gunaratne1, Ivan Garibay1, Nguyen Dang2 1 Complex Adaptive Systems Lab, Department of Industrial Engineering and Management Systems, University of Central Florida, Orlando, Florida, USA 2 AI group, School of Computer Science, University of St Andrews, St Andrews, UK * Corressponding authors Email: [email protected] (CG) Email: [email protected] (IG) Abstract Agent-based modeling of artificial societies offers a platform to test human-interpretable, causal explanations of human behavior that generate society-scale phenomena. However, parameter calibration is insufficient to conduct an adequate data-driven exploration of the importance of causal factors that constitute agent rules, resulting in models with limited causal accuracy and robustness. We introduce evolutionary model discovery, a framework that combines genetic programming and random forest regression to evaluate the importance of a set of causal factors hypothesized to affect the individual's decision-making process. We investigated the farm plot seeking behavior of the ancestral Pueblo of the Long House Valley simulated in the Artificial Anasazi model our proposed framework. We evaluated the importance of causal factors not considered in the original model that we hypothesized to have affected the decision-making process. Contrary to the original model, where closeness was the sole factor driving farm plot selection, selection of higher quality land and desire for social presence are shown to be more important. In fact, model performance is improved when agents select farm plots further away from their failed farm plot. Farm selection strategies designed using these insights into the socio-agricultural behavior of the ancestral Pueblo significantly improved the model's accuracy and robustness. Introduction Exposing the mechanics of the human decision-making processes that cause complex, society-scale phenomena is a difficult endeavor. These decision-making processes are often driven by multiple causal factors [1] with researchers having no direct means of measuring how these factors contribute to society-scale phenomena. Abductive reasoning via data-driven modeling and simulation techniques can overcome these issues by 'growing' artificial societies [2] and adjusting their configurations until adequate matches between simulation results and real world data are achieved. Agent-based modeling (ABM) in particular, offers the benefit of representing behaviors as human-interpretable rules. These rules are driven by the agent's autonomous evaluation of a variety of factors that are hypothesized by the modeler to be important in the decision-making process being modeled, indicated by the ability of August 21, 2019 1/16 ABM to simulate real world observations. However, a particular behavior rule only represents a single hypothetical decision-making process contained within a large space of possible, alternate decision-making processes. Exploring this vast space of rules requires the repeated re-implementation of multiple versions of the same ABM with different embedded decision-making processes [3]; this is a tedious task, involving the comparison of a massive number of combinations of causal factors. Thus, researchers often resort to modeling the most intuitive decision-making processes, if not process, which risks a subjective and inaccurate representation of the actual individual behavior [4]. The current standard of ABM exploration, parameter calibration, is a black-box technique and does not perform white-box rule exploration. Parameter calibration works on the assumption of the correctness of a predefined rule and fine tunes the coefficients of its constituent factors, but cannot it easily experiment between different structures and operators through which these factors combine. Unless the importance of factors and how they are structured in the behavior rule are established, the underlying behavior rule merely remains an untested hypothesis of the actual individual behavior [2 -- 4]. Parameter calibration tools are readily available for ABM frameworks, such as BehaviorSearch [5] for NetLogo [6], and OptQuest [7] for AnyLogic [8]. Inductive games have been used to infer the decision-making of societies via game theory [9], yet no established methodology exists for ABMs. As ABM rules are implemented as program instructions, genetic programming [10] is a highly suitable technique for model discovery. However, research into using genetic programming with ABMs for exploration of causality has been limited [11, 13, 14]. To meet this need, we introduce evolutionary model discovery, a technique for agent rule exploration and causal factor importance measurement, which combines the automated program generation capability of genetic programming [10] with the factor importance evaluation capability of random forest regression [15 -- 20]. Unlike current standard techniques like pattern-oriented modeling [3] and model selection [21], evolutionary model discovery has the advantage of avoiding manual and repetitive re-implementation of models through automated program generation, resulting in a greatly reduced risk of implementation errors. Agent rules generated through genetic programming consist of functions of primitives that are easily comparable, as they follow a common representation. Using this representation, differences between candidate models are isolated to the code implementing the decision under scrutiny, to facilitate factor analysis and to avoid the need to compare two completely different implementations. The comparability of candidate models is important in drawing insights into the causes of the society-level phenomena being simulated. The stochasticity of genetic programming allows for the exploration of a vast space of possible agent rules, while selection of fitter models for breeding the next generation of rules ensures the exploitation of stronger factor interactions. Assumptions on agent behavior can be relaxed and rules with deeper factor interactions evaluated. Genetic programming, random forest training, and factor importance evaluation are all easily parallelizable techniques, which is important considering the large search space that can result even from a simple factor set. In this study, we employ evolutionary model discovery to discover plausible interactions of causal factors by comparing their importance in the farm plot selection of the ancestral Pueblo community modeled in the Artificial Anasazi [22, 24]. The ABM simulates the population dynamics of the Long House Valley between the years 800 AD to 1400 AD during which there was a sudden population collapse around 1350 AD. The original model demonstrated that this collapse was not caused by environmental factors alone. The model is data driven and simulations attempt to match the annual population time-series measuring households in the valley, which August 21, 2019 2/16 was estimated through data gathered from archaeological digs [22]. The agents in the model represent households, and are dependent on the agricultural success of their farm plot for sustenance and reproduction. The farm plot selection strategy originally implemented dictates that upon depletion of a household's current farm plot, the agent moves to the next closest available plot of land. In other words, the sole factor influencing this decision is the minimization of distance over the complete set of available plots of land in the valley. We argue that this behavior is not entirely human-like, and could have been influenced by factors other than distance. Instead, we hypothesize that nine different factors and four different social structures governing information flow may have driven the farm plot seeking behavior of the modeled Pueblo society. Specifically, we hypothesize that the following factors could have had significant importance: distance (FDist), dryness of the farm-land (FDry), quality of farm-land (FQual), yield of the land in the previous year (FY ield), water availability (FW ater), social presence near the potential farm land (FSoc), homophily by age (FAge), homophily by agricultural success (FAgri), and inter-zone migration (FM ig ), under the following possible social connectivity configurations: full information of the valley (SAll), information provided by family immediate family members (SF am), information provided by the most productive households (SP erf ), or information from the nearest neighbors (SN hbr). We consider the coefficients of these factors in the evolved agent behavior rules as the factors' 'presence' in that particular behavior rule. Each factor's presence is then analyzed for its importance at predicting the ABM's fitness through feature importance analysis on a random forest trained on data generated by the genetic program. Utilizing a random forest for this purpose allowed us to measure both main effects of the factors' presence and the joint contributions of factors towards the ABM's fitness. After identifying the most important factors, we determined the optimal presence for them. With these insights we were able to construct causally accurate and robust farm selection procedures. Our results falsify the original assumption [22, 24] that closeness was the sole causal factor governing farm plot selection of the ancestral Pueblo society. Instead, evolutionary model discovery reports the most important factors as quality, social presence, migration from zone, distance, and dryness in order of decreasing importance. In particular, the selection of higher quality land that either had a higher social presence or was located in a different zone was shown to be more likely behavior and versions of the Artificial Anasazi with these farm selection strategies were significantly more robust against random initialization of parameters. Our results indicate that the farm selection strategy was likely more human-like than that implemented in original version of the model [22, 24]. Methodology Farm Plot Selection in the Artificial Anasazi The Artificial Anasazi is an agent-based model of the Kayenta Anasazi during the years of 800 AD to 1350 AD [22, 24]. This model was initially developed as part of a larger effort to study the ancestral Pueblo civilization that occupied the Long House Valley region. The ABM is implemented in NetLogo [6, 24]. Archaeological excavations provide annual population time series data as estimated counts of households that existed in the valley during the period of study. Annual data on water sources and estimated soil dryness (Adjusted Palmer Drought Severity Index) for each grid location on the map are provided. The model used a normal distribution to map relative quality of soil over the map. The agent-based model simulates the rise and fall August 21, 2019 3/16 of households over a geographic map of the valley over time and produces a time series of annual household count. The original purpose of the Artificial Anasazi was to test if environmental factors could have triggered the sudden disappearance of the Anasazi from the Long House Valley around 1350 AD. Critics of the Artificial Anasazi have argued that the agent-based model itself is but a single candidate explanation of the social phenomenon at hand, the rise and fall of the Anasazi population over time [4]. However, we view this as an advantage as the Artificial Anasazi can be used as a test-bed to discover multiple plausible explanations of the population dynamics of the Long Valley at the time. Testing combinations of hypothesized factors that may have influenced actual decision-making processes of the individuals results in a vast search space of plausible Artificial Anasazi behavior results. We concentrated on a particular sub-model of the Artificial Anasazi: the farm plot selection strategy. The households perform farm plot selection under two conditions: 1) when a new child household is hatched by a household that has enough resources to increase its family size, or 2) when the current farm plot is unable to produce enough yield to satisfy the nutrition needs of the household anymore. The original model, hypothesizes that the households simply selected the next closest available farm plot to the household's current farm plot during farm plot selection, i.e., minimizing over distance. A patch must be free of farms or households and not be located inside a water body to be available. Consequently, the original farm selection strategy ignores other sensory data available to the households regarding the land and the state of other households in the valley. Hypothesized Alternate Factors Influencing Farm Plot Selection Human social behavior is rarely entirely rational. Accordingly, our hypothesis proposed that the farm selection decisions of the ancestral Pueblo were complex, and took into account the state of the potential farm plots available to them and the social influences of other households around them. Agent Zero [2] models the human decision making process into three dimensions: social, emotional and rational. Similarly, we defined factors that we hypothesized to influence the farm plot selection process within these dimensions. The social component is expressed through four mutually exclusive social connectivity configurations through which the agent could receive information on a subset of potential farm plots, s, out of the entire set of potential farm plots in the valley, SAll. The received information is then processed through a utility function f (x) defined as a combination of factors and operators, F , which consider both the internal state of the household and the conditions of the farm plot and its surroundings in order to determine the next farm plot x′ ∈s⊂ SAll as in Eq (1). x′ = argmax x∈s⊂SAll f (x) (1) Households in the original Artificial Anasazi model consider a single factor, distance, which we will refer to as FDist, and choose the potential farm plot with minimal distance to their current farm location. No further factors are considered in the decision making process. Furthermore, the original model assumes that the households have complete information of the valley, and every potential farm plot is compared. Therefore, the farm selection process of the original Artificial Anasazi can be represented as in Eq (1). x′ = argmax x∈SAll (−FDist(x)) (2) Arguing that the farm selection decision may have been more complex, considering a variety of other factors, we proposed an extended factor set consisting of four social August 21, 2019 4/16 and five rational factors, namely: homophily by age (FHAge), homophily by agricultural productivity (FHAgri), social presence (FSoc), migration from current zone (FM ig ), comparison of quality (FQual), comparison of dryness (FDry), comparison of yield (FY eild), comparison of water availability (FW ater), and comparison of distance (FDist). Additionally, the numerical operators + and − are included in F , for the aggregation of sub-scores reported by the social/emotional and rational factors. Four hypothesized configurations of social connectivity were included F . These configurations determined the subset of all viable farm plots that were to be considered by the households for comparison. 1) Full information (SAll): Households had complete knowledge of all potential farm plots in the valley. Full information was used by agents in the original version of the model, assuming that each household knew and compared every potential farm plot in the Long House Valley. 2) Family inherited information (SF am): Households solely depended on information available through their 'family'. Families are defined as a household's parent household, sibling households, any surviving grandparents, and the household itself. 3) Nearest-neighbor information (SN eigh): agents only consider the farm plots known to their neighboring households within a fixed radius of their current location. 4) Best performers SP erf : Households only consider potential farm plots known to the best performing households, demonstrating a leadership dynamic. Four social/emotional factors were included in F : two types of homophily (the tendency for social entities to congregate among those with similar traits), need for social presence, and one of fleeing/migration. Each social/emotional factor returned a sub-score representing the desirability of each evaluated farm plot. Sub-scores were normalized within the factors, to lie in the range of 0 to 1, for fair comparison. 1) Homophily by age (FHAge): Households prefer to select farm plots near other households that are of similar age, where age is measured as the number of simulation steps the household has survived since splitting from its parent. 2) Homophily by agricultural productivity (FHAgri): Households tend to select farm plots near other households with a similar corn stock to itself. 3) Social presence (FSoc): Agents score potential farm plots with many nearby households higher than those in isolation. 4) Fleeing/migration (FM ig): Agents score potential farm plots that are in a completely different zone than the current one with a full sub-score, while patches in the same zone receive a sub-score of zero. Five Rational factors considered for the farm selection process were logical comparisons of sensory data on the potential farm plots already available to the households in the original model. Similar to the social/emotional factors, rational factors also returned a normalized sub-score of farm plot desirability between 0 and 1. 1) Comparison of quality (FQual): Higher sub-scores were reported for potential farm plots with higher quality of land. 2) Comparison of dryness (FDry): Higher sub-scores were reported for potential farm plots with higher dryness of land. 3) Comparison of yield (FY eild): Higher sub-scores were reported for potential farm plots that were known to have higher yield in the previous year. 4) Water availability (FW ater): Higher sub-scores were reported for potential farm plots with more nearby water sources. 5) Comparison of distance (FDist): Higher sub-scores were reported for potential farm plots that were closer to the current farm plot location. Evolutionary model discovery Evolutionary model discovery allows agent-based modelers to explore the importance of a hypothesized set of factors affecting individual-level decision making towards a macro, society-level outcome. Accordingly, evolutionary model discovery requires the modeler to identify the particular agent behavior rule being evaluated within the original agent-based model. The modeler must also provide a set of hypothesized August 21, 2019 5/16 factors and combining operators that the modeler hypothesizes to affect the decision-making process represented by the agent behavior rule. A factor Fi ∈ F , where F is the modeler's set of hypothetical factors and operators, is defined as in Eq (3). Where C is the set of commands defined within F that are applied on the n number of input parameters P to produce an output return value R, where the type of each parameter tPj and the type of the return value tR are each an element of the set T of all possible parameter and return types defined by the modeler. A factor is considered an operator if C resembles an operation on one or more factors, which it accepts as parameters, rather than resembling a decision-making step. In order for a factor or operator Fi to accept another Fj as an input, the condition Eq (4) must be met. Fi = (C, R, P tR, tPk ∈ T ∀k = 1...n) ∃k, tRfi = tPfj ,k (3) (4) An agent behavior rule b ∈ B is represented as a tree of factors combined under this condition. Depending on T and the factor definitions, the space of behavior rules B can be infinitely large. To prevent the construction of such undesirably large trees, is we specify a maximum depth for all b. There must be at least one Fi of which tRFi the return type expected by the entire agent behavior rule. Given the ABM and F , evolutionary model discovery performs two stages of analysis. First, models driven by alternate decision making processes consisting of combinations of elements of F are evolved through genetic programming [10, 25, 26]. Genetic programming performs automated program implementation and is a suitable approach towards automating the rule discovery process [11 -- 14]. Genetic programming evolves generations of programs through crossover and mutation operators performed on a representation consisting of primitives and terminals that combine to define program statements. Primitives are defined as a set of functions that encode program statements and may be strongly typed to only accept child and parent primitives that are compatible with the arguments and return statements accepted by its program statement. Primitives with no arguments are considered terminals. The syntax tree representation is perhaps the most common representation used in genetic programming, and arranges the primitives and terminals into a tree structure, a representation compatible with b. Programs in a generation that have a closer fit to data are more likely to be selected for reproduction through crossover and mutation to populate the next generation of programs. Second, factor and factor interaction importance was assessed by random forest feature importance measurement. A random forest regressor was trained on the factor presence to fitness data produced by the genetic program. Random forests are an ensemble learning algorithm consisting of a forest of randomized decision trees [15 -- 17]. The two most common factor importance measurement techniques for random forests are gini importance (or mean decrease in impurity), and permutation importance (or mean decrease in accuracy) [15 -- 17]. However, both gini importance and permutation importance are unable to quantify the importance of factor interactions, as they consider the global importance each factor has for the random forest. Functional analysis of variance [20, 27] is able to quantify the importance of factor interactions, yet lacked precision considering the inherent heteroskedasticity of the data produced by the genetic program, caused by its tendency to explore and test models of higher fitness. Instead, joint contribution [18, 19] was used for this purpose as it has been successfully used to assess the importance of variable interactions in a large number of recent studies [28 -- 34]. August 21, 2019 6/16 Twenty genetic programming runs were executed with the objective of minimizing the (RMSE) between the simulated household count to the actual household count over 550 simulation ticks of the Artificial Anasazi. Details on the RMSE calculation can be found in [14]. In order to ensure robustness of the evolved rules, the parameters of the ABM were randomly initialized with values ±5% about the optimal parameter values found through Stonedahl's calibration of the Artificial Anazasi through a genetic algorithm [5] (ie: water source distance = (10.925, 12.075), death age span = (9.5, 10.5), min fertility = (0.1615, 0.1785), base nutrition need = (175.75, 194.25), fertility span = (0.0285, 0.0315), min fertility ends age = (27.55, 30.45), harvest variance = (0.418, 0.462), harvest adjustment = (0.608, 0.672), maize gift to child = (0.4465, 0.4935), min death age = (38.0, 42.0), fertility ends age span = (4.75, 5.25)). The genetic program was implemented with the Distributed Evolutionary Algorithms in Python library (DEAP) [35] and parallelized by SCOOP [36]. Each genetic program run was executed for 100 generations over populations of 50 individuals. Syntax trees of minimum depth 4 and maximum depth 10 were used to avoid trees exhibiting bloat. The Half-and-Half tree builder was used for initialization [10]. To accommodate the high computational cost, the genetic program runs were distributed across a 48 vcpu Amazon Web Services EC2 instance. The random forest and gini importance algorithm of Scikit-learn [37] were used, while ELI5 [38] was used for permutation accuracy importance, and tree interpreter [19] for joint contribution measurement. Finally, new farm selection strategies were designed taking into account the insights gained through evolutionary model discovery. The robustness of the Artificial Anasazi with these new strategies were tested against the original model by comparing the RMSE of 100 runs of each model under randomized initialization of parameters within the ranges above. Results The resulting best farm selection strategies evolved by the genetic program by run are provided in Table 1 along with their respective RMSE values. 15 of the runs produced RMSE values lower than the current best RMSE in the literature obtained through parameter calibration of the Artificial Anasazi model with the original farm plot selection by closeness (733.6) [39]. All best scoring rules for each run utilized SAll, i.e., the model produced best results when the agents had full information regarding available farm plots as shown in Fig. 1, comparing SAll, SF am, SN eigh, and SP erf over the complete factor presence to fitness data. One-tailed Mann-Whitney U tests comparing the fitness of all rules by their social connectivity configurations confirmed that rules with SAll had significantly (α = 0.05) lower RMSE than the other three configurations: argmaxx∈SAll f (x) (p = 4.856 × 10−154), argmaxx∈SAll f (x) (p = 1.983 × 10−57). Also, rules with SN eigh argmaxx∈SAll were shown to have significantly (alpha=0.05) lower RMSE than those with SF am and SP erf : argmaxx∈SN eigh argmaxx∈SN eigh were shown to have significantly (alpha=0.05) lower RMSE than rules with SP erf : f (x) (p = 0.012). Accordingly, the rest of the argmaxx∈SF am analyses detailed in this paper were performed on rules where the social connectivity configuration was SAll. f (x) < argmaxx∈SN eigh f (x) < argmaxx∈SP erf f (x) (p = 2.339−24). Finally, rules with SF am f (x) < argmaxx∈SF am f (x) (p = 3.535 × 10−14), f (x) < argmaxx∈SF am f (x) (p = 2.045 × 10−113), f (x) < argmaxx∈SP erf f (x) < argmaxx∈SP erf Fig. 2 displays the distribution of RMSE against factor presence, for presence values that were recorded in at least 200 rules across the 20 genetic program runs. Negative correlations to RMSE (higher fitness) are seen between FDist, FQual, FW ater, FY ield, FM ig , FSoc, and FAge, and in general the genetic program favored the positive August 21, 2019 7/16 SNeigh SPerf SFam SAll 1000 1500 Root Mean Squared Error 2000 2500 3000 Fig 1. Best fit to data was obtained under SAll. Comparison of the RMSE produced by the Artificial Anasazi model when agents had full information (SAll), information through family households (SF am), information through the households with most agricultural success (SP erf ), or information through neighboring households (SN eigh). Models that used SAll produced the lowest RMSE overall argmaxx∈SAll argmaxx∈SAll argmaxx∈SAll f (x) (p = 2.045 × 10−113), f (x) (p = 4.856 × 10−154), f (x) (p = 1.983 × 10−57). f (x) < argmaxx∈SF am f (x) < argmaxx∈SN eigh f (x) < argmaxx∈SP erf presence of these factors, and evolved more rules with these factors having a positive effect on farm selection. FDry on the other hand had a negative correlation to RMSE for presence less than 2. FDist -1 0 1 2 3 FWater 4 5 6 -3 -2 FDry -1 0 FYield FQual 1 2 -1 0 1 2 3 4 5 6 8 FMig -1 0 1 FSoc 2 3 -2 -1 0 1 2 3 4 -2 -1 0 FHAge 3 4 6 1 2 FHAgri 3000 2000 1000 2000 1000 2000 1000 r o r r E d e r a u q S n a e M t o o R -1 0 1 2 3 4 5 1 2 0 -1 3 Factor Presence 4 7 -2 -1 0 1 2 3 Fig 2. RMSE vs Factor Presence under SAll. RMSE distributions by factor presence produced by evolutionary model discovery of the farm selection strategy of the Artificial Anasazi under SAll. Only presence values that appeared at least 200 times in the genetic program are displayed. Most factors display negative correlations to RMSE, while FDry shows a positive correlation. The random forest fit the factor presence to fitness data best for a forest of 520 August 21, 2019 8/16 Table 1. The candidate farm selection strategies of models produced by the evolutionary model discovery process along with their best fitness as reported by the genetic programming search. GP Run Best scoring rule (FM ig(x)) 0 (−FDist(x) − FDry(x) + 2 ∗ FM ig(x)) (FY ield(x) + FHAgri(x)) (FM ig(x) − FHAgri(x)) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll argmax x∈SAll (FM ig(x)) (FDist(x)) (FDist(x)) (FY ield(x)) (FM ig(x)) (FQual(x)) (FQual(x)) (FM ig(x)) Best Fitness 753.430820 755.270812 709.502643 738.949931 730.475188 752.519767 728.293210 714.205153 734.249957 701.208243 720.281195 723.633194 687.122260 732.189183 728.772255 706.282521 715.957401 715.468378 701.438522 701.300934 (FDist(x) − FDry(x)) (4 ∗ FDist(x) + FDry(x) + FQual(x) + FW ater(x) + FSoc(x) + FHAge(x)) (FDist(x) + FQual(x) + FW ater(x) − FY ield(x) + FM ig(x) + FSoc(x)) (FDist(x) + FQual(x) + 2 ∗ FY ield(x) + 2 ∗ FM ig(x) + FSoc(x) + FHAgri(x)) (FDist(x) + FSoc(x)) (FDist(x) + 2 ∗ FQual(x) + FY ield(x) + FSoc(x) + 3 ∗ FHAge(x)) (−FDist(x) + FSoc(x) − FHAgri(x)) (FQual(x) + FM ig(x) + FSoc(x)) regression trees, testing from 10 to 1000 trees with a train/test split 90%-10%. Accordingly, a forest of 520 trees was used for factor importance determination. Factor importance under SAll obtained through both the gini importance and permutation accuracy importance techniques can be seen in Fig. 3. Gini importance generally had less precise estimations than permutation accuracy importance. Yet both techniques indicated FQual as the factor of highest importance towards RMSE prediction. FSoc, FM ig, and FDist also scored higher importance values than the other factors hypothesized. Fig. 4 displays the p-values of one-tailed Mann Whitney U tests (alpha=0.05), comparing the permutation importance of each factor A against every other factor B, testing the alternate hypothesis: importance of A ¿ importance of B. According to the results, 7 of the 9 factors showed significant difference and could be ordered in terms of permutation accuracy importance as FQual, FSoc, FDist, FM ig, FW ater, FY ield, FHAgri, FHAge, and FDry. Fig. 5 compares the top ten joint contributions towards RMSE prediction of the random forest by individual factors, and joint contributions of factors considered in August 21, 2019 9/16 Gini Importance Permutation Accuracy Importance FQual FSoc FDist FMig FWater FYield FHAgri FHAge FDry 0.0 0.1 0.2 0.3 0.4 0.4 0.3 0.2 0.1 0.0 Fig 3. FQual, FSoc, FDist, and FM ig have highest Gini and Permutation Accuracy Importance. Gini importance and permutation accuracy importance of the hypothesized factors towards a random forest's ability to predict the models' RMSE. Gini importance results are less decisive than permutation accuracy importance. Both techniques agree that FQual, FSoc, FDist, and FM ig are the most important factors. FDry FHAge FHAgri FYield FWater FMig FDist FSoc FQual 5.2e-01 9.1e-05 9.1e-05 9.1e-05 9.1e-05 B 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 5.2e-01 1.0e+00 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 5.2e-01 1.0e+00 1.0e+00 9.1e-05 9.1e-05 9.1e-05 9.1e-05 9.1e-05 5.2e-01 1.0e+00 1.0e+00 1.0e+00 9.1e-05 9.1e-05 9.1e-05 9.1e-05 5.2e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 9.1e-05 1.6e-04 4.4e-02 5.2e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 9.1e-05 1.1e-03 5.2e-01 9.6e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 9.1e-05 5.2e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 5.2e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 A FQual FSoc FDist FMig FWater FYield FHAgri FHAge FDry Fig 4. Statistical confirmation of the existence of order by importance among causal factors. Results from systematic Mann-Whitney U tests on the permutation accuracy importance results. The cells contain p-values for the alternate hypothesis that A > B (null hypothesis A = B). Green cells indicate agreement of the alternate hypothesis. The results indicate a clear ordering of the factors by importance. pairs and triples. Again, FQual demonstrates far higher importance than any other factor or factor interaction. The factor pairs (FQual, FM ig) and (FQual, FSoc) also demonstrate high importance, followed by (FQual, FM ig, FSoc), (FDry, FQual, FM ig), and (FDist, FQual, FSoc). Overall, FQual is present in all highest scoring joint contributions. Despite FDry having very low individual importance, FDry showed higher importance when considered in combination with FQual and FM ig. August 21, 2019 10/16 n o i t c a r e t n I r o t c a F ['FQual'] ['FQual', 'FMig'] ['FQual', 'FSoc'] ['FQual', 'FMig', 'FSoc'] ['FDry', 'FQual', 'FMig'] ['FDist', 'FQual', 'FSoc'] 0.0 0.2 0.8 1.0 Normalized Contribution to Random Forest Prediction 0.4 0.6 Fig 5. FQual, [FQual,FMig], and [FQualFSoc] have highest joint contribution to farm plot selection. Ordered barchart of highest normalized joint contribution scores of factors and interactions of three or less under SAll. Again, FQual shows a far larger contribution to the random forest's ability to predict model RMSE than other factors and factor interactions, and is present in all of the highest contributing interactions. Interactions [FQual,FM ig ] and [FQual,FSoc] also demonstrate high joint contribution. Considering the evidence of FQual, FSoc, FM ig, FDist, and FDry as important factors, Fig. 6 demonstrates Mann Whitney U tests conducted for each factor Fi, for the alternate hypothesis that RMSE when presence of Fi was A, is less than the RMSE when presence of Fi was B in rules with SAll. Models with positive presence of FQual, FSoc, FDist, and FM ig showed significantly higher fitness (with the exception of when presence of FM ig = -2). Models with strong positive or negative presence of FDry showed lower RMSE overall, most likely a result of FDry's interaction with FQual, FSoc, or FM ig. The lowest median RMSE for (FQual, FSoc) was 985 at presence of FSoc at 5 and presence of FQual at; the lowest median RMSE for (FQual, FM ig) was 997 at presence of FM ig at 3 and presence of FQual at 5. Finally, rules following the three highest joint contributions were constructed using the best values for each factor concerned: argmaxx∈SAll(FQual(x)), argmaxx∈SAll(5FSoc(x) + 6FQual(x)), and argmaxx∈SAll(3FM ig(x) + 5FQual(x)), and RMSE was compared against the original farm selection strategy argmaxx∈SAll(−FDist(x)) for 100 runs each under random initialization of parameters within the ranges specified in section . Fig. 7 shows that all three of these rules derived through evolutionary model discovery have significantly lower RMSE than that of the original farm selection strategy under randomized parameter initialization. Discussion and Conclusion Despite being an excellent tool for the construction and analysis of human-interpretable explanations of social phenomena, ABMs risk premature assumptions when modeling individuals' decision-making processes. Parameter calibration alone cannot adequately explore the causal factors and their possible interactions in order to infer more accurate decision-making processes. This is primarily due to the absence of a systematic method for behavior inference and discovery. We address this issue with the introduction of evolutionary model discovery, which is able to distinguish, out of a hypothesized set, the causal factors that are important to simulate the behavior of interest. By combining automated program generation of genetic programming with feature importance evaluation of random forests, evolutionary model discovery is able to quantify the importance of these factors August 21, 2019 11/16 FQual FSoc FMig A -1 0 1 2 3 4 5 6 8 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.1e-26 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 5.0e-48 1.5e-66 5.0e-01 9.6e-01 1.0e+00 1.0e+00 9.7e-01 8.0e-01 9.9e-01 3.1e-46 2.1e-33 3.7e-02 5.0e-01 1.0e+00 1.0e+00 8.7e-01 6.4e-01 9.8e-01 1.2e-52 1.3e-43 6.9e-08 8.2e-04 5.0e-01 6.1e-01 3.4e-01 2.3e-01 8.2e-01 A 3.1e-44 2.8e-26 8.4e-06 2.3e-03 3.9e-01 5.0e-01 2.8e-01 2.1e-01 7.7e-01 2.7e-24 3.7e-08 3.4e-02 1.3e-01 6.6e-01 7.2e-01 5.0e-01 3.7e-01 8.5e-01 2.6e-14 5.4e-04 2.0e-01 3.6e-01 7.7e-01 7.9e-01 6.3e-01 5.0e-01 8.8e-01 6.7e-19 2.7e-07 5.1e-03 2.1e-02 1.8e-01 2.3e-01 1.5e-01 1.2e-01 5.0e-01 -1 0 4 1 2 -1 0 1 2 3 4 5 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 5.4e-05 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 5.7e-13 4.2e-22 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.9e-18 1.3e-24 4.7e-04 5.0e-01 8.2e-01 8.5e-01 9.3e-01 A 9.4e-14 3.9e-11 1.8e-03 1.8e-01 5.0e-01 5.5e-01 7.4e-01 2.3e-12 1.9e-09 2.8e-03 1.5e-01 4.5e-01 5.0e-01 6.9e-01 7.6e-12 7.3e-09 1.4e-03 7.0e-02 2.6e-01 3.1e-01 5.0e-01 -2 -1 0 1 2 3 4 6 5.0e-01 1.1e-03 3.7e-01 9.3e-01 9.8e-01 1.0e+00 9.8e-01 9.9e-01 1.0e+00 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 6.3e-01 5.7e-12 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 7.2e-02 3.8e-26 1.5e-27 5.0e-01 9.9e-01 1.0e+00 9.2e-01 9.8e-01 1.7e-02 2.8e-25 8.2e-16 1.1e-02 5.0e-01 9.5e-01 7.3e-01 9.1e-01 2.0e-03 1.3e-19 3.6e-10 7.8e-04 4.5e-02 5.0e-01 3.4e-01 5.7e-01 2.1e-02 8.0e-09 6.5e-04 7.6e-02 2.7e-01 6.6e-01 5.0e-01 6.7e-01 6.4e-03 4.7e-11 2.5e-05 1.5e-02 9.3e-02 4.3e-01 3.3e-01 5.0e-01 3 B 5 6 8 -1 0 1 FDry A -3 -2 -1 0 1 2 5.0e-01 7.0e-01 1.8e-01 7.6e-02 8.3e-03 9.4e-01 3.0e-01 5.0e-01 7.9e-04 1.3e-06 5.0e-10 9.5e-01 8.2e-01 1.0e+00 5.0e-01 2.1e-02 6.4e-06 1.0e+00 9.2e-01 1.0e+00 9.8e-01 5.0e-01 8.5e-05 1.0e+00 9.9e-01 1.0e+00 1.0e+00 1.0e+00 5.0e-01 1.0e+00 5.8e-02 4.8e-02 5.4e-05 1.2e-06 4.5e-09 5.0e-01 2 B A -1 0 1 2 3 4 5 6 3 4 5 -2 -1 0 1 2 3 4 6 B FDist 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.5e-13 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 5.2e-22 7.2e-13 5.0e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.9e-32 6.3e-25 2.6e-08 5.0e-01 5.9e-01 7.2e-01 8.9e-01 9.8e-01 1.6e-23 1.6e-12 2.9e-05 4.1e-01 5.0e-01 6.5e-01 8.5e-01 9.7e-01 1.5e-18 4.5e-09 2.0e-04 2.8e-01 3.5e-01 5.0e-01 7.7e-01 9.5e-01 4.2e-13 8.4e-07 4.8e-04 1.1e-01 1.5e-01 2.3e-01 5.0e-01 8.2e-01 7.9e-11 1.8e-06 2.0e-04 2.0e-02 3.3e-02 5.3e-02 1.8e-01 5.0e-01 -3 -2 -1 B 0 1 2 -1 0 1 2 3 4 5 6 B Fig 6. Optimal presence scores for causal factors with highest importance. Results from systematic one-tailed Mann-Whitney U tests between presence values of the 5 most important factors for the alternate hypothesis: RMSE for presence A < RMSE for presence B (null hypothesis: RMSE for presence A = RMSE for presence B) for α = 0.05. Green cells indicate agreement of the alternate hypothesis. Results indicate that for FQual, FSoc, FM ig, and FDist RMSE is generally lower for higher, positive presence. For FDry, both negative and higher positive presence may provide low RMSE scores. Fig 7. Models designed through evolutionary model discovery insights are significantly more robust. Comparison between the RMSE of 100 runs of three models with farm selection strategies designed taking into consideration the insights from evolutionary model discovery, 1) argmaxx∈SAll(FQual(x)), 2) argmaxx∈SAll(5FSoc(x) + 6FQual(x)), and 3) argmaxx∈SAll(3FM ig(x) + 5FQual(x)), against 100 runs of the original farm selection strategy argmaxx∈SAll(−FDist(x)) in [22 -- 24, 39], under random initialization of parameters. The three farm selection strategies derived from evolutionary model discovery are far more robust under random parameter initialization and show significantly better RMSE scores compared to the original model. to the decision-making process that result in society-level phenomena simulated by the ABM. This allows for the construction of agent rules that more accurately represent the actual decision-making process of individuals and result in more robust models. August 21, 2019 12/16 Applying evolutionary model discovery on the Artificial Anasazi we show that the socio-agricultural behavior of the ancestral Pueblo of the Long House Valley was more deliberative and informed than originally assumed. Our results indicate that, contrary to the original farm selection behavior, where households would select the next closest possible plot of land once their present farm was depleted, the households most likely selected potential farming land with higher soil quality (FQual). Further, it was highly likely that the households had good knowledge of the potential arable land throughout the valley, since SAll was the best social connectivity configuration for information spread. Also, the desire to congregate into communities was indicated, as positive desire for social presence (FSoc) was the second most important factor, and acting on information on arable land known to neighboring households (SN eigh) was the second most successful social connectivity configuration. Further, instead of choosing closer potential farm plots (−FDist), choosing farm plots that were further away from the households current farm plot (FDist) or moving to a completely different zone in the region (FM ig ) was found to be a more likely behavior. Finally, versions of the Artificial Anasazi where farm plot selection was driven by seeking higher quality land, higher quality land with more social presence, and higher quality land in different zones, all proved to be significantly more robust than the decision to move to the next closest available plot of land (Fig. 7). Supporting information S1 File. EvolutionaryModelDiscoveryArtificialAnasazi.zip This archive contains the Evolutionary Model Discovery Python source code. This Python package is also being actively maintained at: https://github.com/chathika/evolutionarymodeldiscovery and documentation is available at https://evolutionarymodeldiscovery.readthedocs.io/en/latest/. NetLogo models of the Artificial Anasazi with the best . S1 Table. FactorScores.csv Factor presence to model fitness data produced by the 20 genetic programming runs on the Artificial Anasazi. Acknowledgments This research was supported by DARPA SocialSim program HR001117S0018 and Amazon AWS Research Grant. References 1. Epstein JM. Agent Zero: Toward Neurocognitive Foundations for Generative Social Science. Princeton, NJ: Princeton University Press; 2013. 2. Epstein JM. Agent-based computational models and generative social science. Complexity. 1999;4(5): 41 -- 60. 3. Grimm V, Revilla E, Berger U, Jeltsch F, Mooij WM, Railsback SF, et al. Pattern-oriented modeling of agent-based complex systems: lessons from ecology. Science. 2005;310(5750): 987 -- 991. 4. Grune-Yanoff T. The explanatory potential of artificial societies. Synthese. 2009;169(3): 539 -- 555. August 21, 2019 13/16 5. Stonedahl F, Wilensky U. BehaviorSearch [computer software]. Center for Connected Learning and Computer Based Modeling, Northwestern University, Evanston, IL Available online: http://www.behaviorsearch.org. 2010. 6. Wilensky U. NetLogo [computer software]. Center for Connected Learning and Computer-Based Modeling, Northwestern University, Evanston, IL. Available online: http://ccl.northwestern.edu/netlogo/. 1999. 7. Laguna M, Marti R. The OptQuest callable library. Optimization software class libraries. Springer; 2003. pp. 193 -- 218. 8. Borshchev A. The big book of simulation modeling: multimethod modeling with AnyLogic 6. AnyLogic North America Chicago; 2013. 9. DeDeo S, Krakauer DC, Flack JC. Inductive game theory and the dynamics of animal conflict. PLoS computational biology. 2010;6(5): e1000782. 10. Koza JR. Genetic programming: on the programming of computers by means of natural selection. vol. 1. MIT press; 1992. 11. Manson SM. Agent-based modeling and genetic programming for modeling land change in the Southern Yucatan Peninsular Region of Mexico. Agriculture, ecosystems & environment. 2005;111(1): 47 -- 62. 12. Manson SM. Bounded rationality in agent-based models: experiments with evolutionary programs. International Journal of Geographical Information Science. 2006;20(9): 991 -- 1012. 13. Zhong J, Luo L, Cai W, Lees M. Automatic rule identification for agent-based crowd models through gene expression programming. In: Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems. International Foundation for Autonomous Agents and Multi-agent Systems; 2014. pp. 1125 -- 1132. 14. Gunaratne C, Garibay I. Alternate social theory discovery using genetic programming: towards better understanding the artificial anasazi. In: Proceedings of the Genetic and Evolutionary Computation Conference. ACM; 2017. pp. 115 -- 122. 15. Breiman L. Manual on setting up, using, and understanding random forests v3. 1. Statistics Department University of California Berkeley, CA, USA. 2002;1. 16. Genuer R, Poggi JM, Tuleau-Malot C. Variable selection using random forests. Pattern Recognition Letters. 2010;31(14): 2225 -- 2236. 17. Louppe G. Understanding Random Forests: From Theory to Practice. Ph.D. Thesis, The Universit´e de Li`ege, Li`ege, Belgique; 2014. 18. Saabas A. Interpreting random forests. In: Diving into data, 24 [Internet] Available online: https://blog.datadive.net/interpreting-random-forests/. 2014. 19. Saabas A. Treeinterpreter [computer software]. Available online: https://github.com/andosa/treeinterpreter. 2019. 20. Hutter F, Hoos H, Leyton-Brown K. An efficient approach for assessing hyperparameter importance. In: International Conference on Machine Learning; 2014. p. 754 -- 762. August 21, 2019 14/16 21. Stratton RJ. Automated theory selection using agent based models. Ph.D. Thesis, King's College London, UK; 2015. 22. Dean JS, Gumerman GJ, Epstein JM, Axtell RL, Swedlund AC, Parker MT, et al. Understanding Anasazi culture change through agent-based modeling. Dynamics in human and primate societies: Agent-based modeling of social and spatial processes. 2000;pp. 179 -- 205. 23. Janssen M A. Understanding artificial anasazi. Journal of Artificial Societies and Social Simulation. 2009;12(4): 13 24. Stonedahl F, Wilensky U. Artificial Anasazi model [computer software] Available from: http://ccl.northwestern.edu/netlogo/models/ArtificialAnasazi. 2010. 25. Langdon WB, Harman M. Optimizing existing software with genetic programming. IEEE Transactions on Evolutionary Computation. 2014;19(1): 118 -- 135. 26. Petke J, Haraldsson SO, Harman M, Langdon WB, White DR, Woodward JR. Genetic improvement of software: a comprehensive survey. IEEE Transactions on Evolutionary Computation. 2017;22(3): 415 -- 432. f 27. Dang N, De Causmaecker P. Analysis of algorithm components and parameters: some case studies. In: International Conference on Learning and Intelligent Optimization. Springer. 2018; pp. 288 -- 303 2018;. 28. Lundberg SM, Erion GG, Lee SI. Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:180203888. 2018;. 29. Smith EM, Nantes A, Hogue A, Papas I. Forecasting customer behaviour in constrained E-commerce platforms. In: 8th International Conference of Pattern Recognition Systems (ICPRS 2017). IET; 2017. pp. 1 -- 8. 30. Beillevaire M. Inside the Black Box: How to Explain Individual Predictions of a Machine Learning Model : How to automatically generate insights on predictive model outputs, and gain a better understanding on how the model predicts each individual data point. KTH, School of Electrical Engineering and Computer Science (EECS); 2018. 31. Rea C, Erickson K, Granetz R, Johnson R, Eidietis N, Montes K, et al. Initial Results of a Machine Learning-based Real Time Disruption Predictor on DIII-D. In: Proc. 45th EPS Conf. on Plasma Physics, Europhysics Conf. Abstracts. vol. 42; 2018. 32. Granetz R, Rea C, Montes K, Tinguely R, Eidietis N, Meneghini O, et al. Machine learning for disruption warning on ALCATOR C-MOD, DIII-D, and EAST tokamaks. In: Proc. 27th IAEA Fusion Energy Conference, IAEA, Vienna; 2018. 33. Morice-Atkinson X, Hoyle B, Bacon D. Learning from the machine: interpreting machine learning algorithms for point-and extended-source classification. Monthly Notices of the Royal Astronomical Society. 2018;481(3): 4194 -- 4205. 34. Bastrakova E. Improving interpretability of complex predictive models. Universitat Polit`ecnica de Catalunya; 2017. August 21, 2019 15/16 35. Fortin FA, De Rainville FM, Gardner MA, Parizeau M, Gagn´e C. DEAP: Evolutionary Algorithms Made Easy. Journal of Machine Learning Research. 2012;13: 2171 -- 2175. 36. Hold-Geoffroy Y, Gagnon O, Parizeau M. Once you SCOOP, no need to fork. In: Proceedings of the 2014 Annual Conference on Extreme Science and Engineering Discovery Environment. ACM; 2014. p. 60. 37. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. 2011;12: 2825 -- 2830. 38. Korobov M, Lopuhin K. Permutation Importance; Available from: https://eli5.readthedocs.io/en/latest/blackbox/permutation_importance.html. 2019. 39. Stonedahl F, Wilensky U. Evolutionary Robustness Checking in the Artificial Anasazi Model. In: AAAI Fall Symposium: Complex Adaptive Systems; 2010. pp. 120 -- 129. August 21, 2019 16/16
1301.2678	2	1301	2013-01-22T09:22:30	Verification of Agent-Based Artifact Systems	[ "cs.MA", "cs.AI", "cs.LO" ]	Artifact systems are a novel paradigm for specifying and implementing business processes described in terms of interacting modules called artifacts. Artifacts consist of data and lifecycles, accounting respectively for the relational structure of the artifacts' states and their possible evolutions over time. In this paper we put forward artifact-centric multi-agent systems, a novel formalisation of artifact systems in the context of multi-agent systems operating on them. Differently from the usual process-based models of services, the semantics we give explicitly accounts for the data structures on which artifact systems are defined. We study the model checking problem for artifact-centric multi-agent systems against specifications written in a quantified version of temporal-epistemic logic expressing the knowledge of the agents in the exchange. We begin by noting that the problem is undecidable in general. We then identify two noteworthy restrictions, one syntactical and one semantical, that enable us to find bisimilar finite abstractions and therefore reduce the model checking problem to the instance on finite models. Under these assumptions we show that the model checking problem for these systems is EXPSPACE-complete. We then introduce artifact-centric programs, compact and declarative representations of the programs governing both the artifact system and the agents. We show that, while these in principle generate infinite-state systems, under natural conditions their verification problem can be solved on finite abstractions that can be effectively computed from the programs. Finally we exemplify the theoretical results of the paper through a mainstream procurement scenario from the artifact systems literature.	cs.MA	cs	Verification of Agent-Based Artifact Systems Francesco Belardinelli Laboratoire Ibisc, Universit´e d'Evry, France Alessio Lomuscio Department of Computing, Imperial College London, UK Fabio Patrizi Dipartimento di Ingegneria Informatica, Automatica e Gestionale "A. Ruberti" Universit`a di Roma "La Sapienza", Italy [email protected] [email protected] [email protected] Abstract Artifact systems are a novel paradigm for specifying and implementing business processes described in terms of interacting modules called artifacts. Artifacts consist of data and lifecy- cles, accounting respectively for the relational structure of the artifacts' states and their possible evolutions over time. In this paper we put forward artifact-centric multi-agent systems, a novel for- malisation of artifact systems in the context of multi-agent systems operating on them. Differently from the usual process-based models of services, the semantics we give explicitly accounts for the data structures on which artifact systems are defined. We study the model checking problem for artifact-centric multi-agent systems against speci- fications written in a quantified version of temporal-epistemic logic expressing the knowledge of the agents in the exchange. We begin by noting that the problem is undecidable in general. We then identify two noteworthy restrictions, one syntactical and one semantical, that enable us to find bisimilar finite abstractions and therefore reduce the model checking problem to the instance on fi- nite models. Under these assumptions we show that the model checking problem for these systems is EXPSPACE-complete. We then introduce artifact-centric programs, compact and declarative representations of the programs governing both the artifact system and the agents. We show that, while these in principle generate infinite-state systems, under natural conditions their verification problem can be solved on finite abstractions that can be effectively computed from the programs. Finally we exemplify the theoretical results of the paper through a mainstream procurement sce- nario from the artifact systems literature. 1. Introduction Much of the work in the area of reasoning about knowledge involves the development of formal techniques for the representation of epistemic properties of rational actors, or agents, in a multi- agent system (MAS). The approaches based on modal logic are often rooted on interpreted sys- tems (Parikh & Ramanujam, 1985), a computationally grounded semantics (Wooldridge, 2000) used for the interpretation of several temporal-epistemic logics. This line of research was thoroughly ex- plored in the 1990s leading to a significant body of work (Fagin, Halpern, Moses, & Vardi, 1995). Further significant explorations have been conducted since then; a recent topic of interest has fo- cused on the development of automatic techniques, including model checking (Clarke, Grumberg, & Peled, 1999), for the verification of temporal-epistemic specifications for the autonomous agents in a MAS (Gammie & van der Meyden, 2004; Kacprzak, Nabialek, Niewiadomski, Penczek, P´olrola, Szreter, Wozna, & Zbrzezny, 2008; Lomuscio, Qu, & Raimondi, 2009). This has led to develop- ments in a number of areas traditionally outside artificial intelligence, knowledge representation 1 and MAS, including security (Dechesne & Wang, 2010; Ciobaca, Delaune, & Kremer, 2012), web- services (Lomuscio, Solanki, Penczek, & Szreter, 2010) and cache-coherence protocols in hardware design (Baukus & van der Meyden, 2004). The ambition of the present paper is to offer a simi- lar change of perspective in the area of artifact systems (Cohn & Hull, 2009), a growing topic in Service-Oriented Computing (SOC). Artifacts are structures that "combine data and process in an holistic manner as the basic building block[s]" (Cohn & Hull, 2009) of systems' descriptions. Artifact systems are services constituted by complex workflow schemes based on artifacts which the agents interact with. The data component is given by the relational databases underpinning the artifacts in a system, whereas the workflows are described by "lifecycles" associated with each artifact schema. While in the standard services paradigm services are made public by exposing their processes interface, in artifact systems both the data structures and the lifecycles are advertised. Services are composed in a "hub" where op- erations on the artifacts are executed. Implementations of artifact systems, such as the IBM engine BARCELONA (Heath, Hull, & Vacul´ın, 2011), provide a hub where the service choreography and service orchestratation (Alonso, Casati, Kuno, & Machiraju, 2004) are carried out. While artifact systems are beginning to drive new application areas, such as case management systems (Marin, Hull, & Vacul´ın, 2012), we identify two shortcomings in the present state-of-the- art. Firstly, the artifact systems literature (Bhattacharya, Gerede, Hull, Liu, & Su, 2007; Deutsch, Hull, Patrizi, & Vianu, 2009; Hull, 2008; Nooijen, Fahland, & Dongen, 2012) focuses exclusively on the artifacts themselves. While there is obviously a need to model and implement the artifact infrastructure, importantly we also need to account for the agents implementing the services acting on the artifact system. This is of particular relevance given that artifact systems are envisaged to play a leading role in information systems. We need to be able to reason not just about the artifact states but also about what actions specific participants are allowed and not allowed to do, what knowledge they can or cannot derive in a system run, what system state they can achieve in coordination with their peers, etc. In other words, we need to move from the description of the artifact infrastructure to one that encompasses both the agents and the infrastructure. Secondly, there is a pressing demand to provide the hub with automatic choreography and or- chestration capabilities. It is well-known that choreography techniques can be leveraged on auto- matic model checking techniques; orchestration can be recast as a synthesis problem, which, in turn, can also benefit from model checking technology. However, while model checking and its applications are relatively well-understood in the plain process-based modelling, the presence of data makes these problems much harder and virtually unexplored. Additionally, infinite domains in the underlying databases lead to infinite state-spaces and undecidability of the model checking problem. The aim of this paper is to make a concerted contribution to both problems above. Firstly, we provide a computationally grounded semantics to systems comprising the artifact infrastructure and the agents operating on it. We use this semantics to interpret a temporal-epistemic language with first-order quantifiers to reason about the evolution of the hub as well as the knowledge of the agents in the presence of evolving, structured data. We observe that the model checking problem for these structures is undecidable in general and analyse two notable decidable fragments. In this context, a contribution we make is to provide finite abstractions to infinite-state artifact systems, thereby presenting a technique for their effective verification for a class of declarative agent-based, artifact-centric programs that we here define. We evaluate this methodology by studying its compu- 2 tational complexity and by demonstrating its use on a well-known scenario from the artifact systems literature. 1.1 Artifact-Centric Systems Service-oriented computing is concerned with the study and development of distributed applica- tions that can be automatically discovered and composed by means of remote interfaces. A point of distinction over more traditional distributed systems is the interoperability and connectedness of ser- vices and the shared format for both data and remote procedure calls. Two technology-independent concepts permeate the service-oriented literature: orchestration and choreography (Alonso et al., 2004; Singh & Huhns, 2005). Orchestration involves the ordering of actions of possibly different services, facilitated by a controller or orchestrator, to achieve a certain overall goal. Choreogra- phy concerns the distributed coordination of different actions through publicly observable events to achieve a certain goal. A MAS perspective (Wooldridge, 2001) is known to be particularly helpful in service-oriented computing in that it allows us to ascribe information states and private or com- mon goals to the various services. Under this view the agents of the system implement the services and interact with one another in a shared infrastructure or environment. A key theoretical problem in SOC is to devise effective mechanisms to verify that service com- position is correct according to some specification. Techniques based on model checking (Clarke et al., 1999) and synthesis (Berardi, Cheikh, Giacomo, & Patrizi, 2008) have been put forward to solve the composition and orchestration problem for services described and advertised at inter- face level through finite state machines (Calvanese, Giacomo, Lenzerini, Mecella, & Patrizi, 2008). More recently, attention has turned to services described by languages such as WS-BPEL (Alves et al., 2007), which provide potentially unbounded variables in the description of the service pro- cess. Again, model checking approaches have successfully been used to verify complex service compositions (Bertoli, Pistore, & Traverso, 2010; Lomuscio, Qu, & Solanki, 2012). While WS-BPEL provides a model for services with variables, the data referenced by them is non-permanent. The area of data-centric workflows (Hull, Narendra, & Nigam, 2009; Nigam & Caswell, 2003) evolved as an attempt to provide support for permanent data, typically present in the form of underlying databases. Although usually abstracted away, permanent data is of central importance to services, which typically query data sources and are driven by the answers they obtain; see, e.g., (Berardi, Calvanese, Giacomo, Hull, & Mecella, 2005). Therefore, a faithful model of a service behavior cannot, in general, disregard this component. In response to this, proposals have been made in the workflows and service communities in terms of declarative specifications of data-centric services that are advertised for automatic discovery and composition. The artifact- centric approach (Cohn & Hull, 2009) is now one of the leading emerging paradigms in the area. As described in (Hull, 2008; Hull, Damaggio, De Masellis, Fournier, Gupta, Heath, Hobson, Linehan, Maradugu, Nigam, Sukaviriya, & Vaculin, 2011) artifact-centric systems can be presented along four dimensions. Artifacts are the holders of all structured information available in the system. In a business- oriented scenario this may include purchase orders, invoices, payment records, etc. Artifacts may be created, amended, and destroyed at run time; however, abstract artifact schemas are provided at design time to define the structure of all artifacts to be manipulated in the system. Intuitively, external events cause changes in the system, including in the value of artifact attributes. 3 The evolution of artifacts is governed by lifecycles. These capture the changes that an artifact may go through from creation to deletion. Intuitively, a purchase order may be created, amended and operated on by several events before it is fullfilled and its existence in the system terminated: a lifecycle associated with a purchase order artifact formalises these transitions. Services are seen as the actors operating on the artifact system. They represent both human and software actors, possibly distributed, that generate events on the artifact system. Some services may "own" artifacts, and some artifacts may be shared by several services. However, not all artifacts, or parts of artifacts, are visible to all services. Views and windows respectively determine which parts of artifacts and which artifact instances are visible to which service. An artifact hub is a system that maintains the artifact system and processes the events generated by the services. Services generate events on the artifact system according to associations. Typically these are declarative descriptions providing the precondition and postconditions for the generation of events. These generate changes in the artifact system according to the artifact lifecycles. Since events may trigger changes in several artifacts in the system, events are processed by a well-defined seman- tics (Damaggio, Hull, & Vacul´ın, 2011; Hull et al., 2011) that governs the sequence of changes an artifact-system may undertake upon consumption of an event. Such a semantics, based on the use of Prerequisite-Antecedent-Consequent (PAC) rules, ensures acyclicity and full determinism in the updates on the artifact system. GSM is a declarative language that can be used to describe artifact systems. BARCELONA is an engine that can be used to run a GSM-based artifact-centric system (Heath et al., 2011). The above is a partial and incomplete description of the artifact paradigm. We refer to (Cohn & Hull, 2009; Hull, 2008; Hull et al., 2011) for more details. As it will be clear in the next section, in line with the agent-based approach to services, we will use agent-based concepts to model services. The artifact-system will be represented as an environ- ment, constituted by evolving databases, upon which the agents operate; lifecycles and associations will be modelled by local and global transition functions. The model is intended to incorporate all artifact-related concepts including views and windows. In view of the above in this paper we address the following questions. How can we give a transition-based semantics for artifacts and agents operating on them? What syntax should we use to specify properties of the agents and the artifacts themselves? Can we verify that an artifact system satisfies certain properties? As this will be shown to be undecidable, can we find suitable fragments on which this can actually be carried out? If so, what is the resulting complexity? Lastly, can we provide declarative specifications for the agent programs so that these can be verified by model checking? Can this technique be used on mainstream scenarios from the SOC literature? This paper intends to contribute answering these questions. 1.2 Related Work As stated above, virtually all current literature on artifact-centric systems focuses on properties and implementations of the artifact-system as such. Little or no attention is given to the actors on the system, whether they are human or artificial agents. A few formal techniques have, however, been put forward to verify the core, non-agent aspects of the system; in the following we briefly compare these to this contribution. To our knowledge the verification of artifact-centric business processes was first discussed in (Bhattacharya et al., 2007), where reachability and deadlocks are phrased in the context of 4 artifact-centric systems and complexity results for the verification problem are given. The present contribution differs markedly from (Bhattacharya et al., 2007) by employing a more expressive specification language, even if the agent-related aspects are not considered, and by putting forward effective abstraction procedures for verification. In (Gerede & Su, 2007) a verification technique for artifact-centric systems against a variant of computation-tree logic is put forward. The decidability of the verification problem is proven for the language considered under the assumption that the interpretation domain is bounded. Decidability is also shown for the unbounded case by making restrictions on the values that quantified variables can range over. In the work here presented we also work on unbounded domains, but do not require the restrictions present in (Gerede & Su, 2007): we only insist on the fact that the number of distinct values in the system does not exceed a given threshold at any point in any run. Most importantly, the interplay between quantification and modalities here considered allows us to bind and use variables in different states. This is a major difference as this feature is very expressive and known to lead to undecidability. A related line of research is followed in (Deutsch et al., 2009; Damaggio, Deutsch, & Vianu, 2012), where the verification problem for artifact systems against two variants of first-order linear- time temporal logic is considered. Decidability of the verification problem is retained by imposing syntactic restrictions on both the system descriptions and the specifications to check. This effec- tively limits the way in which new values introduced at every computational step can be used by the system. Properties based on arithmetic operators are considered in (Damaggio et al., 2012). While there are elements of similarity between these approaches and the one we put forward here, includ- ing the fact that the concrete interpretation domain is replaced by an abstract one, the contribution here presented has significant differences from these. Firstly, our setting is branching-time and not linear-time thereby resulting in different expressive power. Secondly, differently from (Deutsch et al., 2009; Damaggio et al., 2012), we impose no constraints on nested quantifiers. In contrast, (Damaggio et al., 2012) admits only universal quantification over combinations of quantifier-free first-order formulas. Thirdly, the abstraction results we present here are given in general terms on the semantics of declarative programs and do not depend on a particular presentation of the system. More closely related to the present contribution is (Hariri, Calvanese, Giacomo, Deutsch, & Montali, 2012), where conditions for the decidability of the model checking problem for data- centric dynamic systems, e.g., dynamic systems with relational states, are given. In this case the specification language used is a first-order version of the µ-calculus. While our temporal fragment is subsumed by the µ-calculus, since we use indexed epistemic modalities as well as a common knowledge operator, the two specification languages have different expressive power. To retain decidability, like we do here, the authors assume a constraint on the size of the states. However, differently from the contribution here presented, (Hariri et al., 2012) assume limited forms of quan- tification whereby only individuals persisting in the system evolution can be quantified over. In this contribution we do not make this restriction. Irrespective of what above, the most important feature that characterises our work is that the set-up is entirely based on epistemic logic and multi-agent systems. We use agents to represent the autonomous services operating in the system and agent-based concepts play a key role in the modelling, the specifications, and the verification techniques put forward. Differently from all ap- proaches presented above we are not only concerned with whether the artifact-system meets a par- ticular specification. Instead, we also wish to consider what knowledge the agents in the system acquire by interacting among themselves and with the artifact-system during a system run. Ad- 5 ditionally, the abstraction methodology put forward is modular with respect to the agents in the system. These features enable us to give constructive procedures for the generation of finite abstrac- tions for artifact-centric programs associated with infinite models. We are not aware of any work in the literature tackling any of these aspects. Relation to previous work by the authors. This paper combines and expands preliminary results originally discussed in (Belardinelli, Lomuscio, & Patrizi, 2011a), (Belardinelli, Lomuscio, & Patrizi, 2011b), (Belardinelli, Lomuscio, & Patrizi, 2012a), and (Belardinelli, Lomuscio, & Pa- trizi, 2012b). In particular, the technical set up of artifacts and agents is different from that of our preliminary studies and makes it more natural to express artifact-centric concepts such as views. Differently from our previous attempts we here incorporate an operator for common knowledge and provide constructive methods to define abstractions for all notions of bisimulation. We also con- sider the complexity of the verification problem, previously unexplored, and evaluate the technique in detail on a case study. 1.3 Scheme of the Paper The rest of the paper is organised as follows. In Section 2 we introduce Artifact-centric Multi- Agent Systems (ACMAS), the semantics we will be using throughout the paper to describe agents operating on an artifact system. In the same section we put forward FO-CTLK, a first-order logic with knowledge and time to reason about the evolution of the knowledge of the agents and the artifact system. This enables us to propose a satisfaction relation based on the notion of bounded quantification, define the model checking problem, and highlight some properties of isomorphic states. An immediate result we will explore concerns the undecidability of the model checking problem for ACMAS in their general setting. Section 3 is concerned with synctactical restrictions on FO- CTLK that enable us to guarantee the existence of finite abstractions of infinite-state ACMAS, thereby making the model checking problem feasible by means of standard techniques. Section 4 tackles restrictions orthogonal to those of Section 3 by focusing on a subclass of ACMAS that admits a decidable model checking problem when considering full FO-CTLK specifi- cations. The key finding here is that bounded and uniform ACMAS, a class identified by studying a strong bisimulation relation, admit finite abstractions for any FO-CTLK specification. The section concludes by showing that under these restrictions the model checking problem is EXPSPACE- complete. We turn our attention to artifact programs in Section 6 by defining the concept of artifact-centric programs. We define them through natural, first-order preconditions and postconditions in line with the artifact-centric approach. We give a semantics to them in terms of ACMAS and show that their generated models are precisely those uniform ACMAS studied earlier in the paper. It follows that, under some boundedness conditions, which can be naturally expressed, the model checking problem for artifact-centric programs is decidable and can be executed on finite models. Section 7 reports a scenario from the artifact systems literature. This is used to exemplify the technique by providing finite abstractions that can be effectively verified. We conclude in Section 8 where we consider the limitations of the approach and point to further work. 6 2. Artifact-Centric Multi-Agent Systems In this section we formalise artifact-centric systems and state their verification problem. As data and databases are important constituents of artifact systems, our formalisation of artifacts relies on them as underpinning concepts. However, as discussed in the previous section, we here give prominence to agent-based concepts. As such, we define our systems as comprising both the artifacts in the system as well as the agents that interact with the system. A standard paradigm for logic-based reasoning about agent systems is interpreted systems (Parikh & Ramanujam, 1985; Fagin et al., 1995). In this setting agents are endowed with private local states and evolve by performing actions according to an individual protocol. As data play a key part, as well as to allow us to specify properties of the artifact system, we will define the agents' local states as evolving database instances. We call this formalisation artifact-centric multi-agent systems (AC- MAS). AC-MAS enable us to represent naturally and concisely concepts much used in the artifact paradigm such as the one of view discussed earlier. Our specification language will include temporal-epistemic logic but also quantification over a domain so as to represent the data. This is an usual verification setting, so we will formally define the model checking problem for this set up. 2.1 Databases and First-Order Logic As discussed above, we use databases as the basic building blocks for defining the states of the agents and the artifact system. We here fix the notation and terminology used. We refer to (Abite- boul, Hull, & Vianu, 1995) for more details on databases. Definition 2.1 (Database Schemas) A (relational) database schema is a set D = {P1/q1, . . . , Pn/qn} of relation symbols Pi, each associated with its arity qi ∈ N. Instances of database schemas are defined over interpretation domains. Definition 2.2 (Database Instances) Given an interpretation domain U and a database schema D, a D-instance over U is a mapping D associating each relation symbol Pi ∈ D with a finite qi-ary relation over U, i.e., D(Pi) ⊆ U qi. The set of all D-instances over an interpretation domain U is denoted by D(U ). We simply refer to "instances" whenever the database schema D is clear by the context. The active domain of an instance D, denoted as adom(D), is the set of all individuals in U occurring in some tuple of some predicate interpretation D(Pi). Observe that, since D contains a finite number of relation symbols and each D(Pi) is finite, so is adom(D). To fix the notation, we recall the syntax of first-order formulas with equality and no function symbols. Let V ar be a countable set of individual variables and C be a finite set of individual constants. A term is any element t ∈ V ar ∪ C. Definition 2.3 (FO-formulas over D) Given a database schema D, the formulas ϕ of the first- order language LD are defined by the following BNF grammar: ϕ ::= t = t′ Pi(t1, . . . , tqi) ¬ϕ ϕ → ϕ ∀xϕ where Pi ∈ D, t1, . . . , tqi is a qi-tuple of terms and t, t′ are terms. 7 We assume "=" to be a special binary predicate with fixed obvious interpretation. To summarise, LD is a first-order language with equality over the relational vocabulary D with no function symbols and with finitely many constant symbols from C. Observe that considering a finite set of constants is not a limitation. Indeed, since we will be working with finite sets of formulas, C can always be defined so as to be able to express any formula of interest. In the following we use the standard abbreviations ∃, ∧, ∨, and 6=. Also, free and bound variables are defined as standard. For a formula ϕ we denote the set of its variables as vars(ϕ), the set of its free variables as free(ϕ), and the set of its constants as const(ϕ). We write ϕ(~x) to list explicitly in arbitrary order all the free variables x1, . . . , xℓ of ϕ. By slight abuse of notation, we treat ~x as a set, thus we write ~x = free(ϕ). A sentence is a formula with no free variables. Given an interpretation domain U such that C ⊆ U, an assignment is a function σ : V ar 7→ U. For an assignment σ, we denote by σ(cid:0)x u(cid:1)(x) = u; and (ii) σ(cid:0)x u(cid:1)(x′) = σ(x′), for every x′ ∈ V ar different from x. For convenience, we extend assignments to constants so that σ(t) = t, if t ∈ C; that is, we assume a Herbrand interpretation of constants. We can now define the semantics of LD. u(cid:1) the assignment such that: (i) σ(cid:0)x Definition 2.4 (Satisfaction of FO-formulas) Given a D-instance D, an assignment σ, and an FO-formula ϕ ∈ LD, we inductively define whether D satisfies ϕ under σ, written (D, σ) = ϕ, as follows: (D, σ) = Pi(t1, . . . , tqi) (D, σ) = t = t′ (D, σ) = ¬ϕ (D, σ) = ϕ → ψ (D, σ) = ∀xϕ iff iff iff iff iff hσ(t1), . . . , σ(tqi)i ∈ D(Pi) σ(t) = σ(t′) it is not the case that (D, σ) = ϕ (D, σ) = ¬ϕ or (D, σ) = ψ A formula ϕ is true in D, written D = ϕ, iff (D, σ) = ϕ, for all assignments σ. for all u ∈ adom(D), we have that (D, σ(cid:0)x u(cid:1)) = ϕ Observe that we adopt an active-domain semantics, that is, quantified variables range only over the active domain of D. Also notice that constants are interpreted rigidly; so, two constants are equal if and only if they are syntactically the same. In the rest of the paper, we assume that every interpretation domain includes C. Also, as a usual shortcut, we write (D, σ) 6= ϕ to express that it is not the case that (D, σ) = ϕ. Finally, we introduce the ⊕ operator on D-instances that will be used later in the paper. Let the n/qn} obtained from D primed version of a database schema D be the schema D′ = {P ′ by syntactically replacing each predicate symbol Pi with its primed version P ′ 1/q1, . . . , P ′ i of the same arity. Definition 2.5 (⊕ Operator) Given two D-instances D and D′, we define D ⊕ D′ as the (D ∪ D′)- instance such that D ⊕ D′(Pi) = D(Pi) and D ⊕ D′(P ′ i ) = D′(Pi). Intuitively, the ⊕ operator defines a disjunctive join of the two instances, where relation symbols in D are interpreted according to D, while their primed versions are interpreted according to D′. 2.2 Artifact-Centric Multi-Agent Systems In the following we introduce the semantic structures that we will use throughout the paper. We define an artifact-centric multi-agent system as a system comprising an environment representing all interacting artifacts in the system and a finite set of agents interacting with such environment. 8 As agents have views of the artifact state, i.e., projections of the status of particular artifacts, we assume the building blocks of their private local states also to be modelled as database instances. In line with the interpreted systems semantics (Fagin et al., 1995) not everything in the agents' states needs to be present in the environment; a portion of it may be entirely private and not replicated in other agents' states. So, we start by introducing the notion of agent. Definition 2.6 (Agent) Given an interpretation domain U, an agent is a tuple A = hD, L, Act, P ri, where: • D is the local database schema; • L ⊆ D(U ) is the set of local states; • Act is the finite set of action types of the form α(~p), where ~p is the tuple of abstract parame- ters; • P r : L 7→ 2Act(U ) is the local protocol function, where Act(U ) is the set of ground actions of the form α(~u) where α(~p) ∈ Act and ~u ∈ U ~p is a tuple of ground parameters. Intuitively, at a given time each agent A is in some local state l ∈ D(U ) that represents all the information agent A has at its disposal. In this sense we follow (Fagin et al., 1995) but require that this information is structured as a database. Again, following standard literature we assume that the agents are autonomous and proactive and perform the actions in Act according to the protocol function P r. In the definition above we distinguish between "abstract parameters" to denote the language in which particular action parameters are given, and their concrete values or "ground parameters". We assume that the agents interact among themselves and with an environment comprising all artifacts in the system. As artifacts are entities involving both data and process, we can see them as collections of database instances paired with actions and governed by special protocols. Without loss of generality we can assume the environment state to be a single database instance including all artifacts in the system. From a purely formal point of view this allows us to represent the environment as a special agent. Of course, in any specific instantiation the environment and the agents will be rather different, exactly in line with the standard propositional version of interpreted systems. We can therefore define the synchronous composition of agents with the environment. Definition 2.7 (Artifact-Centric Multi-Agent Systems) Given an interpretation domain U and a set Ag = {A0, . . . , An} of agents Ai = hDi, Li, Acti, P rii defined on U, an artifact-centric multi- agent system (or AC-MAS) is a tuple P = hS, U, s0, τ i where: • S ⊆ L0 × · · · × Ln is the set of reachable global states; • U is the interpretation domain; • s0 ∈ S is the initial global state; • τ : S × Act(U ) 7→ 2S is the global transition function, where Act(U ) = Act0(U ) × · · · × Actn(U ) is the set of global (ground) actions, and τ (hl0, . . . , lni, hα0(~u0), . . . , αn(~un)i) is defined iff αi(~ui) ∈ P ri(li) for every i ≤ n. 9 As we will see in later sections, AC-MAS are the natural extension of interpreted systems to the first order to account for environments constituted of artifact-centric systems. They can be seen as a specialisation of quantified interpreted systems (Belardinelli & Lomuscio, 2012), a general extension of interpreted systems to the first-order case. In the formalisation above the agent A0 is referred to as the environment E. The environment includes all artifacts in the system as well as additional information to facilitate communication between the agents and the hub, e.g., messages in transit etc. At any given time an AC-MAS is described by a tuple of database instances, representing all the agents in the system as well as the artifact system. A single interpretation domain for all database schemas is given. Note that this does not break the generality of the representation as we can always extend the domain of all agents and the environment before composing them into a single AC-MAS. The global transition function defines the evolution of the system through synchronous composition of actions for the environment and all agents in the system. Much of the interaction we are interested in modelling involves message exchanges with pay- load, hence the action parameters, between agents and the environment, i.e., agents operating on the artifacts. However, note that the formalisation above does not preclude us from modelling agent-to- agent interactions, as the global transition function does not rule out successors in which only some agents change their local state following some actions. Also observe that essential concepts such as views are naturally expressed in AC-MAS by insisting that the local state of an agent includes part of the environment's, i.e., the artifacts the agent has access to. Not all AC-MAS need to have views defined, so it is also possible for the views to be empty. Other artifact-based concepts such as lifecycles are naturally expressed in AC-MAS. As artifacts are modelled as part of the environment, a lifecycle is naturally encoded in AC-MAS simply as the sequence of changes induced by the transition function τ on the fragment of the environment representing the lifecycle in question. We will show an example of this in Section 7. Some technical remarks now follow. To simplify the notation, we denote a global ground action as ~α(~u), where ~α = hα0(p0), . . . , αn(pn)i and ~u = h~u0, . . . , ~uni, with each ~ui of appropriate size. We define the transition relation → on S × S such that s → s′ if and only if there exists a ~α(~u) ∈ Act(U ) such that s′ ∈ τ (s, ~α(~u)). If s → s′, we say that s′ is a successor of s. A run r . from s ∈ S is an infinite sequence s0 → s1 → · · · , with s0 = s. For n ∈ N, we take r(n) = sn. A state s′ is reachable from s if there exists a run r from the global state r(0) = s such that r(i) = s′, for some i ≥ 0. We assume that the relation → is serial. This can be easily obtained by assuming that each agent has a skip action enabled at each local state and that performing skip induces no changes in any of the local states. We consider S to be the set of states reachable from the initial state s0. For convenience we will use also the concept of temporal-epistemic (t.e., for short) run. Formally a t.e. run r from a state s ∈ S is an infinite sequence s0 ❀ s1 ❀ . . . such that s0 = s and si → si+1 or si ∼k si+1, for some k ∈ Ag. A state s′ is said to be temporally-epistemically reachable (t.e. reachable, for short) from s if there exists a t.e. run r from the global state r(0) = s such that for some i ≥ 0 we have that r(i) = s′. Obviously, temporal-epistemic runs include purely temporal runs as a special case. As in plain interpreted systems (Fagin et al., 1995), we say that two global states s = hl0, . . . , lni ni are epistemically indistinguishable for agent Ai, written s ∼i s′, if li = l′ and s′ = hl′ i. Differently from interpreted systems the local equality is evaluated on database instances. Also, notice that we admit U to be infinite, thereby allowing the possibility of the set of states S to be 0, . . . , l′ 10 infinite. AC-MAS. Indeed, unless we specify otherwise, we will assume to be working with infinite-state Finally, for technical reasons it is useful to refer to a global database schema D = D0 ∪ · · · ∪ Dn of an AC-MAS. Every global state s = hl0, . . . , lni is associated with the (global) D-instance Ds ∈ D(U ) such that Ds(Pi) = Sj∈Ag lj(Pi), for Pi ∈ D. We omit the subscript s when s is clear from the context and we write adom(s) for adom(Ds). Notice that for every s ∈ S, the Ds associated with s is unique, while the converse is not true in general. 2.3 Model Checking We now define the problem of verifying an artifact-centric multi-agent system against a specification of interest. By following the artifact-centric model, we wish to give data the same prominence as processes. To deal with data and the underlying database instances, our specification language needs to include first-order logic. Further, we require temporal logic to describe the system execution. Lastly, we use epistemic logic to express the information the agents have at their disposal. Hence, we define a first-order temporal epistemic specification language to be interpreted on AC-MAS. The specification language will be used in Section 6 to formalise properties of artifact-centric programs. Definition 2.8 (The Logic FO-CTLK) The first-order CTLK (or FO-CTLK) formulas ϕ over a database schema D are inductively defined by the following BNF: ϕ ::= φ ¬ϕ ϕ → ϕ ∀xϕ AXϕ AϕU ϕ EϕU ϕ Kiϕ Cϕ where φ ∈ LD and 0 < i ≤ n. The notions of free and bound variables for FO-CTLK extend straightforwardly from LD, as well as functions vars, free, and const. As usual, the temporal formulas AXϕ and AϕU ϕ′ (resp. EϕU ϕ′) are read as "for all runs, at the next step ϕ" and "for all runs (resp. some run), ϕ until ϕ′". The epistemic formulas Kiϕ and Cϕ intuitively mean that "agent Ai knows ϕ" and "it is common knowledge among all agents that ϕ" respectively. We use the abbreviations EXϕ, AF ϕ, AGϕ, EF ϕ, and EGϕ as standard. Observe that free variables can occur within the scope of modal op- erators, thus allowing for the unconstrained alternation of quantifiers and modal operators, thereby allowing us to refer to elements in different modal contexts. We consider also a number of frag- ments of FO-CTLK. The sentence atomic version of FO-CTLK without epistemic modalities, or SA-FO-CTL, is the language obtained from Definition 2.8 by removing the clauses for epistemic operators and restricting atomic formulas to first-order sentences, so that no variable appears free in the scope of a modal operator: ϕ ::= φ ¬ϕ ϕ → ϕ AXϕ AϕU ϕ EϕU ϕ where φ ∈ LD is a sentence. We will consider also the language FO-ECTLK, i.e., the existential fragments of FO-CTLK, defined as follows: ϕ ::= φ ϕ ∧ ϕ ϕ ∨ ϕ ∀xϕ ∃xϕ EXϕ EϕU ϕ ¯Kiϕ ¯Cϕ, where φ ∈ LD, with ∧ and ∨ the standard abbreviations, ¯Kiϕ ≡ ¬Ki¬ϕ, and ¯Cϕ ≡ ¬C¬ϕ. The semantics of FO-CTLK formulas is defined as follows. 11 Definition 2.9 (Satisfaction for FO-CTLK) Consider an AC-MAS P, an FO-CTLK formula ϕ, a state s ∈ P, and an assignment σ. We inductively define whether P satisfies ϕ in s under σ, written (P, s, σ) = ϕ, as follows: (P, s, σ) = ϕ (P, s, σ) = ¬ϕ (P, s, σ) = ϕ → ϕ′ (P, s, σ) = ∀xϕ (P, s, σ) = AXϕ (P, s, σ) = AϕU ϕ′ (P, s, σ) = EϕU ϕ′ (P, s, σ) = Kiϕ (P, s, σ) = Cϕ iff iff iff iff iff iff iff iff iff for all u ∈ adom(s), (P, s, σ(cid:0)x u(cid:1)) = ϕ (Ds, σ) = ϕ, if ϕ is an FO-formula it is not the case that (P, s, σ) = ϕ (P, s, σ) = ¬ϕ or (P, s, σ) = ϕ′ for all runs r, if r(0) = s, then (P, r(1), σ) = ϕ for all runs r, if r(0) = s, then there is k ≥ 0 s.t. (P, r(k), σ) = ϕ′, and for all j, 0 ≤ j < k implies (P, r(j), σ) = ϕ for some run r, r(0) = s and there is k ≥ 0 s.t. (P, r(k), σ) = ϕ′, and for all j, 0 ≤ j < k implies (P, r(j), σ) = ϕ for all s′, s ∼i s′ implies (P, s′, σ) = ϕ for all s′, s ∼ s′ implies (P, s′, σ) = ϕ where ∼ is the transitive closure of S1...n ∼i. A formula ϕ is said to be true at a state s, written (P, s) = ϕ, if (P, s, σ) = ϕ for all assignments σ. Moreover, ϕ is said to be true in P, written P = ϕ, if (P, s0) = ϕ. A key concern in this paper is to explore the model checking of AC-MAS against first-order temporal-epistemic specifications. Definition 2.10 (Model Checking) Model checking an AC-MAS P against an FO-CTLK formula ϕ amounts to finding an assignment σ such that (P, s0, σ) = ϕ. It is easy to see that whenever U is finite the model checking problem is decidable as P is a finite- state system. In general this is not the case. Theorem 2.11 The model checking problem for AC-MAS w.r.t. FO-CTLK is undecidable. Proof (sketch). This can be proved by showing that every Turing machine T whose tape contains an initial input I can be simulated by an artifact system PT,I. The problem of checking whether T terminates on that particular input can be reduced to checking whether PT,I = ϕ, where ϕ encodes the termination condition. The detailed construction is similar to that of Theorem 4.10 of (Deutsch, Sui, & Vianu, 2007). Given the general setting in which the model checking problem is defined above, the negative result is not surprising. In the following we identify syntactic and semantic restrictions for which the problem is decidable. 2.4 Isomorphisms We now investigate the concept of isomorphism on AC-MAS. This will be needed in later sections to produce finite abstractions of infinite-state AC-MAS. In what follows let P = hS, U, s0, τ i and P ′ = hS ′, U ′, s′ 0, τ i be two AC-MAS. Definition 2.12 (Isomorphism) Two local states l, l′ ∈ D(U ) are isomorphic, written l ≃ l′, iff there exists a bijection ι : adom(l) ∪ C 7→ adom(l′) ∪ C such that: 12 l(P1) b a b d l(P2) a l′(P1) b c b e l′(P2) c l′′(P1) d f d e l′′(P2) f Figure 1: Examples of isomorphic and non-isomorphic local states. (i) ι is the identity on C; (ii) for every Pi ∈ D, ~u ∈ U qi, we have that ~u ∈ l(Pi) iff ι(~u) ∈ l′(Pi). When this is the case, we say that ι is a witness for l ≃ l′. Two global states s ∈ S and s′ ∈ S ′ are isomorphic, written s ≃ s′, iff there exists a bijection ι : adom(s) ∪ C 7→ adom(s′) ∪ C such that for every j ∈ Ag, ι is a witness for lj ≃ l′ j. Notice that isomorphisms preserve the constants in C as well as predicates in the local states up to renaming of the corresponding terms. Any function ι as above is called a witness for s ≃ s′. Obviously, the relation ≃ is an equivalence relation. Given a function f : U 7→ U ′ defined on adom(s), f (s) denotes the interpretation in D(U ′) obtained from s by renaming each u ∈ adom(s) as f (u). If f is also injective (thus invertible) and the identity on C, then f (s) ≃ s. Example. For an example of isomorphic states, consider an agent with local database schema D = {P1/2, P2/1}, let U = {a, b, c, . . .} be an interpretation domain, and fix the set C = {b} of constants. Let l be the local state such that l(P1) = {ha, bi, hb, di} and l(P2) = {a} (see Figure 1). Then, the local state l′ such that l′(P1) = {hc, bi, hb, ei} and l′(P2) = {c} is isomorphic to l. This can be easily seen by considering the isomorphism ι, where: ι(a) = c, ι(b) = b, and ι(d) = e. On the other hand, the state l′′ where l′′(P1) = {hf, di, hd, ei} and l′′(P2) = {f } is not isomorphic to l. Indeed, although a bijection exists that "transforms" l into l′′, it is easy to see that none can be such that ι′(b) = b. Note that, while isomorphic states have the same relational structure, two isomorphic states do not necessarily satisfy the same FO-formulas as satisfaction depends also on the values assigned to free variables. To account for this, we introduce the following notion. Definition 2.13 (Equivalent assignments) Given two states s ∈ S and s′ ∈ S ′, and a set of vari- ables V ⊆ V ar, two assignments σ : V ar 7→ U and σ′ : V ar 7→ U ′ are equivalent for V w.r.t. s and s′ iff there exists a bijection γ : adom(s) ∪ C ∪ σ(V ) 7→ adom(s′) ∪ C ∪ σ′(V ) such that: (i) γadom(s)∪C is a witness for s ≃ s′; (ii) σ′V = γ ◦ σV . Intuitively, equivalent assignments preserve both the (in)equalities of the variables in V and the constants in s, s′ up to renaming. Note that, by definition, the above implies that s, s′ are isomorphic. We say that two assignments are equivalent for an FO-CTLK formula ϕ, omitting the states s and s′ when it is clear from the context, if these are equivalent for free(ϕ). We can now show that isomorphic states satisfy exactly the same FO-formulas. Proposition 2.14 Given two isomorphic states s ∈ S and s′ ∈ S ′, an FO-formula ϕ, and two assignments σ and σ′ equivalent for ϕ, we have that (Ds, σ) = ϕ iff (Ds′, σ′) = ϕ 13 Proof. The proof is by induction on the structure of ϕ. Consider the base case for the atomic formula ϕ ≡ P (t1, . . . , tk). Then (Ds, σ) = ϕ iff hσ(t1), . . . , σ(tk)i ∈ Ds(P ). Since σ and σ′ are equivalent for ϕ, and s ≃ s′, this is the case iff hσ′(t1), . . . , σ′(tk)i ∈ Ds′(P ), that is, (Ds′, σ′) = ϕ. The base case for ϕ ≡ t = t′ is proved similarly, by observing that the satisfaction of ϕ depends only on the assignments, and that the function γ of Def. 2.13 is a bijection, thus all the (in)equalities between the values assigned by σ and σ′ are preserved. This is sufficient to guarantee that σ(t) = σ(t′) iff σ′(t) = σ′(t′). The inductive step for the propositional connectives is straightforward. Finally, if ϕ ≡ ∀yψ, then (Ds, σ) = ϕ iff for all u ∈ adom(s), (Ds, σ(cid:0)y u(cid:1)) = ψ. Now consider the witness ι = γadom(s)∪C for s ≃ s′, where γ is as in Def. 2.13. We have that σ(cid:0)y u(cid:1) and σ′(cid:0) y u(cid:1)) = ψ iff (Ds′, σ′(cid:0) y ι(u)(cid:1)) = ψ. Since ι is a bijection, this is the case iff for all u′ ∈ adom(s′), (Ds′, σ′(cid:0) y u′(cid:1)) = ψ, i.e., (Ds′, σ′) = ϕ. ι(u)(cid:1) are equivalent for ψ. By induction hypothesis (Ds, σ(cid:0)y This leads us to the following result. Corollary 2.15 Given two isomorphic states s ∈ S and s′ ∈ S ′ and an FO-sentence ϕ, we have that Ds = ϕ iff Ds′ = ϕ Proof. From right to left. Suppose, by contradiction, that Ds 6= ϕ. Then there exists an assignment σ s.t. (Ds, σ) 6= ϕ. Since free(ϕ) = ∅, if ι is a witness for s ≃ s′, then the assignment σ′ = ι ◦ σ is equivalent to σ for s and s′. By Proposition 2.14 we have that (Ds′, σ′) 6= ϕ, that is, Ds′ 6= ϕ. The case from left to right can be shown similarly. Thus, isomorphic states cannot be distinguished by FO-sentences. This enables us to use this notion when defining simulations as we will see in the next section. 3. Abstractions for Sentence Atomic FO-CTL In the previous section we have observed that model checking AC-MAS against FO-CTLK is un- decidable in general. So, it is clearly of interest to identify decidable settings. In what follows we introduce two main results. The first, presented in this section, identifies restrictions on the language; the second, presented in the next section, focuses on semantic constraints. While these cases are in some sense orthogonal to each other, we show that they both lead to decidable model checking problems. They are also both carried out on a rather natural subclass of AC-MAS that we call bounded, which we identify below. Our goal for proceeding in this manner is to identify finite abstractions of infinite-state AC-MAS so that verification of programs, that admit AC-MAS as models, can be conducted on them, rather than on infinite-state AC-MAS. We will see this in detail in Section 6. Given our aims we begin by defining a first notion of bisimulation in the context of AC-MAS. Bisimulations will be used to show that all bounded AC-MAS admit a finite, bisimilar, abstraction that satisifies the same SA-FO-CTL specifications as the original AC-MAS. Also in what follows we assume that P = hS, U, s0, τ i and P ′ = hS ′, U ′, s′ 0, τ ′i. Definition 3.1 (Simulation) A relation R ⊆ S × S ′ is a simulation iff hs, s′i ∈ R implies: 1. s ≃ s′; 14 2. for every t ∈ S, if s → t then there exists t′ ∈ S ′ s.t. s′ → t′ and ht, t′i ∈ R. Definition 3.1 presents the standard notion of simulation applied to the case of AC-MAS. The dif- ference from the propositional case is that we here insist on the states being isomorphic, a generali- sation from the usual requirement for propositional valuations to be equal (Blackburn, de Rijke, & Venema, 2001). As in the standard case, two states s ∈ S and s′ ∈ S ′ are said to be similar, written s (cid:22) s′, if there exists a simulation relation R s.t. hs, s′i ∈ R. It can be proven that the similarity relation (cid:22) is a simulation itself, and in particular the largest one w.r.t. set inclusion, and that it is transitive and reflexive. Finally, we say that P ′ simulates P, written P (cid:22) P ′, if s0 (cid:22) s′ 0. We extend the above to bisimulations. Definition 3.2 (Bisimulation) A relation B ⊆ S × S ′ is a bisimulation iff both B and B−1 = {hs′, si hs, s′i ∈ B} are simulations. We say that two states s ∈ S and s′ ∈ S ′ are bisimilar, written s ≈ s′, if there exists a bisimulation B s.t. hs, s′i ∈ B. Similarly to simulations, it can be proven that the bisimilarity relation ≈ is the largest bismulation. Further, it is an equivalence relation. Finally, P and P ′ are said to be bisimilar, written P ≈ P ′, if s0 ≈ s′ 0. Since, as shown in Proposition 2.15, the satisfaction of FO-sentences is invariant under iso- morphisms, we can now extend the usual bisimulation result from the propositional case to that of SA-FO-CTL. We begin by showing a result on bisimilar runs. Proposition 3.3 Consider two AC-MAS P and P ′ such that P ≈ P ′, s ≈ s′, for some s ∈ S, s′ ∈ S ′, and a run r of P such that r(0) = s. Then there exists a run r′ of P ′ such that: (i) r′(0) = s′; (ii) for all i ≥ 0, r(i) ≈ r′(i). Proof. We show by induction that such run r′ in P ′ exists. For i = 0, let r′(0) = s′. Obviously, r(0) ≈ r′(0). Now, assume, by induction hypothesis, that r(i) ≈ r′(i). Let r(i) → r(i + 1). Since r(i) ≈ r′(i), by Def. 3.1, there exists t′ ∈ S ′ such that r′(i) → t′ and r(i + 1) ≈ t′. Let r′(i + 1) = t′; hence we obtain r(i + 1) ≈ r′(i + 1). By definition r′ is a run of P ′. This enables us to show that bisimilar AC-MAS preserve SA-FO-CTL formulas. This is an extension of analogous results on propositional CTL. Lemma 3.4 Consider the AC-MAS P and P ′ such that P ≈ P ′, s ≈ s′, for some s ∈ S, s′ ∈ S ′ and an SA-FO-CTL formula ϕ. Then, (P, s) = ϕ iff (P ′, s′) = ϕ Proof. The proof is by induction on the structure of ϕ. Observe first that since ϕ is sentence- atomic, its satisfaction does not depend on assignments. We report the proof for the left-to-right part of the implication; the converse can be shown similarly. The base case for an FO-sentence ϕ follows from Prop. 2.15. The inductive cases for proposi- tional connectives are straightforward. 15 For ϕ ≡ AXψ, assume for contradiction that (P, s) = ϕ and (P ′, s′) 6= ϕ. Then, there exists a run r′ s.t. r′(0) = s′ and (P ′, r′(1)) 6= ψ. By Def. 3.2 and 3.1 there exists a t ∈ S s.t. s → t and t ≈ r′(1). Further, by seriality of →, s → t can be extended to a run r s.t. r(0) = s and r(1) = t. By the induction hypothesis we obtain that (P, r(1)) 6= ψ. Hence, (P, r(0)) 6= AXψ, which is a contradiction. For ϕ ≡ EψU φ, let r be a run with r(0) = s such that there exists k ≥ 0 such that (P, r(k)) = φ, and for every j, 0 ≤ j < k implies (P, r(j)) = ψ. By Prop. 3.3 there exists a run r′ s.t. r′(0) = s′ and for all i ≥ 0, r′(i) ≈ r(i). By the induction hypothesis we have that for each i ∈ N, (P, r(i)) = ψ iff (P ′, r′(i)) = ψ, and (P, r(i)) = φ iff (P ′, r′(i)) = φ. Therefore, r′ is a run s.t. r′(0) = s′, (P ′, r′(k)) = φ, and for every j, 0 ≤ j < k implies (P ′, r′(j)) = ψ, i.e., (P ′, s′) = EψU φ. For ϕ ≡ AψU φ, assume for contradiction that (P, s) = ϕ and (P ′, s′) 6= ϕ. Then, there exists a run r′ s.t. r′(0) = s′ and for every k ≥ 0, if (P ′, r′(k)) = φ, then there exists j s.t. 0 ≤ j < k and (P ′, r′(j)) 6= ψ. By Prop. 3.3 there exists a run r s.t. r(0) = s and for all i ≥ 0, r(i) ≈ r′(i). Further, by the induction hypothesis we have that (P, r(i)) = ψ iff (P ′, r′(i)) = ψ and (P, r(i)) = φ iff (P ′, r′(i)) = φ. But then r is s.t. r(0) = s and for every k ≥ 0, if (P, r(k)) = φ, then there exists j s.t. 0 ≤ j < k and (P, r(j)) 6= ψ. That is, (P, s) 6= AψU φ, which is a contradiction. By applying the result above to the case of s = s0 and s′ = s′ 0, we obtain the following. Theorem 3.5 Consider the AC-MAS P and P ′ such that P ≈ P ′, and an SA-FO-CTL formula ϕ. We have P = ϕ iff P ′ = ϕ In summary we have proved that bisimilar AC-MAS validate the same SA-FO-CTL formulas. In the next section we use this result to reduce, under additional assumptions, the verification of an infinite-state AC-MAS to that of a finite-state one. 3.1 Finite Abstractions of Bisimilar AC-MAS We now define a notion of finite abstraction for AC-MAS. We prove that abstractions are bisimilar to the corresponding concrete model. We are particularly interested in finite abstraction; so we operate on a special class of infinite models that we call bounded. Definition 3.6 (Bounded AC-MAS) An AC-MAS P is b-bounded, for b ∈ N, if for all s ∈ S, adom(s) ≤ b. An AC-MAS is b-bounded if none of its reachable states contains more than b distinct elements. Observe that bounded AC-MAS may be defined on infinite domains U. Furthermore, note that a b- bounded AC-MAS may contain infinitely many states, all bounded by b. So b-bounded systems are infinite-state in general. Notice also that the value b bounds only the number of distinct individuals in a state, not the size of the state itself, i.e., the amount of memory required to accommodate the individuals. Indeed, the infinitely many elements of U need an unbounded number of bits to be represented (e.g., as finite strings), so, even though each state is guaranteed to contain at most b distinct elements, nothing can be said about how large the actual space required by such elements is. On the other hand, it should be clear that memory-bounded AC-MAS are finite-state (hence b-bounded, for some b). 16 Thus, seen as programs, b-bounded AC-MAS are in general memory-unbounded. Therefore, for the purpose of verification, they cannot be trivially checked by generating all their executions -- as it would be the case if they were memory-bounded -- like standard model checking techniques typically do. However, we will show later that any b-bounded infinite-state ACMAS admits a finite abstraction which can be used to verify it. We now introduce abstractions in a modular manner by first introducing a set of abstract agents from a concrete AC-MAS. Definition 3.7 (Abstract agent) Let A = hD, L, Act, P ri be an agent defined on the interpretation domain U. Given a set U ′ of individuals, we define the abstract agent A′ = hD′, L′, Act′, P r′i on U ′ such that: 1. D′ i = Di; i ⊆ D′ 2. L′ 3. Act′ i(U ′); i = Acti; 4. α(~u′) ∈ P r′ i(l′ i) iff there exist li ∈ Li and α(~u) ∈ P ri(li) s.t. l′ i ≃ li, for some witness ι, and ~u′ = ι′(~u), for some bijection ι′ extending ι to ~u. Given a set Ag of agents defined on U, let Ag′ be the set of the corresponding abstract agents on U ′. We remark that A′, as defined in Definition 3.7, is indeed an agent and complies with Defini- tion 2.6. Notice that the protocol of A′ is defined on the basis of its corresponding concrete agent A and requires the existence of a bijection between the elements in the local states and the action parameters. Thus, in order for a ground action of A to have a counterpart in A′, the last requirement of Definition 3.7 constrains U ′ to contain a sufficient number of distinct values. As it will become apparent later, the size of U ′ determines how closely an abstract system can simulate its concrete counterpart. We can now formalize the notion of abstraction that we will use in this section. Definition 3.8 (Abstraction) Let P be an AC-MAS over Ag and Ag′ the set of agents obtained as in Definition 3.7, for some U ′. The AC-MAS P ′ defined over Ag′ is said to be an abstraction of P iff: • s′ 0 ≃ s0; • t′ ∈ τ ′(s′, ~α(~u′)) for some ~α(~u′) ∈ Act(U ′) iff there exist s, t ∈ S and ~α(~u) ∈ Act(U ), such that t ∈ τ (s, ~α(~u)), s ≃ s′ and t ≃ t′ for some witness ι, and ~u′ = ι′(~u) for some ι′ extending ι. Notice that abstractions have initial states isomorphic to their concrete counterparts. The con- dition in Definition 3.8 means that whenever s ≃ s′ for some witness ι, ~u′ = ι(~u), t ∈ τ (s, α(~u)) and t′ ∈ τ (s′, α(~u′)), then t ≃ t′. This constraint means that action are data-independent. So, for example, a copy action in the concrete model has a corresponding copy action in the abstract model regardless of the data that are copied. Crucially, this condition requires that the domain U ′ contains enough elements to simulate the concrete states and action effects as the following result makes precise. In what follows we take NAg = NAg′ = PAi∈Ag maxα(~p)∈Acti{~p}, i.e., NAg is the sum of the maximum numbers of parameters contained in the action types of each agent in Ag. 17 Theorem 3.9 Consider a b-bounded AC-MAS P over an infinite interpretation domain U, an SA- FO-CTLK formula ϕ, and a finite interpretation domain U ′ such that C ⊆ U ′ and U ′ ≥ b + C + NAg. Any abstraction P ′ of P is bisimilar to P. 0i ∈ R. Observe first that s′ 0 ≃ s0, so hs0, s′ Proof. Define a relation R as R = {hs, s′i ∈ S × S ′ s ≃ s′}. We show that R is a bisimulation such that hs0, s′ 0i ∈ R. Next, consider s ∈ S and s′ ∈ S ′ such that s ≃ s′ (i.e., hs, s′i ∈ R), and assume that s → t, for some t ∈ S. Then, there exists α(~u) ∈ Act(U ) s.t. t ∈ τ (s, α(~u)). We show next that there exists t′ ∈ S ′ s.t. s′ → t′ and t ≃ t′. To this end, observe that, since U ′ ≥ b + C and adom(t) ≤ b, we can define an injective function f : adom(t) ∪ C 7→ U ′ such that f (t) ≃ t. We take t′ = f (t); it remains to prove that s′ → t′. By the condition on the cardinality of U ′ we can extend f to ~u as well, and set ~u′ = f (~u). Then, by the definition of P ′ we have that t′ ∈ τ ′(s′, α(~u′)). Hence, s′ → t′. So, R is a simulation relation between P and P ′. Since R−1 can similarly be shown to be a simulation, it follows that P and P ′ are bisimilar. By combining this result with Lemma 3.4, we can easily derive the main result of this section. Theorem 3.10 If P is a b-bounded AC-MAS over an infinite interpretation domain U, and P ′ an abstraction of P over a finite interpretation domain U ′ such that C ⊆ U ′ and U ′ ≥ b + C + NAg, then for every SA-FO-CTLK formula ϕ, we have that P = ϕ iff P ′ = ϕ. This result states that we can reduce the verification of an infinite AC-MAS to the verification of a finite one. Given the fact that checking a finite AC-MAS is decidable, this is a noteworthy result. Note, however, that we do not have a constructive definition for the construction of an abstract AC-MAS P ′ from a concrete AC-MAS P. This is of no consequence though, as in practice any concrete artifact-system will be defined by a program, e.g., in the language GSM, as discussed in the introduction. Of importance, instead, is to be able to derive finite abstractions not just for arbitrary AC-MAS but for those that are models of concrete programs. We will do this in Section 6 where we will use the result above. Observe that an abstract AC-MAS as in Definition 3.8 depends on the set Ag′ of abstract agents defined in Definition 3.7. However, other abstract AC-MAS defined on different sets of agents, exist. This is a standard outcome when defining modular abstractions, as the same system can be obtained by considering different agent components. 4. Abstractions for FO-CTLK In the previous section we showed that syntactical restrictions on the specification language lead to finite abstractions for bounded AC-MAS. A natural question that arises is whether the limitation to sentence-atomic specifications can be removed. Doing so would enable us to check any agent-based FO-CTLK specification not on an infinite-state AC-MAS, but on its finite abstraction. The key concept we identify in this section that enables us to achieve the above is that of unifor- mity. As we will see later uniform AC-MAS are systems for which the behaviour does not depend on the actual data present in the states. This means that the system contains all possible transitions that are enabled according to parametric action rules, thereby resulting in a rather "full" transition 18 relation. This notion corresponds to that of genericity in databases (Abiteboul et al., 1995). We use the term "uniformity" as we refer to transition systems and not databases. To achieve finite abstractions we proceed as follows. We first introduce a notion of bisimulation stronger than the one discussed in the previous section. In Subsection 4.1 we show that this new bisimulation relation guarantees that uniform AC-MAS satisfy the same formulas in FO-CTLK. We use this result to show that bounded, uniform systems admit finite abstractions (Subsection 4.2). In the rest of the section we let P = hS, U, s0, τ i and P ′ = hS ′, U ′, s′ and assume, unless stated differently, that s = hl0, . . . , lni ∈ S, and s′ = hl′ 0, τ ′i be two AC-MAS 0, . . . , l′ ni ∈ S ′. 4.1 ⊕-Bisimulation Plain bisimulations are known to be satisfaction preserving in a modal propositional setting (Black- burn et al., 2001). In the following we explore the conditions under which this applies to AC-MAS as well. We begin by using a notion of bisimulation which is also based on isomorphism, but it is stronger than the one discussed in Section 3 and later explore its properties in the context of uniform AC-MAS. Definition 4.1 (⊕-Simulation) A relation R on S × S ′ is a ⊕-simulation if hs, s′i ∈ R implies: 1. s ≃ s′; 2. for every t ∈ S, if s → t then there exists t′ ∈ S ′ s.t. s′ → t′, s ⊕ t ≃ s′ ⊕ t′, and ht, t′i ∈ R; 3. for every t ∈ S, for every 0 < i ≤ n, if s ∼i t then there exists t′ ∈ S ′ s.t. t ∼i t′, s ⊕ t ≃ s′ ⊕ t′, and ht, t′i ∈ R. Observe that Definition 4.1 differs from Definition 3.1 not only by adding a condition for the epis- temic relation, but also by insisting that s ⊕ t ≃ s′ ⊕ t′. This condition ensures that the ⊕-similar transitions in AC-MAS have isomorphic disjoint unions. Two states s ∈ S and s′ ∈ S ′ are said to be ⊕-similar, iff there exists an ⊕-simulation R s.t. hs, s′i ∈ R. Note that all ⊕-similar states are isomorphic as condition 2. above ensures that t ≃ t′. We use the symbol (cid:22) both for similarity and ⊕-similarity, as the context will disambiguate. Also ⊕-similarity can be shown to be the largest ⊕-simulation, reflexive, and transitive. Further, we say that P ′ ⊕-simulates P if s0 (cid:22) s′ 0. ⊕-simulations can naturally be extended to ⊕-bisimulations. Definition 4.2 (⊕-Bisimulation) A relation B on S × S ′ is a ⊕-bisimulation iff both B and B−1 = {hs′, si hs, s′i ∈ B} are ⊕-simulations. Two states s ∈ S and s′ ∈ S ′ are said to be ⊕-bisimilar iff there exists an ⊕-bisimulation B such that hs, s′i ∈ B. Also for bisimilarity and ⊕-bisimilarity, we use the same symbol, ≈, and can prove that ≈ is the largest ⊕-bisimulation, and an equivalence relation. We say that P and P ′ are ⊕-bisimilar, written P ≈ P ′ iff so are s0 and s′ 0. While we observed in the previous section that bisimilar, hence isomorphic, states in bisimilar systems preserve sentence atomic formulas, it is instructive to note that this is not the case when full FO-CTLK formulas are considered. Example. Consider Figure 2, where C = ∅ and P and P ′ are given as follows. For the number 0(P ) = {1}; n of agents equal to 1, we define D = D′ = {P/1} and U = N; s0(P ) = s′ 19 3 4 5 P 1 P ′ 1 2 2 Figure 2: ⊕-Bisimilar AC-MAS not satisfying the same FO-CTLK formulas. τ = {hs, s′i s(P ) = {i}, s′(P ) = {i + 1}}; τ ′ = {hs, s′i s(P ) = {i}, s′(P ) = {(i + 1) mod 2}}. Notice that S ⊆ D(N) and S ′ ⊆ D(N). Clearly we have that P ≈ P ′. Now, consider the constant-free FO-CTLK formula ϕ = AG(∀x(P (x) → AXAG¬P (x))). It can be easily seen that P = ϕ while P ′ 6= ϕ. The above shows that ⊕-bisimilarity is not a sufficient condition to guarantee preservation of the satisfaction of FO-CTLK formulas. Intuitively, this is a consequence of the fact that ⊕-bisimilar AC-MAS do not preserve value associations along runs. For instance, the value 1 in P ′ is infinitely many times associated with the odd values occurring in P. By quantifying across states we are able to express this fact and are therefore able to distinguish the two structures. This is a difficulty as, intuitively, we would like to use ⊕-bisimulations to demonstrate the existence of finite abstractions. Indeed, as we will show later, this happens for the class of uniform AC-MAS, defined below. Definition 4.3 (Uniformity) An AC-MAS P is said to be uniform iff for every s, t, s′ ∈ S, t′ ∈ D(U ), 1. if t ∈ τ (s, ~α(~u)) and s ⊕ t ≃ s′ ⊕ t′ for some witness ι, then for every constant-preserving bijection ι′ that extends ι to ~u, we have that t′ ∈ τ (s′, ~α(ι′(~u))); 2. if s ∼i t and s ⊕ t ≃ s′ ⊕ t′, then s′ ∼i t′. This definition captures the idea that actions take into account and operate only on the relational structure of states and action parameters, irrespectively of the actual data they contain (apart from a finite set of constants). Intuitively, it says that if t can be obtained by executing α(~u) in s, and we replace in s, ~u and t, the same element v with v′, obtaining, say, s′, ~u′ and t′, then t′ can be obtained by executing α(~u′) in s′. In terms of the underlying Kripke structures this means that the systems are "full" up to ⊕, i.e., in all uniform AC-MAS the points t′ identified above are indeed part of the system and reachable from s′. A similar condition is required on the epistemic relation. A useful property of uniform systems is the fact that the latter requirement is implied by the former, as shown by the following result. Proposition 4.4 If an AC-MAS P satisfies req. 1 in Def. 4.3 and adom(s0) ⊆ C, then req. 2 is also satisfied. Proof. If s ⊕ t ≃ s′ ⊕ t′, then there is a witness ι : adom(s) ∪ adom(t) ∪ C 7→ adom(s′) ∪ adom(t′) ∪ C that is the identity on C (hence on adom(s0)). Assume s ∼i t, thus li(s) = li(t), and li(s′) = ι(li(s)) = ι(li(t)) = li(t′). Notice that this does not guarantee that s′ ∼i t′, 20 as we need to prove that t′ ∈ S. This can be done by showing that t′ is reachable from s0. Since t is reachable from s0, there exists a run s0 → s1 → . . . → sk s.t. sk = t. Extend now ι to a total and injective function ι′ : adom(s0) ∪ · · · ∪ adom(sk) ∪ C 7→ U. This can always be done because U ≥ adom(s0) ∪ · · · ∪ adom(sk) ∪ C. Now consider the sequence ι′(s0), ι′(s1), . . . , ι′(sk). Since adom(s0) ⊆ C then ι(s0) = s0 and, because ι′ extends ι, we have that ι′(s0) = ι(s0) = s0. Further, ι′(sk) = ι(t) = t′. By repeated applications of req. 1 we can show that ι′(sm+1) ∈ τ (ι′(sm), ~α(ι′(~u))) whenever sm+1 ∈ τ (sm, ~α(~u)), for m < k. Hence, the sequence is actually a run from s0 to t′. Thus, t′ ∈ S, and s′ ∼i t′. Thus, as long as adom(s0) ⊆ C, to check whether an AC-MAS is uniform, it is sufficient to take into account only the transition function. A further distinctive feature of uniform systems is that all isomorphic states are ⊕-bisimilar. Proposition 4.5 If an AC-MAS P is uniform, then for every s, s′ ∈ S, s ≃ s′ implies s ≈ s′. Proof. We prove that B = {hs, s′i ∈ S × S s ≃ s′} is a ⊕-bisimulation. Observe that since ≃ is an equivalence relation, so is B. Thus B is symmetric and B = B−1. Therefore, proving that B is a ⊕-simulation proves also that B−1 is a ⊕-simulation; hence, that B is a ⊕-bisimulation. To this end, let hs, s′i ∈ B, and assume s → t for some t ∈ S. Then, t ∈ τ (s, α(~u)) for some α(~u) ∈ Act(U ). Consider a witness ι for s ≃ s′. By cardinality considerations ι can be extended to a total and injective function ι′ : adom(s) ∪ adom(t) ∪ {~u} ∪ C 7→ U. Consider ι′(t) = t′; it follows that ι′ is a witness for s ⊕ t ≃ s′ ⊕ t′. Since P is uniform, t′ ∈ τ (s′, α(ι′(~u))), that is, s′ → t′. Moreover, ι′ is a witness for t ≃ t′, thus ht, t′i ∈ B. Next assume that hs, s′i ∈ B and s ∼i t, for some t ∈ S. By reasoning as above we can find a witness ι for s ≃ s′, and an extension ι′ of ι s.t. t′ = ι′(t) and ι′ is a witness for s ⊕ t ≃ s′ ⊕ t′. Since P is uniform, s′ ∼i t′ and ht, t′i ∈ B. This result intuitively means that submodels generated by isomorphic states are ⊕-bisimilar. Next we prove some partial results, which will be useful in proving our main preservation theo- rem. The first two results guarantee that under appropriate cardinality constraints the ⊕-bisimulation preserves the equivalence of assignments w.r.t. a given FO-CTLK formula. Lemma 4.6 Consider two ⊕-bisimilar and uniform AC-MAS P and P ′, two ⊕-bisimilar states s ∈ S and s′ ∈ S ′, and an FO-CTLK formula ϕ. For every assignments σ and σ′ equivalent for ϕ w.r.t. s and s′, we have that: 1. for every t ∈ S s.t. s → t, if U ′ ≥ adom(s) ∪ adom(t) ∪ C ∪ σ(free(ϕ)), then there exists t′ ∈ S ′ s.t. s′ → t′, t ≈ t′, and σ and σ′ are equivalent for ϕ w.r.t. t and t′. 2. for every t ∈ S s.t. s ∼i t, if U ′ ≥ adom(s) ∪ adom(t) ∪ C ∪ σ(free(ϕ)), then there exists t′ ∈ S ′ s.t. s′ ∼i t′, t ≈ t′, and σ and σ′ are equivalent for ϕ w.r.t. t and t′. Proof. To prove (1), let γ be a bijection witnessing that σ and σ′ are equivalent for ϕ w.r.t. s and s′. Suppose that s → t. Since s ≈ s′, by definition of ⊕-bisimulation there exists t′′ ∈ S ′ s.t. s′ → t′′, s ⊕ t ≃ s′ ⊕ t′′, and t ≈ t′′. Now, define Domj . = adom(s) ∪ adom(t) ∪ C, and partition it into: • Domγ . = adom(s) ∪ C ∪ (adom(t) ∩ σ(free(ϕ)); • Domι′ . = adom(t) \ Domγ. 21 Let ι′ : Domι′ 7→ U ′ \ Im(γ) be an invertible (total) function. Observe that Im(γ) = adom(s′) ∪ C ∪ σ′(free(ϕ)) = adom(s) ∪ C ∪ σ(free(ϕ)), thus from the fact that U ′ ≥ adom(s) ∪ adom(t) ∪ C ∪ σ(free(ϕ)) we have U ′ \ Im(γ) ≥ Dom(ι′), which guarantees the existence of ι′. Next, define j : Domj 7→ U ′ as follows: j(u) = (cid:26) γ(u), if u ∈ Domγ ι′(u), if u ∈ Domι′ Obviously, j is invertible. Thus, j is a witness for s ⊕ t ≃ s′ ⊕ t′, where t′ = j(t). Since s ⊕ t ≃ s′ ⊕ t′′ and ≃ is an equivalence relation, s′ ⊕ t′ ≃ s′ ⊕ t′′. Thus, s′ → t′, as P ′ is uniform. Moreover, σ and σ′ are equivalent for ϕ w.r.t. t and t′, by construction of t′. To check that t ≈ t′, observe that, since t′ ≃ t′′ and P ′ is uniform, by Prop. 4.5 it follows that t′ ≈ t′′. Thus, since t ≈ t′′ and ≈ is transitive, we obtain that t ≈ t′. The proof for (2) has an analogous structure and is omitted. It can be proven that this result is tight, i.e., that if the cardinality requirement is violated, there exist cases where assignment equivalence is not preserved along temporal or epistemic transitions. Lemma 4.6 easily generalizes to t.e. runs. Lemma 4.7 Consider two ⊕-bisimilar and uniform AC-MAS P and P ′, two ⊕-bisimilar states s ∈ S and s′ ∈ S ′, an FO-CTLK formula ϕ, and two assignments σ and σ′ equivalent for ϕ w.r.t. s and s′. For every t.e. run r of P, if r(0) = s and for all i ≥ 0, U ′ ≥ adom(r(i)) ∪ adom(r(i + 1)) ∪ C ∪ σ(free(ϕ)), then there exists a t.e. run r′ of P ′ s.t. for all i ≥ 0: (i) r′(0) = s′; (ii) r(i) ≈ r′(i); (iii) σ and σ′ are equivalent for ϕ w.r.t. r(i) and r′(i). (iv) for every i ≥ 0, if r(i) → r(i + 1) then r′(i) → r′(i + 1), and if r(i) ∼j r(i + 1), for some j, then r′(i) ∼j r′(i + 1). Proof. Let r be a t.e. run s.t. U ′ ≥ adom(r(i)) ∪ adom(r(i + 1)) ∪ C ∪ σ(free(ϕ)) for all i ≥ 0. We inductively build r′ and show that the conditions above are satisfied. For i = 0, let r′(0) = s′. By hypothesis, r is s.t. U ′ ≥ adom(r(0)) ∪ adom(r(1)) ∪ C ∪ σ(free(ϕ)). Thus, since r(0) ❀ r(1), by Lemma 4.6 there exists t′ ∈ S ′ s.t. r′(0) ❀ t′, r(1) ≈ t′, and σ and σ′ are equivalent for ϕ w.r.t. r(1) and t′. Let r′(1) = t′. Lemma 4.6 guarantees that the transitions r′(0) ❀ t′ and r(0) ❀ r(1) can be chosen so that they are either both temporal or both epistemic with the same index. The case for i > 0 is similar. Assume that r(i) ≈ r′(i) and σ and σ′ are equivalent for ϕ w.r.t. r(i) and r′(i). Since r(i) ❀ r(i+1) and U ′ ≥ adom(r(i))∪adom(r(i+1))∪C∪σ(free(ϕ)), by Lemma 4.6 there exists t′ ∈ S ′ s.t. r′(i) ❀ t′, σ and σ′ are equivalent for ϕ w.r.t. r(i + 1) and t′, and r(i + 1) ≈ t′. Let r′(i + 1) = t′. It is clear that r′ is a t.e. run in P ′, and that, by Lemma 4.6, the transitions of r′ can be chosen so as to fulfill requirement (iv). We can now prove the following result, which states that FO-CTLK formulas cannot distin- guish ⊕-bisimilar and uniform AC-MAS. This is in marked contrast with the earlier example in this section which operated on ⊕-bisimilar but non-uniform AC-MAS. 22 Theorem 4.8 Consider two ⊕-bisimilar and uniform AC-MAS P and P ′, two ⊕-bisimilar states s ∈ S and s′ ∈ S ′, an FO-CTLK formula ϕ, and two assignments σ and σ′ equivalent for ϕ w.r.t. s and s′. If 1. for every t.e. run r s.t. r(0) = s, for all k ≥ 0 we have U ′ ≥ adom(r(k)) ∪ adom(r(k + 1)) ∪ C ∪ σ(free(ϕ)) + vars(ϕ) \ free(ϕ); and 2. for every t.e. run r′ s.t. r′(0) = s′, for all k ≥ 0 we have U ≥ adom(r′(k)) ∪ adom(r′(k + 1)) ∪ C ∪ σ′(free(ϕ)) + vars(ϕ) \ free(ϕ); then (P, s, σ) = ϕ iff (P ′, s′, σ′) = ϕ. Proof. The proof is by induction on the structure of ϕ. We prove that if (P, s, σ) = ϕ then (P ′, s′, σ′) = ϕ. The other direction can be proved analogously. The base case for atomic formulas follows from Prop. 2.14. The inductive cases for propositional connectives are straightforward. For ϕ ≡ ∀xψ, assume that x ∈ free(ψ) (otherwise consider ψ, and the corresponding case), and no variable is quantified more than once (otherwise rename the other variables). Let γ be a bijection witnessing that σ and σ′ are equivalent for ϕ w.r.t. s and s′. For u ∈ adom(s), consider the assignment σ(cid:0)x free(ϕ)∪{x}; so σ(cid:0)x u(cid:1). By definition, γ(u) ∈ adom(s′), and σ′(cid:0) x γ(u)(cid:1) are equivalent for ψ w.r.t. s and s′. Moreover, σ(cid:0)x γ(u)(cid:1) is well-defined. Note that free(ψ) = u(cid:1)(free(ψ)) ≤ σ(free(ϕ)) + 1, as u may not occur in σ(free(ϕ)). The same considerations apply to σ′. Further, vars(ψ) \ free(ψ) = vars(ϕ) \ free(ϕ) − 1, as vars(ψ) = vars(ϕ), free(ψ) = free(ϕ) ∪ {x}, and x /∈ free(ϕ). Thus, both hypotheses 1. and 2. remain satisfied if we replace ϕ with ψ, σ u(cid:1) and σ′(cid:0) x u(cid:1)) = ψ then with σ(cid:0)x (P ′, s′, σ′(cid:0) x u(cid:1), and σ′ with σ′(cid:0) x γ(u)(cid:1)) = ψ. Since u ∈ adom(s) is generic and γ is a bijection, the result follows. γ(u)(cid:1). Therefore, by the induction hypothesis, if (P, s, σ(cid:0)x For ϕ ≡ AXψ, assume by contradiction that (P, s, σ) = ϕ but (P ′, s′, σ′) 6= ϕ. Then, there exists a run r′ s.t. r′(0) = s′ and (P ′, r′(1), σ′) 6= ψ. By Lemma 4.7, which applies as vars(ϕ) \ free(ϕ) ≥ 0, there exists a run r s.t. r(0) = s, for all i ≥ 0, r(i) ≈ r′(i) and σ and σ′ are equivalent for ψ w.r.t. r(i) and r′(i). Since r is a run s.t. r(0) = s, it satisfies hypothesis 1. Moreover, the same hypothesis is necessarily satisfied by all the t.e. runs r′′ s.t., , for some i ≥ 0, r′′(0) = r(i) (otherwise, the t.e. run r(0) · · · r(i)r′′(1)r′′(2) · · · would not satisfy the hypothesis); the same considerations apply w.r.t hypothesis 2 and for all the t.e. runs r′′′ s.t. r′′′(0) = r′(i), for some i ≥ 0. In particular, these hold for i = 1. Thus, we can inductively apply the Lemma, by replacing s with r(1), s′ with r′(1), and ϕ with ψ (observe that vars(ϕ) = vars(ψ) and free(ϕ) = free(ψ)). But then we obtain (P, r(1), σ) 6= ψ, thus (P, r(0), σ) 6= AXψ. This is a contradiction. For ϕ ≡ EψU φ, assume that the only variables common to ψ and φ occur free in both formulas (otherwise rename the quantified variables). Let r be a run s.t. r(0) = s, and there exists k ≥ 0 s.t. (P, r(k), σ) = φ, and (P, r(j), σ) = ψ for 0 ≤ j < k. By Lemma 4.7 there exists a run r′ s.t. r′(0) = s′, and for all i ≥ 0, r′(i) ≈ r(i), and σ and σ′ are equivalent for ϕ w.r.t. r′(i) and r(i). From each bijection γi witnessing that σ and σ′ are equivalent for ϕ w.r.t. r′(i) and r(i), define the bijections γi,ψ = γiadom(r(i))∪C∪σ(free(ψ)) and γi,φ = γiadom(r(i))∪C∪σ(free(φ)). Since free(ψ) ⊆ free(ϕ), free(φ) ⊆ free(ϕ), it can be seen that γi,ψ and γi,φ witness that σ and σ′ are equivalent for respectively ψ and φ w.r.t. r′(i) and r(i). By the same argument used for the AX case above, 23 hypothesis 1 holds for all the t.e. runs r′′ s.t. r′′(0) = r(i), for some i ≥ 0, and hypothesis 2 holds for all the t.e. runs r′′′ s.t. r′′′(0) = r′(i). Now observe that σ(free(φ)), σ(free(ψ)) ≤ σ(free(ϕ)). Moreover, by the assumption on the common variables of ψ and φ, (vars(ϕ)\free(ϕ)) = (vars(ψ)\ free(ψ))⊎(vars(φ)\free(φ)), thus vars(ϕ) \ free(ϕ) = (vars(ψ) \ free(ψ)+(vars(φ) \ free(φ), hence (vars(ψ) \ free(ψ), (vars(φ) \ free(φ) ≤ vars(ϕ) \ free(ϕ). Therefore hypotheses 1 and 2 hold also with ϕ uniformly replaced by ψ or φ. Then, the induction hypothesis applies for each i, by replacing s with r(i), s′ with r′(i), and ϕ with either ψ or φ. Thus, for each i, (P, r(i), σ) = ψ iff (P ′, r′(i), σ′) = ψ, and (P, r(i), σ) = φ iff (P ′, r′(i), σ′) = φ. Therefore, r′ is a run s.t. r′(0) = s′, (P ′, r′(k), σ′) = φ, and for every j, 0 ≤ j < k implies (P ′, r′(j), σ′) = ψ, i.e., (P ′, s′, σ′) = EψU φ. For ϕ ≡ AψU φ, assume by contradiction that (P, s, σ) = ϕ but (P ′, s′, σ′) 6= ϕ. Then, there exists a run r′ s.t. r′(0) = s′ and for every k ≥ 0, either (P ′, r′(k), σ′) 6= φ or there exists j s.t. 0 ≤ j < k and (P ′, r′(j), σ′) 6= ψ. By Lemma 4.7 there exists a run r s.t. r(0) = s, and for all i ≥ 0, r(i) ≈ r′(i) and σ and σ′ are equivalent for ϕ w.r.t. r(i) and r′(i). Similarly to the case of EψU φ, it can be shown that σ and σ′ are equivalent for ψ and φ w.r.t. r(i) and r′(i), for all i ≥ 0. Further, assuming w.l.o.g. that all variables common to ψ and φ occur free in both formulas, it can be shown, as in the case of EψU φ, that the induction hypothesis holds on every pair of runs obtained as suffixes of r and r′, starting from their i-th state, for every i ≥ 0. Thus, (P, r(i), σ) = ψ iff (P ′, r′(i), σ′) = ψ, and (P, r(i), σ) = φ iff (P ′, r′(i), σ′) = φ. But then r is s.t. r(0) = s and for every k ≥ 0, either (P, r(k), σ) 6= φ or there exists j s.t. 0 ≤ j < k and (P, r(j), σ) 6= ψ, that is, (P, s, σ) 6= AψU φ. This is a contradiction. For ϕ ≡ Kiψ, assume by contradiction that (P, s, σ) = ϕ but (P ′, s′, σ′) 6= ϕ. Then, there exists s′′ s.t. s′ ∼i s′′ and (P ′, s′′, σ′) 6= ψ. By Lemma 4.7 there exists s′′′ s.t. s′′′ ≈ s′′, s ∼i s′′′, and σ and σ′ are equivalent for ψ w.r.t. s′′ and s′′′. Thus, by an argument analogous to that used for the case of AX, we can apply the induction hypothesis, obtaining (P, s′′′, σ) 6= ψ. But then (P, s, σ) 6= Kiψ, which is a contradiction. Finally, for ϕ ≡ Cψ, assume by contradiction that (P, s, σ) = ϕ but (P ′, s′, σ′) 6= ϕ. Then, there exists an s′′ s.t. s′ ∼ s′′ and (P ′, s′′, σ′) 6= ψ. Again by Lemma 4.7 there exists s′′′ s.t. s′′′ ≈ s′′, s ∼ s′′′, and σ and σ′ are equivalent for ψ w.r.t. s′′ and s′′′. Thus, by an argument analogous to that used for the case of Ki, we can apply the induction hypothesis, obtaining (P, s′′′, σ) 6= ψ. But then (P, s, σ) 6= Cψ, which is a contradiction. We can now easily extend the above result to the model checking problem for AC-MAS. Theorem 4.9 Consider two ⊕-bisimilar and uniform AC-MAS P and P ′, and an FO-CTLK formula ϕ. If 1. for all t.e. runs r s.t. r(0) = s0, and for all k ≥ 0, U ′ ≥ adom(r(k)) ∪ adom(r(k + 1)) ∪ C + vars(ϕ), and 2. for all t.e. runs r′ s.t. r′(0) = s′ 0, and for all k ≥ 0, U ≥ adom(r′(k)) ∪ adom(r′(k + 1)) ∪ C + vars(ϕ) then P = ϕ iff P ′ = ϕ. 24 Proof. Equivalently, we prove that if (P, s0, σ) 6= ϕ for some σ, then there exists a σ′ such that (P ′, s′ 0, σ′) 6= ϕ, and viceversa. To this end, observe that hypotheses 1. and 2. imply, respectively, hypotheses 1. and 2. of Theorem 4.8. Further, notice that, by cardinality considerations, given the assignment σ : V ar 7→ U, there exists an assignment σ′ : V ar 7→ U ′ s.t. σ and σ′ are equivalent for ϕ w.r.t. s0 and s′ 0. Thus, by applying Theorem 4.8 we have that if there exists an assignment σ 0, σ′) 6= ϕ. The converse can be s.t. (P, s0, σ) 6= ϕ, then there exists an assignment σ′ s.t. (P ′, s′ proved analogously, as the hypotheses are symmetric. This result shows that uniform AC-MAS can in principle be verified by model checking a ⊕- bisimilar one. Note that this applies to infinite AC-MAS P as well. In this case the results above enable us to show that the verification question can be posed on the corresponding, possibly finite, P ′ as long as U ′, as defined above, is sufficiently large for P ′ to ⊕-bisimulate P. A noteworthy class of infinite systems for which these results prove particularly powerful is that of bounded AC-MAS, which, as discussed in the next subsection, always admit a finite abstraction. 4.2 Finite Abstractions We now combine the notion of uniformity explored so far in this section with the assumption on boundedness made in Section 3.1. Our aim remains to identify conditions under which the verifi- cation of an infinite AC-MAS can be reduced to the verification of a finite one. Differently from Section 3.1 we here operate on the full FO-CTLK specification language. The main result here is given by Corollary 4.14 which guarantees that, in the context of bounded AC-MAS, uniformity is a sufficient condition for ⊕-bisimilar finite abstractions to be satisfaction preserving. In the following we assume that any AC-MAS P is such that adom(s0) ⊆ C. If this is not the case, C can be extended so as to include all the (finitely many) elements in adom(s0). Further, we recall that NAg is the sum of the maximum numbers of parameters contained in the action types of We start by formalizing the notion of ⊕-abstraction. each agent in Ag, i.e., NAg = PAi∈Ag maxα(~x)∈Acti{~x}. Definition 4.10 (⊕-Abstraction) Let P = hS, U, s0, τ i be an AC-MAS over Ag, and Ag′ the set of abstract agents obtained as in Definition 3.7, for some domain U ′. The AC-MAS P ′ = hS ′, U ′, s′ 0, τ ′i over Ag′ is said to be an ⊕-abstraction of P iff: • s′ 0 = s0; • t′ ∈ τ ′(s′, ~α(~u′)) iff there exist s, t ∈ S and ~α(~u) ∈ Act(U ), such that s ⊕ t ≃ s′ ⊕ t′, for some witness ι, t ∈ τ (s, ~α(~u)), and ~u′ = ι′(~u) for some bijection ι′ extending ι to ~u. Notice that P ′ is indeed an AC-MAS as it satisfies the relevant conditions on protocols and transitions in Definition 2.7. Indeed, if t′ ∈ τ ′(s′, ~α(~u′)), then there exist s, t ∈ S, and ~α(~u) such that t ∈ τ (s, ~α(~u)), s ⊕ t ≃ s′ ⊕ t′ for some witness ι, and ~u = ι′(~u′) for some bijection ι′ extending ι. This means that αi(~ui) ∈ P ri(li) for i ≤ n. By definition of P r′ i(l′ i) for i ≤ n. Further, if U ′ has finitely many elements, then S ′ has finitely many states. Observe that by varying U ′ we obtain different ⊕-abstractions. i we have that αi(~u′ i) ∈ P r′ Next, we investigate the relationship between an AC-MAS and its ⊕-abstractions. A first useful result states that every finite ⊕-abstraction is uniform, independently of the properties of the AC- MAS they abstract. 25 Lemma 4.11 Every ⊕-abstraction P ′ of an AC-MAS P is uniform. Proof. Consider s, t, s′ ∈ S ′, t′ ∈ D(U ′), and ~α(~u) ∈ Act′(U ′) s.t. t ∈ τ ′(s, ~α(~u)) and s ⊕ t ≃ s′ ⊕ t′, for some witness ζ. We need to show that P ′ admits a transition from s′ to t′. Since P ′ is an ⊕-abstraction of P, given the definition of τ ′, there exist s′′, t′′ ∈ S and ~α(~u′′) ∈ Act(U ) s.t. t′′ ∈ τ (s′′, ~α(~u′′)), s′′ ⊕ t′′ ≃ s ⊕ t, for some witness ι, and ~u = ι′(~u′′), for some constant- preserving bijection ι′ extending ι to ~u′′. Consider ~u′ ∈ U ′~u such that ~u′ = ζ ′(~u), for some constant-preserving bijection ζ ′ extending ζ to ~u. Obviously, the composition ζ ′ ◦ ι′ is a constant- preserving bijection such that ~u′ = ζ ′(ι′(~u′′)). Moreover, it can be easily restricted to a witness for s′′ ⊕ t′′ ≃ s′ ⊕ t′. But then, since P ′ is an ⊕-abstraction of P, this implies that t′ ∈ τ ′(s′, ~α(~u′)). Thus, P ′ is uniform. The second result below guarantees that every b-bounded AC-MAS is bisimilar to any of its ⊕-abstractions, provided these are built over a sufficiently large interpretation domain. Lemma 4.12 Consider a uniform, b-bounded AC-MAS P over an infinite interpretation domain U, and an interpretation domain U ′ such that C ⊆ U ′. If U ′ ≥ 2b + C + NAg, then any ⊕-abstraction P ′ of P over U ′ is bisimilar to P. 0, then s0 ≃ s′ 0, and hs0, s′ Proof. Let B = {hs, s′i ∈ S × S ′ s ≃ s′}. We prove that B is a ⊕-bisimulation such that hs0, s′ 0i ∈ B. We start by proving that B is a ⊕-simulation relation. To this end, observe that since 0i ∈ B. Next, consider hs, s′i ∈ B, thus s ≃ s′. Assume that s0 = s′ s → t, for some t ∈ S. Then, there must exist ~α(~u) ∈ Act(U ) such that t ∈ τ (s, ~α(~u)). Moreover, since U ′ ≥ 2b + C + NAg, PAi∈Ag ~ui ≤ NAg, and adom(s) ∪ adom(t) ≤ 2b, the witness ι for s ≃ s′ can be extended to SAi∈Ag ~ui as a bijection ι′. Now let t′ = ι′(t). By the way ι′ has been defined, it can be seen that s ⊕ t ≃ s′ ⊕ t′. Further, since P ′ is an ⊕-abstraction of P, we have that t′ ∈ τ ′(s′, ~α(~u′)) for ~u′ = ι′(~u), that is, s′ → t′ in P ′. Therefore, there exists t′ ∈ S ′ such that s′ → t′, s ⊕ t ≃ s′ ⊕ t′, and ht, t′i ∈ B. As regards the epistemic relation, assume s ∼i t for some i ∈ {1, . . . , n} and t ∈ S. By definition of ∼i, li(s) = li(t). Since U ′ ≥ 2b + C, any witness ι for s ≃ s′ can be extended to a witness ι′ for s ⊕ t ≃ s′ ⊕ t′, where t′ = ι′(t). Obviously, li(s′) = li(t′). Thus, to prove that s′ ∼i t′, we need to show that t′ ∈ S ′, i.e., that t′ is reachable in P ′ from s′ 0 = s0. To this end, observe that since t ∈ S, there exists a purely temporal run r such that r(0) = s0 and r(k) = t, for some k ≥ 0. Thus, there exist also ~α1(~u1) . . . , ~αk(~uk) such that r(j + 1) ∈ τ (r(j), ~αj+1(~uj+1)), for 0 ≤ j < k. Since U ′ ≥ 2b + C, we can define, for 0 ≤ j < k, a function ιj that is a witness for r(j)⊕ r(j + 1) ≃ ιj(r(j))⊕ ιj (r(j + 1)). In particular, this can be done starting from j = k − 1, defining ιk−1 so that ιk−1(r(k)) = ιk−1(t) = t′, and proceeding backward to j = 0, guaranteeing that, for 0 ≤ j < k, ιj(r(j + 1)) = ιj+1(r(j + 1)). Observe that since adom(s0) ⊆ C, necessarily i0(r(0)) = i0(s0) = s0 = s′ 0. Moreover, as U ′ ≥ 2b + C + NAg, each ιj can be extended to a bijection ι′ j, to the elements occurring in ~uj+1. Thus, given that P ′ is an ⊕-abstraction of P, for 0 ≤ j < k, we have that ι′ j(r(j + 1)) ∈ k−1(r(k)) is a run of P ′, and, τ (ι′ k−1(r(k)), t′ is reachable in P ′. Therefore s′ ∼i t′. Further, since t ≃ t′, by definition since t′ = ι′ of B, it is the case that ht, t′i ∈ B, hence B is a ⊕-simulation. 0(r(0)) → · · · → ι′ j(r(j)), ~α(ι′ j(~uj+1))). Hence, the sequence ι′ To prove that B−1 is a ⊕-simulation, given hs, s′i ∈ B (thus s ≃ s′), assume that s′ → t′, for some t′ ∈ S ′. Obviously, there exists ~α(~u′) ∈ Act(U ′) such that t′ ∈ τ ′(s′, ~α(~u′)). Because P ′ is an ⊕-abstraction of P, there exist s′′, t′′ ∈ S and ~α(~u′′) ∈ Act(U ) such that s′′ ⊕ t′′ ≃ s′ ⊕ t′, for some witness ι, and t′′ ∈ τ (s′′, α(~u′′)), with ~u′′ = ι′(~u′), for some bijection ι′ extending ι to ~u′. 26 Observe that s′ ≃ s′′, thus, by transitivity of ≃, we have s ≃ s′′. The fact that there exists t ∈ S such that s → t easily follows from the uniformity of P. Thus, since t′ ≃ t, we have ht, t′i ∈ B. For the epistemic relation, assume s′ ∼i t′, for some t′ ∈ S ′ and 0 < i ≤ n. Let ι be a witness for s′ ≃ s, and let ι′ be an extension of ι that is a witness for s′ ⊕ t′ ≃ s ⊕ t. For t = ι′(t′), it can be seen that li(s) = li(t). Observe that t′ ∈ S ′. Using an argument essentially analogous to the one above, but exploiting the fact that P is uniform, that P ′ is certainly b-bounded, and that U > 2b + C + NAg as U is infinite, we show that t ∈ S by constructing a run r of P such that r(k) = t, for some k ≥ 0. Then s ∼i t. Further, since t′ ≃ t, we have ht, t′i ∈ B. Therefore, B−1 is a ⊕-simulation. So, P and P ′ are bisimilar. This result allows us to prove our main abstraction theorem. Theorem 4.13 Consider a b-bounded and uniform AC-MAS P over an infinite interpretation do- main U, an FO-CTLK formula ϕ, and an interpretation domain U ′ such that C ⊆ U ′. If U ′ ≥ 2b + C + max{vars(ϕ), NAg}, then for any ⊕-abstraction P ′ of P over U ′, we have that: P = ϕ iff P ′ = ϕ. Proof. By Lemma 4.11, P ′ is uniform. Thus, by the hypothesis on the cardinalities of U and U ′, Lemma 4.12 applies, so P and P ′ are bisimilar. Obviously, also P ′ is b-bounded. Thus, since P and P ′ are b-bounded, and by the cardinality hypothesis on U and U ′, Theorem 4.9 applies. In particular, notice that for every temporal-epistemic run r s.t. r(0) = s0, and for all k ≥ 0, we have that U ′ ≥ adom(r(k))∪adom(r(k+1))∪C+vars(ϕ), as adom(r(k)) ≤ b, by b-boundedness. Therefore, P = ϕ iff P ′ = ϕ. Note that the theorem above does not require U ′ to be infinite. So, by using a sufficient number of abstract values in U ′, we can in principle reduce the verification of an infinite, bounded, and uniform AC-MAS to the verification of a finite one. The following corollary to Theorem 4.13 states this clearly. Corollary 4.14 Given a b-bounded and uniform AC-MAS P over an infinite interpretation domain U, and an FO-CTLK formula ϕ, there exists an AC-MAS P ′ over a finite interpretation domain U ′ such that P = ϕ iff P ′ = ϕ. It should also be noted that U ′ can simply be taken to be any finite subset of U satisfying the cardinality requirement above. By doing so, the finite ⊕-abstraction P ′ can be defined simply as the restriction of P to U ′. Thus, every infinite, b-bounded and uniform AC-MAS is bisimilar to a finite subsystem which satisfies the same formulas. Note that, similarly to what noted at page 18 we are not concerned in the actual construction of the finite abstraction. This is because we intend to construct it directly from an artifact-centric program, as we will do in Section 6. Before that we explore the complexity of the model checking problem. 5. The Complexity of Model Checking Finite AC-MAS against FO-CTLK Specifications We now analyse the complexity of the model checking problem for finite AC-MAS with respect to FO-CTLK specifications. The input of the problem consists of an AC-MAS P on a finite domain U 27 and an FO-CTLK formula ϕ; the solution is an assignment σ such that (P, s0, σ) = ϕ. Hereafter we follow (Grohe, 2001) for basic notions and definitions. To encode an AC-MAS P we use a tuple EP = hU, D, s0, Φτ i, where U is the (finite) interpretation domain, D is the global database schema, s0 is the initial state, and Φτ = {φα1, . . . , φαm } is a set of FO-formulas, each capturing the transitions associated with a ground action αi. Since U is finite, so is the set of ground actions, thus Φτ . Each ϕαi is a FO-formula over local predicate symbols, in both normal and "primed" form, that is, φα can mention both P and P ′. For the semantics of Φτ , we have that s′ ∈ τ (s, α) iff s ⊕ s′ = φα, for s, s′ ∈ D(U ). It can be proved that every transition relation τ can be represented . in this way, and that, given EP, the size P = S + τ of the corresponding AC-MAS P is at most doubly exponential in EP of Pk. In particular, P = S + τ ≤ 23·2EP = U + D + Φτ , where D = PPk∈D qk, for qk the arity . 4 . We consider the combined complexity of the input, that is, EP + ϕ. In particular, we say that the combined complexity of model checking finite AC-MAS against FO-CTLK specifications is EXPSPACE-complete if the problem is in EXPSPACE, i.e., there is a polynomial p(x) and an algorithm solving the problem in space bound by 2p(EP +ϕ). We say it is EXPSPACE-hard if every EXPSPACE problem can be reduced to model checking finite AC-MAS against FO-CTLK specifications. We now state the following complexity result. Theorem 5.1 The complexity of the model checking problem for finite AC-MAS against FO-CTLK specifications is EXPSPACE-complete. Proof. To show that the problem is in EXPSPACE, recall that P is at most doubly exponential w.r.t. the size of the input, thus so is S. We describe an algorithm that works in NEXPSPACE, which combines the algorithm for model checking the first-order fragment of FO-CTLK and the temporal epistemic fragment. Since NEXPSPACE = EXPSPACE, the result follows. Given an AC- MAS P and an FO-CTLK formula ϕ, we guess an assignment σ. Given such σ, we check whether (P, s0, σ) = ϕ. This can be done by induction according to the structure of ϕ. If ϕ is atomic, this check can be done in polynomial time w.r.t. the size of the state it is evaluated on, that is exponential time w.r.t. EP . If ϕ is of the form ∀xψ, then we can apply the algorithm for model checking first- order (non-modal) logic, which works in PSPACE. Finally, if the outmost operator in ϕ is either a temporal or epistemic modality, then we can extend the automata-based algorithm to model check propositional CTL in (Kupferman, Vardi, & Wolper, 2000), which works in logarithmic space in S. However, we remarked above that S is generally doubly exponential in EP . Thus, if the main operator in ϕ is either a temporal or epistemic modality, then this step can be performed in space singly exponential in EP . All these steps can be performed in time polynomial in the size of ϕ. As a result, the total combined complexity of model checking finite AC-MAS is in NEXPSPACE = EXPSPACE. To prove that the problem is EXPSPACE-hard we show a reduction from any problem in EXPSPACE. We assume standard definitions of Turing machines and reductions (Papadimitriou, 1994). If A is a problem in EXPSPACE, then there exists a deterministic Turing machine TA = hQ, Σ, q0, F, δi, where Q is the finite set of states, Σ the machine alphabet, q0 ∈ Q the initial state, F the set of accepting states, and δ the transition function, that solves A using at most space 2p(in) on a given input in, for some polynomial function p. As standard, we assume δ to be a relation on (Q × Σ × Q × Σ × D), with D = {L, R}, and hq, c, q′, c′, di ∈ δ representing a transition from state q to state q′, with characters c and c′ read and written respectively , and head direction d ((L)eft and (R)ight). Without loss of generality, we assume that TA uses only the righthand half of the tape. 28 From TA and in, we build an encoding EP = hD, U, s0, Φτ i of an AC-MAS P induced by a single (environment) agent AE = hDE, LE, ActE, P rEi defined on U = Σ ∪ Q ∪ {0, 1}, where: (i) DE = {P/p(in) + 1, Q/1, H/p(in), F/1}; (ii) LE = DE(U ); (iii) ActE is the singleton {αE}, with αE parameter-free; (iv) αE ∈ P rE(lE) for every lE ∈ D(U ). Intuitively, the states of P correspond to configurations of TA, while τ mimics δ. To define EP , we let D = DE. The intended meaning of the predicates in D is as follows: the first p(in) elements of a P -tuple encode (in binaries) the position of a non-blank cell, and the (p(in) + 1)-th element contains the symbol appearing in that cell; Q contains the current state q of TA; H contains the position of the cell the head is currently on; F contains the final states of TA, i.e., F = F. The initial state s0 represents the initial configuration of TA, that is, for in = in0 · · · inℓ: s(Q) = {q0}; s(H) = {h0, . . . , 0i}; and s(P ) = {hBIN(i), inii i ∈ {0, . . . , ℓ}}, where BIN(i) stands for the binary encoding in p(in) bits of the integer i. Observe that p(in) bits are enough to index the (at most) 2p(in) cells used by TA. As to the transition relation, we define Φτ = {φαE }, where: φαE = _hq,c,q′,c′,di∈δ (∀xF (x) ↔ F ′(x)) ∧ Q(q) ∧ (∀xQ(x) → x = q) ∧ Q′(q′) ∧ (∀xQ′(x) → x = q′) ∧ ∃~p(H(~p) ∧ (∀xH(x) → x = ~p) ∧ (P (~p, c) ∨ (c = ✷ ∧ ¬∃xP (~p, x)))) ∧ ∃~p′(d = R → SUCC(~p, ~p′)) ∧ (d = L → SUCC(~p′, ~p)) ∧ H ′(~p′) ∧ (∀xH ′(x) → x = ~p′) ∧ (P ′(~p, c′) ↔ (c′ 6= ✷)) ∧ (∀xP ′(~p, x) → x = c′) ∧ (∀~x, y(P (~x, y) ∧ (~x 6= ~p) → P ′(~x, y)) ∧ (∀~x, yP ′(~x, y) → (P (~x, y) ∨ (~x = ~p ∧ y = c′)))) The symbol ✷ represents the content of blank cells, while SUCC(~x, ~x′) = Vp(in) i = 1 ↔ ((x′ i = 1) ∧ (x′ j=1 xj = 1))) is a formula capturing that ~x′ is the successor of ~x, for ~x and ~x′ interpreted as p(in)-bit binary encodings of integers (observe that {0, 1} ∈ U). Such a formula can obviously be written in polynomial time w.r.t. p(in), as well as EP , and in particular s0 and φαE . i = 1 ∧ ¬Vi−1 i = 0 ∧Vi−1 j=1 xj = 1) ∨ (x′ i=1 (x′ i = 0∨x′ As it can be seen by analyzing Φτ , the obtained transition function is such that τ (s, αE) = s′ iff, for δ(q, c) = (q′, c′, d) in TA, we have that: s′(P ) is obtained from s(P ) by overwriting with c′ (if not blank) the symbol in position (p(in) + 1) of the tuple in s(P ) beginning with the p(in)-tuple s(H) (that is, c by definition of φαE ); by updating s(H) according to d, that is by increasing or decreasing the value it contains; and by setting s′(Q) = {q′}. The predicate F does not change. Observe that cells not occurring in P are interpreted as if containing ✷ and that when ✷ is to be written on a cell, the cell is simply removed from P . It can be checked that, starting with s = s0, by iteratively generating the successor state s′ according to Φτ , i.e., s′ s.t. s ⊕ s′ = φαE , one obtains a (single) P-run that is a representation of the computation of TA on in, where each pair of consecutive P-states corresponds to a computation step. In particular, at each state, Q contains the current state of TA. It should be clear that ϕ = EF (∃xQ(x) ∧ F (x)) holds in P iff TA accepts in. Thus, by checking ϕ, we can check whether TA accepts in. This completes the proof of EXPSPACE-hardness. 29 Note that the result above is given in terms of the "data structures" in the model, i.e., U and D, and not the state space S itself. This accounts for the high complexity of model checking AC-MAS, as the state space is doubly exponential in the size of data. While EXPSPACE-hardness indicates intractability, we note that this is to be expected given that we are dealing with quantified structures which are in principle prone to undecidability. Recall also from Section 4.2 that the size of the interpretation domain U ′ of the abstraction P ′ is linear in the bound b, the number of constants in C, the size of φ, and NAg. Hence, model checking bounded and uniform AC-MAS is EXPSPACE-complete with respect to these elements, whose size will generally be small. Thus, we believe than in several cases of practical interest model checking AC-MAS may be entirely feasible. We now conclude the section with some observations on the verification of bounded and un- bounded systems. Observe that the results presented in Sections 3.1 and 4.2 apply to infinite but bounded AC-MAS, i.e., whose global states never exceed a certain size in any run. It is however worth noting that existential fragments of the specification languages considered so far need not be examined with respect to the whole AC-MAS. Indeed in bounded model checking for CTLK submodels are iteratively explored until a witness for an existential specification is found (Penczek & Lomuscio, 2003). If that happens, we can deduce that the existential specification holds on the full model as well. As we show below, we can extend these result to the case of infinite AC-MAS. To begin, define the b-restriction Pb of an AC-MAS P as follows. Definition 5.2 (b-Restriction) Given an AC-MAS P = hS, U, s0, τ i and b ∈ N such that b ≥ adom(s0) ∪ C, the b-restriction Pb = hSb, U, s0, τbi of P is such that • Sb = {s ∈ S adom(s) ≤ b}; • s′ ∈ τb(s, α(~u)) iff s′ ∈ τ (s, α(~u)) and s, s′ ∈ Sb. Notice that s0 ∈ Sb by construction and τb is the restriction of τ to Sb; the interpretation domain U is the same in P and Pb. The result below demonstrates that if a FO-ECTLK formula holds on the b-restriction, then the formula holds on the whole AC-MAS. Theorem 5.3 Consider an AC-MAS P and its b-restriction Pb, for b ∈ N. For any formula φ in FO-ECTLK, we have that: Pb = φ ⇒ P = φ Proof. By induction on the construction of φ. The base case for atomic formulas and the inductive cases for propositional connectives are trivial, as the interpretation of relation symbols for states in Sb is the same as in S. As to the existential operators EX and EU, it suffices to remark that if r is a run in Pb satisfying either EXψ or EψU ψ′, then r belongs to P as well by definition of Pb. The cases for the epistemic modalities ¯Ki and ¯C are similar: if (Pb, s, σ) = ¯Kiφ, then there exists s′ ∈ Sb such that s ∼i s′ and (Pb, s′, σ) = φ. In particular, s′ ∈ S and therefore (P, s, σ) = ¯Kiφ. For ¯Cφ the proof is similar by considering the transitive closure of the epistemic relations. Finally, the case of quantifiers follows from the fact that the active domain for each state is the same in P and Pb. Observe that there are specifications in FO-CTLK that are not preserved from Pb to P. For instance, consider the specification ϕb = AG∀x1, . . . , xb+1Wi6=j(xi = xj) in SA-FO-CTL, which 30 expresses the fact that every state in every run contains at most b distinct elements. The formula ϕb is clearly satisfied by Pb but not in P, whenever P is unbounded. Theorem 5.3 can in principle form the basis for an incremental iterative procedure for checking an existential specification φ on an infinite AC-MAS P. We can begin by taking a reasonable bound b and check Pb = φ. If that holds we can deduce P = φ; if not we can increase the bound and repeat. The procedure is sound but clearly not complete. As mentioned earlier, this is in spirit of bounded model checking (Biere, Cimatti, Clarke, Strichman, & Zhu, 2003). Here, however, the bound is on the size of the states, rather than the length of the runs. 6. Model Checking Artifact-Centric Programs We have so far developed a formalism that can be used to specify and reason about temporal- epistemic properties of models representing artifact-centric systems. We have identified two notable classes that admit finite abstractions. As we remarked in the introduction, however, artifact-centric systems are typically implemented through declarative languages such as GSM (Hull et al., 2011). It is therefore of paramount interest to investigate the verification problem, not just on a Kripke semantics such as AC-MAS, but on concrete programs. As discussed, while GSM is a mainstream declarative language for artifact-centric environments, alternative declarative approaches exist. In what follows for the sake of generality we ground our discussion on a very wide class of declarative languages and define the notion of artifact-centric program. Intuitively, an artifact-centric program (or AC program) is a declarative description of a whole multi-agent system, i.e., a set of services, that interact with the artifact system (see discussion in the Introduction). Since artifact systems are also typically implemented declaratively (see (Heath et al., 2011)) in what follows AC programs will be used to encode both the artifact system itself and the agents in the system. This also enables us to import into the formalism the previously discussed features of views and windows typical in GSM and other languages. This section is organised as follows. Firstly, we define AC programs and give their semantics in terms of AC-MAS. Secondly, we show that any AC-MAS that results from an AC program is uniform. This enables us to state that, as long as the generated AC-MAS is bounded, any AC program admits an AC-MAS as its finite model. In this context it is actually important to give constructive procedures for the generation of the finite abstraction; we provide such a procedure here. This enables us to state that, under the assumptions we identify, AC programs admit decidable verification by means of model checking their finite model. We start by defining the abstract syntax of AC programs. Definition 6.1 (AC Program) An artifact-centric program (or AC program) is a tuple ACP = hD, U, Σi, where: • D is the program's database schema; • U is the program's interpretation domain; • Σ = {Σ0, . . . , Σn} is the set of agent programs Σi = hDi, li0, Ωii, where: -- Di ⊆ D is agent i's database schema, s.t. Di ∩ Dj = ∅, for i 6= j; -- li0 ∈ Di(U ) is agent i's initial state (as a database instance); 31 -- Ωi is the set of local action descriptions in terms of preconditions and postconditions of the form α(~x) . = hπ(~y), ψ(~z)i, where: ∗ α(~x) is the action signature and ~x = ~y ∪ ~z is the set of its parameters; ∗ π(~y) is the action precondition, i.e., an FO-formula over Di; ∗ ψ(~z) is the action postcondition, i.e., an FO-formula over D ∪ D′. Recall that local database schemas and instances were introduced in Definition 2.6. Observe that AC programs are defined modularly by giving the agents' programs including preconditions and postconditions as well as those of the environment. Notice that preconditions use relation symbols from the local database only, while postcondi- tions can use any symbol from the whole D. This accounts for the intuition formalised in AC-MAS as well as present in temporal-epistemic logic literature that agents' actions may change the envi- ronment and the state of other agents. For an action α(~x), we let const(α) = const(π) ∪ const(ψ), vars(α) = vars(π) ∪ vars(ψ), and free(α) = ~x. An execution of α(~x) with ground parameters ~u ∈ U ~x is the ground action α(~u) = hπ(~v), ψ( ~w)i, where ~v (resp. ~w) is obtained by replacing each yi (resp. zi) with the value occurring in ~u at the same position as yi (resp. zi) in ~x. Such replacements make both π(~v) and ψ( ~w) ground. Finally, we define the set CACP of all constants mentioned in ACP , i.e., CACP = Sn i=1(cid:0)adom(Di0) ∪Sα∈Ωi const(α)(cid:1). The semantics of a program is given in terms of the AC-MAS induced by the agents that the program implicitly defines. Formally, this is captured by the following definition. Definition 6.2 (Induced Agents) Given an AC program ACP = hD, U, Σi, an agent induced by ACP is a tuple Ai = hDi, Li, Acti, P rii on the interpretation domain U such that, for Σi = hDi, li0, Ωii: • Li ⊆ Di(U ) is the set of the agent's local states; • Acti = {α(~x) α(~x) ∈ Ωi} is the set of local actions; • The protocol P ri(li) is defined by α(~u) ∈ P ri(li) iff li = π(~v) for α(~u) = hπ(~v), ψ( ~w)i. Note that the definition of induced agent is in line with the definition of Agents (Definition 2.6). Agents induced as above are composed to give an AC-MAS associated with an AC program. Definition 6.3 (Induced AC-MAS) Given an AC program ACP and the set Ag = {A0, . . . , An} of agents induced by ACP , the AC-MAS induced by ACP is the tuple PACP = hS, U, s0, τ i, where: • S ⊆ L0 × · · · × Ln is the set of reachable states; • s0 = hl00, . . . , ln0i is the initial global state; • U is the interpretation domain; • τ is the global transition function defined by the following condition: s′ ∈ τ (s, hα1(~u1), . . . , αn(~un)i), with s = hl0, . . . , lni and αi(~ui) = hπi(~vi), ψi( ~wi)i (i ∈ {0, . . . , n}), iff the following conditions are satisfied: -- for every i ∈ {0, . . . , n}, li = πi(~vi); 32 -- adom(s′) ⊆ adom(s) ∪Si=0,...,n ~wi ∪ const(ψi); -- Ds ⊕ Ds′ = ψi( ~wi), where Ds and Ds′ are obtained from s and s′ as discussed on p. 11. Given an AC program, the induced AC-MAS is the Kripke model representing the whole execution tree for the AC program and representing all the data in the system. Observe that all actions per- formed are enabled by the respective protocols and that transitions can introduce only a bounded number of new elements in the active domain, those bound to the action parameters. It follows from the above that AC programs are parametric with respect to the interpretation domain, i.e., by replacing the interpretation domain we obtain a different AC-MAS. For simplicity, we assume that for every postcondition ψ in a program, if a predicate does not occur in the postcondition, it is left unchanged by the relevant transitions. Formally, this means that we implicitly add a conjunct of the form ∀~xP (~x) ↔ P ′(~x) (∗) to the postcondition whenever P is not mentioned in ψ. Further, we assume that every program induces an AC-MAS whose transition relation is serial, i.e., AC-MAS states always have successors. These are basic requirements that can be easily fulfilled, for instance, by assuming that each agent has a skip action with an empty precondition and a postcondition of the form (∗) for every P ∈ D. In the next section we present an example of one such program. A significant feature of AC programs is that they induce uniform AC-MAS. Lemma 6.4 Every AC-MAS P induced by an AC program ACP is uniform. Proof. By Prop. 4.4, it is sufficient to consider only the temporal transition relation →, as adom(s0) ⊆ CACP . Consider s, s′, s′′ ∈ S and s′′′ ∈ L0 × · · · × Ln, such that s ⊕ s′ ≃ s′′ ⊕ s′′′ for some witness ι. Also, assume that there exists ~α(~u) = hα1(~u1), . . . , αn(~un)i ∈ Act(U ) such that s′ ∈ τ (s, ~α(~u)). We need to prove that for every constant-preserving bijection ι′ that extends ι to ~u, we have that s′′′ ∈ τ (s′′, ~α(ι′(~u))). To this end, we remark that any witness ι for s ⊕ s′ ≃ s′′ ⊕ s′′′ can be extended to an injective function ι′ on Si∈Ag ~ui. Obviously, U contains enough distinct elements for ι′ to exist, as every ~ui takes values from U. Now, by an argument analogous to that of Proposition 2.14, it can be seen that for any F O-formula ϕ and equivalent assignments σ and σ′, we have that (s ⊕ s′, σ) = ϕ iff (s′′ ⊕ s′′′, σ′) = ϕ. But then, this holds, in particular, for σ′ obtained from σ by applying ι′ to the values assigned to each parameter, i.e., ι′(~u), and α(~u) −−−→ s′. Thus, we have for the pre- and postconditions of all actions involved in the transition s s′′′ ∈ τ (s′′, ~α(ι′(~u))), i.e., P is uniform. We can now define what it means for an AC program to satisfy a specification, by referring to its induced AC-MAS. Definition 6.5 Given an AC program ACP , a FO-CTLK formula ϕ, and an assignment σ, we say that ACP satisfies ϕ under σ, written (ACP, σ) = ϕ, iff (PACP , s0, σ) = ϕ. It follows that the model checking problem for an AC program against a specification φ is defined in terms of the model checking problem for the AC-MAS PACP against φ. The following result allows us to reduce the verification of any AC program with an infinite interpretation domain U1, that induces a b-bounded AC-MAS, to the verification of an AC program over a finite U2. To show how it can be done, we let NACP = Pi∈{1,...,n} maxα(~x)∈Ωi{~x} be the maximum number of different parameters that can occur in a joint action of ACP . 33 Lemma 6.6 Consider an AC program ACP1 = hD, U1, Σi operating on an infinite interpretation domain U1 and assume its induced AC-MAS PACP1 = hS1, U1, s10, τ1i is b-bounded. Consider a finite interpretation domain U2 such that CACP1 ⊆ U2 and U2 ≥ 2b + CACP1 + NACP1 and the AC program ACP2 = hD, U2, Σi. Then, the AC-MAS PACP2 = hS2, U2, s20, τ2i induced by ACP2 is a finite abstraction of PACP1. i ⊆ D′ Proof. Let Ag1 and Ag2 be the set of agents induced respectively by ACP1 and ACP2, according to Def. 6.2. First, we prove that the set of agents Ag1 and Ag2 satisfy Def. 3.7, for Ag = Ag1 and Ag′ = Ag2. To this end, observe that because ACP1 and ACP2 differ only in U, by Def. 6.2, D = D′, L′ i(U ′), and Act′ = Act. Thus, only requirement 4 of Def. 3.7 still needs to be proved. To see it, fix i ∈ {1, . . . , n} and assume that α(~u) ∈ P ri(li). By Def. 6.2, we have that li = π(~v), for α(~u) = hπ(~v), ψ( ~w)i. By the assumption on U2, since const(α) ⊆ CACP1 ⊆ U2, ~u ≤ NACP1, and adom(li) ≤ b, we can define an injective function ι : adom(li) ∪ ~u ∪ CACP1 7→ U2 that is the identity on CACP1. Thus, for l′ i = ι(li), we can easily extract from ι a witness for i. Moreover, it can be seen that ~v and ~v′ are equivalent for π. Then, by applying Prop. 2.14 to li ≃ l′ li and l′ i, we conclude that l′ i). So, we have shown the right-to-left part of requirement 4. The left-to-right part can be shown similarly and in a simplified way as U1 is infinite. i = π(~v′), for ~v′ = ι(~v). Hence, by Def. 6.2, α(~u′) ∈ P r′ i(l′ 1 = hl′ 1 ∈ τ1(s1, ~α(~u)). Consider s1⊕s′ Thus, we have proven that Ag = Ag1 and Ag′ = Ag2 are obtained as in Def. 3.7. Hence, the assumption on Ag and Ag′ in Def. 4.10 is fulfilled. We prove next that also the remaining requirements of Def. 4.10 are satisfied. Obviously, since Σ is the same for ACP1 and ACP2, by Def. 6.3, s10 = s20, so the initial states of PACP1 and PACP2 are the same. It remains to show that the requirements on τ1 and τ2 are satisfied. We prove the right-to-left part. To this end, take two states s1 = hl10, . . . , l1ni, s′ 1ni in S1 and a joint action ~α(~u) = hα0(~u0), . . . , αn(~un)i ∈ Act(U ) such that s′ 1. By the assumptions on U2, there exists an injective function ι : adom(s1) ∪ adom(s′ 1) ∪ ~u ∪ CACP1 7→ U2 that is the identity on CACP1 (recall that adom(s1), adom(s′ 1) ≤ b). Then, for s2 = hι(l10), . . . , ι(l1n)i, s′ 2 = hι(l′ 2. Moreover, it can be seen that for every πi and ψi in ~αi(~xi) = hπi(~yi), ψi(~zi)i, ~u and ~u′ = ι(~u) are equivalent with respect to s1 ⊕ s′ 2. Now, consider Def. 6.3 and recall that both PACP1 and PACP2 are AC-MAS induced by ACP1, ACP2, respectively. By applying Prop. 2.14, we have that, for i ∈ {0, . . . , n}: ι(l1i) = πi(ι(~vi)) iff l1i = πi(~vi); Ds2 ⊕Ds′ = ψi( ~wi). In addition, by the definition of ι, adom(s′ 2) ⊆ adom(s2)∪Si=0,...,n ι( ~wi)∪const(ψi). But then, it is the case that s′ 2, ~α(ι(~u0), . . . , ι(~un))). So we have proved the right-to-left part of the second requirement of Def. 4.10. The other direction follows similarly. Therefore, PACP2 is an abstraction of PACP1. 1) ⊆ adom(s1) ∪Si=0,...,n ~wi ∪ const(ψi) iff adom(s′ 1n)i in S2, we can extract, from ι, a witness for s1 ⊕ s′ 1 and s2 ⊕ s′ 10), . . . , ι(l′ 1 ≃ s2 ⊕ s′ 10, . . . , l′ = ψi(ι( ~wi)) iff Ds1 ⊕Ds′ 1 2 2 ∈ τ2(s′ Intuitively, Lemma 6.6 shows that the following diagram commutes, where [U1/U2] stands for the replacement of U1 by U2 in the definition of ACP1. Observe that since U2 is finite, one can actually apply Def. 6.3 to obtain PACP2, while this cannot be done for ACP1, as U1 is infinite. ACP1 Def. 6.3 PACP1 [U1/U2] Def. 4.10 ACP2 Def. 6.3 / PACP2 34 / / / The following result, a direct consequence of Lemma 4.12 and Lemma 6.6, is the key conclusion of this section. Theorem 6.7 Consider an FO-CTLK formula ϕ, an AC program ACP1 operating on an infinite interpretation domain U1 and assume its induced AC-MAS PACP1 is b-bounded. Consider a finite interpretation domain U2 such that CACP1 ⊆ U2 and U2 ≥ 2b+CACP +max{NACP , vars(ϕ)}, and the AC program ACP2 = hD, U2, Σi. Then we have that: ACP1 = ϕ iff ACP2 = ϕ. Proof. By Lemma 6.6 PACP2 is a finite abstraction of PACP1. Moreover, U2 ≥ 2b + CACP + max{NACP , vars(ϕ)} implies U2 ≥ 2b+CACP +vars(ϕ). Hence, we can apply Lemma 4.12 and the result follows. The above is the key result in this section. It shows that if the generated AC-MAS model is bounded, then any AC program can be verified by model checking its finite ⊕-abstraction, i.e., a ⊕-bisimilar AC-MAS defined on a finite interpretation domain. Note that in this case the procedure is entirely constructive: given an AC program ACP1 = hD, U1, Σi on an infinite domain U1 and an FO-CTLK formula ϕ, to check whether ACP1 satisfies the specification ϕ, we first consider the finite "abstraction" ACP2 = hD, U2, Σi defined on a finite domain U2 satisfying the requirement on cardinality in Theorem 6.7. Since U2 is finite, also the induced AC-MAS PACP2 is finite, hence we can apply standard model checking techniques to verify whether PACP2 satisfies ϕ. Finally, by definition of satisfaction for AC programs and Theorem 6.7, we can transfer the result obtained to decide the model checking problem for the original infinite AC program ACP1 and ϕ. Also observe that in the finite abstraction considered above the abstract interpretation domain U2, depends on the number of distinct variables that the specification ϕ contains. Thus, in principle, to check the same AS program against a different specification ϕ′, one should construct a new 2, and then check ϕ′ against it. However, abstraction PACP ′ it can be seen that if the number of distinct variables of ϕ′ does not exceed that of ϕ, the abstraction PACP2, used to check ϕ, can be re-used for ϕ′. Formally, let FO-CTLKk be the set of all FO-CTLK formulas containing at most k distinct variables. We have the following corollary to Theorem 6.7. using a different interpretation domain U ′ 2 Corollary 6.8 If U2 ≥ 2b + CACP + max{NACP , k}, then, for every FO-CTLKk formula ϕ, ACP1 = ϕ iff ACP2 = ϕ. This result holds in particular for k = NACP ; thus for FO-CTLKNACP formulas, we have an abstraction procedure that is specification-independent. Theorem 6.7 requires the induced AC-MAS to be bounded, which may seem a difficult condition to check a priori. Note however that AC programs are declarative. As such it is straightforward to give postconditions that enforce that no transition will generate states violating the boundedness requirement. The scenario in the next section will exemplify this. 7. The Order-to-Cash Scenario In this section we exemplify the methodology presented so far in the context of a business process inspired by an IBM customer use-case (Hull et al., 2011). The order-to-cash scenario describes the actions performed by a number of agents in an e-commerce situation relating to the purchase 35 createP O prepared submitP O pending pay paid shipP O shipped deleteP O (a) Purchase Order lifecyle createM O preparation doneM O acceptM O submitted accepted shipM O shipped rejectM O rejected deleteM O deleteM O (b) Material Order lifecyle Figure 3: Lifecycles of the artifacts involved in the order-to-cash scenario. and delivery of a product. The agents in the system consist of a manufacturer, some customers, and some suppliers. The process begins when a customer prepares and submits a purchase order (PO), i.e., a list of products the customer requires, to the manufacturer. Upon receiving a PO, the manufacturer prepares a material order (MO), i.e., a list of components needed to assemble the requested products. The manufacturer then selects a supplier and forwards him the relevant material order. Upon receipt a supplier can either accept or reject a MO. In the former case he then proceeds to deliver the requested components to the manufacturer. In the latter case he notifies the manufacturer of his rejection. If an MO is rejected, the manufacturer can delete it and then prepare and submit new MOs. When the components required have been delivered to the manufacturer, he assembles the product and, provided the order has been paid for, he delivers it to the customer. Any order which is directly on indirectly related to a PO can be deleted only after the PO is deleted. We can encode the order-to-cash business process as an artifact-centric program ACPotc, where the artifact data models are represented as database schemas and its evolution is characterised by an appropriate set of operations. It is natural to identify 2 classes of artifacts, representing the PO and the MO, each corresponding to the respective orders by the agents. An intuitive representation of the artifact lifecycles, i.e., the evolution of some key records in the artifacts' states, capturing only the dependence of actions from the artifact statuses, is shown in Fig. 3. Note that this is an incomplete representation of the business process, as the interaction between actions and the artifact data content is not represented. Next, we encode the whole system as an AC program, where the artifact data models are repre- sented as a relational database schema, and the corresponding lifecycles are formally characterised by an appropriate set of actions. We reserve a distinguished relation for each artifact class. In addition, we introduce static relations to store product and material information. For the sake of presentation we assume to be dealing with three agents only: one customer c, one manufacturer m and one supplier s. The database schema Di for each agent i ∈ {c, m, s} can therefore be given as: • Customer c: Dc = {Products(prod code, budget), PO (id, prod code, offer , status)}; • Manufacturer m: Dm = {PO (id, prod code, offer , status), MO (id, prod code, price, status)}; • Supplier s: Ds = {Materials(mat code, cost), MO (id, prod code, price, status)}. The relations Products and Materials, as well as PO and MO are self-explanatory. Note the presence of the attribute status in the relations corresponding to artifacts. 36 As interpretation domain, we consider the infinite set Uotc of alphanumeric strings. Also, we assume that in the initial state the only non-empty relations are Products and Materials, which contain background information, such as the catalogue of available products. Hence, the artifact-centric program ACPotc corresponding to the order-to-cash scenario can be given formally as follows: Definition 7.1 The artifact-centric program ACPotc is a tuple hDotc, Uotc, Σotci, where: • the program's database schema Dotc and interpretation domain Uotc are introduced as above, i.e., Dotc = Dc ∪ Dm ∪ Ds = {Products/2, PO /4, MO /4, Materials /2} and Uotc is the set of all alphanumeric strings. • Σ = {Σc, Σm, Σs} is the set of agent specifications for the customer c, the manufacturer m and the supplier s. Specifically, for each i ∈ {c, m, s}, Σi = hDi, li0, Ωii is such that: -- Di ⊆ D is agent i's database schema as detailed above, i.e., Dc = {Products /2, PO /4}, Dm = {PO /4, MO /4}, and Ds = {MO/4, Materials /2}. -- lc0, lm0, and ls0 are database instances in Dc(Uotc), Dm(Uotc), and Ds(Uotc) respec- tively s.t. lc0(Products) and ls0(Materials) are not empty, i.e., they contain some back- ground information, while lc0(PO), lm0(PO ), lm0(MO) and ls0(MO) are empty. -- We assume that Ωc contains the actions createPO(prod code,offer), submitPO(po id), pay(po id), deletePO(po id). Similarly, Ωm = {createMO (po id, price), doneMO (mo id), shipPO (po id), deleteMO (mo id)} and Ωs = {acceptMO (mo id), rejectMO (mo id), shipMO (mo id)}. System actions capture legal operations on the underlying database and, thus, on artifacts. In Table 1 we report some of their specifications. Variables (from V ) and constants (from U) are distinguished by fonts v and c, respectively. From Section 6 we adopt the convention that an action affects only those relations whose name occurs in ψ. Consider, for instance, the action createPO performed by the customer c, whose purpose is the creation of a PO artifact instance related to a given prod code. Its precondition requires that the action parameter prod code refers to an actual product in the P roducts database; while the postcondition guarantees that the offer value in PO is set equal to budget as well as the id of the new PO is unique. As regards the action createMO, performed by the manufacturer m and meant to create instances of MO artifacts, its precondition requires that po id is the identifier of some existing PO. Its postcondition states that, upon execution, the MO relation contains exactly one additional tuple, with identifier attribute set to id, with attribute status set to preparation and asking price set to price. As an example of action triggering an artifact's status transition, consider the action doneMO performed also by the manufacturer m. doneMO is executable only if the MO artifact is in status preparation; its effect is to set the status attribute to submitted. Finally, as an example of an action triggered by a choice, consider the action acceptMO performed by the supplier s. It is triggered only if the entries for the product code pc and the price p have matching values in the Materials database. The action outcome is to set the status attribute to accepted. Notice that although actions are typically conceived to manipulate artifacts of a specific class their preconditions and postconditions may depend on artifact instances of different classes. For example note that the action createMO manipulates MO artifacts, but its preconditions and post- conditions may depend on artifact instances originating from different classes (e.g. createMO's 37 Table 1: Specification of the actions affecting the artifacts PO and MO in the order-to-cash scenario. • createP O(prod code) = hπ(prod code), ψ(prod code)i, where: -- π(prod code) ≡ ∃b P roducts(prod code, b) -- ψ(prod code) ≡ ∃id, b (P O′(id, prod code, b, prepared)∧ P roducts(prod code, b)∧ ∀id′, pc, o, s (P O(id′, pc, o, s) → id 6= id′)) • createM O(po id, price) = hπ(po id, price), ψ(po id, price)i, where: -- π(po id, price) ≡ ∃pc, o (P O(po id, pc, o, prepared) -- ψ(po id, price) ≡ (M O′(po id, pc, price, preparation)∧ ∃oP O(po id, pc, o, prepared)∧ ∀id′, pc, pr, s (M O(id′, pc, pr, s) → id 6= id′)) • doneM O(mo id) = hπ(mo id), ψ(mo id)i, where: -- π(mo id) ≡ ∃pc, p M O(mo id, pc, pr, preparation) (M O(mo id, pc, p, s) → (M O′(mo id, pc, p, submitted)∧ -- ψ(mo id) ≡ ∀w, pc, p, s (cid:0)(w 6= mo id → (M O(w, pc, p, s) ↔ M O′(w, pc, p, s)))∧ (s 6= submitted → ¬M O′(mo id, pc, p, s))))(cid:1) • acceptM O(mo id) = hπ(mo id), ψ(mo id)i, where: -- π(mo id) ≡ ∃mo id, pc, p M O(mo id, pc, pr, submitted) ∧ M aterial(pc, p) (M O(mo id, pc, p, s) → (M O′(mo id, pc, p, accepted)∧ -- ψ(mo id) ≡ ∀w, pc, p, s (cid:0)(w 6= mo id → (M O(w, pc, p, s) ↔ M O′(w, pc, p, s)))∧ (s 6= accepted → ¬M O′(mo id, pc, p, s))))(cid:1) 38 precondition depends on PO artifacts). We stress that action executability depends not only on the status attribute of an artifact, but on the data content of the whole database, i.e., of all other artifacts. Similarly, action executions affect not only status attributes. Most importantly, by using first-order formulas such as φb = ∀x1, . . . , xb+1Wi6=j(xi = xj) in the postcondition ψ, we can guarantee that the AC program in question is bounded and is therefore amenable to the abstraction methodology of Section 6. We now define the agents induced by the AC program ACPotc given above according to Defi- nition 6.2. Definition 7.2 Given the AC program ACPotc = hDotc, Uotc, Σotci, the agents Ac, Am and As induced by ACPotc are defined as follows: • Ac = hDc, Lc, Actc, P rci, where (i) Dc is as above; (ii) Lc = Dc(Uotc); (iii) Actc = Ωc = {createPO (prod code, offer ), submitPO (po id), pay(po id), deletePO (po id)}; and (iv) α(~u) ∈ P rc(lc) iff lc = π(~v) for α(~u) = hπ(~v), ψ( ~w)i. • Am = hDm, Lm, Actm, P rmi, where (i) Dm is as above; (ii) Lm = Dm(Uotc); (iii) Actm = Ωm = {createMO (po id, price), doneMO (mo id), shipPO (po id), deleteMO (mo id)}; and (iv) α(~u) ∈ P rm(lm) iff lm = π(~v) for α(~u) = hπ(~v), ψ( ~w)i. • As = hDs, Ls, Acts, P rsi, where (i) Ds is as above; (ii) Ls = Ds(Uotc); (iii) Acts = Ωs = {acceptMO (mo id), rejectMO (mo id), shipMO (mo id)}; and (iv) α(~u) ∈ P rs(ls) iff lm = π(~v) for α(~u) = hπ(~v), ψ( ~w)i. By the definition of Am we can see that createM O(po id, price) ∈ P rm(lm) if and only if the interpretation lm(P O) of the relation P O in the local state lm contains a tuple hpo id, pc, o, preparedi for some product pc and offer o; while doneM O(mo id) ∈ P rm(lm) iff lm(M O) contains a tuple in the interpretation lm(M O) with id mo id and status preparation. It can also be checked that, in line with our discussion in Section 2, a full version of the function τotc given above can easily encode the artifacts' lifecycles as given in Figure 3. We can now define the AC-MAS generated by the set of agents {Ac, Am, As} according to Definition 6.3. Definition 7.3 Given the AC program ACPotc and the set Ag = {Ac, Am, As} of agents induced by ACPotc, the AC-MAS induced by ACPotc is the tuple Potc = hSotc, Uotc, s0 otc, τotci, where: • Sotc ⊆ Lc × Lm × Ls is the set of reachable states; • Uotc is the interpretation domain; • s0 otc = hlc0, lm0, ls0i is the initial global state, where the only non-empty relation are Products and Materials; • τotc is the global transition function defined according to Def. 6.3. As an example we give a snippet of the transition function τotc by considering the global action α(~u) = hcreateP O(pc), doneM O(m), acceptM O(m′)i enabled by the respective protocols in a global state s. By the definition of the actions createP O(pc), doneM O(m), and acceptM O(m′) 39 we have that li(s) ∈ P ri for i ∈ {c, m, s} implies that the Products relation contains infor- mation about the product pc. Also, the interpretation of the relation M O contains the tuples hm, p, pr, preparationi and hm′, p′, pr′, submittedi for some products p and p′. By the definition of τotc it follows that for every s′ ∈ Sotc, s ψcreateP O(pc) ∧ ψdoneM O(m) ∧ ψacceptM O(m′), that is, α(~u) −−−→ s′ implies that Ds ⊕ Ds′ = Ds ⊕ Ds′ = ∃id, b (P O′(id, pc, b, prepared) ∧ P roducts(pc, b) ∧ ∀id′, p, o, s (P O(id′, p, o, s) → id 6= id′)) ∧ (M O(m, p, pr, s) → (M O′(m, p, pr, submitted) ∧ ∀w, p, pr, s (cid:0)(w 6= m → (M O(w, p, pr, s) ↔ M O′(w, p, pr, s))) ∧ (s 6= submitted → ¬M O′(m, p, pr, s))))(cid:1) ∧ ∀w, p, pr, s (cid:0)(w 6= m′ → (M O(w, p, pr, s) ↔ M O′(w, p, pr, s))) ∧ (s 6= accepted → ¬M O′(m′, p, pr, s))))(cid:1) (M O(m′, p, pr, s) → (M O′(m′, p, pr, accepted) ∧ Hence, the interpretation of the relation PO in Ds′ extends Ds(P O) with the tuple hid, pc, b, preparedi, where id is a fresh id. The tuples for the material orders m and m′ are updated in Ds′(M O) by becoming hm, p, pr, submittedi and hm′, p′, pr′, acceptedi, respectively. In view of the second condition on τotc in Definition 6.3, no other elements are changed in the transition. Finally, notice that these extensions are indeed the interpretations of PO and MO in Ds′. Thus, the operational semantics satisfies the intended meaning of actions. We can now investigate properties of the AC program ACPotc by using specifications in FO- CTLK. For instance, the following formula specifies that the manufacturer m knows that each ma- terial order MO has to match a corresponding purchase order PO: ϕmatch = AG ∀id, pc (∃pr, s M O(id, pc, pr, s) → Km∃o, s′P O(id, pc, o, s′)) The next specification states that given a material order MO, the customer will eventually know that the corresponding PO will be shipped. ϕfulfil = AG ∀id, pc (∃pr, s M O(id, pc, pr, s) → EF Kc∃o P O(id, pc, o, shipped)) Further, we may be interested in checking whether budget and costs are always kept secret from the supplier s and the customer c respectively, and whether the customer (resp., the supplier) knows this fact: ϕbudget = Kc ∀pc AG ¬∃b Ks P roducts(pc, b) ϕcost = Ks ∀mc AG ¬∃c Kc M aterials(mc, c) Other interesting specifications describing properties of the artifact system and the agents oper- ating in it can be similarly formalised in FO-CTLK, thereby providing the engineer with a valuable tool to assess the implementation. We now proceed to exploit the methodology of Section 6 to verify the AC program ACPotp. We use ϕmatch as an example specification; analogous results can be obtained for other formulas. Observe that according to Definition 6.3 the AC-MAS induced by ACPotp has infinitely many states. 40 We assume two interpretations for the relations Products and Materials, which determine an initial state D0. Consider the maximum number max of parameters and the constants CΩ in the operations in Ωc, Ωm and Ωs. In the case under analysis we have that max = 2. We earlier remarked that formulas such as φb in the postcondition of actions force the AC-MAS Potc corresponding to ACPotc is bounded. Here we have that Potc is b-bounded. According to Corollary 4.14, we can therefore consider a finite domain U ′ such that and such that U ′ ⊇ D0 ∪ CΩ ∪ const(ϕmatch) D0(P roducts) ∪ D0(M aterials) ∪ CΩ U ′ ≥ 2b + D0 + CΩ + const(ϕmatch) + max = 2b + D0 + CΩ + 2 For instance, we can consider any subset U ′ of Uotc satisfying the conditions above. Given that U ′ satisfies the hypothesis of Theorem 6.7, it follows that the AC program ACPotc over Uotc satisfies ϕmatch if and only if ACPotc over U ′ does. But the AC-MAS induced by the latter is a finite-state system, which can be constructively built by running the AC program ACPotc on the elements in U ′. Thus, ACPotc = ϕmatch is a decidable instance of model checking that can be therefore solved by means of standard techniques. A manual check on the finite model indeed reveals that ϕmatch, ϕbudget and ϕcost are satisfied in the finite model, whereas ϕfulfil is not. By Corollary 4.14 the AC-MAS Potc induced by ACPotp satisfies the same specifications. Hence, in view of Definition 6.5, we conclude that the artifact- centric program ACPotp satisfies ϕmatch, ϕbudget and ϕcost but does not satisfy ϕfulfil . This is entirely in line with our intuitions of the scenario. 8. Conclusions and Future Work In this paper we put forward a methodology for verifying agent-based artifact-centric systems. We proposed AC-MAS, a novel semantics incorporating first-order features, that can be used to rea- son about multi-agent systems in an artifact-centric setting. We observed that the model checking problem for these structures against specifications given in a first-order temporal-epistemic logic is undecidable and proceeded to identify suitable fragments for which decidability can be retained. We identified two orthogonal solutions to this issue. In the former we operated a restriction to the specification language and showed that, by limiting ourselves to sentence-atomic temporal- epistemic specifications, infinite-state, bounded AC-MAS admit finite abstractions. In the latter we kept the full first-order temporal-epistemic logic but identified the noteworthy subset of uniform AC-MAS. In this setting we showed that bounded uniform AC-MAS admit finite abstractions. The abstractions we identified in each setting depend on novel notions of bisimulation at first-order that we proposed. We explored the complexity of the model checking problem in this context and showed this to be EXPSPACE-complete. While this is obviously a hard problem, we need to consider that these are first-order structures which normally lead to undecidable problems. We were also reassured by the fact that the abstract interpretation domain is actually linear in the size of the bound considered. 41 Mindful of the practical needs for verification in artifact-centric systems, we then explored how finite abstractions can actually be built. To this end, rather than investigating one specific data- centric language, we defined a general class of declarative artifact-centric programs. We showed that these systems admit uniform AC-MAS as their semantics. Under the assumption of bounded systems we showed that model checking these multi-agent system programs is decidable and gave a constructive procedure operating on bisimilar, finite models. While the results are general, they can be instantiated for various artifact-centric languages. For instance (Belardinelli et al., 2012b) explores finite abstractions of GSM programs by using these results. We exemplified the methodology put forward on a use-case consisting of several agents pur- chasing and delivering products. While the system has infinitely many states we showed it admits a finite abstraction that can be used to verify a variety of specifications on the system. A question left open in the present paper is whether the uniform condition we provided is tight. While we showed this to be a sufficient condition, we did not explore whether this is necessary for finite abstractions or whether more general properties can be given. In this context it is of interest that artifact-centric programs generate uniform structures. Also, it will be worthwhile to explore whether a notion related to uniformity can be applied to other domains in AI, for example to retain decidability of specific calculi. This would appear to be the case as preliminary studies in the Situation Calculus demonstrate (De Giacomo, Lesp´erance, & Patrizi, 2012). On the application side, we are also interested in exploring ways to use the results of this paper to build a model checker for artifact-centric MAS. Previous efforts in this area, including (Gonzalez, Griesmayer, & Lomuscio, 2012), are limited to finite state systems. It would therefore be of great interest to construct finite abstractions on the fly to check practical e-commerce scenarios such as the one here discussed. References Abiteboul, S., Hull, R., & Vianu, V. (1995). Foundations of Databases. Addison-Wesley. Alonso, G., Casati, F., Kuno, H. A., & Machiraju, V. (2004). Web Services - Concepts, Architectures and Applications. Data-Centric Systems and Applications. Springer. Alves et al. (2007). Web Services Business Process Execution Language Version 2.0. http://docs.oasis-open.org/wsbpel/2.0/wsbpel-v2.0.pdf. Baukus, K., & van der Meyden, R. (2004). A Knowledge Based Analysis of Cache Coherence. In Proc. of the 6th International Conference on Formal Engineering Methods (ICFEM'04), pp. 99 -- 114. Belardinelli, F., & Lomuscio, A. (2012). Interactions between Knowledge and Time in a First-Order Logic for Multi-Agent Systems: Completeness Results. Journal of Artificial Intelligence Re- search, 45, 1 -- 45. Belardinelli, F., Lomuscio, A., & Patrizi, F. (2011a). A Computationally-Grounded Semantics for Artifact-Centric Systems and Abstraction Results. In Proc. of the 22nd International Joint Conference on Artificial Intelligence (IJCAI'12), pp. 738 -- 743. Belardinelli, F., Lomuscio, A., & Patrizi, F. (2011b). Verification of Deployed Artifact Systems via Data Abstraction. In Proc. of the 9th International Conference on Service-Oriented Comput- ing (ICSOC'11), pp. 142 -- 156. 42 Belardinelli, F., Lomuscio, A., & Patrizi, F. (2012a). An Abstraction Technique for the Verification of Artifact-Centric Systems. In Proc. of the 13th International Conference on Principles of Knowledge Representation and Reasoning (KR'12), pp. 319 -- 328. Belardinelli, F., Lomuscio, A., & Patrizi, F. (2012b). Verification of GSM-Based Artifact-Centric Systems through Finite Abstraction. In Proc. of the 10th International Conference on Service- Oriented Computing (ICSOC'12), pp. 17 -- 31. Berardi, D., Calvanese, D., Giacomo, G. D., Hull, R., & Mecella, M. (2005). Automatic Com- In Proc. of the 31st position of Transition-based Semantic Web Services with Messaging. International Conference on Very Large Data Bases (VLDB'05), pp. 613 -- 624. Berardi, D., Cheikh, F., Giacomo, G. D., & Patrizi, F. (2008). Automatic Service Composition via Simulation. International Journal of Foundations of Computer Science, 19(2), 429 -- 451. Bertoli, P., Pistore, M., & Traverso, P. (2010). Automated Composition of Web Services via Plan- ning in Asynchronous Domains. Artificial Intelligence, 174(3-4), 316 -- 361. Bhattacharya, K., Gerede, C. E., Hull, R., Liu, R., & Su, J. (2007). Towards Formal Analysis of Artifact-Centric Business Process Models. In Proc. of the 5th International Conference on Business Process Management (BPM'07), pp. 288 -- 304. Biere, A., Cimatti, A., Clarke, E. M., Strichman, O., & Zhu, Y. (2003). Bounded Model Checking. Advances in Computers, 58, 118 -- 149. Blackburn, P., de Rijke, M., & Venema, Y. (2001). Modal Logic, Vol. 53 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press. Calvanese, D., Giacomo, G. D., Lenzerini, M., Mecella, M., & Patrizi, F. (2008). Automatic Service Composition and Synthesis: the Roman Model. IEEE Data Engineering Bulletin, 31(3), 18 -- 22. Ciobaca, S., Delaune, S., & Kremer, S. (2012). Computing Knowledge in Security Protocols Under Convergent Equational Theories. Journal of Automated Reasoning, 48(2), 219 -- 262. Clarke, E. M., Grumberg, O., & Peled, D. A. (1999). Model Checking. The MIT Press, Cambridge, Massachusetts. Cohn, D., & Hull, R. (2009). Business Artifacts: A Data-Centric Approach to Modeling Business Operations and Processes. IEEE Data Engineering Bulletin, 32(3), 3 -- 9. Damaggio, E., Deutsch, A., & Vianu, V. (2012). Artifact Systems with Data Dependencies and Arithmetic. ACM Transactions on Database Systems, 37(3), 22:1 -- 22:36. Damaggio, E., Hull, R., & Vacul´ın, R. (2011). On the Equivalence of Incremental and Fixpoint Semantics for Business Artifacts with Guard-Stage-Milestone Lifecycles. In Proc. of the 9th International Conference on Business Process Management (BPM'11). De Giacomo, G., Lesp´erance, Y., & Patrizi, F. (2012). Bounded Situation Calculus Action Theories and Decidable Verification. In Proc. of the 13th International Conference on Principles of Knowledge Representation and Reasoning (KR'12), pp. 467 -- 477. Dechesne, F., & Wang, Y. (2010). To Know or not to Know: Epistemic Approaches to Security Protocol Verification. Synthese, 177(Supplement-1), 51 -- 76. 43 Deutsch, A., Hull, R., Patrizi, F., & Vianu, V. (2009). Automatic Verification of Data-centric Busi- ness Processes. In Proc. of the 12th International Conference on Database Theory (ICDT'09), pp. 252 -- 267. Deutsch, A., Sui, L., & Vianu, V. (2007). Specification and Verification of Data-Driven Web Appli- cations. Journal of Computer and System Sciences, 73(3), 442 -- 474. Fagin, R., Halpern, J. Y., Moses, Y., & Vardi, M. Y. (1995). Reasoning About Knowledge. The MIT Press. Gammie, P., & van der Meyden, R. (2004). MCK: Model Checking the Logic of Knowledge. In Proc. of 16th International Conference on Computer Aided Verification (CAV'04), pp. 479 -- 483. Gerede, C. E., & Su, J. (2007). Specification and Verification of Artifact Behaviors in Business Process Models. In Proc. of the 5th International Conference on Service-Oriented Computing (ICSOC'07), pp. 181 -- 192. Gonzalez, P., Griesmayer, A., & Lomuscio, A. (2012). Verifying GSM-Based Business Artifacts. In Proc. of the 19th IEEE International Conference on Web Services (ICWS'12), pp. 25 -- 32. Grohe, M. (2001). Generalized Model-Checking Problems for First-Order Logic. In Proc. of the 18th Annual Symposium on Theoretical Aspects of Computer Science (STACS'01), pp. 12 -- 26. Hariri, B. B., Calvanese, D., Giacomo, G. D., Deutsch, A., & Montali, M. (2012). Verification of Relational Data-Centric Dynamic Systems with External Services. CoRR, abs/1203.0024. Heath, F. T., Hull, R., & Vacul´ın, R. (2011). Barcelona: A Design and Runtime Environment for Modeling and Execution of Artifact-centric Business Processes (demo). In Proc. of the 9th International Conference on Business Process Management Demo Track (BPM'11). Hull, R. (2008). Artifact-Centric Business Process Models: Brief Survey of Research Results and Challenges. In Proc. (part II) of Confederated International Conferences, CoopIS, DOA, GADA, IS, and ODBASE 2008 (On the Move to Meaningful Internet Systems: OTM'08), pp. 1152 -- 1163. Hull, R., Damaggio, E., De Masellis, R., Fournier, F., Gupta, M., Heath, III, F. T., Hobson, S., Line- han, M., Maradugu, S., Nigam, A., Sukaviriya, P. N., & Vaculin, R. (2011). Business Artifacts with Guard-Stage-Milestone Lifecycles: Managing Artifact Interactions with Conditions and Events. In Proc. of the 5th ACM International Conference on Distributed Event-Based Sys- tems (DEBS'11), pp. 51 -- 62. Hull, R., Narendra, N. C., & Nigam, A. (2009). Facilitating Workflow Interoperation Using Artifact- Centric Hubs. In Proc. of the 7th International Conference on Service-Oriented Computing (ICSOC-ServiceWave '09), pp. 1 -- 18. Kacprzak, M., Nabialek, W., Niewiadomski, A., Penczek, W., P´olrola, A., Szreter, M., Wozna, B., & Zbrzezny, A. (2008). VerICS 2007 - a Model Checker for Knowledge and Real-Time. Fundamenta Informaticae, 85(1-4), 313 -- 328. Kupferman, O., Vardi, M. Y., & Wolper, P. (2000). An Automata-Theoretic Approach to Branching- Time Model Checking. Journal of the ACM, 47(2), 312 -- 360. 44 Lomuscio, A., Qu, H., & Raimondi, F. (2009). MCMAS: A Model Checker for the Verification of Multi-Agent Systems. In Proc. of the 21st International Conference on Computer Aided Verification (CAV'09), pp. 682 -- 688. Lomuscio, A., Qu, H., & Solanki, M. (2012). Towards Verifying Contract Regulated Service Com- position. Autonomous Agents and Multi-Agent Systems, 24(3), 345 -- 373. Lomuscio, A., Solanki, M., Penczek, W., & Szreter, M. (2010). Runtime Monitoring of Contract Regulated Web Services. In Proc. of the 9th International Conference on Autonomous Agents and Multiagent Systems (AAMAS'10), pp. 1449 -- 1450. Marin, M., Hull, R., & Vacul´ın, R. (2012). Data Centric BPM and the Emerging Case Management Standard: A Short Survey. In Proc. of the 1st (BPM) International Workshop on Adaptive Case Management. Nigam, A., & Caswell, N. S. (2003). Business Artifacts: An Approach to Operational Specification. IBM Systems Journal, 42(3), 428 -- 445. Nooijen, E., Fahland, D., & Dongen, B. V. (2012). Automatic Discovery of Data-Centric and Artifact-Centric Processes. In Proc. of the 1st (BPM) Workshop on Data- & Artifact-centric BPM (DAB'12). Papadimitriou, C. H. (1994). Computational complexity. Addison-Wesley. Parikh, R., & Ramanujam, R. (1985). Distributed Processes and the Logic of Knowledge. In Proc. of Logics of Programs, Conference, pp. 256 -- 268. Penczek, W., & Lomuscio, A. (2003). Verifying Epistemic Properties of Multi-agent Systems via Bounded Model Checking. Fundamenta Informaticae, 55(2), 167 -- 185. Singh, M. P., & Huhns, M. N. (2005). Service-Oriented Computing: Semantics, Processes, Agents. Wiley. Wooldridge, M. (2000). Computationally Grounded Theories of Agency. In Proc. of the 4th Inter- national Conference on Multi-Agent Systems (ICMAS'00), pp. 13 -- 22. Wooldridge, M. (2001). Introduction to Multiagent Systems. John Wiley & Sons, Inc. 45
1905.08750	2	1905	2019-06-01T10:05:01	Adaptation and learning over networks under subspace constraints -- Part I: Stability Analysis	[ "cs.MA", "eess.SP" ]	This paper considers optimization problems over networks where agents have individual objectives to meet, or individual parameter vectors to estimate, subject to subspace constraints that require the objectives across the network to lie in low-dimensional subspaces. This constrained formulation includes consensus optimization as a special case, and allows for more general task relatedness models such as smoothness. While such formulations can be solved via projected gradient descent, the resulting algorithm is not distributed. Starting from the centralized solution, we propose an iterative and distributed implementation of the projection step, which runs in parallel with the stochastic gradient descent update. We establish in this Part I of the work that, for small step-sizes $\mu$, the proposed distributed adaptive strategy leads to small estimation errors on the order of $\mu$. We examine in the accompanying Part II [2] the steady-state performance. The results will reveal explicitly the influence of the gradient noise, data characteristics, and subspace constraints, on the network performance. The results will also show that in the small step-size regime, the iterates generated by the distributed algorithm achieve the centralized steady-state performance.	cs.MA	cs	Adaptation and learning over networks under subspace constraints -- Part I: Stability Analysis Roula Nassif†, Member, IEEE, Stefan Vlaski†,‡, Student Member, IEEE, 1 Ali H. Sayed†, Fellow Member, IEEE † Institute of Electrical Engineering, EPFL, Switzerland ‡ Electrical Engineering Department, UCLA, USA [email protected] [email protected] [email protected] Abstract This paper considers optimization problems over networks where agents have individual objectives to meet, or individual parameter vectors to estimate, subject to subspace constraints that require the objectives across the network to lie in low-dimensional subspaces. This constrained formulation includes consensus optimization as a special case, and allows for more general task relatedness models such as smoothness. While such formulations can be solved via projected gradient descent, the resulting algorithm is not distributed. Starting from the centralized solution, we propose an iterative and distributed implementation of the projection step, which runs in parallel with the stochastic gradient descent update. We establish in this Part I of the work that, for small step-sizes µ, the proposed distributed adaptive strategy leads to small estimation errors on the order of µ. We examine in the accompanying Part II [2] the steady-state performance. The results will reveal explicitly the influence of the gradient noise, data characteristics, and subspace constraints, on the network performance. The results will also show that in the small step-size regime, the iterates generated by the distributed algorithm achieve the centralized steady-state performance. Distributed optimization, subspace projection, gradient noise, stability analysis. Index Terms 9 1 0 2 n u J 1 ] A M . s c [ 2 v 0 5 7 8 0 . 5 0 9 1 : v i X r a This work was supported in part by NSF grant CCF-1524250. A short conference version of this work appears in [1]. 2 I. INTRODUCTION Distributed inference allows a collection of interconnected agents to perform parameter estimation tasks from streaming data by relying solely on local computations and interactions with immediate neighbors. Most prior literature focuses on consensus problems, where agents with separate objective functions need to agree on a common parameter vector corresponding to the minimizer of the aggregate sum of the individual costs, namely, wo = arg min w Jk(w), N(cid:88)k=1 (1) where Jk(·) is the cost function at agent k, N is the number of agents in the network, and w ∈ CL is the global parameter vector, which all agents need to agree upon -- see Fig. 1 (middle). Each agent seeks to estimate wo through local computations and communications among neighboring agents without the need to know any of the costs besides their own. Among many useful strategies that have been proposed in the literature [3] -- [10], diffusion strategies [3] -- [5] are particularly attractive since they are scalable, robust, and enable continuous learning and adaptation in response to drifts in the location of the minimizer. However, there exist many network applications that require more complex models and flexible algorithms than consensus implementations since their agents may involve the need to estimate and track multiple distinct, though related, objectives. For instance, in distributed power system state estimation, the local state vectors to be estimated at neighboring control centers may overlap partially since the areas in a power system are interconnected [11], [12]. Likewise, in monitoring applications, agents need to track the movement of multiple correlated targets and to exploit the correlation profile in the data for enhanced accuracy [13], [14]. Problems of this kind, where nodes need to infer multiple, though related, parameter vectors, are referred to as multitask problems. Existing strategies to address multitask problems generally exploit prior knowledge on how the tasks across the network relate to each other [11] -- [29]. For example, one way to model relationships among tasks is to formulate convex optimization problems with appropriate co-regularizers between neighboring agents [13], [16] -- [19]. Graph spectral regularization can also be used in order to leverage more thoroughly the graph spectral information and improve the multitask network performance [20]. In other applications, it may happen that the parameter vectors to be estimated at neighboring agents are related according to a set of linear equality constraints [12], [21] -- [26]. However, in this paper, and the accompanying Part II [2], we consider multitask inference problems where each agent seeks to minimize an individual cost (expressed as the expectation of some loss function), and where the collection of parameter vectors to be estimated across the network is required to lie in a low-dimensional subspace -- see Fig. 1 (left). That is, we let wk ∈ CMk denote some parameter vector at node k and let W = col{w1, . . . , wN} denote the collection of parameter vectors from across the network. We associate with each agent k a differentiable convex cost Jk(wk) : CMk → R, which is expressed as the expectation of some loss function Qk(·) and written as Jk(wk) = EQk(wk; xk), where xk denotes the random data. The expectation is computed over the distribution of the data. Let M =(cid:80)N k=1 Mk. We consider constrained problems of the form: o = arg min W W J glob(W) (cid:44) N(cid:88)k=1 Jk(wk), 3 (2) subject to W ∈ R(U), where R(·) denotes the range space operator, and U is an M × P full-column rank matrix with P (cid:28) M. Each agent k is interested in estimating the k-th Mk × 1 subvector wo N}. In order to solve (2) iteratively, the gradient projection method can be applied [30]: k of Wo = col{wo 1, . . . , wo Wi = PU(cid:16)Wi−1 − µ col(cid:8)∇w∗ k Jk(wk,i−1)(cid:9)N k=1(cid:17) , i ≥ 0, (3) where Wi = col{w1,i, . . . , wN,i} with wk,i the estimate of wo parameter, ∇w∗ of wk), and PU is the projector onto the P -dimensional subspace of CM spanned by the columns of U: k at iteration i and agent k, µ > 0 is a small step-size Jk(·) is the (Wirtinger) complex gradient [4, Appendix A] of Jk(·) relative to w∗k (complex conjugate k PU = U(U∗U)−1U∗, (4) where we used the fact that U is a full-column rank matrix. We are particularly interested in solving the problem in the stochastic setting when the distribution of the data xk is generally unknown. This means that the risks Jk(·) and their gradients ∇w∗ Jk(·) are unknown. As such, approximate gradient vectors need to be employed. A common construction in stochastic approximation theory is k to employ the following approximation at iteration i: (cid:92) ∇w∗ k Jk(wk) = ∇w∗ k Qk(wk; xk,i), (5) where xk,i represents the data observed at iteration i. The difference between the true gradient and its approximation is called gradient noise. This noise will seep into the operation of the algorithm and one main challenge is to show that despite its presence, agent k is still able to approach wo k asymptotically. Although the gradient update in (3) and (5) can be performed locally at agent k, the projection operation requires a fusion center. To see this, let us introduce an intermediate variable ψk,i at node k: ψk,i = wk,i−1 − µ∇w∗ k Jk(wk,i−1). (6) After evaluating ψk,i locally, each agent at each iteration needs to send its estimate ψk,i to a fusion center, which performs the projection operation in (3) by computing Wi = PUcol{ψ1,i, . . . , ψN,i}, and then sends the resulting estimates wk,i back to the agents. While centralized solutions can be powerful, decentralized solutions are more attractive since they are more robust and respect the privacy policy at each agent [4]. Thus, a second challenge we face in this paper is how to carry out the projection through a distributed network where each node performs local computations and exchanges information only with its neighbors. We propose in Section II an adaptive and distributed iterative algorithm allowing each agent k to converge, in k of (2), for sufficiently small µ. Conditions on the the mean-square-error sense, within O(µ) from the solution wo 4 network topology and signal subspace ensuring the feasibility of a distributed implementation are provided. We also show how some well-known network optimization problems, such as consensus optimization [3] -- [5] and multitask smooth optimization [16], can be recast in the form (2) and addressed with the strategy proposed in this paper. The analysis in Section III of this Part I shows that, for sufficiently small µ, the proposed adaptive strategy leads to small estimation errors on the order of the small step-size. Building on the results of this Part I, we shall derive in Part II [2] a closed-form expression for the steady-state network mean-square-error performance. This closed form expression will reveal explicitly the influence of the data characteristics (captured by the second-order properties of the costs and second-order moments of the gradient noises) and subspace constraints (captured by U), on the network performance. The results will also show that, in the small step-size regime, the iterates generated by the distributed implementation achieve the centralized steady-state performance. For illustration purposes, distributed sub-optimal beamforming is considered in Section IV of this Part I. Notation: All vectors are column vectors. Random quantities are denoted in boldface. Matrices are denoted in capital letters while vectors and scalars are denoted in lower-case letters. We use the symbol (·)(cid:62) to denote matrix transpose, the symbol (·)∗ to denote matrix complex-conjugate transpose, and the symbol Tr(·) to denote trace operator. The symbol diag{·} forms a matrix from block arguments by placing each block immediately below and to the right of its predecessor. The operator col{·} stacks the column vector entries on top of each other. The symbol ⊗ denotes the Kronecker product. The M × M identity matrix is denoted by IM . II. DISTRIBUTED INFERENCE UNDER SUBSPACE CONSTRAINTS We move on to propose and study a distributed solution for solving (2) with a continuous adaptation mechanism. The solution must rely on local computations and communications with immediate neighborhood, and operate in real-time on streaming data. To proceed with the analysis, one of the challenges we face is that the projection in (3) requires non-local exchange of information. Our strategy is to replace the M × M projection matrix PU in (3) by an M × M matrix A that satisfies the following conditions: (cid:40) lim i→∞Ai = PU, Ak(cid:96) = [A]k(cid:96) = 0, if (cid:96) /∈ Nk and k (cid:54)= (cid:96), (7) (8) where [A]k(cid:96) denotes the (k, (cid:96))-th block of A of dimension Mk × M(cid:96) and Nk denotes the neighborhood of agent k, i.e., the set of nodes connected to agent k by an edge. The sparsity condition (8) characterizes the network topology and ensures local exchange of information at each time instant i. By replacing the projector PU in (3) by A and the true gradients ∇w∗ Jk(·) by their stochastic approximations, we obtain the following distributed adaptive solution at each agent k: k ψk,i = wk,i−1 − µ (cid:92) ∇w∗ wk,i = (cid:80)(cid:96)∈Nk Ak(cid:96)ψ(cid:96),i, k Jk(wk,i−1), (9)  where we used condition (8), and where ψk,i is an intermediate estimate and wk,i is the estimate of wo k at agent k and iteration i. As we shall see in Section III, condition (7) helps ensure convergence toward the optimum. Necessary and sufficient conditions for the matrix equation (7) to hold are given in the following lemma. 5 Lemma 1. (Necessary and sufficient conditions for (7)) The matrix equation (7) holds, if and only if, the following conditions on the projector PU and the matrix A are satisfied: APU = PU, PUA = PU, ρ(A − PU) < 1, (10) (11) (12) where ρ(·) denotes the spectral radius of its matrix argument. It follows that any A satisfying condition (7) has one as an eigenvalue with multiplicity P , and all other eigenvalues are strictly less than one in magnitude. Proof. See Appendix A. The arguments are along the lines developed in [31] for distributed averaging with proper adjustments to handle general subspace constraints. Note that conditions (10) -- (12) appeared previously (with proof omitted) in the context of distributed denoising in wireless sensor networks [29]. In such problems, the N sensors are observing N-dimensional signal, with each entry of the signal corresponding to one sensor. Using the prior knowledge that the observed signal belongs to a low-dimensional subspace, the sensor task is to denoise the corresponding entry of the signal by projecting in a distributed iterative manner onto the signal subspace in order to improve the error variance. However, in this work, we consider the more general problem of distributed inference over networks. If we replace PU by (4), multiply both sides of (10) by U, and multiply both sides of (11) by U∗, conditions (10) and (11) reduce to: AU = U, U∗A = U∗. (13) (14) Conditions (13) and (14) state that the P columns of U are right and left eigenvectors of A associated with the eigenvalue 1. Together with these two conditions, condition (12) means that A has P eigenvalues at one, and that all other eigenvalues are strictly less than one in magnitude. In the following, we discuss how some well-known network optimization problems can be recast in the form (2) and addressed with strategies in the form of (9). Remark 1. (Distributed consensus optimization). Let Mk = L for all agents. If we set in (2) P = L and U = (1N ⊗ IL) where 1N is the N × 1 vector of all ones, then solving problem (2) will be equivalent to solving 1√N the well-known consensus problem (1). Different algorithms for solving (1) over strongly-connected networks have been proposed [3] -- [9]. By picking any N × N doubly-stochastic matrix A = [ak(cid:96)] satisfying: ak(cid:96) ≥ 0, A1N = 1N , 1(cid:62)N A = 1(cid:62)N , ak(cid:96) = 0 if (cid:96) /∈ Nk and k (cid:54)= (cid:96) (15) 6 Fig. 1. Inference under subspace constraints. (Left) Illustrative scheme of problem (2): Each agent k in the network has an individual wk to estimate, subject to subspace constraints that enforce the objectives across the network to lie in R(U). (Middle) Consensus optimization (1): Agents in the network seek to estimate an L× 1 common vector w corresponding to the minimizer of the aggregate sum of individual costs. (Right) Coupled optimization [23]: Different agents generally seek to estimate different, but overlapping, parameter vectors. the diffusion strategy for instance takes the form [3] -- [5]: ψk,i = wk,i−1 − µ (cid:92) ∇w∗ wk,i = (cid:80)(cid:96)∈Nk ak(cid:96)ψ(cid:96),i. Jk(wk,i−1), k  (16) Observe that this strategy can be written in the form of (9) with Ak(cid:96) = ak(cid:96)IL and A = A ⊗ IL. It can be verified that, when A satisfies (15) over a strongly connected network, the matrix A will satisfy (8), (13), (14), and (12). Remark 2. (Distributed coupled optimization). Similarly, with a proper selection of U, multitask inference problems with overlapping parameter vectors [22] -- [24] can also be recast in the form (2). This scenario is illustrated in Fig. 1 (right). In this example, agent k is influenced by only a subset of the entries of a global w = [w1, w2, w3] and seeks to estimate wk = [w2, w3]. For a given variable w(cid:96) and any two arbitrary agents containing w(cid:96) in their costs, it is assumed that the network topology is such that there exists at least one path linking one agent to the other [23]. By properly selecting the matrix U, the network vector W = col{w1, . . . , wN} can be written as W = Uw and, therefore, distributed coupled optimization can be recast in the form (2). It can be verified that the coupled diffusion strategy proposed in [23] for solving this problem can be written in the form of (9) and that the (doubly-stochastic) matrix A in [23] satisfies conditions (7) and (8). Remark 3. (Distributed optimization under affine constraints). Several existing works consider (distributed or offline) variations of the following problem [26] -- [28]: o = arg min W W Jk(wk), (17) N(cid:88)k=1 subject to D∗W = d, where D is an M × (M − P ) full-column rank matrix and d is an (M − P ) × 1 column vector. It turns out that the online distributed strategy proposed in this work can be used to solve (17) for general constraints that are not 123456kJ1(w1)J2(w2)J3(w3)J4(w4)J5(w5)J6(w6)Jk(wk)NkW=col{w1,...,wN}∈R(U)W=col{w1,...,wN}∈R(U)123456kJ1(w)J2(w)J3(w)J4(w)J5(w)J6(w)Jk(w)W=(1N⊗IL)w∈R(U),U=1√N(1N⊗IL)W=(1N⊗IL)w∈R(U),U=1√N(1N⊗IL)123456kJ1(w1,w2)J2(w1,w3)J3(w1,w2,w3)Jk(w2,w3)J6(w3)J5(w2,w3)J4(w1)U=⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣100010100001100010001.........⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦W=Uw∈R(U),w=[w1,w2,w3]W=Uw∈R(U),w=[w1,w2,w3]necessarily local. To see this, we first note that the gradient projection method can be applied to solve (17) [30]: 7 Wi = PD(cid:16)Wi−1 − µ col(cid:8)∇w∗ k Jk(wk,i−1)(cid:9)N k=1(cid:17) + dD, i ≥ 0, (18) (cid:44) IM − D(D∗D)−1D∗ and dD (cid:44) D(D∗D)−1d. Since PD is a projection matrix, it can be decomposed where PD as PD =(cid:80)P p=1 upu∗p with {up} the orthonormal eigenvectors of PD associated with the P eigenvalues at one, and thus Pd can be replaced by PU = UU∗ with U = [u1, . . . , uP ]. Therefore, solution (18) has a form similar to the earlier solution (3) with the rightmost term dD in (18) absent from (3). Following the same line of reasoning that led to (9), we can similarly obtain the following distributed adaptive solution for solving (17): ψk,i = wk,i−1 − µ (cid:92) ∇w∗ wk,i = (cid:80)(cid:96)∈Nk Ak(cid:96)ψ(cid:96),i + dD,k, k Jk(wk,i−1),  (19) where dD,k is the k-th sub-vector of (IM − A)dD corresponding to node k (see Appendix B), and A = [Ak(cid:96)] is a properly selected matrix satisfying conditions (7) and (8). Although algorithm (19) is different than (9) due to the presence of the constant term dD,k, the mean-square-error analyzes of both algorithms are the same, as we shall see in Section III. In Section IV, we shall apply (19) to solve linearly constrained beamforming [32], [33]. Remark 4. (Distributed inference under smoothness). Let Mk = L for all agents. In such problems, each agent k in the network has an individual cost Jk(wk) to minimize subject to a smoothness condition over the graph. The smoothness requirement softens the transition in the tasks {wk} among neighboring nodes and can be measured in terms of a quadratic form of the graph Laplacian [16]: S(W) = W(cid:62)LcW = 1 2 N(cid:88)k=1 (cid:88)(cid:96)∈Nk ck(cid:96)(cid:107)wk − w(cid:96)(cid:107)2, (20) where Lc = Lc ⊗ IL with Lc = diag{C1N} − C denoting the graph Laplacian. The matrix C = [ck(cid:96)] is an N × N symmetric weighted adjacency matrix with ck(cid:96) ≥ 0 if (cid:96) ∈ Nk and ck(cid:96) = 0 otherwise. The smaller S(W) is, the smoother the signal W on the graph is. Since Lc is symmetric positive semi-definite, it can be decomposed as Lc = V ΛV (cid:62) where Λ = diag{λ1, . . . , λN} with λm the non-negative eigenvalues ordered as 0 = λ1 ≤ λ2 ≤ . . . ≤ λN and V = [v1, . . . , vN ] is the matrix of orthonormal eigenvectors. When the graph is 1N [34]. Using the eigenvalue connected, there is only one zero eigenvalue with corresponding eigenvector v1 = 1√N decomposition L = (V ΛV (cid:62)) ⊗ IL, S(W) can be written as: S(W) = W(cid:62)(Λ ⊗ IL)W = λm(cid:107)wm(cid:107)2, (21) N(cid:88)m=1 where W = (V (cid:62) ⊗ IL)W and wm = (v(cid:62)m ⊗ IL)W. Given that λm ≥ 0, the above expression shows that W is considered to be smooth if (cid:107)wm(cid:107)2 corresponding to large λm is negligible. Thus, for a smooth W, S(W) will be equal to (cid:80)p m=1 λm(cid:107)wm(cid:107)2 with p (cid:28) N. By choosing U = U ⊗ IL where U = [v1, . . . , vp], the smooth signal W will be in the range space of U since it can be written as W = Us with s = col{w1, . . . , wp}. Therefore, distributed inference problems under smoothness can be recast in the form (2). 8 Before proceeding, note that, in some cases, one may find a family of matrices A satisfying conditions (12), (13), and (14) under the sparsity constraints (8). For example, in consensus optimization described in Remark 1 where (1N ⊗ IL), by ensuring that the underlying graph is strongly connected and by choosing any doubly- U = 1√N stochastic A satisfying the sparsity constraints, the resulting matrix A = A⊗ IL will satisfy the required conditions. The same observation holds for coupled optimization problems described in Remark 2. Several policies for designing locally doubly-stochastic matrices have been proposed in the literature [3] -- [5]. For more general U, designing an A satisfying conditions (7) and (8) can be written as the following feasibility problem: find A such that AU = U, U∗A = U∗, (22) ρ(A − PU) < 1, [A]k(cid:96) = 0, if (cid:96) /∈ Nk and (cid:96) (cid:54)= k, which is challenging in general. Not all network topologies satisfying (8) guarantee the existence of an A satisfying condition (7). The higher the dimension of the signal subspace is, the greater the graph connectivity has to be. In the works [1], [29], it is assumed that the sparsity constraints (8) and the signal subspace lead to a feasible problem. That is, it is assumed that problem (22) admits at least one solution. As a remedy for the violation of such assumption, one may increase the network connectivity by increasing the transmit power of each node, i.e., adding more links [29]. In the accompanying Part II [2], we shall relax the feasibility assumption by considering the problem of finding an A that minimizes the number of edges to be added to the original topology while satisfying the constraints (12), (13), and (14). In this case, if the original topology leads to a feasible solution, then no links will be added. Otherwise, we assume that the designer is able to add some links to make the problem feasible. In the following section, we consider that a feasible A (topology) is computed by the designer and that its blocks {Ak(cid:96)}(cid:96)∈Nk are provided to agent k in order to run algorithm (9). We shall study the performance of (9) in the mean-square-error sense. We shall consider the general complex case, in addition to the real case, since complex-valued combination matrix A and data xk,i are important in several applications, as will be the case in the distributed beamforming application considered later in Section IV. III. STABILITY ANALYSIS relative to wo In this Part I, we shall establish mean-square-error stability by showing that, for each agent k, the error variance k− wk,i(cid:107)2 = O(µ). Then, building on this result, we will assess in the accompanying Part II [2] the size of this mean-square error by k enters a bounded region whose size is in the order of µ, namely, lim supi→∞ E(cid:107)wo deriving closed-form expression for the network mean-square-deviation (MSD) defined by [4]: MSD (cid:44) µ lim µ→0(cid:18)lim sup i→∞ 1 µ E(cid:18) 1 N (cid:107)W o − Wi(cid:107)2(cid:19)(cid:19) , (23) where Wi (cid:44) col{wk,i}N k=1. In this way, we will be able to conclude that distributed strategies of the form (9) with small step-size are able to lead to reliable performance even in the presence of gradient noise. We will be able also to conclude that the iterates generated by the distributed implementation achieve the centralized steady-state performance. As explained in [4, Chap. 8], in the general case where Jk(wk) are not necessarily quadratic in the (complex) variable wk, we need to track the evolution of both quantities wk,i and (w∗k,i)(cid:62) in order to examine how the network is performing. Since Jk(wk) is real valued, the evolution of the complex conjugate iterates (w∗k,i)(cid:62) is given by: 9 Representations (9) and (24) can be grouped together into a single set of equations by introducing extended vectors of dimensions 2Mk × 1 as follows: k ∇w(cid:62)  (cid:0)ψ∗k,i)(cid:62) = (cid:0)w∗k,i−1(cid:1)(cid:62) − µ (cid:92) (cid:0)w∗k,i(cid:1)(cid:62) = (cid:80)(cid:96)∈Nk (A∗k(cid:96))(cid:62)(cid:0)ψ∗(cid:96),i(cid:1)(cid:62). (cid:0)w∗k,i−1)(cid:62)  − µ (cid:0)ψ∗k,i)(cid:62)  = (A∗k(cid:96))(cid:62)   (cid:0)w∗k,i)(cid:62)  = (cid:88)(cid:96)∈Nk wk,i−1 ψk,i wk,i Ak(cid:96) 0 0    Jk(wk,i−1), (cid:92) ∇w∗ (cid:92) ∇w(cid:62) k k Jk(wk,i−1) Jk(wk,i−1)  (cid:0)ψ∗(cid:96),i)(cid:62)  . ψ(cid:96),i (24) (25) (26) Therefore, when the data is complex, extended vectors and matrices need to be introduced in order to analyze the network evolution. The arguments and results presented in the analysis are applicable to both cases of real and complex data through the use of data-type variable h defined in Table I. When the data is real-valued, the complex conjugate transposition should be replaced by the real transposition. Table I lists a couple of variables and symbols that will be used in the sequel for both real and complex data cases. The superscript "e" is used to refer to extended quantities. Although in the real data case no extended quantities should be introduced, we use the superscript "e" for both data cases for compactness of notation. A. Modeling conditions We analyze (9) under conditions (8), (12), (13), and (14) on A, and the following assumptions on the risks {Jk(·)} and on the gradient noise processes {sk,i(·)} defined as: Jk(w) − sk,i(w) (cid:44) ∇w∗ k (cid:92) ∇w∗ k Jk(w). Before proceeding, we introduce the Hermitian Hessian matrix functions [4, Appendix B]: Hk(wk) (cid:44) ∇2 wk Jk(wk), = k ∇w(cid:62) [∇wkJk(wk)], [∇wkJk(wk)] [∇wkJk(wk)] (∇w(cid:62) (∇w∗ H(W) (cid:44) diag{H1(w1), . . . , HN (wN )} ,  ∇w∗ ∇w(cid:62) k k k k [∇wkJk(wk)])∗ [∇wkJk(wk)])(cid:62)  (hM × hM ). (27) (hMk × hMk) when the data is real (Mk × Mk) when the data is complex (2Mk × 2Mk) (28) 10 DEFINITION OF SOME VARIABLES USED THROUGHOUT THE ANALYSIS. I IS A PERMUTATION MATRIX DEFINED BY (31). TABLE I Variable Real data case Complex data case Data-type variable h Gradient vector Error vector (cid:101)we k,i 1 k Jk(wk) ∇w(cid:62) (cid:101)wk,i from (45) Gradient noise se k,i(w) sk,i(w) from (26) Bias vector be k bk from (48) (k, (cid:96))-th block of Ae Matrix U e Matrix J e Ak(cid:96) U J from (41) Matrix V e R, VR, from (41) Matrix (V e L,)∗ V∗L, from (41) Jk(wk) k (s∗k,i(w))(cid:62) 2 (b∗k)(cid:62) ∇w∗  (cid:101)wk,i ((cid:101)w∗k,i)(cid:62)  sk,i(w)  bk  Ak(cid:96)  U  J  VR,  V∗L, 0 0 0 0 I(cid:62) 0 0       I (A∗k(cid:96))(cid:62) 0 (U∗)(cid:62) 0 (J ∗ )(cid:62) 0 (V∗R,)(cid:62) 0 V(cid:62)L, I(cid:62) Note that, when J glob(W) =(cid:80)N ∇2 where I is a permutation matrix given by: k=1 Jk(wk), we have: WJ glob(W) = when the data is real H(W), IH(W)I(cid:62), when the data is complex I (cid:44)  IM1 0 0 0 0 0 0 0 0 IM1 0 0 0 IM2 0 0 0 0 0 0 0 0 IM2 0 0 . . . 0 . . . 0 ... . . . 0 IMN . . . 0 . . . 0 ... 0 . . . 0 0 0 0 0 0 0 0 0 IMN .   (29) (30) This matrix consists of 2N × 2N blocks with (m, n)-th block given by: [I]mn (cid:44) IMk, IMk, 0, if m = k, n = 2(k − 1) + 1 if m = k + N, n = 2k otherwise for m, n = 1, . . . , 2N and k = 1, . . . , N. Assumption 1. (Conditions on aggregate and individual costs). The individual costs Jk(wk) ∈ R are assumed to be twice differentiable and convex such that: νk h IhMk ≤ Hk(wk) ≤ δk h IhMk, where νk ≥ 0 for k = 1, . . . , N. It is further assumed that, for any W, H(W) satisfies: 0 < ν h IhP ≤ (U e)∗H(W)U e ≤ δ h IhP , for some positive parameters ν ≤ δ. The data-type variable h and the matrix U e are defined in Table I. Condition (33) ensures that problem (2), which can be rewritten as: W o = Uso, with so = arg min s f (s) (cid:44) J glob(Us), has a unique minimizer Wo. This is due to the fact that the Hessian of f (s), which is given by: ∇2 sf (s) = WJ glob(W)(cid:3)W=U s U, U(cid:62)(cid:2)∇2 diag(cid:8)U∗,U(cid:62)(cid:9)(cid:2)∇2 (33) WJ glob(W)(cid:3)W=U s diag(cid:8)U, (U∗)(cid:62)(cid:9) , IhP > 0, (29) ν h is positive definite under condition (33). = (U e)∗ H(Us)U e ≥ (real data case) (complex data case) Assumption 2. (Conditions on gradient noise). The gradient noise process defined in (26) satisfies for any w ∈ F i−1 and for all k, (cid:96) = 1, . . . , N: 11 (31) (32) (33) (34) (35) (36) (37) (38) (39) E[sk,i(w)F i−1] = 0, E[sk,i(w)s∗(cid:96),i(w)F i−1] = 0, E[sk,i(w)s(cid:62)(cid:96),i(w)F i−1] = 0, k (cid:54)= (cid:96), k (cid:54)= (cid:96), E[(cid:107)sk,i(w)(cid:107)2F i−1] ≤ (βk/h)2(cid:107)w(cid:107)2 + σ2 s,k, k ≥ 0, σ2 s,k ≥ 0, and where F i−1 denotes the filtration generated by the random processes {w(cid:96),j} for for some β2 all (cid:96) = 1, . . . , N and j ≤ i − 1. As explained in [3] -- [5], these conditions are satisfied by many objective functions of interest in learning and adap- tation such as quadratic and logistic risks. Condition (36) essentially states that the gradient vector approximation 12 should be unbiased conditioned on the past data, which is a reasonable condition to require. Condition (39) states that the second-order moment of the gradient noise process should get smaller for better estimates, since it is bounded by the squared-norm of the iterate. Conditions (37) and (38) state that the gradient noises across the agents are uncorrelated and second-order circular. Without loss of generality, we shall introduce the following assumption on the matrix U 1. Assumption 3. (Condition on U). The full-column rank matrix U is assumed to be semi-unitary, i.e., its column vectors are orthonormal and U∗U = IP . Before proceeding, we introduce an N × N block matrix Ae whose (k, (cid:96))-th block is defined in Table I. This matrix will appear in our subsequent study. Observe that in the real data case, Ae = A, and that in the complex data case, Ae can be seen as an extended version of the combination matrix A. The next statement exploits the eigen-structure of Ae that will be useful for establishing the mean-square stability. Lemma 2. (Jordan canonical decomposition). Under Assumption 3, the M × M combination matrix A satisfying conditions (13), (14), and (12) admits a Jordan canonical decomposition of the form: A (cid:44) VΛV, Λ = IP 0 0 J  , V =(cid:104) U VR, (cid:105) , V−1 = U∗ V∗L,  , where J is a Jordan matrix with the eigenvalues (which may be complex but have magnitude less than one) on the diagonal and > 0 on the super-diagonal. It follows that the hM × hM matrix Ae defined in Table I admits a Jordan decomposition of the form: with: with Ae (cid:44) V e Λe (V e )−1, Λe = IhP 0 0 J e  , V e = [U e V e R,], (V e )−1 = (U e)∗ (V e L,)∗  where U e,J e ,V e R,, and (V e L,)∗ are defined in Table I. Since (V e )−1V e = IhM , the following relations hold: (U e)∗U e = IhP , Proof. See Appendix C. (V e L,)∗V e R, = Ih(M−P ), (U e)∗V e R, = 0, (V e L,)∗U e = 0. 1This assumption is not restrictive since for any full-column rank matrix U(cid:48) = [u(cid:48)1, . . . , u(cid:48)P ] with P ≤ M, we can generate by using, for example, the Gram-Schmidt process [35, pp. 15], a semi-unitary matrix U = [u1, . . . , uP ] that spans the same P -dimensional subspace of CM as U(cid:48), i.e., R(U) = R(U(cid:48)). (40) (41) (42) (43) (44) B. Network error vector recursion Let (cid:101)wk,i denote the error vector at node k:  where: k (cid:92) ∇w∗ (cid:92) ∇w(cid:62) k k − wk,i. (cid:101)wk,i (cid:44) wo Jk(wk,i−1)  = −H k,i−1(cid:101)we H k,i−1 (cid:44)(cid:90) 1 0 ∇2 Jk(wk,i−1) Jk(wo wk k,i−1 + be k − se k,i(wk,i−1) k − t(cid:101)wk,i−1)dt, k,i, se k,i(wk,i−1), and be and (cid:101)we k are defined in Table I with: bk (cid:44) ∇w∗ k Jk(wo k). 13 (45) (46) (47) (48) (49) (50) (51) (52) (53) (54) (55) (56) Consider first the complex data case. Using (26) and the mean-value theorem [36, pp. 24], [4, Appendix D], we can express the stochastic gradient vectors appearing in (25) as follows: Subtracting (wo k)e = col{wo k, ((wo k)∗)(cid:62)} from both sides of (25) and by introducing the following extended vectors and matrices, which collect quantities from across the network: i N,i(cid:9) , (cid:101)We 1,i, . . . ,(cid:101)we Hi−1 (cid:44) diag{H 1,i−1, . . . , H N,i−1} , Bi−1 (cid:44) Ae(IhM − µHi−1), 1,i(w1,i−1), . . . , se 1, . . . , be N} , N,i(wN,i−1)(cid:9) , se i (cid:44) col(cid:8)(cid:101)we (cid:44) col(cid:8)se be (cid:44) col{be we can show that the network weight error vector (cid:101)We i = Bi−1(cid:101)We (cid:101)We (cid:88)(cid:96)∈Nk since Wo is the solution of problem (2), and thus: i + µAebe where Ae is defined in Table I and where we used (46) and the fact that i−1 − µAese Ak(cid:96)wo (cid:96) = wo k, AW o = APUW o (10) = PUW o = W o. i in (49) evolves according to the following dynamics: For real data, the model can be simplified since we do not need to track the evolution of the complex conjugate (w∗k,i)(cid:62). Although we use the notation "e" for the quantities in the above recursion, it is to be understood that the extended quantities {(cid:101)we k,i, be k, se k,i,Ae} should be replaced by the quantities {(cid:101)wk,i, bk, sk,i,A} as in Table I. 14 The stability analysis of recursion (54) is facilitated by transforming it to a convenient basis using the Jordan )−1 and introducing the decomposition of Ae in Lemma 2. Multiplying both sides of (54) from the left by (V e transformed iterates and variables: (V e )−1Aese i = )−1(cid:101)We i = )−1Aebe = µ(V e µ(V e i i We i W(cid:86)e i  (cid:44) (U e)∗(cid:101)We L,)∗(cid:101)We  (cid:44) L,)∗Aebe  (cid:44) (V e µ (U e)∗Aese L,)∗Aese µ (V e µ (U e)∗Aebe µ (V e i i se i s(cid:86)e i  ,  , (cid:86)e  , 0 b i = (IhP − D11,i−1)We We − D22,i−1)W(cid:86)e W(cid:86)e i = (J e i−1 − D12,i−1W(cid:86)e i−1 − D21,i−1We i−1 − se i , i−1 − s(cid:86)e (cid:86)e i + b , we obtain from Lemma 2: where (57) (58) (59) (60) (61) (62) (63) (64) (65) (66) (67) (68) (69) D11,i−1 = µ (U e)∗Hi−1U e, D12,i−1 = µ (U e)∗Hi−1V e D21,i−1 = µJ e D22,i−1 = µJ e L,)∗Hi−1U e, L,)∗Hi−1V e R,, R,. (V e (V e Recursions (60) and (61) can be written more compactly as: IhP − D11,i−1 −D21,i−1 −D12,i−1 − D22,i−1 J e  We W(cid:86)e i−1 i−1  We i W(cid:86)e i   = − se i s(cid:86)e i  + 0 (cid:86)e  . b The zero entry in (59) is due to the fact that since the constrained optimization problem (2) can be written alternatively as: (U e)∗Ae be (42),(44) = (U e)∗be = 0, minimize W J glob(W) = Jk(wk) N(cid:88)k=1 subject to (IM − PU)W = 0. The Lagrangian associated with problem (68) is given by: L(W; γ) = J glob(W) + hRe{γ∗(IM − PU)W} 15 o) be 0 γ  +(IM − PU)γ = 0, where γ is the M × 1 vector of Lagrange multipliers. From the optimality conditions, we obtain the following condition on Wo: ∇W(cid:62)J glob(W (cid:125) (cid:123)(cid:122) (cid:124) ∇W(cid:62)J glob(Wo)   ∇W∗J glob(Wo) + (cid:125) (cid:123)(cid:122) (cid:124) where we used the fact that(cid:80)N k=1 Jk(wk) is real valued and where be and I are given by (53) and (31), respectively. In the real data case, by multiplying both sides of the previous relation by (U e)∗ = U(cid:62), we obtain (U e)∗be = 0. For complex data, by multiplying both sides of the previous equation by (U e)∗I(cid:62) with U e defined in Table I, we obtain (U e)∗be = 0. Now, considering both real and complex data cases, we arrive at (67). (IM − PU)(cid:62)  (γ∗)(cid:62)  = 0, (when the data is complex) (when the data is real) IM − PU 0 Remark 5. Regarding algorithm (19), it can be verified that the weight error vector (cid:101)We according to recursion (54). The constant driving terms {dD,k} will disappear when subtracting wo of (19) since wo k satisfies the following relation: i will end up evolving k from both sides (70) Ibe (cid:96) + dD,k, where we used the fact that the optimal solution Wo in (17) verifies: Ak(cid:96)wo wo k = (cid:88)(cid:96)∈Nk W o = PUW o + dD o + dD) + (IM − A)dD (10) = A(PUW = AW o + (IM − A)dD. (71) (72) By rewriting the constraint in (17) as (IM − PU)W = dD and repeating similar arguments as (68) -- (70), we can show that (U e)∗be = 0. Therefore, the transformed iterates We i in (57) will continue to evolve according to recursions (60) and (61). i and W(cid:86)e In the following, we shall establish the mean-square-error stability of algorithm (9). In the accompanying Part II, we will derive a closed-form expression for the network MSD defined by (23). The derivation is demanding. However, the arguments are along the lines developed in [4, Chaps. 9 -- 11] for standard diffusion (16) with proper adjustments to handle possibly complex valued block matrices {Ak(cid:96)} satisfying conditions (7) and (8) and the subspace constraints. C. Mean-square-error stability Theorem 1. (Network mean-square-error stability). Consider a network of N agents running the distributed strategy (9) with a matrix A satisfying conditions (13), (14), and (12) and U satisfying Assumption 3. Assume the individual costs, Jk(wk), satisfy the conditions in Assumption 1. Assume further that the first and second-order moments of the gradient noise process satisfy the conditions in Assumption 2. Then, the network is mean-square- error stable for sufficiently small step-sizes, namely, it holds that: E(cid:107)wo k − wk,i(cid:107)2 = O(µ), lim sup i→∞ k = 1, . . . , N, (73) 16 for small enough µ. Proof. See Appendix D. IV. DISTRIBUTED LINEARLY CONSTRAINED MINIMUM VARIANCE (LCMV) BEAMFORMER Consider a uniform linear array (ULA) of N = 14 antennas, as shown in Fig. 2. A desired narrow-band signal s0(i) ∈ C from far field impinges on the array from known direction of arrival (DOA) θ0 = 30◦ along with two uncorrelated interfering signals {s1(i), s2(i)} ∈ C from DOAs {θ1 = −60◦, θ2 = 60◦}, respectively. We assume that the DOA of s3(i) is roughly known. The received signal at the array is therefore modeled as: xi = a(θ0)s0(i) + a(θn)sn(i) + vi (74) 2(cid:88)n=1 (cid:124) zi (cid:123)(cid:122) (cid:125) where xi = col{x1(i), . . . , xN (i)} is an N × 1 vector that collects the received signals at the antenna elements, n=0 are N × 1 array manifold vectors (steering vectors) for the desired and interference signals, and {a(θn)}2 vi = col{v1(i), . . . , vN (i)} is the additive noise vector at time i. With the first element as the reference point, the N × 1 array manifold vector a(θn) is given by a(θn) = col(cid:8)1, e−jτn, e−j2τn, . . . , e−j(N−1)τn(cid:9) [32], with λ sin(θn) where d denotes the spacing between two adjacent antenna elements, and λ denotes the wavelength τn = 2πd of the carrier signal. The antennas are assumed spaced half a wavelength apart, i.e., d = λ/2. Beamforming problems generally deal with the design of a weight vector h = col{h1, . . . , hN} ∈ CN×1 in order to recover the desired signal s0(i) from the received data xi [32], [33]. The narrowband beamformer output can be expressed as y(i) = h∗xi. Among many possible criteria, we use the linearly-constrained-minimum-variance (LCMV) design, namely, ho = arg min h Eh∗xi2 = h∗Rxh subject to D∗h = b, (75) where D is an N × P matrix and b is a P × 1 vector, in order to suppress the influence of the perturbation zi on the output y(i) while preserving the signal component. Since the DOAs of s0(i) is known and the DOA of s2(i) is roughly known, matrix D can be chosen as D = [a(30◦) a(58.5◦) a(61.5◦)], and the vector b as b = col{1, 0.01, 0.01}. In this way, we set unit response to the direction of the desired signal so that s0(i) passes through the array without distortion. In a distributed setting, the objective of agent (antenna element) k is to estimate ho k, the k-th component of ho in (75). Neighboring agents are allowed to exchange their observations x(cid:96)(i). To each agent k, we associate a neighborhood set Nk, an Mk × 1 parameter vector wk, and an Mk × 1 regression vector uk,i, defined in Table II depending on the node location on the array. Observe that the parameter ν controls the network topology. For DISTRIBUTED BEAMFORMING SETTINGS FOR UNIFORM LINEAR ARRAYS OF N ANTENNAS (1 ≤ ν ≤ N − 1). TABLE II 17 Neighboring set Nk {max{1, k − ν}, . . . , min{k + ν, N}} Parameter vector wk col{hm}min{k+ν,N} m=max{1,k−ν} col{Nm− 1 Regressor uk,i 2 xm(i)}min{k+ν,N} m=max{1,k−ν} example, ν = N − 1 corresponds to a fully connected network setting. We associate with each agent k a cost Jk(wk) (cid:44) w∗k E[uk,iu∗k,i]wk. Instead of solving (75), we propose to solve: W o = arg min W J glob(W) = N(cid:88)k=1 subject to D∗W = d, Jk(wk) (76) where the equality constraint D∗W = d merges the equality constraint in (75) and the equality constraints that need to be imposed on the parameter vectors at neighboring nodes in order to achieve equality between common entries (see Table II). Let E denote the binary connection matrix with [Ek(cid:96)] = 1 if (cid:96) ∈ Nk, and 0 otherwise. Under the consensus constraints, it can be shown that: J glob(W) = Jk(wk) = h∗(F ◦ Rx)h N(cid:88)k=1 where F is an N ×N matrix with [F ]k(cid:96) = [E2]k(cid:96) and ◦ is the element-wise product. Therefore, collecting obser- √NkN(cid:96) vations from neighboring nodes allows partial covariance matrix computation, which will be used in optimization. For the partial covariance F ◦Rx to converge to the true covariance Rx in (75), we need to set ν = N −1 in order to have F = 1N 1(cid:62)N . Note that, two main classes of distributed beamforming appear in the literature [37]. In the first class, which is considered here, the covariance matrix is approximated to form distributed implementations [37] -- [40] (77) leading to sub-optimal beamformers. In the second class, the proposed beamformers obtain statistical optimality but do so at the expense of restricting the topology of the underlying network [41]. Different from [38], the current distributed solution preserves convexity and is scalable since nodes exchange and compute M(cid:96)×1 sub-vectors {w(cid:96)} with M(cid:96) = N(cid:96) < N instead of N × 1 vectors. Algorithm (9) can be applied to solve (76). The signals {sn(i)}2 n=0 are i.i.d. zero-mean complex Gaussian s,n = 1,∀n. The additive noise vi is zero-mean complex Gaussian with covariance random variables with variance σ2 Eviv∗i = σ2 vIN (σv = 0.7). We set ν = 4. The complex combination matrix A is set as the solution of the feasibility problem (22) with the constraint ρ(A − PU) < 1 replaced by ρ(A − PU) ≤ 1 − ( = 0.01) and the constraint A = A∗ added2. The resulting problem is solved via CVX package [42]. Note that the distributed implementation is feasible in this example. We set µ = 0.005. The output signal-to-interference-plus-noise ratio (SINR) given by E(cid:104) σ2 vIN is illustrated in Fig. 2 (right). The dashed black curve is the beampattern obtained by the centralized, also known as the constrained LMS [33], algorithm (µ = 0.001). The (cid:105) with Rz =(cid:80)2 s,na(θn)a∗(θn) + σ2 s,0h∗ h∗ i Rzhi i a(θ0) n=1 σ2 2These changes make the problem convex -- see [2, Sec. 3] for further details. 18 Fig. 2. (Left) Uniform linear array of N antennas. (Right) Comparison of output SINR. results are averaged over 1000 Monte-Carlo runs. We observe that the distributed solution performs well compared to the centralized implementation. V. CONCLUSION In this work, we considered inference problems over networks where agents have individual parameter vectors to estimate subject to subspace constraints that require the parameters across the network to lie in low-dimensional subspaces. Based on the gradient projection algorithm, we proposed an iterative and distributed implementation of the projection step, which runs in parallel with the stochastic gradient descent update. We showed that, for small step-size parameter, the network is able to approach the minimizer of the constrained problem to arbitrarily good accuracy levels. APPENDIX A PROOF OF LEMMA 1 First we prove sufficiency by proving that if PU is a projection matrix and A satisfies conditions (10), (11), and (12), then the matrix equation (7) holds. If A satisfies (10) and (11), then: Ai − PU (10) = = = (10),(11) = (10) = Ai − AiPU Ai(I − PU) Ai(I − PU)i (A(I − PU))i (A − PU)i (78) where we used the fact that (I −PU) = (I −PU)i since (I −PU) is a projector. Applying condition (12) and using the fact that for any matrix B, limi→∞ Bi = 0 if and only if ρ(B) < 1, we obtain the desired convergence (7). !hNh3h2h1Arrayoutputy(i)x3(i)x2(i)x1(i)xN(i)ddFarfieldsignalsources0(i)θ0dsin(θ0)(wavelengthλ)θnInterferencesignalsn(i)PatternformingnetworkIteration,i#10400.20.40.60.811.21.41.61.82OutputSINRindB-6-4-202468101214DistributedsolutionCentralizedsolutionTo prove necessity, we shall prove that every time we have (7), we will have PU a projection matrix and conditions (10), (11), and (12) on A satisfied. We use the fact that limi→∞ Ai exists if, and only if, there is a non singular matrix V such that [43]: A = V IK 0 0 J V−1, where the spectral radius of J is less than one. Let v1, . . . , vM be the columns of V and y∗1, . . . , y∗M be the rows of V−1. Then, we have: lim i→∞Ai = lim 0 0 IK 0 i→∞V = V K(cid:88)m=1 IK 0 J i V−1 0 V−1 = K(cid:88)m=1 vmy∗m= V 0 vmy∗m. IK 0 0 V−1. From the left hand-side of (7) and (80), we obtain: PU = lim i→∞Ai = Observe from (79) that one is an eigenvalue of A with multiplicity K and {vm, ym}K and left eigenvectors. Thus, from (81), we obtain: m=1 are the associated right 19 (79) (80) (81) (82) (83) (84) (85) and equations (10) and (11) hold. Moreover, from (79) and (81), we obtain: which is condition (12). Finally, from (81), we have: vmy∗m = PUA = APU = A vmy∗m = PU vmy∗m = PU K(cid:88)m=1 K(cid:88)m=1 vmy∗mA = ρ (A − PU) = ρV U = V K(cid:88)m=1 K(cid:88)m=1 V−1 = ρ(J ) < 1, 0 V−1 = PU. 0 0 0 J IK 0 P 2 0 Thus, PU is a projector, which completes the necessity proof. Since each vmy∗m is a rank-one matrix and their sum (cid:80)M m=1 vmy∗m = VV−1 = I has rank M, the matrix (cid:80)K m=1 vmy∗m must have rank K. Since the rank of PU is equal to P , we obtain from (81) K = P . Thus, the matrix A has P = K eigenvalues at one and all other eigenvalues are strictly less than one. 20 APPENDIX B DRIVING TERM IN ALGORITHM (19) Let W0 denote an M × 1 vector distributed across the network. In order to justify the choice of (I − A)dD in (19), we consider the problem of finding the projection Wo = PUW0 + dD in a distributed and iterative manner through a linear iteration of the form: Wi = AWi−1 + BdD, (86) where A satisfies (7), (8) and B is a properly chosen matrix ensuring convergence toward Wo. Starting from W0 and iterating the above recursion, we obtain: If we let i → ∞ on both sides of (87), we find: For W∞ to be equal to Wo, B in (86) must be chosen such that (cid:80)∞j=0 AjB = I. In the following, we show that B = I − A + PU ensures convergence. From the Jordan canonical form of A introduced in (41), we have: Wi = Ai W0 + i−1(cid:88)j=0 AjBdD. W0 + ∞(cid:88)j=0 AjBdD. Wi = lim W∞ = lim i→∞ i→∞Ai (cid:124) (cid:123)(cid:122) (cid:125)PU Aj = V I − A + PU = V Aj(I − A + PU) = V IP 0 ∞(cid:88)j=0 IP 0 0 J i IP 0 I − J , . V−1 V−1 (I − J) V−1 0 = I, (87) (88) (89) (90) (91) If we multiply both terms and compute the infinite sum in (88), we obtain: where we used the fact that(cid:80)∞j=0 J i Now, since PU = PD and PDdD = 0, we obtain BdD = (I − A)dD, which justifies the choice in (19). 0 (cid:80)∞j=0 J i = I − J since ρ(J) < 1. APPENDIX C PROOF OF LEMMA 2 We start by noting that the M × M matrix A satisfying conditions (13), (14), and (12) admits a Jordan canonical decomposition of the form: A = VΛV−1, Λ = IP 0 0 J  , V =(cid:104) U VR (cid:105) , V−1 = U∗ V∗L  (92) where the matrix J consists of Jordan blocks, with each one of them having the form (say for a Jordan block of size 3 × 3): , (93)  λ 1 0 0 λ 1 0 0 λ  21 (94) (95) (97) (98) (99) (100) (101) (102) where the eigenvalue λ may be complex but has magnitude less than one. Let E = diag{IP , , 2, . . . , M−P} with > 0 any small positive number independent of µ. The matrix A in (92) can be written alternatively as: where = VΛV−1 A = VE(cid:124)(cid:123)(cid:122)(cid:125)V  , IP 0 0 J V −1 E−1ΛE E−1V−1 (cid:124) (cid:123)(cid:122) (cid:125)Λ (cid:124) (cid:123)(cid:122) (cid:125) V =(cid:104) U VR, (cid:105) , Λ = V−1 = U∗ V∗L,  , and where the matrix J consists of Jordan blocks, with each one of them having a form similar as (93) with > 0 appearing on the upper diagonal instead of 1, and where the eigenvalue λ may be complex but has magnitude less than one. Obviously, since V−1 V = IM , it holds that: U∗VR, = 0, V∗L,U = 0, V∗L,VR, = IM−P , (96) and where U∗U = IP from Assumption 3. Now, let us consider the extended version of the matrix A, namely, Ae, which is an N × N block matrix whose (k, (cid:96))-th block is defined in Table I. In the real data case, we have Ae = A. In the complex data case, it can be verified that Ae is similar to the 2 × 2 block diagonal matrix: 0 Ad = A 0 (A∗)(cid:62)  according to: Ad = IAeI(cid:62) where I is the permutation matrix defined by (31). Using (94), we can rewrite the second block in (97) as: (A∗)(cid:62) = (V∗ )(cid:62)(Λ∗ )(cid:62)((V−1 )∗)(cid:62) = (V∗ )(cid:62)(Λ∗ )(cid:62)((V∗ )(cid:62))−1 where (Λ∗ )(cid:62) = diag{IP , (J ∗ )(cid:62)} and (V∗ )(cid:62) =(cid:104) (U∗)(cid:62) (V∗R,)(cid:62) (cid:105) , ((V−1 )∗)(cid:62) = U(cid:62) V(cid:62)L,  . Now, by replacing (94) and (99) into (97), and by introducing the extended 2 × 2 block diagonal matrices: = diag(cid:110)V, (V∗ )(cid:62)(cid:111) , V d = diag(cid:110)Λ, (Λ∗ )(cid:62)(cid:111) , Λd 22 we find that the 2M × 2M matrix Ad has a Jordan decomposition of the form: Ad = V d Λd (V d )−1. Let us again introduce a permutation matrix I(cid:48) given by: 0 IP  I(cid:48) (cid:44) 0 0 0 0 IM−P 0 0 IP 0 0 0 0 0 IM−P .  (103) (104) (105) The matrix Ad in (103) can be written alternatively as: I(cid:48)Λd I(cid:48)(cid:62) (cid:124) (cid:123)(cid:122) (cid:125) (cid:44)Λe where the matrix Λe that the matrix Ae has a Jordan decomposition of the form: I(cid:48)(cid:62) Ad = V d (cid:124) (cid:123)(cid:122) (cid:125)(cid:44) V d(cid:48) I(cid:48)(V d )−1 (cid:124) (cid:123)(cid:122) (V d(cid:48) )−1 (cid:125) is block diagonal defined in (43). Returning now to Ae, and using (98) and (105), we find Ae = V e Λe )−1, (V e where V e and (V e )−1 are defined by: (cid:44) I(cid:62)V d I(cid:48)(cid:62), V e )−1 (cid:44) I(cid:48)(V d (V e )−1I (106) (107) in terms of the permutation matrices I and I(cid:48) in (30), (104) and the block diagonal matrix V d evaluating these matrices, we arrive at (43). in (101). By properly APPENDIX D PROOF OF THEOREM 1 We consider the transformed variables We in (57). Conditioning both sides on F i−1, computing the conditional second-order moments, using the conditions from Assumption 2 on the gradient noise process, and i and W(cid:86)e i computing the expectations again, we get: and E(cid:107)We i(cid:107)2 = E(cid:13)(cid:13)(IhP − D11,i−1)We i(cid:107)2 = E(cid:13)(cid:13)(cid:13)(J e − D22,i−1)W(cid:86)e E(cid:107)W(cid:86)e i−1 − D12,i−1W(cid:86)e i(cid:107)2 i−1(cid:13)(cid:13)2 + E(cid:107)se (cid:86)e(cid:13)(cid:13)(cid:13) 2 i−1 − D21,i−1We i−1 + b + E(cid:107)s(cid:86)e i(cid:107)2 . (108) (109) 23 By applying Jensen's inequality to the convex function (cid:107)x(cid:107)2, we can bound the first term on the RHS of (108) as follows: E(cid:13)(cid:13)(IhP − D11,i−1)We i−1 − D12,i−1W(cid:86)e i−1(cid:13)(cid:13)2 = E(cid:13)(cid:13)(cid:13)(cid:13)(1 − t) ≤ ≤ 1 1 − t 1 1 − t 2 1 t 1 1 − t (IhP − D11,i−1)We E(cid:13)(cid:13)(IhP − D11,i−1)We i−1(cid:13)(cid:13)2 + E(cid:104)(cid:107)IhP − D11,i−1(cid:107)2(cid:13)(cid:13)We i−1(cid:13)(cid:13)2(cid:105) + i−1− t 1 t i−1(cid:13)(cid:13)(cid:13)(cid:13) D12,i−1W(cid:86)e i−1(cid:13)(cid:13)2 E(cid:13)(cid:13)D12,i−1W(cid:86)e E(cid:104)(cid:107)D12,i−1(cid:107)2(cid:13)(cid:13)W(cid:86)e 1 t i−1(cid:13)(cid:13)2(cid:105) (110) for any arbitrary positive number t ∈ (0, 1). By Assumption 1, the Hermitian matrix H k,i−1 defined in (47) can be bounded as follows: νk h 0 ≤ IhMk ≤ H k,i−1 ≤ δk h IhMk. (111) Using the fact that the integral of a matrix is the matrix of the integrals, and the linear property of integration, the Hermitian block D11,i−1 in (62) can be rewritten as: 0 (cid:104)(U e)∗diag(cid:8)∇2 D11,i−1= µ(cid:90) 1 and, therefore, from Assumption 1, D11,i−1 can be bounded as follows: k − t(cid:101)wk,i−1)(cid:9)N k=1 U e(cid:105)dt Jk(wo (112) wk 0 < µ ν h IhP ≤ D11,i−1 ≤ µ δ h IhP , (113) for some positive constants ν and δ that are independent of µ and i. In terms of the 2−induced matrix norm (i.e., maximum singular value), we obtain: and, therefore, (cid:107)IhP − D11,i−1(cid:107) = ρ(IhP − D11,i−1) ≤ max(cid:26)(cid:12)(cid:12)(cid:12)(cid:12)1 − µ δ h(cid:12)(cid:12)(cid:12)(cid:12) ,(cid:12)(cid:12)(cid:12)1 − µ (cid:107)IhP − D11,i−1(cid:107)2 ≤ (1 − µσ11)2, ν h(cid:12)(cid:12)(cid:12)(cid:27) , (114) (115) for some positive constant σ11 that is independent of µ and i. Similarly, using the 2−induced matrix norm (i.e., maximum singular value), we can bound (cid:107)D12,i−1(cid:107)2 as follows: (cid:107)D12,i−1(cid:107)2 (63) R,(cid:107)2 = (cid:107)µ (U e)∗Hi−1V e ≤ µ2(cid:107)(U e)∗(cid:107)2(cid:107)Hi−1(cid:107)2(cid:107)V e R,(cid:107)2 ≤ µ2(cid:18) max 1≤k≤N (cid:107)H k,i−1(cid:107)2(cid:19)(cid:107)V e R,(cid:107)2 1≤k≤N(cid:26) δ2 h2(cid:27) R,(cid:107)2 max ≤ µ2(cid:107)V e = µ2σ2 12, k (111) for some positive constant σ12 and where we used the fact that (cid:107)(U e)∗(cid:107) = σmax((U e)∗) =(cid:112)λmax(U e(U e)∗) = 1. Substituting (110) into (108), and using (115), (116), we get: E(cid:107)We i(cid:107)2 ≤ (1 − σ11µ)2 1 − t E(cid:107)We i−1(cid:107)2 + µ2σ2 12 t E(cid:107)W(cid:86)e i−1(cid:107)2 + E(cid:107)se i(cid:107)2 (117) (116) 24 We select t = σ11µ (for sufficiently small µ). Then, the previous inequality can be written as: E(cid:107)We i(cid:107)2 ≤ (1 − σ11µ)E(cid:107)We i−1(cid:107)2 + µσ2 12 σ11 E(cid:107)W(cid:86)e i−1(cid:107)2 + E(cid:107)se i(cid:107)2. (118) We repeat similar arguments for the second variance relation (109). Using Jensen's inequality again, we obtain: − D22,i−1)W(cid:86)e E(cid:107)(J e 1 E(cid:107)J e W(cid:86)e i−1(cid:107)2 + ≤ t (ρ(J) + )2 E(cid:107)W(cid:86)e ≤ = (ρ(J) + )E(cid:107)W(cid:86)e i−1 − D21,i−1We E(cid:107)D22,i−1W(cid:86)e 1 1 − t i−1(cid:107)2 + i−1(cid:107)2 + 1 1 − t 1 1 − ρ(J) − (b) (a) t (cid:86)e (cid:107)2 i−1 + b i−1 + D21,i−1We E(cid:107)D22,i−1W(cid:86)e (cid:86)e (cid:107)2 i−1 − b i−1 + D21,i−1We i−1 − b i−1 + D21,i−1We (cid:86)e E(cid:107)D22,i−1W(cid:86)e (119) (cid:107)2, (cid:107)2 i−1 − b (cid:86)e for any arbitrary positive number t ∈ (0, 1). In (a) we used the fact that the block diagonal matrix J e Table I satisfies: defined in (cid:107)J e (cid:107)2 ≤ (ρ(J) + )2. (120) Expression (120) can be established by using similar arguments as in [4, pp. 516] and the fact that λmax(cid:0)J (cid:62) (J ∗ )(cid:62)(cid:1) = λmax(cid:0)(J ∗ J)(cid:62)(cid:1) = λmax(J ∗ J). In (b), we used the fact that ρ(J) ∈ (0, 1), and thus, can be selected small enough to ensure ρ(J) + ∈ (0, 1). We then selected t = ρ(J) + . Using Jensen's inequality, the second term on the RHS of (119) can be bounded as follows: E(cid:107)D22,i−1W(cid:86)e i−1 + D21,i−1We i−1 − b (cid:86)e (cid:107)2 ≤ 3E[(cid:107)D22,i−1(cid:107)2(cid:107)W(cid:86)e i−1(cid:107)2] + 3E[(cid:107)D21,i−1(cid:107)2(cid:107)We (cid:86)e i−1(cid:107)2] + 3(cid:107)b Following similar arguments as in (116), we can show that: (cid:107)D21,i−1(cid:107)2 ≤ µ2σ2 21, (cid:107)D22,i−1(cid:107)2 ≤ µ2σ2 22, (121) (cid:107)2 (122) for some positive constants σ21 and σ22. Substituting (121) into (119) and (119) into (109), and using (122), we obtain: E(cid:107)W(cid:86)e i(cid:107)2 ≤(cid:18)ρ(J) + + 1 − ρ(J) − (cid:19) E(cid:107)W(cid:86)e 1 − ρ(J) − (cid:19) E(cid:107)We 1 − ρ(J) − (cid:19)(cid:107)b i−1(cid:107)2 +(cid:18) i−1(cid:107)2 +(cid:18) (cid:107)2 + E(cid:107)s(cid:86)e 3µ2σ2 22 3µ2σ2 21 (cid:86)e 3 i(cid:107)2. (123) From (59), we have: (cid:86)e (cid:107)b where we used the fact that (V e gradient ∇w∗ For the noise terms E(cid:107)se Jk(wo k (cid:107)2 = (cid:107)µJ e (V e L,)∗Ae = J e L,)∗be(cid:107)2 ≤ µ2(ρ(J) + )2(cid:107)(V e (V e L,)∗(cid:107)2(cid:107)be(cid:107)2 k) and since Jk(wk) is twice differentiable, then (cid:107)be(cid:107)2 is bounded and we obtain (cid:107)b L,)∗ from Lemma 2. Since be in (53) is defined in terms of the (cid:107)2 = O(µ2). (cid:86)e (124) i(cid:107)2 in (118) and E(cid:107)s(cid:86)e i(cid:107)2 + E(cid:107)s(cid:86)e E(cid:107)se i(cid:107)2 in (123), we have: = E(cid:107)µ(V e i(cid:107)2 (58) )−1Aese i(cid:107)2 ≤ µ2v2 1 E(cid:107)se i(cid:107)2, (125) where v1 is a positive constant independent of µ and given by v1 (cid:44) (cid:107)(V e i(cid:107)2 =(cid:80)N variances of the individual noise processes, E(cid:107)sk,i(cid:107)2, we have E(cid:107)se each sk,i(wk,i−1), we have from Assumption 2 and Jensen's inequality: k(cid:107)2 + σ2 k − wk,i−1 + wo k/h2)(cid:107)wo E(cid:107)sk,i(wk,i−1)(cid:107)2 ≤ (β2 ≤ 2(β2 = ¯β2 k k=1 s,k k(cid:107)2 + σ2 s,k )−1Ae(cid:107) = (cid:107)Λe (V e )−1(cid:107). In terms of the E(cid:107)sk,i(cid:107)2. For k=1 k,i(cid:107)2 = 2(cid:80)N E(cid:107)se where ¯β2 k (cid:44) 2(β2 k/h2) and ¯σ2 s,k i(cid:107)2 can thus be bounded as follows: (cid:44) 2(β2 s,k ¯β2 k k(cid:107)2 + σ2 k/h2)(cid:107)wo k/h2)E(cid:107)wo k/h2)E(cid:107)(cid:101)wk,i−1(cid:107)2 + 2(β2 E(cid:107)(cid:101)wk,i−1(cid:107)2 + ¯σ2 s,k. The term E(cid:107)se N(cid:88)k=1 N(cid:88)k=1 E(cid:107)(cid:101)wk,i−1(cid:107)2 + 2 E(cid:107)si(cid:107)2 ≤ 2 E(cid:107)V e )−1(cid:101)We i−1(cid:107)2 + σ2 ≤ β2 (V e (cid:107)2E(cid:107)(V e )−1(cid:101)We max(cid:107)V e ≤ β2 2(cid:2)E(cid:107)We i−1(cid:107)2 + E(cid:107)W(cid:86)e = β2 maxv2 s,k, and v2 (cid:44) (cid:107)V e 2[E(cid:107)We 1β2 maxv2 max s ¯σ2 s,k s i−1(cid:107)2 + σ2 i−1(cid:107)2(cid:3) + σ2 s (cid:107). Substituting into (125), we get: i−1(cid:107)2 + E(cid:107)W(cid:86)e i−1(cid:107)2] + µ2v2 1σ2 s . max ¯β2 k, σ2 where β2 (cid:44) max1≤k≤N E(cid:107)se (cid:44) 2(cid:80)N i(cid:107)2 + E(cid:107)s(cid:86)e i(cid:107)2 ≤ µ2v2 Using this bound in (118) and (123), we obtain: k=1 ¯σ2 s E(cid:107)W(cid:86)e E(cid:107)We i(cid:107)2 ≤(cid:18) +(cid:18) i(cid:107)2 ≤(1 − σ11µ + µ2v2 1β2 maxv2 3µ2σ2 21 + µ2v2 1β2 maxv2 1 − ρ(J) − 3 1 − ρ(J) − (cid:19)(cid:107)b (cid:86)e (cid:107)2 + µ2v2 1σ2 s . 12 σ11 + µ2v2 2)E(cid:107)We 2(cid:19) E(cid:107)We i−1(cid:107)2 +(cid:18) µσ2 i−1(cid:107)2 +(cid:18)ρ(J) + + 1β2 maxv2 2(cid:19) E(cid:107)W(cid:86)e 3µ2σ2 22 1 − ρ(J) − 25 (126) (127) (128) (129) i−1(cid:107)2 + µ2v2 1σ2 s , + µ2v2 1β2 maxv2 2(cid:19) E(cid:107)W(cid:86)e i−1(cid:107)2 (130) (131) (132) We can combine (129) and (130) into a single inequality recursion: where Γ is given by:  Γ (cid:44) E(cid:107)We i(cid:107)2 E(cid:107)W(cid:86)e i(cid:107)2  (cid:22) Γ c d  = a b E(cid:107)We E(cid:107)W(cid:86)e e i−1(cid:107)2 i−1(cid:107)2  + e + f  ρ(J) + + O(µ2)  . O(µ) 1 − O(µ) O(µ2) and where a = 1 − O(µ), b = O(µ), c = O(µ2), d = ρ(J) + + O(µ2), e = O(µ2), and f = O(µ2). Now, using the property that the spectral radius of a matrix is upper bounded by its 1−norm norm, we obtain: ρ(Γ) ≤ max(cid:8)1 − O(µ) + O(µ2), ρ(J) + + O(µ) + O(µ2)(cid:9) (133) 26 Since ρ(J) < 1 is independent of µ, and since and µ are small positive numbers that can be chosen arbitrarily small and independently of each other, it is clear that the RHS of the above expression can be made strictly smaller than one for sufficiently small and µ. In that case ρ(Γ) < 1 so that Γ is stable. Moreover, it holds that: Now, by iterating (131) we arrive at: E(cid:107)We i(cid:107)2 E(cid:107)W(cid:86)e from which we conclude that lim supi→∞ lim sup E(cid:107)We i→∞  E(cid:107)(cid:101)We lim sup i→∞ O(1/µ) O(1) (I − Γ)−1 = i(cid:107)2  (cid:22) (I − Γ)−1 E(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  V e 2(cid:2)E(cid:107)We O(µ) O(1)  . e + f  = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) i(cid:107)2 = lim sup i→∞ ≤ lim sup i→∞ We i W(cid:86)e i i(cid:107)2 + E(cid:107)W(cid:86)e i(cid:107)2 = O(µ) and lim supi→∞ v2 e 2 i(cid:107)2(cid:3) = O(µ). O(µ) O(µ2)  E(cid:107)W(cid:86)e i(cid:107)2 = O(µ2). Therefore, (134) (135) (136) REFERENCES [1] R. Nassif, S. Vlaski, and A. H. Sayed, "Distributed inference over networks under subspace constraints," in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process., Brighton, UK, May 2019, pp. 1 -- 5. [2] R. Nassif, S. Vlaski, and A. H. Sayed, "Adaptation and learning over networks under subspace constraints -- Part II: Performance analysis," Submitted for publication., May 2019. [3] A. H. Sayed, S. Y. Tu, J. Chen, X. Zhao, and Z. J. Towfic, "Diffusion strategies for adaptation and learning over networks," IEEE Signal Process. Mag., vol. 30, no. 3, pp. 155 -- 171, 2013. [4] A. H. Sayed, "Adaptation, learning, and optimization over networks," Foundations and Trends in Machine Learning, vol. 7, no. 4-5, pp. 311 -- 801, 2014. [5] A. H. Sayed, "Adaptive networks," Proc. IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014. [6] D. Bertsekas, "A new class of incremental gradient methods for least squares problems," SIAM J. Optim., vol. 7, no. 4, pp. 913 -- 926, 1997. [7] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi-agent optimization," IEEE Trans. Autom. Control, vol. 54, no. 1, pp. 48 -- 61, Jan. 2009. [8] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, "Gossip algorithms for distributed signal processing," Proc. IEEE, vol. 98, no. 11, pp. 1847 -- 1864, Nov. 2010. [9] S. Chouvardas, K. Slavakis, and S. Theodoridis, "Adaptive robust distributed learning in diffusion sensor networks," IEEE Trans. Signal Process., vol. 59, no. 10, pp. 4692 -- 4707, Oct. 2011. [10] P. Braca, S. Marano, and V. Matta, "Enforcing consensus while monitoring the environment in wireless sensor networks," IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3375 -- 3380, Jul. 2008. [11] G. N. Korres, "A distributed multiarea state estimation," IEEE Trans. Power Syst., vol. 26, no. 1, pp. 73 -- 84, Feb. 2011. [12] V. Kekatos and G. B. Giannakis, "Distributed robust power system state estimation," IEEE Trans. Signal Process., vol. 28, no. 2, pp. 1617 -- 1626, 2013. [13] J. Chen, C. Richard, and A. H. Sayed, "Multitask diffusion adaptation over networks," IEEE Trans. Signal Process., vol. 62, no. 16, pp. 4129 -- 4144, 2014. 27 [14] S. Khawatmi and A. M. Zoubir, "Decentralized partitioning over adaptive networks," in Proc. IEEE Int. Workshop Mach. Learn. Signal Process., Vietri sul Mare, Italy, Sep. 2016, pp. 1 -- 6. [15] J. Plata-Chaves, A. Bertrand, M. Moonen, S. Theodoridis, and A. M. Zoubir, "Heterogeneous and multitask wireless sensor networks -- Algorithms, applications, and challenges," IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 3, pp. 450 -- 465, 2017. [16] R. Nassif, S. Vlaski, and A. H. Sayed, "Distributed inference over multitask graphs under smoothness," in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun., Kalamata, Greece, Jun. 2018, pp. 1 -- 5. [17] A. Koppel, B. M. Sadler, and A. Ribeiro, "Proximity without consensus in online multi-agent optimization," in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process., Shanghai, China, May 2016, pp. 3726 -- 3730. [18] D. Hallac, J. Leskovec, and S. Boyd, "Network LASSO: Clustering and optimization in large graphs," in Proc. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney, Australia, Aug. 2015, pp. 387 -- 396. [19] X. Cao and K. J. R. Liu, "Distributed efficient optimization for general network cost minimization problems," in Proc. IEEE International Conference on Communications, Kansas City, MO, USA, May 2018, pp. 1 -- 6. [20] R. Nassif, S. Vlaski, C. Richard, and A. H. Sayed, "A regularization framework for learning over multitask graphs," IEEE Signal Process. Lett., vol. 26, no. 2, pp. 297 -- 301, Feb. 2019. [21] J. Chen, C. Richard, A. O. Hero, and A. H. Sayed, "Diffusion LMS for multitask problems with overlapping hypothesis subspaces," in Proc. IEEE Int. Workshop Mach. Learn. Signal Process. (MLSP), Reims, France, Sept. 2014, pp. 1 -- 6. [22] J. F. C. Mota, J. M. F. Xavier, P. M. Q. Aguiar, and M. Puschel, "Distributed optimization with local domains: Applications in MPC and network flows," IEEE Trans. Autom. Control, vol. 60, no. 7, pp. 2004 -- 2009, Jul. 2015. [23] S. A. Alghunaim and A. H. Sayed, "Distributed coupled multi-agent stochastic optimization," to appear in IEEE Trans. Automatic Control. Also available as arXiv:1712.08817, 2019. [24] J. Plata-Chaves, N. Bogdanovi´c, and K. Berberidis, "Distributed diffusion-based LMS for node-specific adaptive parameter estimation," IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3448 -- 3460, Jul. 2015. [25] A. K. Sahu, D. Jakoveti´c, and S. Kar, "CIRFE: A distributed random fields estimator," IEEE Trans. Signal Process., vol. 66, no. 18, pp. 4980 -- 4995, Sep. 2018. [26] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Diffusion LMS for multitask problems with local linear equality constraints," IEEE Trans. Signal Process., vol. 65, no. 19, pp. 4979 -- 4993, 2017. [27] S. A. Alghunaim, K. Yuan, and A. H. Sayed, "Dual coupled diffusion for distributed optimization with affine constraints," in Proc. IEEE Conf. Decis. Control, Miami Beach, FL., USA, Dec. 2018, pp. 829 -- 834. [28] F. Hua, R. Nassif, C. Richard, and H. Wang, "Penalty-based multitask estimation with non-local linear equality constraints," in Proc. IEEE Int. Work. on Comp. Adv. in Multi-sensor Adapt. Process., Curacao, Dutch Antilles, Dec. 2017, pp. 433 -- 437. [29] S. Barbarossa, G. Scutari, and T. Battisti, "Distributed signal subspace projection algorithms with maximum convergence rate for sensor networks with topological constraints," in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process., Taipei, Taiwan, Apr. 2009, pp. 2893 -- 2896. [30] D. P. Bertsekas, Nonlinear Programming, Athena Scientific, 1999. [31] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging," Systems & Control Letters, vol. 53, no. 1, pp. 65 -- 78, 2004. [32] H. L. Van Trees, Optimum Array Processing, Wiley, NJ, 2002. [33] O. L. Frost III, "An algorithm for linearly constrained adaptive array processing," Proc. IEEE, vol. 60, no. 8, pp. 926 -- 935, Aug. 1972. [34] F. R. K Chung, Spectral Graph Theory, American Mathematical Society, 1997. [35] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, second edition, 2013. [36] B. T. Polyak, Introduction to Optimization, Optimization Software, New York, 1987. [37] A. I. Koutrouvelis, T. W. Sherson, R. Heusdens, and R. C. Hendriks, "A low-cost robust distributed linearly constrained beamformer for wireless acoustic sensor networks with arbitrary topology," IEEE Audio, Speech, Language Process., vol. 26, no. 8, pp. 1434 -- 1448, Aug. 2018. 28 [38] M. O'Connor and W. B. Kleijn, "Diffusion-based distributed MVDR beamformer," in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process., Florence, Italy, May 2014, pp. 810 -- 814. [39] Y. Zeng and R. C. Hendriks, "Distributed delay and sum beamformer for speech enhancement via randomized gossip," IEEE Audio, Speech, Language Process., vol. 22, no. 1, pp. 260 -- 273, Jan. 2014. [40] M. O'Connor, W. B. Kleijn, and T. Abhayapala, "Distributed sparse MVDR beamforming using the bi-alternating direction method of multipliers," in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process., Shanghai, China, Mar. 2016, pp. 106 -- 110. [41] A. Bertrand and M. Moonen, "Distributed LCMV beamforming in a wireless sensor network with single-channel per-node signal transmission," IEEE Trans. Signal Process., vol. 61, no. 13, pp. 3447 -- 3459, Jul. 2013. [42] M. Grant and S. Boyd, "CVX: Matlab software for disciplined convex programming, version 2.1," http://cvxr.com/cvx, Mar. 2014. [43] R. Oldenburger, "Infinite powers of matrices and characteristic roots," Duke Mathematical Journal, vol. 6, no. 2, pp. 357 -- 361, 1940.
1903.00784	1	1903	2019-03-02T22:42:33	Neural MMO: A Massively Multiagent Game Environment for Training and Evaluating Intelligent Agents	[ "cs.MA", "cs.LG", "stat.ML" ]	The emergence of complex life on Earth is often attributed to the arms race that ensued from a huge number of organisms all competing for finite resources. We present an artificial intelligence research environment, inspired by the human game genre of MMORPGs (Massively Multiplayer Online Role-Playing Games, a.k.a. MMOs), that aims to simulate this setting in microcosm. As with MMORPGs and the real world alike, our environment is persistent and supports a large and variable number of agents. Our environment is well suited to the study of large-scale multiagent interaction: it requires that agents learn robust combat and navigation policies in the presence of large populations attempting to do the same. Baseline experiments reveal that population size magnifies and incentivizes the development of skillful behaviors and results in agents that outcompete agents trained in smaller populations. We further show that the policies of agents with unshared weights naturally diverge to fill different niches in order to avoid competition.	cs.MA	cs	Neural MMO: A Massively Multiagent Game Environment for Training and Evaluating Intelligent Agents Joseph Suarez Yilun Du Phillip Isola Igor Mordatch 9 1 0 2 r a M 2 ] A M . s c [ 1 v 4 8 7 0 0 . 3 0 9 1 : v i X r a Abstract The emergence of complex life on Earth is of- ten attributed to the arms race that ensued from a huge number of organisms all competing for finite resources. We present an artificial intelligence re- search environment, inspired by the human game genre of MMORPGs (Massively Multiplayer On- line Role-Playing Games, a.k.a. MMOs), that aims to simulate this setting in microcosm. As with MMORPGs and the real world alike, our en- vironment is persistent and supports a large and variable number of agents. Our environment is well suited to the study of large-scale multiagent interaction: it requires that agents learn robust combat and navigation policies in the presence of large populations attempting to do the same. Baseline experiments reveal that population size magnifies and incentivizes the development of skillful behaviors and results in agents that out- compete agents trained in smaller populations. We further show that the policies of agents with unshared weights naturally diverge to fill different niches in order to avoid competition. 1. Introduction Life on Earth can be viewed as a massive multiagent compe- tition. The cheetah evolves an aerodynamic profile in order to catch the gazelle, the gazelle develops springy legs to run even faster: species have evolved ever new capabilities in order to outcompete their adversaries. The success of biological evolution has inspired many attempts at creating "artificial life" in silico. In recent years, the field of deep reinforcement learning (RL) has embraced a related approach: train agents by hav- ing them compete in simulated games (Silver et al., 2016; OpenAI, 2018; Jaderberg et al., 2018). Such games are immediately interpretable and provide easy metrics derived from the game's "score" and win conditions. However, pop- ular game benchmarks typically define a narrow, episodic task with a small fixed number of players. In contrast, life on Earth involves a persistent environment, an unbounded number of players, and a seeming "open-endedness", where ever new and more complex species emerge over time, with no end in sight (Stanley et al., 2017). Our aim is to develop a simulation platform (see Figure 1) that captures important properties of life on Earth, while also borrowing from the interpretability and abstractions of human-designed games. To this end, we turn to the game genre of Massively Multiplayer Online Role-Playing Games (MMORPGs, or MMOs for short). These games involve a large, variable number of players competing to survive and prosper in persistent and far-flung environments. Our platform simulates a "Neural MMO" -- an MMO in which each agent is a neural net that learns to survive using RL. We demonstrate the capabilities of this platform through a series of experiments that investigate emergent complexity as a function of the number of agents and species that com- pete in the simulation. We find that large populations act as competitive pressure that encourages exploration of the environment and the development of skillful behavior. In addition, we find that when agents are organized into species (share policy parameters), each species naturally diverges from the others to occupy its own behavioral niche. Upon publication, we will opensource the platform in full. 2. Background and Related Work Artificial Life and Multiagent Reinforcement Learning Research in "Artificial life" aims to model evolution and natural selection in biological life; (Langton, 1997; Ficici & Pollack, 1998). Such projects often consider open-ended skill learning (Yaeger, 1994) and general morphology evo- lution (Sims, 1994) as primary objectives. Similar problems have recently resurfaced within multiagent reinforcement learning where the continual co-adaptation of agents can introduce additional nonstationarity that is not present in single agent environments. While there have been multiple attempts to formalize the surrounding theory (Hern´andez- Orallo et al., 2011; Strannegrd et al., 2018), we primarily consider environment-driven works. These typically con- sider either complex tasks with 2-10 agents (Bansal et al., 2017; OpenAI, 2018; Jaderberg et al., 2018) or much sim- pler environments with tens to upwards of a million agents Neural MMO Figure 1. Our Neural MMO platform provides a procedural environment generator and visualization tools for value functions, map tile visitation distribution, and agent-agent dependencies of learned policies. Baselines are trained with policy gradients over 100 worlds. (Lowe et al., 2017; Mordatch & Abbeel, 2017; Bansal et al., 2017; Lanctot et al., 2017; Yang et al., 2018a; Zheng et al., 2017; Jaderberg et al., 2018). Most such works further focus on learning a specific dynamic, such as predator-prey Yang et al. (2018b) or are more concerned with the study than the learning of behavior, and use hard-coded rewards Zheng et al. (2017). In contrast, our work focuses on large agent populations in complex environments. Game Platforms for Intelligent Agents The Arcade Learning Environment (ALE) (Bellemare et al., 2013) and Gym Retro (Nichol et al., 2018) provide 1000+ limited scope arcade games most often used to test individual re- search ideas or generality across many games. Better per- formance at a large random subset of games is a reasonable metric of quality. However, recent results have brought into question the overall complexity each individual environment (Cuccu et al., 2018), and strong performance in such tasks is not particularly difficult for humans. More recent work has demonstrated success on multiplayer games including Go (Silver et al., 2016), the Multiplayer On- line Battle Arena (MOBA) game DOTA2 (OpenAI, 2018), and Quake 3 Capture the Flag (Jaderberg et al., 2018). Each of these projects has advanced our understanding of a class of algorithms. However, these games are limited to 2-12 players, are episodic, with game rounds on the order of an hour, lack persistence, and lack the game mechanics sup- porting large persistent populations -- there is still a large gap in environment complexity compared to the real world. Role-playing games (RPGs) such as Pokemon and Final Fantasy, are in-depth experiences designed to engage human players for hundreds of hours of persistent gameplay. Like the real world, problems in RPGs have many valid solutions and choices have long term consequences. MMORPGs are the (massively) multiplayer analogs to RPGs. They are typically run across several persistent servers, each of which contains a copy of the environment and supports hundreds to millions of concurrent players. Good MMOs require increasingly clever, team-driven usage of the game systems: players attain the complex skills and knowledge required for the hardest challenges only through a curriculum of content spanning hundreds of hours of game- play. Such a curriculum is present in many game genres, but only MMOs contextualize it within persistent social and economic structures approaching the scale of the real world. 3. Neural MMO We present a persistent and massively multiagent environ- ment that defines foraging and combat systems over pro- cedurally generated maps. The Supplement provides full Neural MMO Figure 2. Our platform includes an animated 3D client and a toolbox used to produce the visuals in this work. Agents compete for food and water while engaging in strategic combat. See the Neural MMO section for a brief overview and the Supplement for full details. environment details and Figure 2 shows a snapshot. The core features are support for a large and variable number of agents, procedural generation of tile-based terrain, a food and water foraging system, a strategic combat system, and inbuilt visualization tools for analyzing learned policies Agents (players) may join any of several servers (environ- ment instances). Each server contains an automatically gen- erated tile-based environment of configurable size. Some tiles, such as food-bearing forest tiles and grass tiles, are traversable. Others, such as water and solid stone, are not. Upon joining a server, agents spawn at a random location along the edges of the environment. In order remain healthy (maintain their health statistic), agents must obtain food and water -- they die upon reaching reaching 0 health. At each server tick (time step), agents may move one tile and make an attack. Stepping on a forest tile or next to a water tile refills a portion of the agent's food or water supply, respec- tively. However, forest tiles have a limited supply of food; once exhausted, food has a 2.5 percent chance to regenerate each tick. This means that agents must compete for food tiles while periodically refilling their water supply from in- finite water tiles. They may attack each other using any of three attack options, each with different damage values and tradeoffs. Precise foraging and combat mechanics are detailed in the Supplement. Agents observe local game state and decide on an action each game tick. The environment does not make any further assumptions on the source of that decision, be it a neural network or a hardcoded algorithm. We have tested the envi- ronment with up to 100 million agent trajectories (lifetimes) on 100 cores in 1 week. Real and virtual worlds alike are open-ended tasks where complexity arises with little direction. Our environment is designed as such. Instead of rewarding agents for achieving particular objectives op- timize only for survival time: they receive reward rt = 1 for each time step alive. Competition for finite resources mandates that agents must learn intelligent strategies for gathering food and water in order to survive. One purpose of the platform is to discover game mechanics that support complex behavior and agent populations that can learn to make use of them. In human MMOs, developers aim to create balanced mechanics while players aim to max- imize their skill in utilizing them. The initial configurations of our systems are the results of several iterations of balanc- ing, but are by no means fixed: every numeric parameter presented is editable within a simple configuration file. of survival rewards: R(τ ) =(cid:80)T 4. Architecture and Training Agents are controlled by policies parameterized by neural networks. Agents make observations ot of the game state st and follow a policy π(ot) → at in order to make ac- tions at. We maximize a return function R over trajectory τ = (ot, at, rt, ..., oT , aT , rT ). This is a discounted sum t γtrt where γ = 0.99, T is the time at death and the survival reward rt equals 1, as motivated previously. The policy π may be different for each agent or shared. Algorithm 1 shows high level training logic. The Supplement details the tile-based game state st and hyperparameters (Table 1). Neural MMO Figure 3. Maximum population size at train time varies in (16, 32, 64, 128). At test time, we merge the populations learned in pairs of experiments and evaluate lifetimes at a fixed population size. Agents trained in larger populations always perform better. Figure 4. Population size magnifies exploration: agents spread out to avoid competition. Figure 5. Populations count (number of species) magnifies niche formation. Visitation maps are overlaid over the game map; different colors correspond to different species. Training a single population tends to produce a single deep exploration path. Training eight populations results in many shallower paths: populations spread out to avoid competition among species. Neural MMO Algorithm 1 Neural MMO logic for one game tick. See Experiments (Technical details) for spawning logic. The al- gorithm below makes two omissions for simplicity. First, we use multiple policies and sample a policy π ∼ π1, . . . , πN from the set of all policies when spawning a new agent. Second, instead of performing a policy gradient update ev- ery game tick, we maintain experience buffers from each environment and perform an update once all buffers are full. for each environment server do if number of agents alive < spawn cap then spawn an agent end if for each agent do i ← population index of the agent Make observation ot, decide action πi(ot) → at Environment processes at, computes rt, and updates agent health, food, etc. if agent is dead then remove agent end if end for Update environment state st+1 → f (st, at) end for Perform a policy gradient update on policies π ∼ π1, . . . , πN using ot, at, rt from all agents across all environment servers Input We set the observation state ot equal to the crop of tiles within a fixed L1 distance of the current agent. This includes tile terrain types and the select properties (such as health, food, water, and position) of occupying agents. Our choice of ot is an equivalent representation of what a human sees on the screen, but our environment supports other choices as well. Note that computing observations does not require rendering. Output Agents output action choices at for the next time step (game tick). Actions consist of one movement and one attack. Movement options are: North, South, East, West, and Pass (no movement). Attack options are labeled: Melee, Range, and Mage, with each attack option applying a specific preset amount of damage at a preset effective dis- tance. The environment will attempt to execute both actions. Invalid actions, (e.g. moving into stone), are ignored. Our policy architecture preprocesses the local environment by embedding it and flattening it into a single fixed length vector. We then apply a linear layer followed by linear output heads for movement and attack decisions. New types of action choices can be included by adding additional heads. We also train a value function to estimate the discounted return. As agents receive only a stream of reward 1, this is equal to a discounted estimate of the agent's time until death. We use a value function baselines policy gradient loss and optimize with Adam. It was possible to obtain good performance without discounting, but training was less stable. We provide full details in the supplements. 5. Experiments We present an initial series of experiments using our plat- form to explore multiagent interactions in large popula- tions. We find that agent competence scales with population size. In particular, increasing the maximum number of con- current players (Nent) magnifies exploration and increas- ing the maximum number of populations with unshared weights (Npop) magnifies niche formation. Agents poli- cies are sampled uniformly from a number of "populations" π ∼ π1, . . . , πN . Agents in different populations have the same architecture but do not share weights. Technical details We run each experiment using 100 worlds. We define a constant C over the set of worlds W . For each world w ∈ W , we uniformly sample a c ∈ (1, 2, ...C). We define "spawn cap" such that if world w has a spawn cap c, the number of agents in w cannot exceed c. In each world w, one agent is spawned per game tick provided that doing so would exceed the spawn cap c of w. To match standard MMOs, we would fix Nent = Npop (humans are independent networks with unshared weights). However, this incurs sample complexity proportional to number of populations. We therefore share parameters across groups of up to 16 agents for efficiency. 5.1. Server Merge Tournaments We perform four experiments to evaluate the effects on foraging performance of training with larger populations and with a greater number of populations. For each experiment, we fix Npop ∈ (1, 2, 4, 8) and a spawn cap (the maximum number of concurrent agents) c = 16 × Npop, such that c ∈ (16, 32, 64, 128). We train for a fixed number of trajectories per population. Evaluating the influence of these variables is nontrivial. The task difficulty is highly dependent on the size and com- petence of populations in the environment: mean agent lifetime is not comparable across experiments. Furthermore, there is no standard procedure among MMOs for evaluating relative player competence across multiple servers. How- ever, MMO servers sometimes undergo merges whereby the player bases from multiple servers are placed within a single server. As such, we propose tournament style evaluation in order to directly compare policies learned in different experiment settings. Tournaments are formed by simply concatenating the player bases of each experiment. Figure 3 shows results: we vary the maximum number of agents at test time and find that agents trained in larger settings consistently outperform agents trained in smaller settings. Neural MMO We observe more interesting policies once we introduce the combat module as an additional learnable mode of vari- ation on top of foraging. With combat, agent actions be- come strongly coupled with the states of other agents. As a sanity check, we also confirm that all of the populations trained with combat handily outperform all of the popula- tions trained with only foraging, when these populations compete in a tournament with combat enabled. To better understand theses results, we decouple our anal- ysis into two modes of variability: maximum number of concurrent players (Nent) and maximum number of pop- ulations with unshared weights (Npop). This allows us to examine the effects of each factor independently. In order to isolate the effects of environment randomization, which also encourages exploration, we perform these experiments on a fixed map. Isolating the effects of these variables pro- duces more immediately obvious results, discussed in the following two subsections: 5.2. Nent: Multiagent Magnifies Exploration In the natural world, competition between animals can in- centivize them to spread out in order to avoid conflict. We observe that overall exploration (map coverage) increases as the number of concurrent agents increases (see Figure 4; the map used is shown in Figure 5). Agents learn to explore only because the presence of other agents provides a natural incentive for doing so. 5.3. Npop: Multiagent Magnifies Niche Formation We find that, given a sufficiently large and resource-rich en- vironment, different populations of agents tend to separate to avoid competing with other populations. Both MMOs and the real world often reward masters of a single craft more than jacks of all trades. From Figure 5, specialization to particular regions of the map increases as number of pop- ulations increases. This suggests that the presence of other populations force agents to discover a single advantageous skill or trick. That is, increasing the number of populations results in diversification to separable regions of the map. As entities cannot out-compete other agents of their own population (i.e. agent's with whom they share weights), they tend to seek areas of the map that contain enough resources to sustain their population. 5.4. Environment Randomized Exploration The trend of increasing exploration with increasing entity number is clear when training on a single map as seen in Figure 4, 5, but it is more subtle with environment ran- domization. From Figure 6, all population sizes explore adequately. It is likely that "exploration" as defined by map coverage is not as difficult a problem, in our environment, as developing robust policies. As demonstrated by the Tourna- ment experiments, smaller populations learn brittle policies that do not generalize to scenarios with more competitive pressure -- even against a similar number of agents. 5.5. Agent-Agent Dependencies We visualize agent-agent dependencies in Figure 7. We fix an agent at the center of a hypothetical map crop. For each position visible to that agent, we show what the value func- tion would be if there were a second agent at that position. We find that agents learn policies dependent on those of other agents, in both the foraging and combat environments. 6. Discussion 6.1. Multiagent competition is a curriculum magnifier Not all games are created equal. Some produce more com- plex and engaging play than others. It is unreasonable to expect pure multiagent competition to produce diverse and interesting behavior if the environment does not support it. This is because multiagent competition is a curriculum magnifier, not a curriculum in and of itself. The initial conditions for formation of intelligent life are of paramount importance. Jungle climates produce more biodiversity than deserts. Deserts produce more biodiversity than the tallest mountain peaks. To current knowledge, Earth is the only planet to produce life at all. The same holds true in sim- ulation: human MMOs mirror this phenomenon. Those most successful garner large and dedicated player bases and develop into complex ecosystems. The multiagent setting is interesting because learning is responsive to the competitive and collaborative pressures of other learning agents -- but the environment must support and facilitate such pressures in order for multiagent interaction to drive complexity. There is room for debate as to the theoretical simplest pos- sible seed environment required to produce complexity on par with that of the real world. However, this is not our objective. We have chosen to model our environment after MMOs, even though they may be more complicated than the minimum required environment class, because they are known to support the types of interactions we are interested in while maintaining engineering and implementation feasi- bility. This is not true of any other class environments we are aware of: exact physical simulations are computationally in- feasible, and previously studied genres of human games lack crucial elements of complexity (see Background). While some may see our efforts as cherrypicking environment design, we believe this is precisely the objective: the pri- mary goal of game development is to create complex and engaging play at the level of human intelligence. The player base then uses these design decisions to create strategies far beyond the imagination of the developers. Neural MMO Figure 6. Exploration maps in the environment randomized settings. From left to right: population size 8, 32, 128. All populations explore well, but larger populations with more species develop robust and efficient policies that do better in tournaments. Figure 7. Agents learn to depend on other agents. Each square map shows the response of an agent of a particular species, located at the square's center, to the presence of agents at any tile around it. Random: dependence map of random policies. Early: "bulls eye" avoidance maps learned after only a few minutes of training. Additional maps correspond to foraging and combat policies learned with automatic targeting (as in tournament results) and learned targeting (experimental, discussed in Additional Insights). In the learned targeting setting, agents begin to fixate on the presence of other agents within combat range, as denoted by the central square patterns. Figure 8. Attack maps and niche formation quirks. Left: combat maps from automatic and learned targeting. The left two columns in each figure are random. Agents with automatic targeting learn to make effective use of melee combat (denoted by higher red density). Right: noisy niche formation maps learned in different combat settings with mixed incentives to engage in combat. Neural MMO 6.2. Additional Insights We briefly detail several miscellaneous points of interest in Figure 8. First, we visualize learned attack patterns of agents. Each time an agent attacks, we splat the attack type to the screen. There are a few valid strategies as per the environment. Melee is intentionally overpowered, as a sanity check. This cautions agents to keep their distance, as the first to strike wins. We find that this behavior is learned from observation of the policies learned in Figure 8. Second, a note on tournaments. We equate number of tra- jectories trained upon as a fairest possible metric of training progress. We experimented with normalizing batch size but found that larger batch size always leads to more stable performance. Batch size is held constant, but experience is split among species. This means that experiments with more species have smaller effective batch size: larger populations outperform smaller populations even though the latter are easier to train. Finally, a quick note on niche formation. Obtaining clean visuals is dependent on having an environment where inter- action with other agents is unfavorable. While we ensure this is the case for our exploration metrics, niche formation may also occur elsewhere, such as in the space of effective combat policies. For this reason, we expect our environment to be well suited to methods that encourage sample diversity such as population-based training (Jaderberg et al., 2017). 7. Future Work Our final set of experiments prescribes targeting to the agent with lowest health. Learned targeting was not required to produce compelling policies: agents instead learn effective attack style selection, strafing and engaging opportunisti- cally at the edge of their attack radius. Another possible experiment is to jointly learn attack style selection and tar- geting. This would require an attentional mechanism to handle the variable number of visible targets. We performed only preliminary experiments with such an architecture, but we still mention them here because even noisy learned tar- geting policies significantly alter agent-agent dependence maps. As shown in Figure 7, the small square shaped re- gions of high value at the center of the dependency maps correspond to the ranges of different attack styles. These appear responsive to the current combat policies of other learning agents. We believe that the learned targeting setting is likely to useful for investigating the effects of concurrent learning in large populations. 8. Conclusion We have presented a neural MMO as a research platform for multiagent learning. Our environment supports a large number of concurrent agents, inbuilt map randomization, and detailed foraging and combat systems. The included baseline experiments demonstrate our platform's capacity for research purposes. We find that population size magni- fies exploration in our setting, and the number of distinct species magnifies niche formation. It is our hope that our environment provides an effective venue for multiagent ex- periments, including studies of niche formation, emergent cooperation, and coevolution. The entire platform will be open sourced, including a performant 3D client and research visualization toolbox. Full technical details of the platform are available in the Supplement. Acknowledgements This research was undertaken in fulfillment of an intern- ship at OpenAI. Thank you to Clare Zhu for substantial contributions to the 3D client code. References Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014. URL http://arxiv. org/abs/1409.0473. Bansal, T., Pachocki, J., Sidor, S., Sutskever, I., and Mor- datch, I. Emergent complexity via multi-agent competi- tion. arXiv preprint arXiv:1710.03748, 2017. Bellemare, M. G., Naddaf, Y., Veness, J., and Bowling, M. The arcade learning environment: An evaluation plat- form for general agents. Journal of Artificial Intelligence Research, 47:253 -- 279, 2013. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., and Zaremba, W. Openai gym. CoRR, abs/1606.01540, 2016. URL http://arxiv. org/abs/1606.01540. Cuccu, G., Togelius, J., and Cudre-Mauroux, P. Playing atari with six neurons. arXiv preprint arXiv:1806.01363, 2018. Ficici, S. G. and Pollack, J. B. Challenges in coevolution- ary learning: Arms-race dynamics, open-endedness, and mediocre stable states. In Proceedings of the sixth inter- national conference on Artificial life, pp. 238 -- 247. MIT Press, 1998. Hern´andez-Orallo, J., Dowe, D. L., Espana-Cubillo, S., Hern´andez-Lloreda, M. V., and Insa-Cabrera, J. On more realistic environment distributions for defining, evaluat- ing and developing intelligence. In International Con- ference on Artificial General Intelligence, pp. 82 -- 91. Springer, 2011. Neural MMO Williams, R. J. Simple statistical gradient-following algo- rithms for connectionist reinforcement learning. Machine learning, 8(3-4):229 -- 256, 1992. Yaeger, L. Computational genetics, physiology, metabolism, neural systems, learning, vision, and behavior or poly In SANTA FE INSTI- world: Life in a new context. TUTE STUDIES IN THE SCIENCES OF COMPLEXITY- PROCEEDINGS VOLUME-, volume 17, pp. 263 -- 263. ADDISON-WESLEY PUBLISHING CO, 1994. Yang, Y., Luo, R., Li, M., Zhou, M., Zhang, W., and Wang, J. Mean field multi-agent reinforcement learning. arXiv preprint arXiv:1802.05438, 2018a. Yang, Y., Yu, L., Bai, Y., Wen, Y., Zhang, W., and Wang, J. A study of ai population dynamics with million-agent reinforcement learning. In Proceedings of the 17th Inter- national Conference on Autonomous Agents and MultiA- gent Systems, pp. 2133 -- 2135. International Foundation for Autonomous Agents and Multiagent Systems, 2018b. Zheng, L., Yang, J., Cai, H., Zhang, W., Wang, J., and Yu, Y. Magent: A many-agent reinforcement learning plat- form for artificial collective intelligence. arXiv preprint arXiv:1712.00600, 2017. Jaderberg, M., Dalibard, V., Osindero, S., Czarnecki, W. M., Donahue, J., Razavi, A., Vinyals, O., Green, T., Dunning, I., Simonyan, K., et al. Population based training of neural networks. arXiv preprint arXiv:1711.09846, 2017. Jaderberg, M., Czarnecki, W. M., Dunning, I., Marris, L., Lever, G., Castaneda, A. G., Beattie, C., Rabinowitz, N. C., Morcos, A. S., Ruderman, A., et al. Human- level performance in first-person multiplayer games with population-based deep reinforcement learning. arXiv preprint arXiv:1807.01281, 2018. Lanctot, M., Zambaldi, V., Gruslys, A., Lazaridou, A., Per- olat, J., Silver, D., Graepel, T., et al. A unified game- theoretic approach to multiagent reinforcement learning. In Advances in Neural Information Processing Systems, pp. 4190 -- 4203, 2017. Langton, C. G. Artificial life: An overview. Mit Press, 1997. Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, P., and Mor- datch, I. Multi-agent actor-critic for mixed cooperative- competitive environments. Neural Information Process- ing Systems (NIPS), 2017. Mordatch, I. and Abbeel, P. Emergence of grounded com- positional language in multi-agent populations. arXiv preprint arXiv:1703.04908, 2017. Nichol, A., Pfau, V., Hesse, C., Klimov, O., and Schulman, J. Gotta learn fast: A new benchmark for generalization in rl. arXiv preprint arXiv:1804.03720, 2018. OpenAI. Openai five. https://blog.openai.com/ openai-five/, 2018. Perlin, K. An image synthesizer. SIGGRAPH Com- ISSN 0097- doi: 10.1145/325165.325247. URL http: put. Graph., 19(3):287 -- 296, July 1985. 8930. //doi.acm.org/10.1145/325165.325247. Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484, 2016. Sims, K. Evolving 3d morphology and behavior by compe- tition. Artificial life, 1(4):353 -- 372, 1994. Stanley, K. O., Lehman, J., and Soros, L. Open- endedness: The last grand challenge youve never heard of. https://www.oreilly.com/ideas/ open-endedness-the-last-grand-challenge-youve-never-heard-of, 2017. Accessed: 2017-09-26. Strannegrd, C., Svangrd, N., Lindstrm, D., Bach, J., and Steunebrink, B. Learning and decision-making in artifi- cial animals. Journal of Artificial General Intelligence, 9: 55 -- 82, 07 2018. doi: 10.2478/jagi-2018-0002. Neural MMO Supplement Figure 9. Procedural 80X80 game maps Figure 10. Example Agent Environment The environment state is represented by a grid of tiles. We generate the game map by thresholding a Perlin (Perlin, 1985) ridge fractal, as shown in Figure 9. Each tile has a particular assigned material with various properties, but it also maintains a set of references to all occupying enti- ties. When agents observe their local environment, they are handed a crop of all visible game tiles, including all visible properties of the tile material and all visible properties of oc- cupying agents. All parameters in the following subsystems are configurable; we provide only sane defaults obtained via multiple iterations of balancing. Tiles We adopt a tile based game state, which is common among MMOs. This design choice is computationally efficient for neural agents and can be made natural for human players via animation smoothing. When there is no need to render the game client, as in during training or test time statistical tests, the environment can be run with no limit on server tick rate. Game tiles are as follows: • Grass: Passable tile with no special properties • Forest: Passable tile containing food. Upon moving into a food tile, the agent gains 5 food and the tile decays into a scrub. • Scrub: Passable tile that has a 2.5 percent probability to regenerate into a forest tile on each subsequent tick • Stone: Impassible tile with no special properties • Water: Passable tile containing water. Upon moving adjacent to a water tile, the agent gains 5 water. • Lava: Passable tile that kills the agent upon contact Agents Input: On each game tick, agents (Figure 10) observe a 15x15 square crop of surrounding game tiles and all occupy- ing agents. We extract the following observable properties: Per-tile properties: • Material: an index corresponding to the tile type • nEnts: The number of occupying entities. This is technically learnable from the list of agents, but this may not be true for all architectures. We include it for convenience here, but may deprecate it in the future. Per-agent properties: • Lifetime: Number of game ticks alive thus far • Health: Agents die at 0 health (hp) • Food: Agents begin taking damage at 0 food or water • Water: Agents begin taking damage at 0 food or water • Position: Row and column of the agent • Position Deltas: Offsets from the agent to the observer • Damage: Most recent amount of damage taken • Same Color: Whether the agent is the same color (and thereby is in the same population) as the observer • Freeze: Whether the agent is frozen in place as a result of having been hit by a mage attack Output: Agents submit one movement and one attack ac- tion request per server tick. The server ignores any actions that are not possible or permissible to fulfil, such as attack- ing an agent that is already dead or attempting to move into stone. Pass corresponds to no movement. Movement: North Attack: Melee Range Mage East West Pass South Neural MMO Figure 11. Example foraging behavior Foraging Foraging (Figure 11) implements gathering based survival: • Food: Agents begin with 32 food, decremented by 1 per tick. Agents may regain food by occupying forest tiles or by making use of the combat system. • Water: Agents begin with 32 water, decremented by 1 per tick. Agents may regain water by occupying tiles adjacent to water or making use of the combat system. • Health: Agents begin with 10 health. If the agent hits 0 food, they lose 1 health per tick. If the agent hits 0 water, they lose 1 health per tick. These effects stack. Figure 12. Example combat behavior aggressive agents can use mage combat to immobilize their target before closing in for the kill. In all cases, the best strategy is not obvious, again imposing an arms race. Technical details: • Attack range is defined by l1 distance: "1 range" is a 3X3 grid centered on the attacker. • Spawn Killing Agents are immune during their first 15 game ticks alive. This prevents an exploit known as "spawn killing" whereby players are repeatedly at- tacked immediately upon entering the game. Human games often contain similar mechanism to prevent this strategy, as it results in uninteresting play. The limited availability of forest (food) tiles produces a carrying capacity. This incurs an arms race of exploration strategies: survival is trivial with a single agent, but it re- quires intelligent exploration in the presence of competing agents attempting to do the same. Combat Combat (Figure 12) enables direct agent-agent confrontation by implementing three different attack "styles": • Melee: Inflicts 10 damage at 1 range • Ranged: Inflicts 2 damage at 1-2 range • Mage: Inflicts 1 damage at 1-3 range and freezes the target in place, preventing movement for two ticks Each point of damage inflicted steals one point of food and water from the target and returns it to the attacker. This serves as an incentive to engage in combat. It is still fully possible for agents to develop primarily foraging based strategies, but they must at least be able to defend them- selves. The combat styles defined impose clear but difficult to optimize trade offs. Melee combat fells the target in one attack, but only if they are able to make their attack before the opponent retaliates in kind. Ranged combat produces less risky but more prolonged conflicts. Mage combat does little damage but immobilizes the target, which allows the attacker to retreat in favor of a foraging based strategy. More API The initial release is bundled with two APIs for running experiments on our platform. All of our experiments are RL based, but the API implementation is intentionally generic. Evolutionary methods and algorithmic baselines should work without modification. Gym Wrapper We provide a minimal extension of the Gym VecEnv API (Brockman et al., 2016) that adds support for variable numbers of agents per world and at any given time. This API distributes environment computation of ob- servations and centralizes training and inference. While this standardization is convenient, MMOs differ significantly from arcade games, which are easier to standardize under a single wrapper. The Neural MMO setting requires support for a large, variable number of agents that run concurrently, with aggregation across many randomly generated environ- ments. The Gym API incurs additional communications overhead that the native API bypasses. Native This is the simplest and most efficient interface. It pins the environment and agents on it to the same CPU core. Full trajectories run locally on the same core as the environment. Interprocess communication is only required infrequently to synchronize gradients across all environ- ments on a master core. We currently do the backwards pass on the same CPU cores because our networks are small, but GPU is fully supported. Neural MMO Client The environment visualizer comes bundled with research tools for analyzing agent policies. In the initial release, we provide both a 2D python client (Figure 13) and a 3D web client (Figure 14, 16). The 3D client has the best support for visualizing agent policies. The 2D client is already deprecated; we include it only because it will likely take a few weeks to fully finish porting all of the research tools. We include visualization tools for producing the following visualization maps; additional documentation is available on the project Github: • Value ghosting • Exploration • Interagent dependence • Combat Policy training and architecture Parameters relevant to policy training are listed in Table 1. The neural net architecture, shown in Figure 15, is a simplest possible fully connected network. It consists of a preproces- sor, main network, and output heads. The preprocessor is as follows: • Embed indicies corresponding to each tile into a 7D vector. Also concatenates with the number of occupy- ing entities. • Flatten the tile embeddings • Project visible attributes of nearby entities to 32D • Max pool over entity embeddings to handle variable number of observations • Concatenate the tile embeddings with the pooled entity embeddings • Return the resultant embedding Figure 13. Example map in the 2D client The main network is a single linear layer. The output heads are also each linear layers; they map the output hidden vector from the main network to the movement and combat action spaces, respectively. Separate softmaxes are used to sample movement and combat actions. Technical details • For foraging experiments, the attack network is still present for convenience, but the chosen actions are ignored. • Note that 1D max pooling is used to handle the variable number of visible entities. Attention (Bahdanau et al., 2014) may appear the more conventional approach, but recently (OpenAI, 2018) demonstrated that simpler and more efficient max pooling may suffice. We are unsure if this is true at our scale, but used max pooling nonetheless for simplicity. Figure 14. Example overhead map view in the 3D client Neural MMO Figure 15. Agents observe their local environment. The model embeds this observations and computes actions via corresponding value, movement, and attack heads. These are all small fully connected networks with 50-100k parameters. Training Algorithm Policy Gradients(Williams, 1992) Adam Parameters Parameter Weight Decay Entropy Bonus Discount Factor Table 1. Training details and parameters for all experiments Value lr=1e-3 1e-5 1e-2 0.99 Notes + Value function baseline Pytorch Defaults Training stability is sensitive to this To stabilize training; possibly redundant No additional trajectory postprocessing Figure 16. Perspective screenshot of 3D environment
1905.08041	1	1905	2019-04-10T09:33:03	Inventory Management - A Case Study with NetLogo	[ "cs.MA" ]	Multi-Agent Systems (MAS) have been applied to several areas or tasks ranging from energy networks controlling to robot soccer teams. MAS are the ideal solution when they provide decision support in situations where human decision and actions are not feasible to operate the system in control and in real-time. Thus, we present a case study that is related to dynamic simulation of an automatic inventory management system. We provide two types of agents, the clients, and the seller agents. Through a system of communication, the agents exchange messages to fulfill their inventory needs. The client agents trade products in quantities according to their needs and rely on seller agents if other clients in the retailer chain cannot provide the needed items. Additionally, it is expected that the trading between a client and the sellers is done through a reverted type of auction. This case study MAS uses BDI and FIPA-ACL in its implementation resulting in a clear simulation of the system. We expect to provide a comparison between two distinct situations. One with only external transactions with providers, and a situation where both internal and external transactions are allowed.	cs.MA	cs	INVENTORY MANAGEMENT - A CASE STUDY WITH NETLOGO Rui Portocarrero Sarmento PRODEI - Faculty of Engineering, University of Porto Rua Dr. Roberto Frias, s/n 4200-465 Porto, Portugal [email protected] May 21, 2019 ABSTRACT Multi-Agent Systems (MAS) have been applied to several areas or tasks ranging from energy net- works controlling to robot soccer teams. MAS are the ideal solution when they provide decision support in situations where human decision and actions are not feasible to operate the system in control and in real-time. Thus, we present a case study that is related to dynamic simulation of an automatic inventory management system. We provide two types of agents, the clients, and the seller agents. Through a system of communication, the agents exchange messages to fulfill their inventory needs. The client agents trade products in quantities according to their needs and rely on seller agents if other clients in the retailer chain cannot provide the needed items. Additionally, it is expected that the trading between a client and the sellers is done through a reverted type of auction. This case study MAS uses BDI and FIPA-ACL in its implementation resulting in a clear simulation of the system. We expect to provide a comparison between two distinct situations. One first situation where there are no transactions between retail stores and only external transactions with providers and other situation where both internal and external transactions are allowed. Keywords MAS · Inventory Management · BDI · FIPA-ACL 1 Introduction We are aware of small size retail chains where the store's managers execute stocks management of several stores but in a manual fashion, not automatically. The stores also have to communicate with the providers when in need of inventory. The manager also does this manually and doesn't use any auction to acquire products. One of the contributions of this work is to establish automatic inventory control for the manager. We wish to provide electronic communication of items needs between supermarket agents. Each supermarket agent communicates with other supermarkets to check if they have an excess of some particular item in stock. Thus, if another agent has excess stock, it can provide the item to the chain agent in need of stock, lowering the costs with external providers and also the costs of item storage. Moreover, we want the retail chain stores to establish communication with external providers that might automatically provide the best price from the provider. For this, we wish to create a reverse auction with providers. Some ad hoc multi-agent systems are often created by researchers and developers without using frameworks. Nonethe- less, some frameworks have arisen recently, and these frameworks implement common standards. Some of these stan- dards like, for example, the FIPA agent system platforms and communication languages, have an important role for the framework user. These frameworks save developers time and also aid in the standardization of MAS development. Relevant surveys were previously written and provide insight over several frameworks available [13, 1]. Another contribution of this work is to provide a contextualization of a hypothetical NetLogo user. BDI and FIPA-ACL are standards that provide a more structured way to program MAS, and this is now possible with NetLogo. Thus, by providing additional knowledge of the NetLogo features the user is empowered to do a better and more professional MAY 21, 2019 work with NetLogo. Additionally, the provided inventory management case study is done by applying introduced concepts in NetLogo. Therefore, there is needed contextualization. Introducing this work structure, we start by providing contextualization in 2. Thus, we provide significant related work in this area. Then, we describe our case study in 3. In 5, we exhibit results obtained from the simulation of the MAS system built and described in 3. In 6 we provide a discussion of the results we gathered from the simulations. Finally, we conclude our work, and we also provide some insight into future developments of this work in 7. 2 Related Work In this section, we provide several milestones and previous work on the subject of MAS for inventory management and also the NetLogo framework. Thus, we introduce the earlier work on this concept and also the relevant features of NetLogo that will provide a solution for the development of a simulation scenario. 2.1 Inventory Management In [21], Signorile emphasizes that "the dynamics of the enterprise and the market make the inventory management difficult: materials do not arrive on time, production facilities fail, and employees are ill, causing deviations from the plans. In some cases, these events may be dealt with locally". In particular, Signorile demonstrated that agents "can be used to control an important part of the supply chain, the inventory of a retail store. In fact, the agent model outperforms the current EOQ (Economic order quantity) model" that the retail institution where the case study took place, is using (2002). In [22], Signorile expands their previous work and their decision support system to the inventory system (excluding supply chain activities) for a medium-sized department store in the United States (2005). In [10], Jiang and Sheng approach this problem in a supply chain with non-stationary customer demand. They use MAS with reinforcement learning proved experimentally to be effective (2009). In [5], Boudouda and Boufaida mention that "the agent-based model proposed consists of a set of subsystems agent, each subsystem is an actor in the SC (Supply Chain); these subsystems are grouped into several levels. They seek to coordinate and synchronize different activities. Each subsystem is structured by three cognitive agents (Purchaser Agent, manager stock agent and delivery agent), these agents interact with each other to accomplish their tasks and can communicate, negotiate and collaborate with other subsystems through a communication interface. Two types of interaction between subsystems are defined: internal interaction for the actors belonging to the same level, the second type external can occur between subsystems of different levels..". Additionally, the authors stress that their model "provides a collaboration and negotiation between different actors in the supply chain. It does not replace existing tools and strategies for the SC, but it can be used as a supplement which improves it" (2012). 2.2 NetLogo NetLogo is a platform independent environment used for modelization of social and natural phenomena as multi-agent systems. It provides means to simulate with a great number of agents. The NetLogo platform was developed by Uri Wilensky in 1999 and is under continuous development at the Northwestern's Center for Connected Learning and Computer-Based Modeling (CCL). The CCL is a research group headed by Prof. Uri Wilensky. The center includes staff and students at Northwest- ern working together and in association with researchers from around the world. The group includes educational researchers, curriculum developers, software engineers, and model builders. The center is funded by Northwestern, the USA National Science Foundation, and a few commercial sponsors. Using the NetLogo language, students and researchers have constructed large numbers of models of complex phe- nomena in the natural and social worlds. The Models Library that comes with NetLogo covers phenomena in biology [9, 23], chemistry [8, 11], informatics [4], economics [3], sports [12], sociology [24, 14, 2, 6], medicine [20, 7] and a variety of other domains. Although the NetLogo platform offers support for reactive agent systems, modeling BDI agents and explicit symbolic message exchange is not supported. Sakellariou et al. [17, 18, 19] have extended NetLogo with libraries to support both and used it for setting coursework in an Intelligent Agents unit. According to Sakellariou et al. [18], in the context of an Agent and Multi-Agent Systems course, the student's demands poses an interesting problem regarding practice procedures. Educators have reported a variety of environments and 2 MAY 21, 2019 techniques they use to increase active learning. The authors report their experience using NetLogo as part of the prac- tical coursework that students need to carry out within an Intelligent Agents course. Also, Sakellariou et al. describe two extra NetLogo libraries provided to students, one for BDI-like agents and one for ACL-like communication. Sakellariou et al., describe the two additional NetLogo libraries with more detail than previously in [17]. One for BDI-like agents and one for ACL-like communication that, according to the authors, provide effortless development of goal-oriented agents, that communicate using FIPA-ACL messages. The authors include one simulation scenario that employs these libraries to provide an implementation in which agents cooperate under a Contract Net protocol. 2.2.1 BDI The BDI library for NetLogo provides several NetLogo reports and procedures to implement belief -- desire -- intention software model. The library is built using the standard programming language provided by NetLogo. The only requirements by the authors of the library are that [15]: 1. ALL complex agents MUST have two declared -own variables:"beliefs" and "intentions". These are the variables to which beliefs and intentions are recorded. So, in the users model if there is a breed of agents which you wish to model as "BDI" agents, then the user should have a BREED-own [beliefs intentions] declaration (along with any other variables the user wishes to include in his model). The authors also stress that when the user creates the variables, they should initially set their values to empty a list ([ ]). 2. The users also must have ticks in his model or timeout facilities of the library will not function. 3. The user's model and interface should include a "switch" in the interface, named "show-intentions". This is necessary since the code of the receive message procedure checks in each cycle whether to output the messages or not. In this section, we introduce and explain the most prevalent and essential code to understand other tasks described in this paper. A belief in the library is a list of two elements: the type and the content [15]. could include any string, e.g. "position", "agent", etc. Types facilitate belief management. • The belief type declares the type of the belief, i.e., indicates a "class" that the belief belongs to. Examples • The belief content is the content of a certain type of belief. It can be any NetLogo structure (integer, string, list, etc.). The authors emphasize that there might be many beliefs of the same type with different content. However, two beliefs of the same type and content cannot be added. The following code creates a new belief. (does not stores it in beliefs memory): 1 to−r e p o r t 2 r e p o r t 3 end c r e a t e −b e l i e f ( l i s t b−t y p e [ b−t y p e c o n t e n t ) c o n t e n t ] To enable adding information to the beliefs structure, the following code is provided by the library: 1 t o add−b e l i e f 2 3 4 end [ b e l ] i f member ? b e l b e l i e f s s e t b e l i e f s f p u t b e l b e l i e f s [ s t o p ] To be possible to remove a belief from the list of beliefs the authors provided the following code: 1 t o remove−b e l i e f 2 3 end s e t b e l i e f s [ b e l ] remove b e l b e l i e f s To update the information of some particular belief, the following code was created by the authors: 1 t o update−b e l i e f remove−b e l i e f 2 add−b e l i e f b e l 3 4 end [ b e l ] read−f i r s t −b e l i e f −of−t y p e b e l i e f −t y p e b e l 3 Regarding intentions, several pieces of code were developed by the authors to provide an abstraction for a BDI-like implementation of MAS in NetLogo. The following code provides the execution of an agent intention: MAY 21, 2019 [ myInt ] 1 t o e x e c u t e−i n t e n t i o n s 2 ; ; ; l o c a l s ; ; 3 i f empty ? i n t e n t i o n s l e t myInt get−i n t e n t i o n 4 run i n t e n t i o n −name myInt 5 6 i f 7 i f 8 end r u n r e s u l t show−i n t e n t i o n s [ s e t f i r s t [ s t o p ] i n t e n t i o n s i n t e n t i o n −done myInt [ remove−i n t e n t i o n myInt ] l a b e l i n t e n t i o n s ] ; ; J u s t f o r debugging . As implemented by the authors, the intentions are stored in a STACK of intentions. The first argument is the intention name that should be some executable procedure the library user encodes in NetLogo. The second argument should be a NetLogo REPORTER that when evaluates to true, the intention is removed (either accomplished or dropped). As the authors designed the library, both values have to be STRINGS. Thus, the following code adds an intention in the intentions list: 1 t o add−i n t e n t i o n [ name done ] 2 3 end i n t e n t i o n s f p u t s e t ( l i s t name done ) i n t e n t i o n s Finally, for the removal of a specific intention from the intention stack, we have the following code: 1 t o remove−i n t e n t i o n [ bdi−l i b ## i n t e n t i o n ] 2 3 end i n t e n t i o n s s e t remove−i t e m ( p o s i t i o n bdi−l i b ## i n t e n t i o n i n t e n t i o n s ) i n t e n t i o n s 2.2.2 FIPA-ACL This library for NetLogo implements FIPA-ACL like messages in NetLogo [16]. In this section, we introduce and explain the most common and important code to understand other tasks described in this paper. The library authors have some requirements for the user of the library in NetLogo: 1. All agents that can communicate MUST have a declared -own variable incoming-queue. This is the variable to which all messages are recorded. So, in the user's model, if there is a breed of agents which the user desires to communicate, then the user should have a BREED-own [incoming-queue] declaration (along with any other variables that the user wishes to include in his model. The authors stress that when the user creates the variables, he should set its values to an empty list ([ ]). 2. The user's model should include a "switch" named "show-messages" in the interface. The authors explain this is necessary since the code of the received message checks in each cycle whether to output the messages or not. The code to enable sending of messages provides the definition of the sender and a receiver(s). In agents simulation, some issues might arise from communications like, for example, if the agent that should receive the message is "killed". One solution to this problem is nothing happens. Other solution could yield an error message. An alternative would be to create a safe send. The following code implements the author solution [16]: r e c i p i e n t s get−r e c e i v e r s msg r e c v 0 l e t l e t f o r e a c h r e c i p i e n t s [ 1 t o send [ msg ] 2 3 4 5 6 s e t i f s t r i n g i s 7 ] 8 end r e c v t u r t l e r e c v != nobody [ w i t h o u t−i n t e r r u p t i o n [ ask r e c v [ r e c e i v e msg ] ] ] ( read−from−s t r i n g ? ) ; ; read−from− r e q u i r e d t o c o n v e r t t h e s t r i n g t o number Message reception deals with updating incoming-queue: 4 MAY 21, 2019 [ msg ] 1 t o r e c e i v e 2 3 4 end i f s e t show messages incoming−queue l p u t msg incoming−queue [ show msg ] The following NetLogo reporter returns the next message in the list and removes it from the queue: 1 to−r e p o r t get−message 2 3 4 5 6 end i f empty ? incoming−queue [ r e p o r t " no message " ] l e t nextmsg f i r s t remove−msg r e p o r t nextmsg incoming−queue The following code represents an explicit remove-msg procedure. This is needed since reporters in NetLogo cannot change a variable's value: 1 t o remove−msg 2 3 end s e t incoming−queue but− f i r s t incoming−queue To enable broadcasting to all agents of breed of type t-breed, the following code was provided by the library authors: a g e n t s o f b r e e d t−b r e e d t−b r e e d [ i f e l s e ? = a g e n t f o r e a c h [ who ] o f 1 ; ; b r o a d c a s t i n g t o a l l 2 t o b r o a d c a s t −t o [ t−b r e e d msg a g e n t ] 3 4 5 6 7 8 9 ] 10 end ; ; ; do n o t h i n g ] [ send add−r e c e i v e r ? msg [ ] NOTE: We introduced the exclusion of the agent that broadcasts the message since, as we will see later in this docu- ment, in the case of a client broadcasting to other clients we were not interested that the client agent emitted messages to himself. To create Messages and add the sender, the following reporter code is essential: 1 to−r e p o r t 2 3 end r e p o r t c r e a t e −message [ p e r f o r m a t i v e ] ( l i s t p e r f o r m a t i v e ( word " s e n d e r : " who ) ) Finally, to provide means to add fields to a message the following pieces of code were created by the authors: [ s e n d e r msg ] s e n d e r t o a message . add a r e c e i v e r r e p o r t add msg " s e n d e r : " r e p o r t add msg " r e c e i v e r : " 1 ; ; ; ADDING FIELDS TO A MESSAGE 2 ; ; Adding a s e n d e r 3 to−r e p o r t add−s e n d e r 4 5 end 6 7 ; ; 8 to−r e p o r t add−r e c e i v e r 9 10 end 11 12 ; ; 13 to−r e p o r t add−m u l t i p l e −r e c e i v e r s 14 15 16 17 18 19 end 20 s e t msg add−r e c e i v e r ? msg f o r e a c h r e c e i v e r s [ ] r e p o r t msg a d d i n g m u l t i p l e r e c i p i e n t s [ r e c e i v e r msg ] r e c e i v e r [ r e c e i v e r s msg ] 5 MAY 21, 2019 [ c o n t e n t msg ] c o n t e n t t o a message r e p o r t add msg " c o n t e n t : " 21 ; ; Adding c o n t e n t 22 to−r e p o r t add−c o n t e n t 23 24 end 25 26 ; ; P r i m i t i v e Add command 27 to−r e p o r t add [ msg f i e l d v a l u e ] 28 29 30 31 end f i e l d = " c o n t e n t : " l p u t v a l u e l p u t i f e l s e [ r e p o r t [ r e p o r t l p u t f i e l d msg ] ( word f i e l d v a l u e ) msg ] We should stress that the previous code also provides means to add a primitive associated with the message. This, simultaneously with the other pieces of code provided by the authors, enables a FIPA-ACL like structure for our messages between agents. 3 Case Study We carried several experiments with the work of Sakellariou et al. [17, 18, 19]. We established a simulation environ- ment with the following characteristics 1: 1. Client Agents (retailer group agents) in their fixed locations and with an item database with the following information for each item (a) current item quantity (b) minimum item quantity (c) maximum item quantity (d) price paid for item in last transaction 2. Seller Agents (providers) with an items database with the following information for each item (a) current item price (b) minimum item price (c) maximum item price With these characteristics, we had an idea of simulating several client and seller agents. Thus, we needed to start the simulation with both type of agents database previously configured. Therefore, we had to state the initial state for each agent. 3.1 Initial State The initial state might be different for each client agent since we randomly select from 3 different item database. Thus, we vary the initial state of the client agents regarding item quantity in stock and price paid for each item in the last transaction. One example of the initial conditions of a client agent database is available in table 1. item id pao leite 1lt bolachas cerveja fraldas peixe carne Table 1. Client Initial State Example stock min stock max stock 120 100 50 250 15 20 30 120 100 50 250 15 20 30 25 10 15 12 5 2 5 buy price 0.12 0.54 0.8 0.35 1.7 2.75 2.1 Additionally, the seller agents might also vary with five different initial conditions. Therefore, we change values from seller to seller, for example, regarding item price and also minimum item price. 1Available Code at https://github.com/Sarmentor/InventoryManagement. 6 Another example of initial conditions, this time for seller agent database is available in the following table: MAY 21, 2019 Table 2. Seller Initial State Example item id pao leite 1lt bolachas cerveja fraldas peixe carne price min price unit max price unit 0.12 0.54 0.7 0.35 1.5 2.5 2.25 0.1 0.45 0.5 0.27 1.3 2.2 1.95 0.15 0.65 0.8 0.45 1.9 3.2 2.73 After selecting initial conditions settings for each agent, we can start the simulation. The agents follow predetermined behaviors we describe in the next subsections. 3.2 Client Agents Succinctly, the Client Agents have the following behaviors: 1. Appear randomly at a location in the environment 2. listen to messages (a) if an "cfp" was received from another client i. checks if there are enough items in stock to provide the other client. If the answer is yes sends a message "propose" to the other client with the item quantity and price. ii. if there are not enough items available then it sends a "refuse" message to the other client. (b) if an "accept-proposal" message was received: i. update stocks and reduce the quantity of the available item in the database (c) if no message was received continues to 3. 3. simulate selling items to virtual buyers and in random quantities 4. check inventory (a) if an item stock reached its possible minimum and internal trading is not allowed, then jumps to 4(a)v. If it is allowed then: i. broadcasts cfps message to other retail clients in the group ii. collects proposals iii. adds beliefs "proposal-from-clients" iv. evaluates beliefs ("proposal-from-clients") A. if the proposal was accepted then: • sends "accept-proposal" message to the client with best proposal. • updates stock by increasing the quantity of the item provided by the other client agent and returns back to 2. B. if all clients sent "refuse" message then jumps to 4(a)v. v. broadcasts cfps message to sellers vi. collects proposals from sellers vii. adds beliefs "proposal" viii. evaluates beliefs ("proposal") A. if the proposal has the best price and was accepted after some specified quantity of auction rounds then: • sends "accept-proposal" message to the winning seller. • updates item stock in database after receiving "success" message from wining seller. B. if the proposal was not accepted then back to 2. (b) if no item stock reached its admissible minimum returns to 2. 7 MAY 21, 2019 3.3 Seller Agents Succinctly, the Seller Agents have the following behaviors: 1. Listen to cfps from clients 2. Evaluate and reply proposal with the price for the item. This price is a random number between the asked price and the seller's minimum admissible price for the item 3. Listen to replies from clients, if its a new cfp then 2 or else: (a) if the proposal was accepted (primitive "accept-proposal") then: i. Sends items to client ii. Updates current price for item (b) if the proposal was not accepted (primitive "reject-proposal") then back to 1. 3.4 Auction System When the client agents need to be provided by external agents, we will need to implement an auction system so the offered price for the items would be the best possible. As the external providers have different rates, it is mandatory to establish an auction when the "cfp" messages are broadcasted. Thus, we choose to create a reverse auction, and the following workflow is executed: 1. Client A needs an item and broadcasts "cfp" messages to all seller agents with the current needed item and its price. Two different situations can happen: (a) If all sellers are not able to provide the item with the requested price they refuse to propose and the auction is interrupted (b) If sellers can provide the item they will propose their best price. This best price is a random value between the asked price in the current auction round and the minimum acceptable price for the provider/seller. 2. Client A then broadcasts yet another set of "cfp" to the sellers with the best price in 1(b). 3. Client A then selects the best offer after a predefined amount of auction rounds or if there is only one seller biding. If there is a tie of price between sellers, Client A selects the first seller to have provided the best price. 4 Metrics for Evaluation There are two possible situations in our MAS system: 1. Only external transactions between clients and sellers 2. Both internal (between clients, i.e., the agents belonging to the retail chain) and external transactions (between retail chain agents, i.e., the clients, and the seller agents) For the evaluation of our inventory management MAS we decided that for the comparison of both situations we should measure several metrics: • Average item price, for the clients, given by the formula: • Average time a transaction takes, from consulting the internal/external agents until stock update if transaction is successful, given by the formula: • Internal/external ratio (Only in case of internal and external buying), given by the formula: IT R = T otalInternalT rades T otalExternalT rades With these measures, we expect to have an idea of how these metrics evolve over the simulation period. These are generic measures that might provide some insight into the system behavior, nonetheless, as we will see, the initial conditions of the client agents influence the evolution of the simulations. 8 (cid:80)k (cid:80)k k k AIP = i=1 ItemT radeP ricei AIT T = i=1 ItemT radeT icksi MAY 21, 2019 5 Results In this section, we will provide the results of our simulations. We choose to simulate with five external providers (sellers) and a retail chain of two agents (the clients). 5.1 External Trading Only The external transactions only mode, for the clients, imply we have no internal transactions. Figure 1 captures the NetLogo interface of such simulation. Figure 1. Interface with external trading mode simulation In this mode, the results we obtained were stable after around 25000 ticks of simulation. This translated into 62 external buying transactions for the client agents. Figure 2 provides evidence of the decreasing of ticks it takes for a client to be provided in this mode. It is clear that the implementation of the reverse auction gets faster as the simulation evolves. Figure 2. Interface with external trading mode simulation This is explained by the fact that the clients register the best prices after the first auction is finished. Additionally, when they need to buy the same product later, they request a lower price to the seller agents which makes further auctions successful in fewer auction rounds, thus, the value of 2.242 ticks when the simulation stabilizes. Regarding the average buying price, the value stabilizes at around 1.284. This value starts higher at the beginning of the simulation but, similarly to the previous results of this mode and for the same reasons, it lowers as the simulation evolves. 5.2 Internal and External Trading Depending on the initial conditions, the simulation has different results and takes a different number of ticks to get to a stable state. For example, with an initial state with one client agent with the maximum number of items available for all the products, and the other agent with low amounts of items, we can expect to have more internal trades than external trades for a long period or many ticks. 9 MAY 21, 2019 Figure 3. Time to Provide and Internal/External Ratio with internal and external trading mode simulation Thus, in these conditions and for approximately 3000 ticks, the client agents did 1238 internal buys and 0 external buys. Additionally, the average number of ticks it takes for a client to be provided is 0 which is the main advantage of having only internal trades. This is explained by our number of client agents which is low and the fact that we do not execute any auction in this situation of only having internal trading. Regarding the average buying price, the value is 1.373. Compared with another mode where we can only have external trades, this value is higher which is expected as we do not have external transactions that rely on the reverse auction. This leads clients to be provided with the lowest price possible, thus lowering the average buying price with external trading. We did additional simulations, this time with other initial conditions. With unbalanced client's stocks, for example, with maximum stock for some items and minimum for others, there is more chance for a balance between internal and external trading. Figure 3 shows that, with 10000 ticks of simulation, ten internal buys, and four external solicitations, an interesting case is simulated. It is visible that as the internal/external ratio decreases the time it takes until the client providing finishes are higher. This means that initially in the simulation, more internal transactions were done than external transactions. Then, further in the simulation, some external transactions occur which increases the time needed to provide the clients with the items they need. At this point, the average number of ticks it takes for a client to be provided is 5.286. Additionally, regarding the average buying price, the value is 0.7. This value is lower than previous simulations; this might be due to the items that were sparse in the initial conditions and their lower price. Thus, with more simulation ticks, this value is expected to increase as more expensive items are bought. Additionally, as the simulation progresses, there are more auctions, the price of the items naturally lowers. Figure 4 and Figure 5 shows this situation. 10 MAY 21, 2019 Figure 5. Fish item price evolution Figure 4. Meat item price evolution Both meat and fish price in this simulation starts higher. This is due to the internal transactions at higher initial prices for both items. As the simulation evolves more and more external transactions occur. Thus, with more auctions happening, there is a natural progression of prices towards the minimum possible price regarding the providers in the simulation. An example of a provider context after several thousands of simulation ticks is provided in the following table 3. Table 3. Seller Final State Example item id pao leite 1lt bolachas cerveja fraldas peixe carne price min price unit max price unit 0.15 0.5 0.6 0.3 1.6 3.1 1.7 0.13 0.5 0.6 0.29 1.45 2.9 1.7 0.14 0.54 0.8 0.37 1.75 3.3 2.8 In this table, taken from a simulation after several external transactions, it is clear that for example the item "leite 1lt" as well as "bolachas" and "carne" reached the lowest price possible in the simulation. Recording the initial prices for this provider, we had, at the beginning of the simulation, the following table 4. Table 4. Seller Initial State Example item id pao leite 1lt bolachas cerveja fraldas peixe carne price min price unit max price unit 0.15 0.5 0.6 0.3 1.6 3.1 1.8 0.13 0.5 0.6 0.29 1.45 2.9 1.7 0.14 0.54 0.8 0.37 1.75 3.3 2.8 11 MAY 21, 2019 Both these tables make us conclude that this provider had already the lowest prices possible for "leite 1lt" as well as "bolachas". Nonetheless, the item meat, designated by "carne" reached its lowest price of 1.7 after the simulation. 6 Discussion With these results, it is clear that initial client states regarding item stocks influence, on a large scale, the evolution of the simulation. Therefore, it is expected that due to the relative randomness of initial states the simulation provides less similar results. Nevertheless, the implemented system provides a testbed for further improvements, and the preliminary results are logical and seem intuitive. In a real situation, other factors should be regarded, like for example, the costs with item storage, the costs of trans- portation of items between clients and also the validity of products. Additionally, the providers (seller agents) should also have stock limitations as in real situations. Thus, the reduction of stock and, therefore, the capacity to provide or not a client agent should be studied. These factors would improve further the realism of the simulations and could provide better decision support to the manager of the retail chain group. 7 Conclusions and Future Work With this work, our goal was to provide an introduction to the inventory management problem and how to approach it with MAS frameworks, more specifically, with the NetLogo and its expanded libraries. These libraries help to implement BDI and FIPA-ACL based MAS. We provided the most important concepts of these libraries. Thus, by using these tools, we developed a trading system where we could focus on the inventory management per- formed by agents that could communicate with each other. We provided results by exposing the results of this communication. We used several metrics to compare two situations of trading, with and without internal trading between agents belonging to the retail chain. Furthermore, we imple- mented a reverse auction when the agents belonging to the retail chain would need to buy from external providers. We conclude that, by using these libraries, we increase comprehensibility and efficiency in the development and implementation of this MAS system. Additionally, by exploring the MAS system, we can easily simulate a complex system. The expected complexity of studying a real and similar system makes it imperative to use models that provide a test bench for different situations. These conditions would be very hard to test in real systems, and this is one of the main advantages of MAS. Acknowledgments This work was fully financed by the Faculty of Engineering of the Porto University. Rui Portocarrero Sarmento also gratefully acknowledges funding from FCT (Portuguese Foundation for Science and Technology) through a Ph.D. grant (SFRH/BD/119108/2016). References [1] Allan, R. (2010). Survey of agent based modelling and simulation tools. http://www.grids.ac.uk/NWGrid/ABMS/. [2] Balaraman, V. and Singh, M. (2014). Exploring norm establishment in organizations using an extended axelrod model and with two new metanorms. In Proceedings of the 2014 Summer Simulation Multiconference, SummerSim '14, pages 39:1 -- 39:9, San Diego, CA, USA. Society for Computer Simulation International. [3] Biondo, A. E., Pluchino, A., and Rapisarda, A. (2013). Return migration after brain drain: A simulation approach. Journal of Artificial Societies and Social Simulation, 16(2):11. [4] Blum, C., Lozano, J. A., and Davidson, P. P. (2015). An artificial bioindicator system for network intrusion detection. Artif. Life, 21(2):93 -- 118. [5] Boudouda, S. and Boufaida, M. (2012). A methodological approach for modeling supply chain management. In Communications and Information Technology (ICCIT), 2012 International Conference on, pages 38 -- 43. 12 MAY 21, 2019 [6] Dickerson, M. (2014). Multi-agent simulation, netlogo, and the recruitment of computer science majors. J. Comput. Sci. Coll., 30(1):131 -- 139. [7] Fekir, A. and Benamrane, N. (2011). Segmentation of medical image sequence by parallel active contour. In Software Tools and Algorithms for Biological Systems, pages 515 -- 522. Springer. [8] Galic, N., Ashauer, R., Baveco, H., Nyman, A.-M., Barsi, A., Thorbek, P., Bruns, E., and Van den Brink, P. J. (2014). Modeling the contribution of toxicokinetic and toxicodynamic processes to the recovery of gammarus pulex populations after exposure to pesticides. Environmental toxicology and chemistry, 33(7):1476 -- 1488. [9] Gheorghe, M., Stamatopoulou, I., Holcombe, M., and Kefalas, P. (2005). Modelling dynamically organised In Banatre, J.-P., Fradet, P., Giavitto, J.-L., and Michel, O., editors, Unconventional colonies of bio-entities. Programming Paradigms, volume 3566 of Lecture Notes in Computer Science, pages 207 -- 224. Springer Berlin Heidelberg. [10] Jiang, C. and Sheng, Z. (2009). Case-based reinforcement learning for dynamic inventory control in a multi-agent supply-chain system. Expert Systems with Applications, 36(3, Part 2):6520 -- 6526. [11] Liu, C., Sibly, R. M., Grimm, V., and Thorbek, P. (2013). Linking pesticide exposure and spatial dynamics: An individual-based model of wood mouse (apodemus sylvaticus) populations in agricultural landscapes. Ecological Modelling, 248:92 -- 102. [12] Martins Ratamero, E. (2015). Modelling peloton dynamics in competitive cycling: A quantitative approach. In Cabri, J., Pezarat Correia, P., and Barreiros, J., editors, Sports Science Research and Technology Support, volume 464 of Communications in Computer and Information Science, pages 42 -- 56. Springer International Publishing. [13] Nikolai, C. and Madey, G. (2009). Tools of the trade: A survey of various agent based modeling platforms. Journal of Artificial Societies and Social Simulation, 12(2):2. [14] Pluchino, A., Garofalo, C., Inturri, G., Rapisarda, A., and Ignaccolo, M. (2014). Agent-based simulation of pedestrian behaviour in closed spaces: A museum case study. Journal of Artificial Societies and Social Simulation, 17(1):16. [15] Sakellariou, I. (2010a). Agents with beliefs and intentions in netlogo. http://users.uom.gr/iliass/projects/NetLogo/AgentsWithBeliefsAndIntentionsInNetLogo.pdf. [16] Sakellariou, I. (2010b). An attempt to simulate fipa acl message passing in netlogo. http://users.uom.gr/iliass/projects/NetLogo/FIPA ACL MessagesInNetLogo.pdf. [17] Sakellariou, I., Kefalas, P., and Stamatopoulou, I. (2008a). Enhancing netlogo to simulate bdi communicating agents. In Darzentas, J., Vouros, G. A., Vosinakis, S., and Arnellos, A., editors, SETN, volume 5138 of Lecture Notes in Computer Science, pages 263 -- 275. Springer. [18] Sakellariou, I., Kefalas, P., and Stamatopoulou, I. (2008b). Teaching intelligent agents using netlogo. Pro. of the ACM-IFIP IEEIII Informatics Education Europe III Conference, Venice, Italy. [19] Sakellariou, I., Kefalas, P., and Stamatopoulou, I. (2009). Mas coursework design in netlogo. [20] Seal, J. B., Alverdy, J. C., Zaborina, O., An, G., et al. (2011). Agent-based dynamic knowledge representation of pseudomonas aeruginosa virulence activation in the stressed gut: Towards characterizing host-pathogen interactions in gut-derived sepsis. Theor Biol Med Model, 8(33):1742 -- 4682. [21] Signorile, R. (2002). Simulation of a multiagent system for retail inventory control: A case study. Simulation, 78(5):304 -- 311. [22] Signorile, R. (2005). Innovations in Applied Artificial Intelligence: 18th International Conference on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems, IEA/AIE 2005, Bari, Italy, June 22- 24, 2005. Proceedings, chapter A Decision Support System for Inventory Control Using Planning and Distributed Agents, pages 192 -- 196. Springer Berlin Heidelberg, Berlin, Heidelberg. [23] Stamatopoulou, I., Sakellariou, I., and Kefalas, P. (2007). Formal modelling for in-silico experiments with social insect colonies. In in Current Trends in Informatics, pages 79 -- 89. [24] Wallentin, G. and Loidl, M. (2015). Agent-based bicycle traffic model for salzburg city. GI Forum -- Journal for Geographic Information Science, 2015(1):558 -- 566. 13
1909.01885	1	1909	2019-09-04T15:33:09	Agent-based model for tumour-analysis using Python+Mesa	[ "cs.MA", "q-bio.TO" ]	The potential power provided and possibilities presented by computation graphs has steered most of the available modeling techniques to re-implementing, utilization and including the complex nature of System Biology (SB). To model the dynamics of cellular population, we need to study a plethora of scenarios ranging from cell differentiation to tumor growth and etcetera. Test and verification of a model in research means running the model multiple times with different or in some cases identical parameters, to see how the model interacts and if some of the outputs would change regarding different parameters. In this paper, we will describe the development and implementation of a new agent-based model using Python. The model can be executed using a development environment (based on Mesa, and extremely simplified for convenience) with different parameters. The result is collecting large sets of data, which will allow an in-depth analysis in the microenvironment of the tumor by the means of network analysis.	cs.MA	cs	AGENT-BASED MODEL FOR TUMOR-ANALYSIS USING PYTHON+MESA Ghazal Tashakor (a), Remo Suppi (a) (a) Universitat Autònoma de Barcelona, Department Computer Architecture & Operating Systems, School of Engineering, Campus Bellaterra, Cerdanyola del Vallès, Barcelona, 08193, Spain (a) [email protected] (a) [email protected] ABSTRACT The potential power provided and possibilities presented by computation graphs has steered most of the available modeling techniques to re-implementing, utilization and including the complex nature of System Biology (SB). To model the dynamics of cellular population, we need to study a plethora of scenarios ranging from cell differentiation to tumor growth and etcetera. Test and verification of a model in research means running the model multiple times with different or in some cases identical parameters, to see how the model interacts and if some of the outputs would change regarding different parameters. In this paper, we will describe the development and implementation of a new agent-based model using Python. The model can be executed using a development environment (based on Mesa, and extremely simplified for convenience) with different parameters. The result is collecting large sets of data, which will allow an in-depth analysis in the microenvironment of the tumor by the means of network analysis. Keywords: Agent-based models, Python, System Biology, Graph, Simulation modeling 1. INTRODUCTION At the cellular level, Systems Biology (SB) is mainly a complex natural phenomenon that today has become important for research in certain areas such as drug development and biotechnological productions and applications. Mathematical models of cellular level are built in bio repeated cycles at a portion of time to arrive at a decision. The primary idea behind the bio iterative cycles is to develop systems by allowing software developers to redesign and translate the cellular metabolism. The objective is bringing the desired decision or result closer to discovery in each iteration; these experiments on multi-scale iterative cycle of computational modeling - besides experimental validation and data analysis- would produce incremental samples for the high-throughput technologies. Also the behavior of the cell emerges at the network- level and requires much integrative analysis. Moreover, due to the size and complexity of intercellular biological Alongside networks, computational model should be a considerate part of the production or application. In the following whole simulation challenges, SB would not be needless of developing integrated frameworks for analysis and data management. the intercellular level, many researches in SB also address the issue of cellular population. In evolutionary versions of these scenarios, modeling the interactions of these type of models across large multi- scale would need agent-based modeling. Each agent has a Boolean network for its own expression, much like a gene. So a proper production or application for SB must support not only a compatible simulation method but also suitable methods for model parameter estimations which represent the experimental data (D.2011; An2008). Therefore, ABMs start with rules and mechanisms to reconstruct through the mathematical or computational form and observe the pattern of data. Processing the heterogeneous behavior of individual agent within a dynamic population of agents -which cannot be controlled by an overall controller- needs a higher-level system parallelism which ABMs supports. Biological systems include random behaviors and ABMs accommodate this via generation of population agents in the agent's rules. ABMs have a level of abstraction to create new cellular states or environmental variables without changing core aspects of the simulation. To aggregate the paradoxical nature of emergent behavior which could be observed from any agent in contrast to a conceptual rules of the model, ABMs reproduce emergent behavior. Emergent behavior has a range of stochasticity similar to real world Systems Biology (Bonabeau2002). Finding software platforms for scientific agent-based models require comparing certain software design parameters such as emulating parallelism, and developing schedulers for multiple iterations, which manage ABM run. Many references reviewed and compared different agent- based modeling toolkits. However, from the perspective of biotechnological application and biotechnologists, most of them share a key weakness. It is using complex languages which are not Python. Perhaps re- implementing ABMs in Python would be a wiser technical strategy since it is becoming the language of can induce important role and pathohistological angiogenesis scientific computing, facilitating the web servers for direct visualization of every model step, debugging and developing an intuition of the dynamics that emerge from the model, also allowing users to create agent-based models using built-in core components such as agent schedulers and spatial grids (Villadsen and Jensen 2013). 2. BACKGROUND Spontaneous tumor, which progresses from the initial lesion to highly metastatic forms are generally profiled by molecular parameters such as prognosis response, characteristics. morphology lymph- Tumors angiogenesis, which plays an in promoting cancer spread. Previous studies have shown that the cancer stem cell (CSC) theory could become a hypothesis for tumor development and progression. These CSCs have the capability of both self-renewal and differentiation into diverse cancer cells, so one small subset of cancer cells has characteristics of stem cells as their parents. Hereditary characteristics play a certain role in malignant proliferation, invasion, metastasis, and tumor recurrence. In recent researches the possible relationship between cancer stem cells, angiogenesis, lymph-angiogenesis, and tumor metastasis is becoming a challenge. Due to many evidences and reviews such as (Li 2014; Weis and Cheresh 2011; Carmeliet, P., & Jain, R. K. 2000), metastasis is defined as the spread of cancer cells from the site of an original malignant primary tumor to one or more other places in the body. More than 90% of cancer sufferings and death is associated with metastatic spread. In 1971, Folkman proposed that tumor growth and metastasis are angiogenesis-dependent, and hence, blocking angiogenesis could be a strategy to intercept tumor growth. His hypothesis later confirmed by genetic approaches. Angiogenesis occurs by migration and proliferation of endothelial cells from original blood vessels (Weis and Cheresh 2011). Accordingly, has contributed to the understanding of the molecular and cellular mechanisms occurring in the tumor and in its microenvironment which causes metastasis and this could present a model relatively similar to physiology of human or at the capability of going through genetic manipulations that bring them closer to humans. Hence high spatiotemporal resolution combined with parametric opportunities has been rapidly applied in technology (Granger and Senchenkova 2010). 2.1. Agent-based Models in Systems Biology A primary tumor model addressing the avascular growth state depends on differential equations. They are classified as "lumped models" to predict the temporal evolution of overall tumor size. Since lumped models just provide a quantitative prediction of tumor size over time with only a few parameters and very low computational results, they would not be enough for an tumor modeling with translational cancer research and least has a explicit investigation of many other events such as spatiotemporal dynamics of oxygen and nutrients or cell to cell interactions; also, stromal cells which play a major role in cancer growth and progression in the interaction with tumor cells. The result is disregarding the mutations in the tumor microenvironment and metastasis. These shortcomings lead us to In Silico models of tumor microenvironment. In Silico (Edelman, Eddy & Price 2010) refers to computational models of biology and it has many applications. It is an expression performed on a computer simulation. In Silico models are divided to three main categories (Thorne, Bailey, & Peirce 2007; Soleimani et al., 2018): 1. Continuum based models which solve the spatiotemporal evolution problems of density and concentration of cellular population in the tumor microenvironment. 2. Discrete or agent based on a set of rules which change the cells' states and manage the cells' interactions tumor microenvironment. within the 3. Adaptive Hybrid models which integrate the above solutions. research on Mathematical and computational models of the tumor should cover different aspects of tumor interactions in its microenvironment. Most of the avascular tumor models address the tumor growth dynamics not taking into account the angiogenesis setting; but computational models have started tumor-induced angiogenesis since the beginning of cancer research. Therefore, the angiogenesis models have a very strong background in experimental observations. The most established researches have explicitly focused on the stromal cells to achieve a close computational model of angiogenesis. For example, according to Monte Carlo simulations and energy minimization, cellular models are expanded and elaborated on cellular automata to allocate more than one lattice site to each cell and describe cell to cell and tumor stroma interactions. Nevertheless, building a bulky model over a range of matrix densities -which covers numerous factors for a large domain size and 3D simulations- is restricted by computational and application costs. Based on a paper written by Soleimani et al. (2018), a minimal coupling of a vascular tumor dynamics to tumor angiogenic factors through agent-based modeling has progressed the experimental studies in the recent years. They have mentioned in both reviews that it is challenging to mathematically simulate all the process of a complex system such as tumor growth, metastasis and tumor treatments, because mathematical modeling is still a simplification of the system biology and the results requires validation. Also, building a predictive experimental plus theoretical application without clinical data requires parametrizing and validating again. The best improvement which is expected would be predictive modeling in both response The main problem with this implementation was the limitations of the execution environment (Java memory limitations) and loss of performance with a high number of metastasis cells. In addition, this model did not allow capturing in detail the interactions between the different parameters at the microenvironment level. investigation interaction towards biology preclinical and clinical states (Thorne, Bailey, & Peirce 2007). 2.2. Complex Networks in Biological Models Another interesting approach is the network models. Networks follow patterns and rules and have a specific topology, which allows scientists to go through with a deeper information extraction. Within the fields of biology and medicine, Protein- protein (PPI) networks, biochemical networks, transcriptional regulation networks, signal transduction or metabolic networks are the highlighted network categories in systems biology which could detect early diagnosis. All these networks need compatible data to be produced experimentally or retrieved from various databases for each type of network; but besides analyzing data structures for computational analysis, several topological models have been built to describe the global structure of a network (Girvan, Newman 2002; 2004). 3. MODELING AND SIMULATION Our approach to the generation of an ABM model considered different temporal and spatial scales, focusing first on mitosis as the main axis of tumor growth. This model and its analysis developed in Netlogo (section 3.1). Then large number of environment limitations were detected with a high number of agents and difficulties to analyze the microenvironment of emergent growth. The second approximation is presented in this paper (section 3.2) used dynamic networks based on a large number of interactive agents. This model makes it possible to help researchers in carrying out more detailed research on intercellular network interactions and metastasis in a multiple scale model (Grimm2005). 3.1. Netlogo Model and Experimental Results Our ABM NetLogo model is designed as a self- organized model that illustrates the growth of a tumor and how it resists chemical treatment. This model in NetLogo (based on Wilensky's tumor model (Wilensky 1999)) permits us to change the parameters that affect tumor progression, immune system response, and vascularization. Outputs included the number of living tumor cells and the strength of the immune system which control cells. In this model the tumor has two kinds of cells: stem cells and transitory cells. In the model, tumor cells are allowed to breed, move, or die. The simulation presented cells' control with different and constant immune responses through killing transitory cells, moving stem cells and original cells. Figure 1 shows the steady state of a tumor metastasis visualization with 6 stem cells and the grow-factor =1.75, replication-factor=high, and apoptosis=low. As it could be seen, the growth of metastasis is more aggressive and through reducing apoptosis, there is a greater number of cells that do not die, amounting to near 200,000 cells (agents) (Tashakor, Luque & Suppi, 2017). Figure 1: stem cells evolution and metastasis visualization with grow-factor=1.75, apoptosis=low and replication-factor=high (near 200,000 cells in steady state). 3.2. Python Model and Experimental Results To solve the aforementioned problems, we started a new development using Python +Mesa. Mesa is the agent- based Python project, which has started recently and it has rapidly found its place among the researchers. Mesa is modular and its modeling, analysis and visualization components are integrated and simple to use. This feature convinced us to re-implement our tumor model with Mesa network structure. Furthermore, it supports multi-agent and multi-scale simulations, which is suitable for creating dynamic agent-based models. This paper documents our work with the Python +Mesa to design an agent model based on the tumor model scenario. At the beginning, Mesa was just a library for agent-based modeling in Python but now it is a new open-source, Apache 2.0 licensed package meant to fill the gap in the modeling complex of dynamical systems. In Mesa, the Model class to define the space where the agents evolve handles the environment of an agent-based model. The environment defines a scheduler which manages the agents at every step. For studying the behavior of model under different conditions, we needed to collect the relevant data of the model while it was running; Data Collector class was defined for this task. The adapting modules of Mesa allow us to make changes in the existing ABMs for the purpose of conforming with the requirements of our future framework. Also, monitoring the data management issues when processing actions happen in parallel seems facile in Mesa, since each module has a Python part which runs on the server and turns a model state into JSON data (Masad & Kazil, 2015). The advantage of Mesa is its browser for visualization which allows the user to see the model while running in the browser. Tumor progression is a complex multistage process and the tumor cells have to acquire several distinct properties either sequentially or in parallel. The first problem using NetLogo framework for simulation was using behavioral space which supports multithreading. In this way, performance is limited to the number of cores at the local infrastructure. To solve the local processing problem, we utilized the parametric simulations using a HPC cluster in order to reduce the necessary time to explore a determinate model data space. Since we needed to distribute our model through the distributed architecture to explore the model states in the microenvironment scale, it was revealed that NetLogo also has scalability issues in designing graph networks. Although there are available extensions for scaling up the model in the form of graph, it still depends on the size of the tumor and it could be a very slow process, taking days to finish experiment considering the checkpoints and performance bottlenecks while using a HPC cluster. We had to translate and move our model to a new framework such as the style and structure of Mesa to facilitate our distributed executions. The newest changes in the ground of the old scenario is increasing the population of agents, obtaining graph- based representation of biological network and consideration of multi-state and multi-scale components (An2008). The initial design aspect of the multi-scale architecture of out tumor growth model in Python comes from the acute inflammation based on the key factors such as angiogenesis. Tumor angiogenesis is critical for tumor growth and maintenance (Kaur et al., 2005). Our new model strategy begins with an initial identification of a minor population of cells with the characteristics of "tumor-initiating" cancer stem cells and these cancer stem cells in the assumed tumor reside in close proximity to the blood vessels. Therefore, we chose an angiogenetic switch for our model which have to balance this dynamic. We had angiogenetic and anti- angiogenetic factors. In the case of anti-angiogenetic, we will have the probability of quiescent tumor. In this agent-based model, tumor cells are affected, inflamed and turn quiescent. Table 1 shows the range of parameters we implemented to control the factors. Based on these factors, we can achieve a simulation evolution of metastasis and measure the tumor volume ratio along with the dynamics of the tumor under the influence of the factors. Table 1: different control factors of tumor growth Factors Low Medium 0.0 0.1 0.1 0.4 0.3 0.5 High 1 1 1 Angiogenesis Recovery Quiescent To select a network topology for tumor model, we selected a random graph. Barabasi-Albert model is a scale-free network and it is one of the most basic models since it describes most of the biological networks - especially in evolutionary models; but since there is a need to manage the cell interactions and stromal cells behavior within the tumor microenvironment and due to the time-dependency of the connections, we developed our Python agent-based model on Erdös-Rényi topology. The visualization of the model is a network of nodes (using the Mesa architecture) that shows the distribution of agents and their links. Considering the architecture of Mesa, the creation of agents occurs based on the assignment of nodes to a graph. A scheduler (time module) activates agents and stores their locations and updates the network. The total operation time is directly related to the number of steps necessary to deploy all the agents. Since each agent changes in three states, the process goes on until the tumor agent's volume appears as metastasis. Thus, for P as a Probability and for edge creation in n number of nodes, if (Ps > log (n)/n)) almost all vertices are connected and this function returns a directed graph as we described in Algorithm 1. Erdös-Rényi model takes a number of vertices N and connecting nodes by selecting edges from the (N (N-1)/2) possible edges randomly. The pseudo-code generating a random network is described in Algorithm 1: Algorithm 1 Input: (n,p, p* ) Output: (True, False) Begin Algorithm For n in enumerate(nodes): For p in enumerate(Ps): If p > p* then p_connected(n, p) = 1 Else return 0 Endif Endfor Endfor End Algorithm The interactive visualization in Mesa helps us to identify insights and generate value from connected data. Considering the data analysis for life sciences such as tumor which is almost about connections and dependencies, the large amount of data makes it difficult for researchers to identify insights. Graph visualization makes large amounts of data more accessible and easier to read, as it has been illustrated in Figure 2. Figure 2: Graph visualization for three states (normal, dead and metastatic) shown in three colors Figure 3: Tumor stabilization challenges for cancer stem cells (CSC) with control factors more than medium Figure 3 shows the reduction of cancer stem cells in 550 initial stem cells which are affected by control factors in the time it took, which was about 8 minutes for this simulation. Tumor growth curves are used for derivation of the Tumor Control Index (TCI), such as tumor progression, rejection and stabilization. We tried to combine both rejection and stabilization challenges to show the potential of Python agent-based modeling with different control factors. According to (Peskin 2009), we need a statistical method that leads to volume ratio measurements of tumor, as it could be seen in figure 3. To calculate tumor volume using (Monga.s 2000) we needed tumor width (W) and tumor length (L) which is presented in formula 1: (cid:1)(cid:2)(cid:3)(cid:4)(cid:5) (cid:7)(cid:4)(cid:8)(cid:2)(cid:3)(cid:9) = (cid:11)(cid:12) × (cid:14) 2 This calculation which works pretty well for clinical issues, was made based upon the assumption that solid tumors are more or less spherical like the version we had before in Netlogo Wilensky's tumor model, but not proper for the metastatic diseases, which due to the phase (1) (cid:3) = (cid:16) N= (cid:23)(cid:24) (cid:20) (cid:21)(cid:22)(cid:16) (cid:12)∑ (cid:18)(cid:19) 2(cid:25), K={3,4,5,6,7,8} (2) transition and spreading dynamics will be defined by graph presentations in the future. According to Rai (2017), the model which have the same size and number of connections as of a given network could maintain the degree sequence of the given network. By generating a random network with a given average degree (K) and initial size of tumor (N), we could construct degree sequence (m) and it is presented in formula 2: For volume calculation, we start by assuming that every cell inside the tumor has three states (normal, dead and metastatic) which corresponds to tumor by edges. With the given K degree from three to eight and N initial nodes, we constructed an ER (Erdös-Rényi) network, whose degree sequence of (m) could lead us to the tumor volume ratio. The pseudo-code using the dynamic of degree sequence under the influence of above factors producing tumor volume in graph network is described in Algorithm 2: Algorithm 2 Input: (N,K=3) Output: (Tumor Volume) Begin Algorithm While p_connected= 1 For N in nodes(G) If P (cid:23)(cid:26) Return m/n Endif Endfor End Algorithm (cid:27)(cid:25) < Factor then In Figure 4, we executed the model with different angiogenesis control factor in (0.1 to 0.9) for 60, 360, 650, 100 and 1200 cancer stem cells (CSCs) to produce the tumor volume analysis based on the density distribution of the graph. Figure 3: Tumor Angiogenesis volume ratio for 60 to 1200 cancer stem cell (CSC) 4. CONCLUSION In summary, agent-based modeling in Python with Mesa project represents a valuable and time-saving simulation technique, which allowed us to re-implement our tumor model across a network based space. Mesa enable a modeler to quickly write down the code and quickly explore the results. Mesa also collect the results easily and help us in data integration which is very important in the case of phase transition and pattern- oriented modeling. We migrated successfully from Netlogo to Python-Mesa, which will allow us to reintegrate our system quicker and reiterate simulations taken on previous system with more data in considerably less time. In the future, we will extend the model to run on cloud infrastructure in parallel. This could be an impressive achievement for fast analysis purposes in clinics, both on the predictive diagnostic and therapeutic side. As advantages of this contribution, we have a scalable model that together with Python can be distributed over an HPC architecture eliminating the limitations of other environments (e.g. Netlogo due to memory limitations of the JVM). When Python + Mesa can be deployed over an HPC architecture, there will be a notable increase in scalability and performance since, in this type of environment, the simulation will not be limited by memory. One of the most important disadvantages is the model visualization and animation for oncologists. Netlogo has a cells visualization and animation that can be useful to medical environments. Mesa only has the visualization of the network of agents and their interconnections, so graphic functions must be developed to see the evolution of the cells and the degree of metastasis/apoptosis and other relevant parameters for the medical analyst. Another important aspect that must be addressed will be the functional validation of the whole model since at this moment only a verification and partial validation of the behavior of the affected cells has been carried out. ACKNOWLEDGMENTS The MINECO Spain under contract TIN2014-53172-P and TIN2017-84875-P have supported this research. REFERENCES Wilensky, U., 1998. NetLogo Tumor model. Center for Connected and Computer-Based Modeling, Northwestern University, Evanston, IL. Project Simulation & Model Environment. National Science Foundation REC- 9814682 and REC-0126227 Integrated Learning Li, S., & Li, S., 2014. Cancer stem cells and tumor metastasis (Review). International Journal of Oncology, 44, 1806-1812. Weis, S. M., & Cheresh, D. A., 2011. Tumor angiogenesis: Molecular pathways and therapeutic targets. Nature Medicine. Carmeliet, P., & Jain, R. K., 2000. Angiogenesis in cancer and other diseases. Nature. Thorne, B. C., Bailey, A. M., & Peirce, S. M., 2007. Combining experiments with multi-cell agent- based modeling tissue patterning. Briefings in Bioinformatics. study biological to Edelman, L. B., Eddy, J. A., & Price, N. D., 2010. In silico models of cancer. Wiley Interdisciplinary Reviews: Systems Biology and Medicine. Soleimani, S., Shamsi, M., Ghazani, M. A., Modarres, H. P., Valente, K. P., Saghafian, M., Sanati-Nezhad, A., 2018. Translational models of tumor angiogenesis: A nexus of in silico and in vitro models. Biotechnology Advances. Tashakor, G., Luque, E., & Suppi, R., 2017. High Performance Computing for Tumor Propagation Agent-based Model. Granger, D. N., & Senchenkova, E, 2010. Inflammation and the Microcirculation. Colloquium Series on Integrated Systems Physiology: From Molecule to Function, 2(1), 1 -- 87. Girvan, M., & Newman, M. E. J., 2002. Community structure in social and biological networks. Proc. Natl. Acad. Sci. USA, 99(12), 7821 -- 7826. Tashakor, G., Suppi, R., 2017. HpcNetlogo. Frontend for the concurrent execution of Netlogo experiments using SGE HPC cluster. Cho, D.-Y., Kim, Y.-A., & Przytycka, T. M., 2012. Chapter 5: Network Biology Approach to Complex Diseases. PLoS Computational Biology, 8(12), e1002820. Masad, D., & Kazil, J., 2015. MESA: An Agent-Based Modeling Framework. Proceedings of the 14th Python in Science Conference.53 -- 60. Xing, F., Eisele, J., Bastian, P., & Heermann, D. W., 2015. Computational complexity of agent-based multi-scale cancer modeling. Simulation Series, 47(10), 351 -- 358. Villadsen, J., Jensen, A. S., Ettienne, M. B., Vester, S., Andersen, K. B., & Frøsig, A., 2013. Reimplementing a multi-agent system in python. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 7837 LNAI, pp. 205 -- 216). An, G. 2008. Introduction of an agent-based multi-scale modular architecture for dynamic knowledge representation of acute inflammation. Theoretical Biology and Medical Modelling. Railsback, S. F., Lytinen, S. L., & Jackson, S. K., 2006. Agent-based Simulation Platforms: Review and Development Recommendations. Simulation, 82(9), 609 -- 623. Bonabeau, E., 2002. Agent-based modeling: methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences, 99(suppl. 3), 7280 -- 7287. Norton, K.-A., McCabe Pryor, M. M., & Popel, A. S., 2015. Multiscale Modeling of Cancer. bioRxiv, 33977. D., M., R.S., C., M., R., E.C., F., B., T., & I., R., 2011. Modeling formalisms in systems biology. AMB Express, 1(1), 1 -- 14. Grimm, V., Revilla, E., Berger, U., Jeltsch, F., Mooij, W. M., Railsback, S. F, DeAngelis, D. L. (2005). Pattern-oriented modeling of agent-based complex systems: Lessons from ecology. Science. ALITALO, K., TAMMELA, T. and PETROVA, T.V. 2005. Lymphangiogenesis in development and human disease. Nature 438: 946-953. Kerbel, Robert S., (2008). Tumor angiogenesis. New England Journal of Medicine 358.19 2039-2049. Peskin, A., Kafadar, K., Santos, A., & Haemer, G. 2009. Robust Volume Calculations of Tumors of Various Sizes. D., Singh, Monga, S. P. S., Wadleigh, R., Sharma, A., Adib, H., Strader, L. 2000. Intratumoral therapy of cisplatin/epinephrine injectable gel for palliation in patients with obstructive esophageal cancer. American Journal of Clinical Oncology: Cancer Clinical Trials, 23(4), 386-392 G. Mishra, Rai, A., Pradhan, P., Nagraj, J., Lohitesh, K., Chowdhury, R., & Jalan, S. 2017. Understanding cancer complexome using networks, spectral graph theory and multilayer framework. Scientific Reports, 7. AUTHORS BIOGRAPHY GHAZAL TASHAKOR is Research Fellow and PhD Candidate for trainee research staff position (PIF) at the Computer Architecture & Operating Systems Department (CAOS) of the University Autònoma de Barcelona. Her include high performance simulation, agent-based modelling and their applications, service systems, integration, resource consumption, and execution time. Her e-mail address is: [email protected] research interests REMO SUPPI received his diploma in Electronic Engineering from the Universidad Nacional de La Plata (Argentina), and the PhD degree in Computer Science from the Universitat Autonoma de Barcelona (UAB) in 1996. At UAB, he spent more than 20 years researching on topics including computer simulation, distributed systems, high performance and distributed simulation applied to ABM or individual-oriented models. He has published several scientific papers on the topics above and has been the associate professor at UAB since 1997, also a member of the High Performance Computing for Efficient Applications and Simulation Research Group (HPC4EAS) is: [email protected]. at UAB. His address email
1706.08501	2	1706	2017-07-25T09:58:24	A Simulator for Hedonic Games	[ "cs.MA", "cs.AI" ]	Hedonic games are meant to model how coalitions of people form and break apart in the real world. However, it is difficult to run simulations when everything must be done by hand on paper. We present an online software that allows fast and visual simulation of several types of hedonic games. http://lukemiles.org/hedonic-games/	cs.MA	cs	A Simulator for Hedonic Games Luke Harold Miles University of Kentucky 1 A Story Mr Holt, the kindergarten teacher, gives his class these instructions: Hello class, The Metropolitan Museum of Art has a sudden shortage of sculptures and needs several new ones to fill its shelves. Please break into groups so that each group can build a Lego tower. The director of the museum will be here in an hour to pick up the towers and put them in the museum with your names on them. Please do the best job you can; you don't want to be professionally embarrassed. Each kindergartener wants to be in a group with her friends, but she also wants her friends to be happy in the group; she doesn't want her friends to be miserable. The graph below is a map of who is friends with whom in the small class. Notice that a would have more friends in the group {a, b, c, d, e} than {a, b, c, d}, but maybe a doesn't want e to be in the group because a knows that would make b, c, and d less happy. Strangely, a prefers {a, b, c, d} to {a, b, c, d, e}. You can imagine that the kindergarteners might try to choose the best group in some other way. The class would split into groups one way, but then people would be unhappy and keep changing their groups. How can we model all this? How could we easily visualize all this? 2 Hedonic Games Below is the original definition of a hedonic game. Hedonic games [Banerjee, Konishi, and Sonmez, 2001] were invented to model the formation and reforma- tion of groups. Definition 1. [Banerjee, Konishi, and Sonmez, 2001] A coalition formation game is a pair G = (N, ((cid:23)i)i∈N ), where N is a finite set of players and for every i ∈ N , (cid:23)i is a reflexive, complete, and transitive binary relation on N i = {C ∈ 2N : i ∈ C}. If C, D ∈ N i and C (cid:23)i D and D (cid:54)(cid:23)i C, then we write C (cid:31)i D. Definition 2. [Banerjee, Konishi, and Sonmez, 2001] A coalition structure Γ = {C1, . . . , Ck} is a partition of N . The coalition containing a player i ∈ N is denoted Γ (i). Any subset of N is called a coalition. That's a very minimal definition, and these most general hedonic games don't have many computationally useful properties. For that reason, several subclasses of hedonic games have been invented and studied. First though, let's look at stability. 2.1 The Core If Mr Holt were assigning groups, instead of letting the kids form their own groups, then he might want a way to predict if a given partition will stick before he actually moves people around. "Will the students stay in their groups or will they form new ones?" There are many ways you can ask the question "Is this coalition formation stable?" Seven good ways are mentioned in [Nguyen, Rey, Rey, Rothe, and Schend, 2016]. One of the most important ways to ask the question (and the focus of the survey [Woeginger, 2013]) is "Is this this coalition formation core stable?". Definition 3. In a hedonic game G with a partition Γ , if there is a nonempty set C ⊆ N where ∀i ∈ C : C (cid:31)i Γ (i), then we say that C blocks Γ , or C is a blocking coalition in Γ . If Γ cannot be blocked, then it is called core stable. The set of core stable partitions for a game G is called the core of G. 3 Varieties of Hedonic Games In the below paragraphs, n = N is the number of players, i is a player in N , and C, D ∈ N i are coalitions which contain i. 3.1 Fractional Hedonic Games [Aziz, Brandt, and Harrenstein, 2014] In fractional hedonic games, i assigns some real value vi(j) to every player j ∈ N . It's assumed that vi(i) = 0.1 We say C (cid:23)FR (C) ≥ uFR (D), where i D if uFR i i (cid:88) j∈C uFR i (C) = vi(j). 1 Raising your own score is equivalent to lowering everyone else's score. Lowering your own score is equivalent to raising everyone else's score. A fractional hedonic game is called simple if ∀i, j ∈ N : vi(j) ∈ {0, 1} and is called symmetric if ∀i, j ∈ N : vi(j) = vj(i). Aziz, Brandt, and Harrenstein show that even in fractional hedonic games which are both simple and symmetric, the core is sometimes empty and that checking core emptiness is Σp 2 -complete. 3.2 Friend and Enemy Oriented Hedonic Games [Dimitrov, Borm, Hendrickx, and Sung, 2006] In both of these kinds of games, i splits the other players in N into a set of friends, Fi, and a set of enemies, Ei. In friend-oriented games, i prefers coalitions with more friends and breaks ties by considering the number of enemies. In other words, C (cid:23)FO i D ⇐⇒ C ∩ Fi > D ∩ Fi ∨ (C ∩ Fi = D ∩ Fi ∧ C ∩ Ei ≤ D ∩ Ei) ⇐⇒ uFO where uFO (C) ≥ uFO (C) = nC ∩ Fi − C ∩ Ei. (D), i i i So if C has 8 of i's friends and 600 of i's enemies and D has 7 of i's friends and 0 of i's enemies, then i would still rather be in C. In enemy-oriented games, i tries to minimize enemies and only considers friends to break a tie. In other words, C (cid:23)EO i D ⇐⇒ C ∩ Ei < D ∩ Ei ∨ (C ∩ Ei = D ∩ Ei ∧ C ∩ Fi ≥ D ∩ Fi) ⇐⇒ uEO where uEO (C) ≥ uEO (C) = C ∩ Fi − nC ∩ Ei. (D), i i i Dimitrov, Borm, Hendrickx, and Sung show that the core is guaranteed to be non-empty in both kinds of games. However, finding a core stable partition is NP-hard in enemy-oriented games2 but polynomial time in friend-oriented games. 3.3 Altruistic Hedonic Games Three levels of altruism are considered. Let avg(S) =(cid:80) [Nguyen, Rey, Rey, Rothe, and Schend, 2016] As in friend and enemy oriented hedonic games, i divides the other players into friends, Fi, and enemies, Ei. The idea is that a player wouldn't want to be in a coalition C where his friends were miserable, even if C had all of his friends and none of his enemies. x∈S x/S denote the average of a multiset of numbers. And, as above, the utilities ui are defined so that C (cid:23)i D ⇐⇒ ui(C) ≥ ui(D). 2 More precisely, if you could always find a core stable coalition structure in polyno- mial time, then you could also find the largest clique in any (undirected, unweighted) graph in polynomial time. In selfish-first altruistic games, a player cares most about his own hap- piness and uses his friends' preferences to break ties. 'Happiness' here means the friend-oriented score. This is distinct from friend-oriented games in that a tightly connected coalition C with 6 friends and 3 enemies is preferred to a sparse coalition D with 6 friends and 3 enemies, because i's friends in C are happier than i's friends in D. uSF i (C) = n5uFO i (C) + avg(uFO j (C) : j ∈ C ∩ Fi). In equal-treatment altruistic games, a player takes his and all his friends' opinions into account equally when evaluating a partition: uEQ i (C) = avg(uFO j (C) : j ∈ C ∩ Fi ∪ {ı}). And in altruistic-treatment altruistic games (i.e., truly altruistic games), a player prefers coalitions where his friends are happy and breaks ties by consid- ering his own happiness. uAL i (C) = uFO i (C) + n5avg(uFO j (C) : j ∈ C ∩ Fi). Nguyen, Rey, Rey, Rothe, and Schend show that selfish-first altruistic games always have an nonempty core. Whether equal-treatment altruistic games and truly altruistic games ever have empty cores are open questions. I suspect that the core is always nonempty in both games. 4 The Simulator I wrote software to simulate hedonic games and put in on the internet. You can draw graphs, choose partitions, choose several different player types, and check the stability of the partition under several different measures. Hopefully this will help others and myself quickly understand different hedonic games and speed up the process of finding stable partitions. http://lukemiles.org/hedonic-games The website works better on laptops than smartphones. Updates may have been made to the website since this arXiv version was uploaded. Bibliography Haris Aziz, Felix Brandt, and Paul Harrenstein. Fractional hedonic games. In Proceedings of the 2014 international conference on Autonomous Agents & Multi-Agent Systems (AAMAS), 2014. Suryapratim Banerjee, Hideo Konishi, and Tayfun Sonmez. Core in a simple coalition formation game. Social Choice and Welfare, 2001. Dinko Dimitrov, Peter Borm, Ruud Hendrickx, and Shao Chin Sung. Simple priorities and core stability in hedonic games. Social Choice and Welfare, 2006. Nhan-Tam Nguyen, Anja Rey, Lisa Rey, Jorg Rothe, and Lena Schend. Altru- istic hedonic games. In Proceedings of the 2016 international conference on Autonomous Agents & Multi-Agent Systems (AAMAS), 2016. Gerhard J Woeginger. Core stability in hedonic coalition formation. In Proceed- ings of SOFSEM 2013: Theory and Practice of Computer Science, 2013.
1610.07901	1	1610	2016-10-25T14:45:24	Avoid or Follow? Modelling Route Choice Based on Experimental Empirical Evidences	[ "cs.MA" ]	Computer-based simulation of pedestrian dynamics reached meaningful results in the last decade, thanks to empirical evidences and acquired knowledge fitting fundamental diagram constraints and space utilization. Moreover, computational models for pedestrian wayfinding often neglect extensive empirical evidences supporting the calibration and validation phase of simulations. The paper presents the results of a set of controlled experiments (with human volunteers) designed and performed to understand pedestrian's route choice. The setting offers alternative paths to final destinations, at different crowding conditions. Results show that the length of paths and level of congestion influence decisions (negative feedback), as well as imitative behaviour of "emergent leaders" choosing a new path (positive feedback). A novel here illustrated model for the simulation of pedestrian route choice captures such evidences, encompassing both the tendency to avoid congestion and to follow emerging leaders. The found conflicting tendencies are modelled with the introduction of a utility function allowing a consistent calibration over the achieved results. A demonstration of the simulated dynamics on a larger scenario will be also illustrated in the paper.	cs.MA	cs	Proceedings of the 8th International Conference on Pedestrian and Evacuation Dynamics (PED2016) Hefei, China - Oct 17 – 21, 2016 Paper No. 84 Avoid or Follow? Modelling Route Choice Based on Experimental Empirical Evidences Luca Crociani1∗ Daichi Yanagisawa2 Giuseppe Vizzari1 Katsuhiro Nishinari2 and Stefania Bandini1,2 1 Complex Systems and Artificial Intelligence research center, University of Milano–Bicocca Viale Sarca 336 – Ed. U14, 20126 Milano (ITALY) {name.surname}@disco.unimib.it 2 Research Center for Advanced Science and Technology, The University of Tokyo 4-6-1 Komaba, Meguro-ku, Tokyo 153-8904 (JAPAN) [email protected]; [email protected] Abstract - Computer-based simulation of pedestrian dynamics reached meaningful results in the last decade, thanks to empirical evidences and acquired knowledge fitting fundamental diagram constraints and space utilization. Moreover, computational models for pedestrian wayfinding often neglect extensive empirical ev- idences supporting the calibration and validation phase of simulations. The paper presents the results of a set of controlled experiments (with human volunteers) designed and performed to understand pedestrian's route choice. The setting offers alternative paths to final destinations, at different crowding conditions. Re- sults show that the length of paths and level of congestion influence decisions (negative feedback), as well as imitative behaviour of "emergent leaders" choosing a new path (positive feedback). A novel here illustrated model for the simulation of pedestrian route choice captures such evidences, encompassing both the tendency to avoid congestion and to follow emerging leaders. The found conflicting tendencies are modelled with the introduction of a utility function allowing a consistent calibration over the achieved results. A demonstration of the simulated dynamics on a larger scenario will be also illustrated in the paper. Keywords: Pedestrian Dynamics, Route Choice, Experiment, Simulation, Validation 1. Introduction The simulation of pedestrians and crowds is a consolidated research and application area that, al- though existing results are currently employed on a daily basis by designers and decision makers, still presents challenges for researchers in different fields and disciplines. Innovations in the field come from diverse contributions, from attempts to integrate automated analysis of crowd behaviour and simula- tion [1] to investigations on the impact of emotional aspects on crowd dynamics [2], and even on the usage of simulation as a real-time support to crowd management [3]. Even if we only consider choices and actions related to walking in normal situations, modelling human decision making activities and actions is a complicated task. Decisions on pedestrian's actions are taken at different levels of abstraction, from path planning to the regulation of distance from other pedestrians (potentially having different kinds of relationships, as recent results on the impact of groups indicate [4]) and obstacles present in the environment. A factor making the evaluation of simulation results more com- plicated is the fact that the measure of success and validity of a model is not the optimality of pedestrian dynamics with respect to some cost function (either on the individual or on the aggregated dynamics), as (for instance) in robotics. The success of a simulation model is associated to its plausibility, the adherence of the simulation results to data that can be acquired by means of observations or experiments. The present research effort is aimed at producing insights on this aspect: an experiment involving pedestrians has been set up to investigate to which extent pedestrians facing a relatively simple choice (i.e. choose one of two available gateways leading to the same target area) in which, however, they can ∗Corresponding author. Figure 1: Configuration of the setting for the experiment. face a trade-off situation between length of the trajectory to be covered and estimated travel time. The closest gateway, in fact, is initially selected by most pedestrians but it is too narrow to allow a smooth passage of so many pedestrians, and it quickly becomes congested. The other choice can therefore become much more reasonable, allowing a higher average walking speed and comparable (if not even lower) travel time. We observed that several pedestrians choose longer paths to preserve high walking speed, and often do so following a first emerging leader. Modelling this kind of choices with current approaches can be problematic. The model presented in this paper represents a step in the direction fitting this kind of evidences but providing a general approach to this kind of pedestrian decisions. The proposed model encompasses, in fact, both a proxemic tendency to avoid congestion, as well as an imitation mechanism: these conflicting tendencies can be calibrated according to empirical evidences. The former aspect represents a sort of negative feedback caused by crowding conditions (i.e. avoid a congested exit) but, on the other hand, the incentive to following another pedestrian that just decided to choose a more distant but less crowded passage represents a positive feedback on this choice. In the following, we will first describe the experiment and the achieved results. The description of the model will be given in Section 3, followed by the discussion of its calibration and achieved results. 2. Experimental Study The experiment has been performed at the University of Tokyo in November 2015. A group of 46 persons has participated, uniformly composed of male students aged around 20 years old. The setting has been configured with the intention to acquire empirical evidences on the influence of crowding conditions on route choice decisions. The setting is designed to describe an elementary choice: it is characterized by a rectangular environment of 7.2 × 12 m2 divided in two areas of equal size; the access is regulated by three gates positioned to create three paths of different lengths. A schematic representation of the scenario is illustrated in Figure 1. The entrance and exits of the environment are aligned on the x-axis, in order to generate a shortest path (patha) and to induce the decisions of the participants. The average lengths of the three possible paths –calculated as the sum of distances between the central points of the crossed openings– is as following: (i) patha 12.08; (ii) pathb 12.85; (iii) pathc 14.76. The difference between Patha and Pathb is relatively small (< 1 m) while Pathc is significantly longer than the shortest path. These differences will be reflected on the achieved results pointing out, as generally known, that the main element influencing the route choice is the distance. After each run, the participants were asked to regroup in the starting area right after the exit point, in order to optimize the procedural times. This has also allowed to balance the potential influence of the cultural elements related to left–right preference of movement, by reversing the direction of flow in the scenario. The openings related to Pathb and Pathc were eventually closed to configure different procedures. Each procedure has been repeated 4 times to achieve a more consistent dataset. The procedures are described in the following: 1. only Patha allowed; 2. Patha and Pathb allowed; 84 - 2 3. Patha and Pathc allowed; 4. all gates are open. The procedure iterations have been performed with a randomized schedule, to possibly avoid bias provided by the learning of participants. Finally, to stimulate the will to minimize the travelling time towards the destination, the participants have been asked to reach quickly the opposite side of the setting. 2.1. Methodology for the Analysis The experiment has been performed under an arcade of the university buildings, due to weather conditions. The lack of safe points to attach any camera to the ceiling did not allow to have a zenithal perspective in the video recordings, which would be useful for an automatic tracking. Hence, four HD cameras positioned on tall tripods (5m circa) have been used to record the experiment. The videos have been synchronized by means of a global chronometer shown to the cameras at the beginning of the recording. Since the frame rate of the two cameras with the highest resolution was slightly variable (a small number of frames was skipped during the recording), some manual adjustment has been done to the video synchronization. This was possible by using visible events in the overlapping areas of multiple camera views and also by using the audio track of the videos. The software AviSynth has been employed to achieve the synchronization and merge the multiple video tracks. The results that will be presented in the following subsection has been achieved by means of manual counting and tracking. 2.2. Results and Discussion The video footages of the performed experiments did not allow us to perform a fine tracking of the exact trajectories followed by the different pedestrians, mainly due to the positioning of cameras. Nonetheless, different types of analysis will be carried out to acquire novel empirical evidences on pedestrian route choice decisions in presence of crowding conditions. So far, we focused on a relatively simple analysis that regarded the number of participants who passed through each gate depending on the experimental procedure, and therefore on the level of congestion. Our aim was to verify if the experimental results supported the conjecture that pedestrians, when the perceived level of congestion makes less appealing the shortest path, choose a longer trajectory to preserve their walking speed. Additional analyses will be carried out, within the limits posed by the vantage point, as a consequence of the results of this first round of integrated analysis and synthesis activities [1]. The Table 1 shows the counting of people passing through each gate for procedures with multiple open gates. As already briefly discussed, procedure 2 and 3 are characterized by a decision among two choices of different lengths; in procedure 2 the difference among the alternatives is quite small (i.e. less than a meter, less than 10% of the shortest trajectory, considering the middle point of the gates for the measurement), whereas in procedure 3 the longer trajectory is more significantly worse than the shortest path (i.e. over 20% longer). Finally, procedure 4 allows pedestrians to choose among all three alternatives. Intermediate gates are relatively small (1 m) but the evidence shown that they still allow the passage of two pedestrians almost at the same time. Moreover, the passage between the starting area and the region before the gates is relatively wide (2.4 m) and here pedestrians are able to walk side by side; this implies that some of them are naturally closer to the best trajectory while for others the worst ones are not that longer. Results are essentially in line with the expectations on the conjecture that pedestrians would dis- tribute among the alternative passages in case of congestion. In fact, in procedure 2, almost half of the pedestrians chooses the slightly longest path to avoid the congested path. The fact that, in procedure 3, the longer alternative (P athc) is more significantly worse, makes this choice appealing to a lower number of pedestrians. Procedure 4, finally, shows that the P atha and P athb are perceived as almost equivalent, but a few pedestrians even choose P athc in order to preserve a high walking speed by avoiding congestion on the best alternatives. In particular, a few general qualitative considerations on all the analyzed procedures can be done: • pedestrians choosing longer paths (P athb and P athc) generally enter the area before the gates on upper side, referring to the map of the scenario in Figure 1; • pedestrians choosing longer paths generally do so after some preceding ones have perceivably cho- sen the best path, and therefore can be considered as potential future competitors either for the 84 - 3 (a) 19 s (b) 21 s (c) 23 s (d) 25 s Figure 2: Video camptures from one camera during procedure 4. The behaviours considered in this paper can be recognized: at 19 s of the video, the crowding of P atha and P athb leads a person to employ P athc. After 2 s, other students have taken the same decision, successively followed by others. In the bottom images, a decision change is apparent: the highlighted person is firstly headed to the central gate, but after a short while he notices that P atha is getting empty again and he employs it. occupation of the gate or for the path leading to it; this is particularly apparent for procedure 3 and 4; • sometimes, when choosing the longest path (P athc) pedestrians seem to follow someone before them that had chosen this trajectory before and, at the same time, avoiding much closer pedestrians that have perceivably chosen other shorter paths. Figure 2 emphasizes these considerations by showing 4 sequential screen-shots from one iteration of procedure 4. While these results are still quite aggregated, they suggest that pedestrians are driven by the expected travel time rather than just the length of the trajectory. Moreover, the observed following dynamics could be explained as a form of cooperative behaviour by which those who are on the back select a preceding pedestrian whose choice is assumed to be promising, at a glance, without actually computing the expected quickest time to the exit. To some extent, this reminds conflicting but simultaneously present behavioural components of cohesion and separation, considered in simulation models [5]. In particular, the modelling approach that will be introduced in the following Section, will consider both the fact that other pedestrians are generally perceived as repulsive but also the fact that the decision of a pedestrian to detour (i.e. change a previous decision on the path to be followed) is a locally perceivable event that might trigger a similar reconsideration by nearby pedestrians. 3. A Novel Model for Pedestrian Wayfinding The proposed agent-based model defines two components of the agent, respectively dedicated to the low level reproduction of the movement towards a target (i.e. the operational level, considering a three level model described in [6]) and to the decision making activities related to the next destination to be pursued (i.e. the wayfinding). Due to lack of space, the mechanisms for the physical reproduction of movement will not be described† and the discussion will be limited to the different representations of the environment composing the knowledge needed for the wayfinding. †For technical details of this aspect it is referred to [7]. 84 - 4 Procedure Patha Pathb Pathc 2 22 23 25 23 24 23 21 23 Average 23.25 22.75 3 Average 4 27 28 30 27 28 18 22 21 22 Average 20.75 0 0 0 0 0 16 19 18 19 18 0 0 0 0 0 19 18 16 19 18 12 5 7 5 7.25 Table 1: Counting of number of people passed through each path for procedures 2, 3 and 4. 3.1. The Representation of the Environment and the Knowledge of Agents The environment [8] is discrete and modelled with a rectangular grid of 40 cm sided square cells, as usually applied in this context. The simulation scenario is drawn by means of several markers. Basic markers of the scenario are start areas, obstacles and final destinations –ultimate targets of pedestrians. To allow the wayfinding, two other markers are introduced. Openings are sets of cells that divide the environment into regions, together with obstacles. These objects constitutes decision elements for the route choice, which will be denoted as intermediate targets. Finally, regions are markers that describe the type of the region: with them it is possible to design particular classes of regions (e.g. stairs, ramps). This model uses the floor fields approach [9], spreading potentials from cells of obstacles and desti- nations to provide information about distances. The two types of floor fields are denoted as path field, spread from each target, and obstacle field, a unique field spread from all obstacle cells. In addition, a dynamic floor field that has been denoted as proxemic field is used to reproduce plausible distance towards the agents in low density situations. The presence of intermediate targets allows the computation of a graph-like representation of the walkable space, based on the concept of cognitive map [10]. The algorithm for the computation of this data structure is defined in [11] and it uses the information of the floor fields associated to openings and final destinations. Recent approaches explores also the modelling of partial and incremental knowledge of the environment by agents (e.g. [12]), but this aspect goes beyond the scope of the current work. The data structure identifies regions (e.g. a room) as nodes of the labelled graph and openings as edges. Overall the cognitive map allows the agents to identify their position in the environment and it constitutes a basis for the generation of an additional knowledge base, which will enable the reasoning for the route calculation. This additional data structure is denoted as Paths Tree and it contains the free flow travel times‡ related to plausible paths towards a final destination, from each intermediate target in the environment. The concept of plausibility of a path is encoded in the algorithm for the computation of the tree, which is thoroughly discussed in [13]. Formally, for the choice of paths the agents access the information of a Paths Tree, generated from a final destination End, with the function P aths(R, End). Given the region R of the agent, this outputs a set of couples {(Pi, tti)}. Pi = {Ωk, . . . , End} is the ordered set describing paths starting from Ωk, belonging to Openings(R) and leading to End. tti is the respective free flow travel time. 3.2. The Route Choice Model of Agents This novel part of the model is inspired by the behaviour observed in the experiment. The aim is to propose an approach that would enable agents to choose their path considering distances as well as the evolution of the dynamics, being able to follow emerging leaders as well as to avoid congestions. At the same time, the model must provide a sufficient variability of the results and a calibration over possible empirical data. The life-cycle of the agent is illustrated in Figure 3. First of all, the agent performs a perception of his situation, aimed at understanding its position in the environment and the perceivable markers from its region. At the very beginning of its life, the agent does not have any information about its location, ‡Calculated with their desired speed of walking. 84 - 5 Figure 3: The life-cycle of the agent, emphasizing the two components of the model. thus the first assignment is its localization, inferring the location in the Cognitive Map. Once the agent knows its region, it loads the Paths Tree and evaluates the possible paths towards its final destination. In this model, the evaluation and choice of paths is designed with the concept of path utility, computed to assign a probability to be chosen. The probabilistic choice outputs a new intermediate target of the agent, which will be followed at the operational level. The functions introduced for the wayfinding model are Evaluate Paths and Choose Paths, which will be now discussed. 3.2.1. The Utility and Choice of Paths The function that computes the probability of choosing a path is exponential with respect to the utility value associated to it. This is analogous to the choice of movement at the operational layer: P rob(P ) = N · exp (κttEvaltt(P ) − κqEvalq(P ) + κf Evalf (P )) (1) The exponential function is used to emphasize differences in the perceived utility values of paths, limiting the choice of relatively bad solutions. The exponent comprises the three observed components influencing the route choice decision, which are aggregated with a weighted sum. In particular, the first element evaluates the expected travel times, the second considers the queuing conditions through the considered path and the last one introduces a positive influence of choices of nearby agents to pursue the associated path (i.e. imitation of emerging leaders). Each of the three functions provide normalized values within the range [0, 1]. 3.2.2. The Evaluation of Travelling Times Distances and travel times are elements that significantly affect our choice of route and this is observed also in the experiment, where the shortest path resulted as the most chosen on average. First of all, the free flow travel time tti of a path Pi is achieved with P aths(R, End) (End is the agent's final destination, identifying the appropriate Paths Tree, and R is the region of the agent) and it is summed with the free flow travel time to reach the first opening Ωk described by each path: TravelTime(Pi) = tt i + P FΩk (x, y) Speed d (2) P FΩk (x, y) is the value of the path field associated to Ωk in the position of the agent and Speed d is its desired speed. The value of the travelling time is then evaluated by means of the following function: Eval tt(P ) = Ntt · min Pi∈Paths(r) (TravelTime(Pi)) TravelTime(P ) (3) Ntt is the normalization factor. The minimum value of travel times provides a range of the function of (0,1], being 1 for the path with minimum travel time and decreasing as the difference with the other 84 - 6 paths increases. This modelling choice, makes this function describe the utility of the route in terms of travel times, instead of its cost. 3.2.3. The Evaluation of Congestion The competition arisen between the participants of the experiment made some of them choose longer paths, in order to avoid the possible queue. The behaviour modelled in the agent reflects this effect, by considering congestion as a negative element for the evaluation of the path. For the evaluation of this component, a function is firstly introduced for denoting agents a(cid:48) that precedes the evaluating agent a in the route towards the opening Ω of a path P : Forward (Ω, a) = {a (cid:48) ∈ Ag\{a} : Dest(a (cid:48) ) = Ω ∧ PF Ω(Pos(a (cid:48) )) < PF Ω(Pos(a))} (4) where P os and Dest indicates respectively the position and current destination of the agent; the fact that PF Ω(Pos(a(cid:48))) < PF Ω(Pos(a)) assures that a(cid:48) is closer to Ω than a,. Each agent is therefore able to perceive the main direction of the others (its current destination). This kind of perception is plausible considering that only preceding agents are counted, but we want to restrict its application when agents are sufficiently close to the next passage. To introduce a way to calibrate this perception, the following function and an additional parameter γ is introduced: PerceiveForward (Ω, a) = Forward (Ω, a), 0, if PF Ω(Pos(a)) < γ otherwise (5) (cid:40) (cid:40) The function Eval q is finally defined with the normalization of PerceiveForward values for all the openings connecting the region of the agent: Eval q(P ) = Nq · PerceiveForward (FirstEl (P ), myself ) width(FirstEl (P )) (6) where FirstEl returns the first opening of a path, myself denotes the evaluating agent and width scales the evaluation over the width of the door. 3.2.4. Propagation of Choices - Following behaviour Another behaviour observed in the experiment described a substantial following of emerging leaders who sudden change their route due to the surrounding crowding conditions. The final component of this model aims at representing the effect of additional stimulus for the agents, generated by decision changes of other persons. An additional grid is then introduced to model this event, whose functioning is similar to the one of a dynamic floor field. The grid, called ChoiceField, is used to spread a gradient from the positions of agents that, at a given time-step, change their plan due to the perception of congestion. The functioning of this field is described by two parameters ρc and τc, which defines the diffusion radius and the time needed by the values to decay. The diffusion of values from an agent a, choosing a new target Ω(cid:48), is performed in the cells c of the grid with Dist(Pos(a), c) ≤ ρc with the following function: Diffuse(c, a) = 1/Dist(Pos(a), c) 1 if Pos(a) (cid:54)= c otherwise (7) The diffused values persist in the ChoiceField grid for τc simulation steps, then they are simply discarded. The index of the target Ω(cid:48) is stored together with the diffusion values, thus the grid contains in each cell a vector of couples {(Ωm, diff Ωm ), . . . , (Ωn, diff Ωn)}, describing values of influence associated to each opening of the related region. While multiple neighbour agents change their choices towards the opening Ω(cid:48), the values of the diffusion are summed up in the respective diff Ω(cid:48). Moreover, after having changed its decision, an agent spreads the gradient in the grid for a configurable amount of time steps represented by an additional parameter τa, thus it influences the choices of its neighbours for τa time. The existence of values diff Ωk > 0 for some opening Ωk implies that the agent is influenced in the evaluation phase by one of these openings, but the probability for which this influence is effective is, after all, regulated by the utility weight κf . In case of having multiple diff Ωk > 0 in the same cell, a individual influence is chosen with a simple probability function based on the normalized weights diff associated to the cell. Hence, for an evaluation performed by an agent a at time-step t, the utility component Evalf can be equal to 1 only for one path P , between the paths having diff Ωk > 0 in the position of a. 84 - 7 (a) Procedure 2 (b) Procedure 3 Figure 4: Comparison between empirical data and simulation results with different calibrations. (c) Procedure 4 4. Calibration of the Model and Conclusions To evaluate the reliability of the proposed model, we employed the observed data of the experiment for performing a comparison with results coming from the simulation of the same setting. A set of 50 simulation iterations per each experimental procedure has been configured and the counting of agents passing through each gate has been gathered. Results are shown in Figure 4. The value for parameters has two kinds of impact on the achieved results: it obviously determines the 84 - 8 number of pedestrians choosing a specific door, but it also affects the variance of these numbers. In fact, moving from the first calibration (κtt = 10, κq = 7, κf = 5) to the second (κtt = 10, κq = 2.5, κf = 0.5), we achieved average results closer to the empirical observation, however the variability was too high. This lead us to considering a final calibration (κtt = 100, κq = 25, κf = 5) that preserved the proportion among the weights, but increasing their values to reduce the effect of randomness, leading therefore to very similar average results but a much lower variability, more similar to the one observed in the experiments. In conclusion, the presented model is effective in reproducing the observed phenomena, but it is of general applicability and we are in the process of evaluating the effects of its usage in larger scale egress scenarios. Acknowledgement The experiment was authorised and supported by The University of Tokyo. The authors would like to thank Francesco Crippa for his support in the development and experimentation of the model. References [1] G. Vizzari and S. Bandini, "Studying pedestrian and crowd dynamics through integrated analysis and synthesis," IEEE Intelligent Systems, vol. 28, no. 5, pp. 56–60, 2013. [2] T. Bosse, M. Hoogendoorn, M. C. A. Klein, J. Treur, C. N. van der Wal, and A. van Wissen, "Modelling collective decision making in groups and crowds: Integrating social contagion and interacting emotions, beliefs and intentions," Autonomous Agents and Multi-Agent Systems, vol. 27, no. 1, pp. 52–84, 2013. [Online]. Available: http://dx.doi.org/10.1007/s10458-012-9201-1 [3] A. Tsiftsis, I. G. Georgoudas, and G. C. Sirakoulis, "Real data evaluation of a crowd supervising system for stadium evacuation and its hardware implementation," IEEE Systems Journal, vol. 10, no. 2, pp. 649–660, 2016. [Online]. Available: http://dx.doi.org/10.1109/JSYST.2014.2370455 [4] N. W. F. Bode, S. Holl, W. Mehner, and A. Seyfried, "Disentangling the impact of social groups on response times and movement dynamics in evacuations," PLoS ONE, vol. 10, no. 3, pp. 1–14, 03 2015. [Online]. Available: http://dx.doi.org/10.1371%2Fjournal.pone.0121227 [5] M. Moussaıd, N. Perozo, S. Garnier, D. Helbing, and G. Theraulaz, "The walking behaviour of pedestrian social groups and its impact on crowd dynamics." PloS one, vol. 5, no. 4, p. e10047, 2010. [6] S. P. Hoogendoorn and P. H. L. Bovy, "Pedestrian route-choice and activity scheduling theory and models," Transportation Research Part B: Methodological, vol. 38, no. 2, pp. 169–190, 2004. [7] S. Bandini, L. Crociani, and G. Vizzari, "An approach for managing heterogeneous speed profiles in cellular automata pedestrian models," Journal of Cellular Automata, (in press). [8] D. Weyns, A. Omicini, and J. Odell, Autonomous Agents Multi-Agent Systems, vol. 14, no. 1, pp. 5–30, 2007. [9] C. Burstedde, K. Klauck, A. Schadschneider, and J. Zittartz, "Simulation of pedestrian dynamics using a two-dimensional cellular automaton," Physica A: Statistical Mechanics and its Applications, vol. 295, no. 3 - 4, pp. 507–525, 2001. [10] E. C. Tolman, "Cognitive maps in rats and men." Psychological review, vol. 55, no. 4, pp. 189–208, 1948. [11] L. Crociani, A. Invernizzi, and G. Vizzari, "A hybrid agent architecture for enabling tactical level decisions in floor field approaches," Transportation Research Procedia, vol. 2, pp. 618–623, 2014. [12] E. Andresen, D. Haensel, M. Chraibi, and A. Seyfried, "Wayfinding and cognitive maps for pedestrian models," in Proceedings of Traffic and Granular Flow 2015 (TGF2015). Springer, (in press). [13] L. Crociani, A. Piazzoni, G. Vizzari, and S. Bandini, "When reactive agents are not enough: Tactical level decisions in pedestrian simulation," Intelligenza Artificiale, vol. 9, no. 2, pp. 163–177, 2015. 84 - 9
1807.08663	1	1807	2018-07-23T15:09:25	Measuring collaborative emergent behavior in multi-agent reinforcement learning	[ "cs.MA" ]	Multi-agent reinforcement learning (RL) has important implications for the future of human-agent teaming. We show that improved performance with multi-agent RL is not a guarantee of the collaborative behavior thought to be important for solving multi-agent tasks. To address this, we present a novel approach for quantitatively assessing collaboration in continuous spatial tasks with multi-agent RL. Such a metric is useful for measuring collaboration between computational agents and may serve as a training signal for collaboration in future RL paradigms involving humans.	cs.MA	cs	Measuring collaborative emergent behavior in multi- agent reinforcement learning Sean L. Barton1, Nicholas R. Waytowich2, Erin Zaroukian1, & Derrik E. Asher1 1Computational & Information Sciences Directorate, U.S. Army Research Laboratory 2Human Research & Engineering Directorate, U.S. Army Research Laboratory learning (RL) has Abstract. Multi-agent reinforcement important implications for the future of human-agent teaming. We show that improved performance with multi-agent RL is not a guarantee of the collaborative behavior thought to be important for solving multi-agent tasks. To address this, we present a novel approach for quantitatively assessing collaboration in continuous spatial tasks with multi-agent RL. Such a metric is useful for measuring collaboration between computational agents and may serve as a training signal for collaboration in future RL paradigms involving humans. Keywords: Multi-agent Reinforcement Learning · Deep Reinforcement Learning · Human-Agent Teaming · Collaboration 1 Introduction Reinforcement learning (RL) is an attractive option for providing adaptive behavior in computational agents because of its theoretical generalizability to complex problem spaces [1, 2]. In particular, deep RL recently produced striking results [3]. Extending RL to the multi-agent domain has received an increasing amount of attention as the need for human-agent teams has increased, especially with regards to training agents to behave collaboratively [1, 4 -- 8]. Unfortunately, the nature of multi-agent RL makes guaranteeing collaboration between agents impossible except in limited provable cases, even when these methods yield better task performance [1, 7]. Thus far, evaluating collaboration in multi-agent learning has been accomplished by measuring performance in tasks where coordination is required. While this may be satisfactory for discretized tasks where cooperative policies are provably optimal [1, 6 -- 8] it is not clear that this methodology generalizes well to more complex and continuous tasks (such as those presented in [4, 5] and here). The question at hand is how to assess performance enhancing coordination (or collaboration) between agents in multi-agent RL tasks. Here, we present a method borrowed from the field of ecology, called convergence cross mapping (CCM), and show how it can be used to measure collaboration between agents during a predator- prey pursuit task. Additionally, we show a striking result: a state-of-the-art multi-agent reinforcement learning algorithm does not exhibit coordinated behavior between agents during a collaborative task even though high task performance is achieved, indicating that performance metrics alone are not sufficient for measuring collaboration. 2 Methods Simulation Environment 2.1 In a modification of the classic predator-prey pursuit task (see Figure 1A-C), three slower predator agents score points each time they make contact with a prey agent in a continuous bounded 2D particle environment. Predator agents were identical in terms of capabilities (i.e., velocity and acceleration). This simulation environment was made available through the OpenAI Gym network [9] and was developed for the multi-agent deep learning algorithm discussed below [4]. Prey agents were capable of 33% greater acceleration and 25% greater maximum velocity than any predator agent, making capture by a solitary predator extremely difficult. All agents had the same mass, with minimal elastic properties to provide a small bump force upon collisions. Agent positions and velocities were randomized at the start of each episode, and acceleration was initially set to zero. Predator agents all received a fixed reward when any one of them made contact with the prey agent. Prey agents received a punishment when they were contacted. 2.2 Agents In order to evaluate our metric's ability to estimate collaboration between predators in the predator-prey task, we utilized two types of agents: learning agents and fixed- strategy (non-learning) agents. Learning agents' behaviors were guided by a multi- agent deep deterministic policy gradient (MADDPG) algorithm [4]. In all cases, prey agents were learning agents and thus utilized the MADDPG algorithm independent of predator behavior. Two types of distinct fixed-strategy predators were implemented to demonstrate upper and lower bounds of coordinated behaviors. The first (termed 'Chaser' predators, see Figure 1A) naively pursued prey agents by maximally accelerating in the instantaneous direction of the prey relative to the predator's own position. As such, Chasers were incapable of coordinating their behavior with each other. The second strategy (termed 'Spring' predators, see Figure 1B) also naively minimized distance to the prey, but predator movements were modified by spring forces which constrained their position and velocity relative to one another. A Spring predator's movement direction was a sum of the spring forces acting on it and its desired vector of movement towards the prey. In this case, predator actions were explicitly coordinated with their partners'. The MADDPG algorithm (Figure 1C and 1E) used to guide learning agent behavior is an extension of a deep deterministic policy gradient (DDPG) algorithm [3] into the multi-agent domain [4]. Like DDPG, MADDPG utilizes an actor-critic model with deep neural networks representing policy and Q-learners (Figure 1E). Multi-agent capabilities are achieved by passing information about each agent's state and actions to each critic network. As such, the learning agents are joint action learners as opposed to independent learners, giving them an advantage when coordinating their behaviors [8]. All conditions (Figure 1A-C) were trained for 100k episodes to ensure model convergence before evaluating collaborative behavior (Figure 1D). Each episode lasted for 25 time steps during learning and 2000 time steps during evaluation. These intervals were selected to match what was shown in literature for sufficient learning [4] and required for analysis [10]. During testing, 10 episodes were recorded without learning for each predator strategy in order to produce a distribution of agent behaviors with the same level of training over various random starting positions. Fig. 1. Experimental design. A-C) Yellow circles denote prey, while others represent predator strategies. Black arrows indicate actions. Panels show A) Chaser, B) Spring (coupling shown as grey dotted lines), and C) MADDPG agent conditions. D) At test time, one predator's behavior was determined by the physics of a double pendulum (grey circle). Agents coordinated with a modified predator should have their goal-seeking actions (grey dashed arrow) augmented by the actions of the modified predator. E) Schematic of MADDPG algorithm. 2.3 Collaboration Metrics In order to measure collaboration, we used predators' positions over time as time-series data to implement a technique called convergent cross mapping (CCM) which examines the causal influence one time-series has on another [10]. The CCM technique embeds a time-series in a high-dimensional attractor space and then uses this embedded data as a model to predict states of the other time-series in its own attractor space. To the extent that this is possible, the original time-series is said to be causally driven by the time- series it attempts to model. Thus, in multi-agent tasks with homogeneous agents, we are defining collaboration to be the amount of causal influence between agents as measured by the CCM. Importantly, this metric can be misleading when two time-series have a mutually causal relationship with a third. For the present task, this is important because the causal influence one predator has on another can be confounded by their mutual relationship with the prey. To address this, at test time we modified the behavior of one predator agent such that its previous behavior was replaced by a secondary behavior that did not pursue the prey1. If a causal influence exists between predators, the movements of the modified predator should change the behavior of the other predators, even though the modified predator no longer pursues the prey (see Figure 1D). 3 Results Table 1 shows the log scaled mean reward per episode for predator and prey agents in the Chaser, Spring, and MADDG cases. It is clear that MADDPG-equipped predators are betters performers of this pursuit task, achieving roughly an order of magnitude more average reward per trial. Table 1. Performance results for the different experimental conditions Chaser Spring MADDPG N 10 10 10 Mean Reward 1.85 4.29 13.70 SD 0.66 0.92 2.57 95% CI 0.47 0.66 1.84 While the capacity for coordinated actions afforded by the MADDPG algorithm coupled with the increased task performance may suggest collaboration between predator agents, the CCM analysis tells a different story (Figure 2). As shown, there is very little difference, in terms of causal influence, between MADDPG-equipped predators and naive Chasers. The causal influence of the modified agent might be marginally stronger for MADDPG predators (hinting at the shared action information afforded by this algorithm), but the effect is too weak to make any strong conclusions about the collaboration between the agents in this case. For Spring predators, on the other hand, there is a strong causal influence from the modified predator. The Spring experimental condition (see Figure 2) shows an example of coupled behaviors in the predator-prey pursuit task. This provides a strong validation for the CCM technique, and illustrates an ecological upper-bound on coordinated actions between learning agents. 1 In this case, the dynamics of a double pendulum were used to specify the movement of the modified predator. Fig. 2. CCM measure for causal influence of modified predator behavior on unmodified predators. True causal influence is indicated by an increase in CCM score (Y axis) as more of a trial (X axis) is used to build the model (so called "convergence" [10]). A clear causal relationship is shown for Spring predators. Minimal causal influence is detectable for MADDPG agents. Coordination for Chaser agents is indistinguishable from zero. 4 Discussion The MADDPG algorithm presented by Lowe et al. [4] is promising, because it suggests that multi-agent deep RL may be able to solve complex problems in continuous tasks involving multiple actors. However, as we demonstrate here the improved performance of the MADDPG algorithm does not necessarily indicate coordinated behavior between agents. Assessing collaboration between agents instead requires a direct measure of the coordination between agent actions. We present such a method here, in the form of CCM, and we validate the metric within the context of a continuous multi-agent task. Though we did not find strong evidence of coordination between the actions of learning predators, the marginal increase in coordination over Chaser predators points to potential for collaboration given sufficient learning pressure. Producing collaborative behavior in computational agents is critical for the future of human-agent teams. Human-factors research has long held that when computational systems fail to adapt to human needs, serious issues in performance and human satisfaction can arise [11]. These issues are often best alleviated by promoting a collaborative relationship between humans and computational agents, rather than forcing humans to act as overseers [12, 13]. Measures like CCM, which can be used to assess (and even promote) collaborative behaviors in multi-agent RL, constitute powerful tools for the future of human-computer interactions. Acknowledgements and Disclosure: This research was sponsored by the Army Research Laboratory and was accomplished under Cooperative Agreement Number W911NF-18-2-0058. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. References 1. Matignon, L., Laurent, G.J., Le Fort-Piat, N.: Independent reinforcement learners in cooperative markov games: A survey regarding coordination problems. The Knowledge Engineering Review. 27, 1 -- 31 (2012). 2. Sen, S., Sekaran, M., Hale, J., others: Learning to coordinate without sharing information. In: AAAI. pp. 426 -- 431 (1994). 3. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., Hassabis, D.: Human-level control through deep reinforcement learning. Nature. 518, 529 -- 533 (2015). 4. Lowe, R., WU, Y., Tamar, A., Harb, J., Pieter Abbeel, O., Mordatch, I.: Multi-Agent Actor- Critic for Mixed Cooperative-Competitive Environments. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.) Advances in Neural Information Processing Systems 30. pp. 6382 -- 6393. Curran Associates, Inc. (2017). 5. Foerster, J., Farquhar, G., Afouras, T., Nardelli, N., Whiteson, S.: Counterfactual Multi- Agent Policy Gradients. arXiv:1705.08926 [cs]. (2017). 6. Matignon, L., Laurent, G., Le Fort-Piat, N.: Hysteretic q-learning: An algorithm for decentralized reinforcement learning in cooperative multi-agent teams. In: IEEE/rsj international conference on intelligent robots and systems, iros'07. pp. 64 -- 69 (2007). 7. Lauer, M., Riedmiller, M.: An algorithm for distributed reinforcement learning in cooperative multi-agent systems. In: In proceedings of the seventeenth international conference on machine learning. Citeseer (2000). 8. Claus, C., Boutilier, C.: The dynamics of reinforcement learning in cooperative multiagent systems. AAAI/IAAI. 1998, 746 -- 752 (1998). 9. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., Zaremba, W.: OpenAI Gym. arXiv:1606.01540 [cs]. (2016). 10. Sugihara, G., May, R., Ye, H., Hsieh, C.-h., Deyle, E., Fogarty, M., Munch, S.: Detecting causality in complex ecosystems. science. 1227079 (2012). 11. Parasuraman, R., Sheriden, T.B., Wickens, C.D.: A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans. 30, 286 -- 297 (2000). 12. Rovira, E., McGarry, K., Parasuraman, R.: Effects of Imperfect Automation on Decision Making in a Simulated Command and Control Task. Human Factors. 49, 76 -- 87 (2007). 13. Klein, G., Woods, D.D., Bradshaw, J.M., Hoffman, R.R., Feltovich, P.J.: Ten challenges for making automation a "team player" in joint human-agent activity. IEEE Intelligent Systems. 19, 91 -- 95 (2004).
1612.04299	1	1612	2016-12-13T17:51:59	Algorithms for Graph-Constrained Coalition Formation in the Real World	[ "cs.MA", "cs.AI" ]	Coalition formation typically involves the coming together of multiple, heterogeneous, agents to achieve both their individual and collective goals. In this paper, we focus on a special case of coalition formation known as Graph-Constrained Coalition Formation (GCCF) whereby a network connecting the agents constrains the formation of coalitions. We focus on this type of problem given that in many real-world applications, agents may be connected by a communication network or only trust certain peers in their social network. We propose a novel representation of this problem based on the concept of edge contraction, which allows us to model the search space induced by the GCCF problem as a rooted tree. Then, we propose an anytime solution algorithm (CFSS), which is particularly efficient when applied to a general class of characteristic functions called $m+a$ functions. Moreover, we show how CFSS can be efficiently parallelised to solve GCCF using a non-redundant partition of the search space. We benchmark CFSS on both synthetic and realistic scenarios, using a real-world dataset consisting of the energy consumption of a large number of households in the UK. Our results show that, in the best case, the serial version of CFSS is 4 orders of magnitude faster than the state of the art, while the parallel version is 9.44 times faster than the serial version on a 12-core machine. Moreover, CFSS is the first approach to provide anytime approximate solutions with quality guarantees for very large systems of agents (i.e., with more than 2700 agents).	cs.MA	cs	A Algorithms for Graph-Constrained Coalition Formation in the Real World FILIPPO BISTAFFA and ALESSANDRO FARINELLI, University of Verona JES ´US CERQUIDES and JUAN RODR´IGUEZ-AGUILAR, IIIA-CSIC SARVAPALI D. RAMCHURN, University of Southampton Coalition formation typically involves the coming together of multiple, heterogeneous, agents to achieve both their individual and collective goals. In this paper, we focus on a special case of coalition formation known as Graph-Constrained Coalition Formation (GCCF) whereby a network connecting the agents constrains the formation of coalitions. We focus on this type of problem given that in many real-world applications, agents may be connected by a communication network or only trust certain peers in their social network. We propose a novel representation of this problem based on the concept of edge contraction, which allows us to model the search space induced by the GCCF problem as a rooted tree. Then, we propose an anytime solution algorithm (CFSS), which is particularly efficient when applied to a general class of characteristic functions called m + a functions. Moreover, we show how CFSS can be efficiently parallelised to solve GCCF using a non-redundant partition of the search space. We benchmark CFSS on both synthetic and realistic scenarios, using a real-world dataset consisting of the energy consumption of a large number of households in the UK. Our results show that, in the best case, the serial version of CFSS is 4 orders of magnitude faster than the state of the art, while the parallel version is 9.44 times faster than the serial version on a 12- core machine. Moreover, CFSS is the first approach to provide anytime approximate solutions with quality guarantees for very large systems of agents (i.e., with more than 2700 agents). Categories and Subject Descriptors: I.2 [Computing Methodologies]: Artificial Intelligence General Terms: Algorithms ACM Reference Format: ACM Trans. Intell. Syst. Technol. V, N, Article A (January YYYY), 23 pages. DOI:http://dx.doi.org/10.1145/0000000.0000000 1. INTRODUCTION Coalition Formation (CF) is one of the key approaches to establishing collaborations in multi-agent systems. It involves the coming together of multiple, possibly heteroge- neous, agents in order to achieve either their individual or collective goals, whenever they cannot do so on their own. Building upon the seminal work of Shehory and Kraus [1998], Sandholm et al. [1999] identify the key computational tasks involved in the CF process: (i) coalitional value calculation: defining a characteristic function which, given a coalition as an argument, provides its coalitional value; (ii) coalition structure generation (CSG): finding a partition of the set of agents (into disjoint coalitions) that maximises the sum of the values of the chosen coalitions; and (iii) payment computa- tion: finding the transfer or payment to each agent to ensure it is fairly rewarded for its contribution to its coalition. Author's addresses: F. Bistaffa and A. Farinelli, Department of Computer Science, University of Verona, Verona, Italy; J. Cerquides and J. Rodr´ıguez-Aguilar, IIIA-CSIC, Barcelona, Spain; S. D. Ramchurn, Elec- tronics and Computer Science, University of Southampton, Southampton, United Kingdom. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is per- mitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permissions may be requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481, or [email protected]. c(cid:13) YYYY ACM 2157-6904/YYYY/01-ARTA $15.00 DOI:http://dx.doi.org/10.1145/0000000.0000000 ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:2 Filippo Bistaffa et al. On the one hand, typical CF approaches assume that the values of all the coalitions are stored in memory, allowing to read each value in constant time. However, this assumption makes the size of the input of the CSG and payment computation prob- lems exponential, as the entire set of coalitions (whose size is 2n for n agents) must be mapped to a value. On the other hand, CSG and payment computation are combinato- rial in nature and most existing solutions do not scale well with the number of agents. In this paper, we focus on the CSG problem to provide solutions that can be applied to real-world problems, which usually involve hundreds or thousands of agents. The computational complexity of the CSG problem is due to the size of its search space,1 which contains every possible subset of agents as a potential coalition. How- ever, in many real-world applications, there are constraints that may limit the forma- tion of some coalitions [Rahwan et al. 2011]. Specifically, we focus on a specific type of constraints that encodes synergies or relationships among the agents and that can be expressed by a graph [Myerson 1977], where nodes represent agents and edges encode the relationships between the agents. In this setting, edges enable connected agents to form a coalition and a coalition is considered feasible only if its members represent the vertices of a connected subgraph. Such constraints are present in several real- world scenarios, such as social or trust constraints (e.g., energy consumers who prefer to group with their friends and relatives in forming energy cooperatives [Hampshire County Council 2014]), physical constraints (e.g., emergency responders may join spe- cific teams in disaster scenarios where only certain routes are available), or communi- cation constraints (e.g., non-overlapping communication loci or energy limitations for sending messages across a network from one agent to another). Hereafter, we shall refer to the CF problem where coalitions are encoded by means of graphs as Graph- Constrained Coalition Formation (GCCF). It is important to note that the addition of these constraints does not lower the complexity of the problem. In particular, Voice et al. [2012a] show that the GCCF problem remains NP-complete. In this work, we are primarily interested in developing CSG solutions for GCCF that are deployable in real-world scenarios involving hundreds or thousands of agents, such as collective energy purchasing [Vinyals et al. 2012; Farinelli et al. 2013] and ridesharing [Bistaffa et al. 2015]. Notice that, since the computation of an optimal solution is often infeasible for large-scale systems, our CSG algorithm should be able to provide anytime approximate solutions with good quality guarantees. Moreover, the memory requirements should scale well with the number of agents. In this context, the works by Voice et al. [2012a; 2012b] represent the state of the art for GCCF. However, there are some drawbacks that hinder their applicability. Voice et al. [2012a] make assumptions that do not hold in most real-world applica- tions (see Section 2.1.4), whereas the memory requirements of the approach in [Voice et al. 2012b] grow exponentially in the number of agents, hence limiting the scalabil- ity. To overcome these drawbacks, in this paper we propose CFSS (Coalition Formation for Sparse Synergies), the first approach for GCCF that computes anytime solutions with theoretical quality guarantees for large systems (i.e., more than 2700 agents). As recently noticed in a survey on CSG by Rahwan et al. [2015], previous approaches in the CF literature have been either applied to small-scale synthetic scenarios, or, in the case of heuristic approaches, cannot provide any theoretical guarantees on the quality of their solutions. Moreover, we provide P-CFSS, a parallelised version of CFSS that exploits multi-core CPUs. Finally, we identify a general class of closed-form functions, denoted as m + a, for which we provide upper bounds, allowing for coalitional values to be computed online (i.e., their storage can be avoided). 1A set of n agents can be partitioned in Ω(( n ln(n) )n) ways, i.e. the nth Bell number [Berend and Tassa 2010]. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:3 In more detail, this paper advances2 the state of the art in the following ways: (1) We provide a new representation for GCCF which, by using edge contractions on the graph, can efficiently build a search tree where each node is a feasible coalition structure, while avoiding redundancy (i.e., each solution appears only once). (2) We identify a general class of characteristic functions, i.e., m + a functions, which are expressive enough to represent a wide range of real-world GCCF problems. (3) We propose CFSS, a branch and bound algorithm that, when applied to CF with m + a functions, can solve the CSG problem for GCCF and can provide anytime approximate solutions with good quality guarantees. (4) We propose P-CFSS, a parallel version of CFSS that is up to 9.44 times faster than the serial version on a 12-core machine. The rest of the paper is organised as follows. Section 2 discusses the relationship be- tween our work and the existing literature, and Section 3 formally defines GCCF. Sec- tion 4 explains how we generate our search space, and Section 6 details the domains used to benchmark CFSS, our branch and bound approach described in Section 5, and Section 7 discusses our empirical evaluation. Finally, Section 8 concludes the paper. 2. RELATED WORK In this section we elaborate on related work in the areas of CF (Section 2.1), team formation (Section 2.2), graph theory (Section 2.3) and optimisation (Section 2.4). 2.1. Coalition Formation 2.1.1. Classic CSG algorithms. A number of algorithms have been developed to solve CSG for the general CF problem where all coalitions can be formed (i.e., non-GCCF). These range from mixed-integer programming to branch and bound techniques [Rah- wan et al. 2009] through Dynamic Programming (DP) [Rahwan and Jennings 2008b]. In particular, Sandholm et al. [1999] and Dang and Jennings [2004] focused on provid- ing anytime solutions with quality guarantees. However, their solutions do not scale (growing in O(nn)) and, as discussed by Voice et al. [2012b], they cannot be employed to solve CSG for GCCF, since assigning artificially low values (such as −∞) to in- feasible coalitions would not be suitable for assessing valid bounds. Finally, Rahwan et al. [2008a; 2009; 2012] developed IDP-IP∗, the state of the art algorithm for classic CSG. However, IDP-IP∗ is limited to tens of agents (30 at most) due to its memory requirements (i.e., Θ (2n)), as such approaches need to store all coalition values. To overcome the intractability due to such memory requirements, a number of works [Ohta et al. 2009; Ueda et al. 2011; Tran-Thanh et al. 2013] have examined alternative function representations, which allow to reduce the computational com- plexity of the associated CF problems. Unfortunately, their models may not be able to capture the realistic nature of functions such as the collective energy purchasing one we consider here. On the one hand, this function cannot be concisely expressed as a MC network, as its MC network would require an exponential amount of memory with respect to the number of agents. On the other hand, the concepts of agent types/skills imply that it is possible to fully characterise the contribution of each agent on the ba- sis of a small set of features, in order to achieve the conciseness of the representation. However, in our scenario each agent is associated to its own energy consumption pro- file, resulting in a number of types/skills equal to the number of agents. Hence, we do not compare against these works, since we are interested in developing techniques that can handle complex functions such as the collective energy purchasing function. 2This paper subsumes the work of Bistaffa et al. [2014b] and the non-archival work of Bistaffa et al. [2014a]. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:4 Filippo Bistaffa et al. 2.1.2. CSG algorithms based on heuristics. Very few heuristic solutions to the CSG prob- lem have been developed over the last few years. For example, Sen and Dutta [2000] propose a solution based on genetic algorithms, Dos Santos and Bazzan [2012] pro- pose an approach based on swarm intelligence (the bee clustering algorithm) for task allocation in the RoboCup Rescue domain, and Farinelli et al. [2013] propose an ap- proach based on hierarchical clustering. Meta-heuristic approaches to CSG have also been investigated, for example Keinanen [2009] proposes a CSG algorithm based on Simulated Annealing, while Di Mauro et al. [2010] use a stochastic local search ap- proach (GRASP) to iteratively build a coalition structure of high quality. Even if these approaches are not able to provide any guarantees on the solution quality, they can compute solutions for large numbers of agents. Hence, in Section 7.5 we compare CFSS against C-Link, since it is the most recent heuristic approach for CSG and it has been tested using the collective energy purchasing function, which we also consider. 2.1.3. Constrained CF. The works discussed above focus on unconstrained CF and can- not be directly used in contexts where constraints of various types may limit the for- mation of some coalitions. In this respect, Shehory and Kraus [1998] first introduced the idea, arising in many realistic scenarios, of restricting the maximum cardinality k of the coalitions in CSG, highlighting that, even though this constraint lowers the number of coalitions from exponential, i.e., 2n, to polynomial, i.e., O(cid:0)nk(cid:1), the problem ity guarantees, which, however, can be used if all O(cid:0)nk(cid:1) coalitions are valid. On the remains NP-hard. Therefore, the authors propose an approximate algorithm with qual- other hand, Rahwan et al. [2011] developed a model of Constrained Coalition Forma- tion (CCF), differing from standard CF due to the presence of constraints that forbid the formation of certain coalitions. However, authors provide an algorithm for optimal CSG only for Basic CCF (BCCF) games, which cannot be used to represent every GCCF problem, as shown in Section A.1 of the Appendix. Finally, in a recent work, Iwasaki et al. [2015] proposed an approach to check the non-emptiness of the core when the grand coalition does not form, hence effectively addressing a CSG problem. Notice that, even though such an approach is tested on 1000 agents, the authors assume that the number of feasible coalitions is less than 10000. This assumption is not reasonable for large-scale scenarios we are interested to solve. For the sake of comparison, the number of feasible coalitions with 50 agents and m = 1 (i.e., the simplest network topology we consider in our tests) is ∼ 150 billions, thus severely limiting the scalability of such an approach on large-scale scenarios due to its memory requirements. 2.1.4. State of the art algorithms for GCCF. Voice et al. [2012a; 2012b] were the first to propose algorithms for the GCCF problem. However, there are some drawbacks that hinder their applicability. First, [Voice et al. 2012a] can only be applied to character- istic functions fulfilling the independence of disconnected members (IDM) property. The IDM property requires that, given two disconnected agents i and j, the presence of agent i does not affect the marginal contribution of agent j to a coalition. This as- sumption is rather strong for real-world applications. As noticed by Shehory and Kraus [1998] considering task allocation, the addition of a new agent to a coalition could re- sult in intra-coalition coordination and communication costs, which increase with the size of the coalition. Hence, realistic functions capturing such costs (such as the ones in Section 6.1) do not satisfy the IDM property, hence this approach cannot be applied. Second, the DyCE algorithm [Voice et al. 2012b] uses DP to find the optimal coalition structure by progressively splitting the current solution into its best partition. DyCE is not an anytime algorithm and requires an exponential amount of memory in the num- ber of agents (i.e., Θ (2n)). Hence, the scalability of this approach is limited to systems consisting of tens of agents (around 30). ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:5 2.2. Team formation The problem of forming groups of agents has also been widely studied in the context of Team Formation, in which several formal definitions of such problem have been pro- posed. As an example, Gaston and desJardins [2005] devise a heuristic to modify the graph connecting the agents based on local autonomous reasoning, without consider- ing any concept of global optimal solution. The problem studied by Lappas et al. [2009] focuses on finding a single group of agents who possess a given set of skills, so as to minimise the communication cost within such a group. Marcolino et al. [2013] focus on forming a single group of agents that has the maximum strength in the set of world states. Finally, Liemhetcharat and Veloso [2014] are interested in modelling the values of the characteristic function, based on observations of the agents. In this paper, we ad- dress the specific group formation problem in which groups must form a partition (into disjoint coalitions) of a given set of agents, with the objective of maximising the sum of the coalitional values. Such problem is equivalent to the complete set partitioning problem [Yun Yeh 1986], i.e., the standard definition adopted in the CF literature. 2.3. Graph theory techniques Our approach enumerates all the feasible partitions of the set of agents by means of the edge contraction operation, a graph theoretic technique known for its application in the algorithm to solve the Min-Cut problem [Karger 1993]. Edge contraction has never been employed in CF [Rahwan et al. 2015], hence we aim at investigating its use in this paper. In this context, the problem of enumerating all the connected subgraphs (corresponding to feasible coalitions in GCCF scenarios) of a given graph has been studied in a number of works [Voice et al. 2012b; Skibski et al. 2014]. Nonetheless, such algorithms can only be used to enumerate feasible coalitions, and cannot be applied to enumerate feasible coalition structures (as CFSS does), which are sets of disjoint feasible coalitions that collectively cover the entire set of agents. 2.4. Submodular-supermodular decomposition Submodular functions have been widely studied in the optimisation literature [Schri- jver 2003] in virtue of their natural diminishing returns property, which makes them suitable for many applications [Nemhauser et al. 1978; Narayanan 1997]. Moreover, Shekhovtsov et al. [2006; 2008] focused on general functions that can be decomposed as the sum of supermodular and submodular components, exploiting such a property to achieve better results in the solution of several optimisation problems. While this approach is similar to the decomposition we propose in Section 6.1, our result holds for superadditive and subadditive functions (cf. Definition 4), which are weaker (i.e., more general) properties with respect to supermodularity and submodu- larity. In fact, it is easy to show that supermodularity (resp. submodularity) implies superadditivity (resp. subadditivity), but the converse is not true [Schrijver 2003]. 3. GCCF PROBLEM DEFINITION The Coalition Structure Generation (CSG) problem [Sandholm et al. 1999; Shehory and Kraus 1998] takes as input a finite set of n agents A and a characteristic function v : 2A → R, that maps each coalition C ∈ 2A to its value, describing how much collec- tive payoff a set of players can gain by forming a coalition. A coalition structure CS is a partition of the set of agents into disjoint coalitions. The set of all coalition structures is Π(A). The value of a coalition structure CS is assessed as the sum of the values of its composing coalitions, i.e., (cid:88) C∈CS V (CS) = v(C). (1) ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:6 Filippo Bistaffa et al. CSG aims at identifying CS∗, the most valuable coalition structure, i.e., CS∗ = arg maxCS∈Π(A) V (CS). Graphs have been used in different scenarios to encode syn- ergies, coordination among players, possible collaborations or cooperation structures [Myerson 1977; Voice et al. 2012b; Meir et al. 2012]. Myerson [1977] and Demange [2004] pioneered the study of graphs to model cooperation structures. Given an undi- rected graph G = (A,E), where E ⊆ A×A is a set of edges between agents, representing the relationships between them, Myerson considers a coalition C to be feasible if all of their members are connected in the subgraph of G induced by C. That is, for each pair of players from a, b ∈ C there is a path in G that connects them without going out of C. Thus, given a graph G the set of feasible coalitions is FC(G) = {C ⊆ A The subgraph induced by C on G is connected}. A Graph-Constrained Coalition Formation (GCCF) problem is a CSG problem together with a graph G, in which a coalition C is considered feasible if C ∈ FC(G). Moreover, a coalition structure CS is considered feasible if each of its coalitions is feasible, i.e., CS(G) = {CS ∈ Π(A) CS ⊆ FC(G)}. A GCCF problem aims at identifying the most valuable coalition structure, defined as CS∗ = arg maxCS∈CS(G) V (CS). on the concept of edge contraction. In the next section, we propose a novel representation of the GCCF problem based 4. A GENERAL ALGORITHM FOR GCCF We now present a general algorithm to solve GCCF by showing that all feasible coali- tion structures induced by G can be modelled as the nodes of a search tree in which each feasible coalition structure is represented only once. Specifically, we first detail how we use edge contractions to represent the GCCF problem and then we provide a depth-first approach to build and traverse the search tree to find the optimal solution. 4.1. Generating feasible coalition structures via edge contractions In this section we show that each CS ∈ CS(G) can be represented by a corresponding graph GCS = (V,F), where V ⊆ 2A and F ⊆ V × V, i.e., each node u ∈ V represents a particular coalition. Notice that in the initial graph G = (A,E) each vertex u ∈ A represents a single agent, and hence, G can be seen as the representation of the feasible coalition structure formed by all the singletons. In what follows, we will show that, for each CS ∈ CS(G), the corresponding GCS can be obtained as the contraction of a set of edges of G, and that each contraction of a set of edges of G represents a feasible coalition structure CS ∈ CS(G). In more detail, let us define an edge contraction as follows. DEFINITION 1. Given a graph G = (V,F), where V ⊆ 2A and F ⊆ V × V, and an edge e = (u, v) ∈ F, the result of the contraction of e is a graph G(cid:48) obtained by removing e and the corresponding vertices u and v, and adding a new vertex w = u ∪ v. Moreover, each edge incident to either u or v in G will become incident to w in G(cid:48), merging the parallel edges (i.e., the edges that are incident to the same two vertices) that may result. Intuitively, one edge contraction represents the merging of the coalitions associated to the incident vertices. Figure 1 shows the contraction of the edge ({A} ,{C}), which results in a new vertex {A, C} connected to vertex {B}. Notice that edge contraction is a commutative operation (i.e., first contracting e and then e(cid:48) results in the same graph as first contracting e(cid:48) and then e). Hence, we can define the contraction of a set of edges as the result of contracting each of the edges of the set in any given order. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:7 Fig. 1. Example of an edge contraction (the dashed edge is contracted). Fig. 2. Example of a 2-coloured edge contraction (the dashed edge is contracted). Remark 4.1. Given a graph G, the graph G(cid:48) resulting from the contraction of any set of edges of G represents a feasible coalition structure, where coalitions correspond to the vertices of G(cid:48). Remark 4.2. Given a graph G, any feasible coalition structure CS can be generated by contracting a set of edges of G. Thus, a possible way of listing all feasible coalition structures is to list the contraction of every subset of edges of the initial graph. However, notice that the number of sub- sets of edges is larger than the number of feasible coalition structures over the graph. For example, in the triangle graph in Figure 1a, the number of subsets of edges is 2E = 23 = 8, but the number of feasible coalition structures is 5 (i.e., {A}{B}{C}, {A, B}{C}, {A, C}{B}, {A}{B, C} and {A, B, C}). This redundancy is due to the fact that the contraction of any two or three edges leads to the same coalition structure, i.e., the grand coalition A = {A, B, C}. Thus, we need a way to avoid listing feasible coalition structures more than once. To avoid such redundancies, we mark each edge of the graph to keep track of the edges that have been contracted so far. Notice that there are only two different alternative actions for each edge: either we contract it, or we do not. If we decide to contract an edge, it will be removed from the graph in all the subtree rooted in the current node, but if we decide not to contract it, we have to mark such edge to make sure that we do not contract it in the future steps of the algorithm. To represent such marking, we will use the notion of 2-coloured graph. DEFINITION 2. A 2-coloured graph Gc = (V,F, c) is composed of a set of vertices V ⊆ 2A and a set of edges F ⊆ V × V, as well as a function c : F → {red, green} that assigns a colour (red or green) to each edge of the graph. In our case, a red edge means that a previous decision not to contract that edge was made. On the one hand, green edges can be still contracted. Figure 2a shows an exam- ple of a 2-colour graph in which edge ({A} ,{D}) is coloured in red (dotted line). Hence, in any subsequent step of the algorithm it is impossible to contract it. On the other hand, all other edges in such graph can still be contracted. In a 2-coloured graph, we define a green edge contraction (e.g., dashed line in Figure 2a) as follows. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. AlgorithmsforGraph-ConstrainedCoalitionFormationintheRealWorldA:7{A}{B}{C}(a)Beforecontraction{A,C}{B}(b)AftercontractionFig.1:Exampleofanedgecontraction(thedashededgeiscontracted).{A}{B}{C}{D}{F}(a)Beforecontraction{B}{A,C}{D}{F}(b)AftercontractionFig.2:Exampleofa2-colourededgecontraction(thedashededgeiscontracted).Remark4.1.GivenagraphG,thegraphG(cid:48)resultingfromthecontractionofanysetofedgesofGrepresentsafeasiblecoalitionstructure,wherecoalitionscorrespondtotheverticesofG(cid:48).Remark4.2.GivenagraphG,anyfeasiblecoalitionstructureCScanbegeneratedbycontractingasetofedgesofG.Thus,apossiblewayoflistingallfeasiblecoalitionstructuresistolistthecontractionofeverysubsetofedgesoftheinitialgraph.However,noticethatthenumberofsub-setsofedgesislargerthanthenumberoffeasiblecoalitionstructuresoverthegraph.Forexample,inthetrianglegraphinFigure1a,thenumberofsubsetsofedgesis2E=23=8,butthenumberoffeasiblecoalitionstructuresis5(i.e.,{A}{B}{C},{A,B}{C},{A,C}{B},{A}{B,C}and{A,B,C}).Thisredundancyisduetothefactthatthecontractionofanytwoorthreeedgesleadstothesamecoalitionstructure,i.e.,thegrandcoalitionA={A,B,C}.Thus,weneedawaytoavoidlistingfeasiblecoalitionstructuresmorethanonce.Toavoidsuchredundancies,wemarkeachedgeofthegraphtokeeptrackoftheedgesthathavebeencontractedsofar.Noticethatthereareonlytwodifferentalternativeactionsforeachedge:eitherwecontractit,orwedonot.Ifwedecidetocontractanedge,itwillberemovedfromthegraphinallthesubtreerootedinthecurrentnode,butifwedecidenottocontractit,wehavetomarksuchedgetomakesurethatwedonotcontractitinthefuturestepsofthealgorithm.Torepresentsuchmarking,wewillusethenotionof2-colouredgraph.DEFINITION2.A2-colouredgraphGc=(V,F,c)iscomposedofasetofverticesV⊆2AandasetofedgesF⊆V×V,aswellasafunctionc:F→{red,green}thatassignsacolour(redorgreen)toeachedgeofthegraph.Inourcase,arededgemeansthatapreviousdecisionnottocontractthatedgewasmade.Ontheonehand,greenedgescanbestillcontracted.Figure2ashowsanexam-pleofa2-colourgraphinwhichedge({A},{D})iscolouredinred(dottedline).Hence,inanysubsequentstepofthealgorithmitisimpossibletocontractit.Ontheotherhand,allotheredgesinsuchgraphcanstillbecontracted.Ina2-colouredgraph,wedefineagreenedgecontraction(e.g.,dashedlineinFigure2a)asfollows.ACMTransactionsonIntelligentSystemsandTechnology,Vol.V,No.N,ArticleA,Publicationdate:JanuaryYYYY.AlgorithmsforGraph-ConstrainedCoalitionFormationintheRealWorldA:7{A}{B}{C}(a)Beforecontraction{A,C}{B}(b)AftercontractionFig.1:Exampleofanedgecontraction(thedashededgeiscontracted).{A}{B}{C}{D}{F}(a)Beforecontraction{B}{A,C}{D}{F}(b)AftercontractionFig.2:Exampleofa2-colourededgecontraction(thedashededgeiscontracted).Remark4.1.GivenagraphG,thegraphG(cid:48)resultingfromthecontractionofanysetofedgesofGrepresentsafeasiblecoalitionstructure,wherecoalitionscorrespondtotheverticesofG(cid:48).Remark4.2.GivenagraphG,anyfeasiblecoalitionstructureCScanbegeneratedbycontractingasetofedgesofG.Thus,apossiblewayoflistingallfeasiblecoalitionstructuresistolistthecontractionofeverysubsetofedgesoftheinitialgraph.However,noticethatthenumberofsub-setsofedgesislargerthanthenumberoffeasiblecoalitionstructuresoverthegraph.Forexample,inthetrianglegraphinFigure1a,thenumberofsubsetsofedgesis2E=23=8,butthenumberoffeasiblecoalitionstructuresis5(i.e.,{A}{B}{C},{A,B}{C},{A,C}{B},{A}{B,C}and{A,B,C}).Thisredundancyisduetothefactthatthecontractionofanytwoorthreeedgesleadstothesamecoalitionstructure,i.e.,thegrandcoalitionA={A,B,C}.Thus,weneedawaytoavoidlistingfeasiblecoalitionstructuresmorethanonce.Toavoidsuchredundancies,wemarkeachedgeofthegraphtokeeptrackoftheedgesthathavebeencontractedsofar.Noticethatthereareonlytwodifferentalternativeactionsforeachedge:eitherwecontractit,orwedonot.Ifwedecidetocontractanedge,itwillberemovedfromthegraphinallthesubtreerootedinthecurrentnode,butifwedecidenottocontractit,wehavetomarksuchedgetomakesurethatwedonotcontractitinthefuturestepsofthealgorithm.Torepresentsuchmarking,wewillusethenotionof2-colouredgraph.DEFINITION2.A2-colouredgraphGc=(V,F,c)iscomposedofasetofverticesV⊆2AandasetofedgesF⊆V×V,aswellasafunctionc:F→{red,green}thatassignsacolour(redorgreen)toeachedgeofthegraph.Inourcase,arededgemeansthatapreviousdecisionnottocontractthatedgewasmade.Ontheonehand,greenedgescanbestillcontracted.Figure2ashowsanexam-pleofa2-colourgraphinwhichedge({A},{D})iscolouredinred(dottedline).Hence,inanysubsequentstepofthealgorithmitisimpossibletocontractit.Ontheotherhand,allotheredgesinsuchgraphcanstillbecontracted.Ina2-colouredgraph,wedefineagreenedgecontraction(e.g.,dashedlineinFigure2a)asfollows.ACMTransactionsonIntelligentSystemsandTechnology,Vol.V,No.N,ArticleA,Publicationdate:JanuaryYYYY.A:8 Filippo Bistaffa et al. Algorithm 1 SOLVEGCCF(Gc) 1: best ← Gc, F ← ∅ 2: F.PUSH(Gc) 3: while F (cid:54)= ∅ do 4: 5: 6: 7: 8: return best node ← F.POP() if V (node) > V (best) then best ← node F.PUSH(CHILDREN (node)) (cid:46) Initialise solution with singletons and search frontier F with empty stack (cid:46) Push Gc as the first node to visit (cid:46) Search loop (cid:46) Get current node (cid:46) Check function value (cid:46) Update current best solution (cid:46) Update frontier F (cid:46) Return optimal solution Algorithm 2 CHILDREN(Gc) 1: G(cid:48) ← Gc, Ch ← ∅ 2: for all e ∈ Gc : c (e) = green do 3: 4: 5: return Ch Ch ← Ch ∪ {GREENEDGECONTRACTION (G(cid:48), e)} Mark edge e with colour red in G(cid:48) (cid:46) Initialise graph G(cid:48) with Gc and empty set of children (cid:46) For all green edges (cid:46) Return the set of children DEFINITION 3. Given a 2-coloured graph G = (V,F, c) and a green edge e ∈ F, the result of the contraction of e is a new graph G(cid:48) obtained by performing the contraction of e on G. Whenever two parallel edges are merged into a single one, the resulting edge is coloured in red if at least one of them is red-coloured, and it is green-coloured otherwise. The rationale behind marking parallel edges in this way is that, whenever we mark an edge e = (u, v) to be red, we want the agents in u and v to be in separate coalitions, hence whenever we merge some edges with e we must mark the new edge as red to be sure that future edge contractions will not generate a coalition that contains both the agents corresponding to nodes u and v. For example, note that in Figure 2 the red edge ({A} ,{D}) (dotted in the figure) and the green edge ({D} ,{C}) are merged as a consequence of the contraction of edge ({A} ,{C}), resulting in an edge ({D} ,{A, C}) marked in red. In this way, we enforce that any possible contraction in the new graph will keep agents A and D in separate coalitions. Having defined how we can use the edge contraction operation to generate feasi- ble coalition structures, we now provide a way to generate the whole search space of feasible coalition structures. 4.2. Generating the entire search space Given the green edge contraction operation defined above, we can generate each feasi- ble coalition structure only once. In more detail, at each point of the generation process, each red edge indicates that it has been discarded for contraction from that point on- wards, and hence its vertices cannot be joined. Observe that the way we defined green edge contraction guarantees that the information in red edges is always preserved. Thus, given a 2-coloured graph, its children can be readily assessed as follows: for each edge in the graph, we generate the graph that results from contracting that edge. Moreover, we colour the selected edge in red so that it cannot be contracted again in subsequent edge contractions. Algorithm 1 implements the depth-first3 generation and traversal of our search tree, in which each feasible coalition structure is evaluated by means of the characteristic function and compared with the best (i.e., the one with the highest value) coalition structure so far, hence computing the optimal solution. 3The DFS strategy allows us to traverse the entire tree with polynomial memory requirements, since at each stage of the search we only need to store the ancestors of the current node. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:9 {A} {B} {D} {C} {A, B} {D} {C} {B} {A, D} {C} {A} {B, C} {D} {A} {B} {C, D} {A, B} {C, D} {A, B, D} {C} {B, C} {A, D} {B, C, D} {A} {A, C, D} {B} {A, B, C} {D} {A, B, C, D} Fig. 3. Search tree for a square graph. As an example, Figure 3 shows the search tree generated starting from a square graph, highlighting each generation step with labels on the edges. We now prove that Algo- rithm 1 visits all feasible coalition structures and each of them is visited only once. PROPOSITION 4.3. Given Gc, the tree generated by Algorithm 1 rooted at Gc con- tains all the coalition structures compatible with Gc, each appearing only once. SKETCH OF PROOF. By induction on the number of green edges. Full proof is pro- vided in the Online Appendix. PROPOSITION 4.4. The complexity of Algorithm 1 is O(CS(G) · E). PROOF. There is a bijection between CS(G) and the nodes visited by Algorithm 1, by direct application of Proposition 4.3 to G with all green edges. The creation of each new node yields a GREENEDGECONTRACTION(G, e) operation, whose complexity is O(E) (Definition 3). Hence, the complexity of creating the search tree is O(CS(G) · E).4 Notice that, even for sparse graphs, the number of feasible coalition structures can be very large, as, in general, the GCCF problem is NP-complete [Voice et al. 2012a]. Hence, in the next section we propose a branch and bound technique that helps prune significant parts of the search space, allowing us to compute the optimal solution for any GCCF problem based on an m + a function by generating only a minimal portion of the solution space (i.e., less than 0.32% in our experiments in Section 7.2). In addition, such a bounding technique is employed in the approximate version of our approach, which can compute solutions with quality guarantees for large-scale systems. It is important to note that, in contrast with the optimal version, our approx- imate approach is not characterised by the above discussed exponential complexity, as the search for the solution is executed only for a given time budget (see Section 5.2). 5. A GENERAL BRANCH AND BOUND ALGORITHM FOR m + a FUNCTIONS We now describe CFSS (Coalition Formation for Sparse Synergies), our branch and bound approach to GCCF when applied to the family of m + a characteristic functions. DEFINITION 4. Given a graph G, a function v : FC(G) → R is superadditive (resp. subadditive) if the value of the union of disjoint coalitions is no less (resp. no greater) than the sum of the coalitions' separate values, i.e., v(S ∪ T ) ≥ (resp. ≤) v(S) + v(T ) for all S, T ⊆ A such that S ∩ T = ∅. 4Notice that, since Coalition Structure Generation (CSG) is a particular case of GCCF (i.e., CSG is a GCCF problem with a complete graph), CS(G) can be, in the worst case, equivalent to the nth Bell number, i.e., ln(n) )n) [Berend and Tassa 2010], where n is the number of agents. Nonetheless, in the problems we Ω(( n consider G is sparse and, hence, CS(G) contains a lower number of feasible coalition structures. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:10 Filippo Bistaffa et al. We also define such properties for the function V : CS(G) → R defined in Equation 1. DEFINITION 5. Given a graph G, a function V : CS(G) → R defined according to Equation 1 is superadditive (resp. subadditive) if the underlying function v : FC(G) → R is superadditive (resp. subadditive). DEFINITION 6. Given a graph G, V : CS(G)→R is an m+a function if it is the sum of a superadditive (i.e., monotonic increasing) function V + : CS(G)→R and a subadditive (i.e., antimonotonic) function V − : CS(G)→R. This family is interesting because it allows us to provide an upper bound that underlies our branch and bound strategy, so as to prune significant portions of the search space and have a computationally affordable solution algorithm. We provide a technique to compute an upper bound for the value assumed by the characteristic function in every coalition structure of the subtree ST (CSi) rooted at a given coalition structure CSi. In order to explain how to compute such an upper bound, we first define the element CSi. DEFINITION 7. Given a feasible coalition structure CSi represented by a 2-coloured graph Gc, we define CSi as the coalition structure obtained by removing all red edges from Gc and then contracting all the remaining green edges. Intuitively, CSi represents the connected components in the graph after the removal of all red edges. THEOREM 5.1. Given an m+a function V : CS(G) → R, then M (CSi) = V − (CSi)+ V +(cid:0)CSi SKETCH OF PROOF. V − (CSi) (resp. V +(cid:0)CSi (cid:1) is an upper bound for the value assumed by such function in every coalition M (CSi) = V − (CSi) + V +(cid:0)CSi (cid:1)) is an upper bound for the subaddi- (cid:1) ≥ max{V (CSj) CSj ∈ ST (CSi)}. structure of the subtree ST (CSi) rooted at CSi, i.e., (2) tive (resp. superadditive) component, hence M (CSi) is an upper bound for the charac- teristic function. Full proof is provided in the Online Appendix. Remark 5.2. Given CSi represented by a 2-coloured graph Gc = (V,F, c), it is pos- sible to compute a more precise upper bound for the edge sum with coordination cost function (see Section 6.1.2) by replacing V +(cid:0)CSi (cid:1) with(cid:80) e∈F :c(e)=green w+(e). Building upon Theorem 5.1, we can efficiently assess an upper bound for the value of the characteristic function in any subtree and prune it, if such a value is smaller than the value of the best solution found so far. Algorithm 3 implements CFSS, our branch and bound approach to solve the GCCF problem. We remark that Algorithm 3 is correct and complete, i.e., it computes the optimal solution regardless of the order in which the children of the current node are visited, namely the operation of the CHILDREN function. However, such an order has a strong influence on the performance of CFSS (as shown in Section 7.3), since it can be used Algorithm 3 CFSS(Gc) 1: best ← Gc, F ← ∅ 2: F.PUSH(Gc) 3: while F (cid:54)= ∅ do 4: 5: 6: 7: 8: return best node ← F.POP() if M (node) > V (best) then if V (node) > V (best) then best ← node F.PUSH(CHILDREN (node)) (cid:46) Initialise solution with singletons and search frontier F with empty stack (cid:46) Push Gc as the first node to visit (cid:46) Branch and bound loop (cid:46) Get current node (cid:46) Check bound value (cid:46) Update current best solution (cid:46) Update frontier F (cid:46) Return optimal solution ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:11 Si C E I C1 F D J H C2 B A G Fig. 4. Example of a partition with a cut-set of 3 edges. to compute an upper bound that better resembles the characteristic function (hence improving the effectiveness of the branch and bound pruning). 5.1. Edge ordering heuristic In this section we propose a heuristic to define a total ordering among the edges of a graph G, in order to guide the traversal of the search tree. This results in a significant speed-up of the algorithm, by means of an improvement of the upper bound. In partic- ular, we notice that the value of M (CSi) = V − (CSi) + V +(cid:0)CSi by the value of V +(cid:0)CSi (cid:1) is heavily influenced (cid:1). In fact, it is possible that CSi = {A} (i.e., the grand coali- tion), when CSi contains enough green edges to connect all the nodes of the graph G. This results in a poor bound, since V + is a superadditive function and it reaches its maximum value for A. On the other hand, if red edges form a cut-set for the 2-coloured graph, the procedure in Definition 7 results in a coalition structure CSi = {C1, C2}, as Fig- ure 4 shows. In this case, our bounding technique produces a lower upper bound M (CSi) = V − (CSi) + v+ (C1) + v+ (C2), since v+ (·) is superadditive and, therefore, v+ (C1) + v+ (C2) ≤ v+ (A) . Notice that, having an upper bound that provides a lower overestimation of the characteristic function is crucial for the performance of CFSS, as the condition at line 5 in Algorithm 3 would be verified less often, hence allowing us to prune bigger portions of the search space. Also, it easy to see that when the value of the characteristic function increases in a non-linear way with respect to the size of the coalitions (such as the functions we consider in this paper), the more C1 and C2 are closer to a bisection of A (i.e., the more C1 and C2 are close to A/2), the more pronounced such improvement is. Following this observation, it is preferable to visit the edges that produce a cut of the graph in the first steps of the algorithm, since they will result in the above-explained improvement once such edges are marked in red. Henceforth, we define a total order- ing among the edges of G, producing an ordered graph Go by means of Algorithm 4. Intuitively, such algorithm computes small5 cut-sets by means of the routine CUT(G), which outputs the subgraphs G1 = (V1,F1) and G2 = (V2,F2) resulting from the cut, and the cut-set F(cid:48). Once the cut-set has been found, we label its edges as the first ones in the ordered graph, recursively applying such procedure for all the subsequent subgraphs resulting at each partitioning, until every edge has been ordered. Remark 5.3. In the worst-case, Algorithm 4 makes E calls to CUT, whose complex- ity is O(E) [Karypis and Kumar 1998]. Hence, its worst-case complexity is O(E2). In addition to this edge ordering heuristic, our bounding technique can be employed to provide anytime approximate solutions, as shown in the next section. 5To traverse the minimum number of edges necessary to partition the graph, we need the smallest cut-set. Unfortunately, such a problem (known as the Minimum Bisection problem) is a well known NP-complete problem [Garey and Johnson 1990]. However, our heuristic does not need an optimal solution, since if a suboptimal cut-set (i.e., bigger than the optimal one) is used, our algorithm will still partition the graph in a higher number of steps, resulting in a slightly smaller improvement. Therefore, we adopt an approximate algorithm implemented with the METIS graph partitioning library [Karypis and Kumar 1998]. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:12 Filippo Bistaffa et al. Algorithm 4 ORDER(G) 1: i ← 1, Go ← G, Q ← ∅ 2: Q.PUSH(G) 3: while Q (cid:54)= ∅ do 4: 5: 6: 7: 8: 9: 10: 11: return Go (cid:104)G1, G2, F(cid:48)(cid:105) ← CUT (Q.POP()) Label in Go each edge ∈ F(cid:48) from i to i + F(cid:48) − 1 i ← i + F(cid:48) if V1 > 1 then Q.PUSH(G1) if V2 > 1 then Q.PUSH(G2) (cid:46) Initialise edge counter, ordered graph, and empty queue (cid:46) Push G as the first graph to partition (cid:46) Partitioning loop (cid:46) Partition current graph (cid:46) Increase edge counter (cid:46) If the first subgraph has at least 2 nodes... (cid:46) ... enqueue it (cid:46) If the second subgraph has at least 2 nodes... (cid:46) ... enqueue it (cid:46) Return ordered graph 5.2. Anytime approximate properties Theorem 5.1 can be directly applied to compute an overall bound of an m + a function, with anytime properties. More precisely, let us consider frontier F in Algorithm 3. When we expand frontier F (Line 9) we keep track of the highest value of V (·) in the visited nodes. Hence, given a frontier F , the bound B(F ) is defined as B(F ) = max{V (best) , max CS∈F (3) Thus, B(F ) is the maximum between the values assumed by V (·) inside the frontier (i.e., V (best)) and an estimated upper bound outside of it (i.e., maxCS∈F M (CS)). No- tice that since each M (CS) is an overestimation of the value of V (·) in the correspond- ing subtree, such a maximisation provides a valid upper bound for V (·) in the portion of search space not visited yet. Furthermore, the quality of B(F ) can only be improved by expanding frontier F . More formally, if F (cid:48) is such an expansion, then M (CS)} B (F ) ≥ B (F (cid:48)) ≥ max{V (CS) CS ∈ CS (G)}. (4) This can be easily verified using the definition of M (·). In fact, each bound resulting from the children of a substituted node u ∈ F must be less or equal to M (u) and, hence, Inequality 4 holds. Intuitively, the larger the search space explored, the better is the bound provided. Finally, notice that the fastest way to compute a bound for V (·) is to consider a frontier formed exclusively by the root (i.e., the coalition structure formed by all singletons). Assessing this bound has the same time complexity of computing M, i.e., O(E), and its quality can be satisfactory, as shown in Section 7.4. After the discussion of our branch and bound algorithm for m + a functions, in the next section we discuss some scenarios in which GCCF can be applied, and, in partic- ular, we present three m + a functions that will be used to evaluate our approach. 6. APPLICATIONS FOR GCCF As previously discussed, GCCF is a well known model in cooperative game theory that can be applied to several realistic scenarios. In what follows, we focus on two real- world scenarios, namely social ridesharing and collective energy purchasing, that can be modelled as GCCF problems. In the ridesharing domain, Ma et al. [2013] adopted an heuristic approach in order to increase the potential passenger coverage of a fleet of taxis, while decreasing the total travel mileage of the system. Later on, Bistaffa et al. [2015] tackled the optimisa- tion problem of arranging one-time shared rides among a set of commuters connected through a social network, with the objective of minimising the overall travel cost. Un- like Ma et al. [2013], Bistaffa et al. [2015] explicitly consider coalitions, showing that such a scenario can be modelled as a GCCF problem where the set of feasible coalitions is restricted by the social network. Intuitively, each group of agents that travel in the ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:13 same car is mapped to a feasible coalition, whose coalitional value is defined as the total travel cost of that particular car, i.e., the cost of driving through its passengers' pick-up and destination points. Bistaffa et al. [2015] show that the adoption of the GCCF model in this scenario leads to a cost reduction of up to −36.22% when applied to realistic datasets for both spatial and social data. In the collective energy purchasing scenario [Vinyals et al. 2012], each agent is char- acterised by an energy consumption profile that represents its energy consumption throughout a day. A profile records the energy consumption of a household at fixed intervals (every half hour in our case). The characteristic function of a coalition of agents is the total cost that the group would incur if they bought energy as a collec- tive in two different markets: the spot market, a short term market (e.g., half hourly, hourly) intended for small amounts of energy; and the forward market, a long term one in which larger amounts of energy (spanning weeks and months) can be bought at cheaper prices [Vinyals et al. 2012]. In the edge sum with coordination cost scenario, every edge is associated to a value that represents how well (or bad) those agents perform together, or the cost of completing a coordination task in a robotic environ- ment [Dasgupta et al. 2012]. In the coalition size with distance cost scenario, the for- mation of coalitions favours bigger groups and maximises the similarity of the opinion among their members. Such application could be employed to cluster public opinion, or to detect the presence of "virtual coalitions" among members of a parliament based on their recorded votes (e.g., the votes by the Democratic and the Republican parties). In addition to such practical motivations, these three scenarios are particularly in- teresting as they are modelled by characteristic functions (Equation 5) part of a large family of functions, i.e., m + a functions. In what follows, we discuss the properties of such functions, showing how they can be exploited to significantly speed-up the solu- tion of the associated GCCF problem (see Section 5). 6.1. Benchmark m + a functions We now present three benchmark functions for GCCF, namely the collective energy purchasing function, the edge sum with coordination cost function and the coalition size with distance cost function. In particular, we are interested in their characterisa- tion as m + a functions, showing that they can be seen as the sums of the superadditive and the subadditive parts [Owen 1995]. Such characteristic functions are particularly interesting as they enable an efficient bounding technique to prune part of the search space during the execution of our branch and bound algorithm, presented in Section 5. 6.1.1. Collective energy purchasing. In the collective energy purchasing scenario, energy(C) where T = 48 is the number of energy measurements in each profile, pS ∈ R− and pF ∈ R− represent the unit prices of energy in the spot and forward market respectively, qF : FC(G) → R− stands for the time unit amount of electricity to buy in the forward S : FC(G) → R− for the amount to buy in the spot market at time slot t.6 market and qt energy : FC(G) → R− represents the total energy cost. 6Unit prices (whose values are reported in Section 7) are negative numbers, i.e., they belong to the set R− = {i ∈ R i ≤ 0}, to reflect the direction of payments. Thus, the values of the characteristic function are negative as well, hence they represent costs that, maintaining the maximisation task, we aim to minimise. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Farinelli et al. [2013] proposed the characteristic function (cid:125) S (C) · pS + T · qF (C) · pF qt (cid:123)(cid:122) v (C) = (cid:88)T (cid:124) t=1 +κ (C) , A:14 Filippo Bistaffa et al. Finally, κ : FC(G) → R− stands for a coalition management cost that depends on the size of the coalition and captures the intuition that larger coalitions are harder to manage. The definition of this cost depends on several low level issues (e.g., the capacity of the power networks connecting the customers in the groups, legal fees, and other costs associated to group contracts, etc.), hence a precise definition of this term goes beyond the scope of this paper. Following Farinelli et al. [2013] we use κ(C) = −Cγ with γ > 1 to introduce a non-linear element that penalises the formation of larger coalitions. Hence, the collective energy purchasing function is defined as (cid:20)(cid:88)T (cid:88) (cid:124) V (CS) = C∈CS t=1 (cid:123)(cid:122) S (C) · pS + T · qF (C) · pF qt V +(CS) κ (C) . (5) (cid:123)(cid:122) C∈CS V −(CS) (cid:125) (cid:21) (cid:125) (cid:88) (cid:124) + PROPOSITION 6.1. The collective energy purchasing function is m + a. SKETCH OF PROOF. The cost of the energy necessary to fulfil the aggregated con- sumption profiles of the coalitions, i.e., V + (CS), is clearly superadditive, while the sum of the coalition management costs, i.e., V − (CS), is subadditive, as they increase when coalition sizes increase. Full proof is provided in the Online Appendix. 6.1.2. Edge sum with coordination cost. In the edge sum with coordination cost function every edge of G is mapped to a real value by a function w : E → R [Deng and Papadim- itriou 1994]. Each coalitional value is the sum of the weights of the edges among its members. In order to have a better description of the management and communication costs in larger coalitions, we also introduce a penalising factor κ (C),7 with the same definition given in the previous section. Hence, we define this function as v (C) = (6) where the function edges : FC(G) → 2E provides the set of all the edges connecting any two members of a given coalition C, i.e., edges (C) = {(v1, v2) ∈ E v1 ∈ C and v2 ∈ C}. In order to characterise this scenario with an m + a function, we rewrite Equation 6 as w(e) + κ (C) , e∈edges(C) (cid:88) (cid:2)w+(e) + w−(e)(cid:3) + κ (C) , v (C) = (cid:88) (cid:26)w(e), if w(e) ≥ 0, e∈edges(C) 0, (cid:26)w(e), if w(e) < 0, where w+(e) = otherwise, w−(e) = In other words,(cid:80) edges in edges (C), while (cid:80) e∈edges(C) w+(e) represents the sum of all the positive weights of the e∈edges(C) w−(e) represents the sum of the negative ones. (cid:20)(cid:88) (cid:88) (cid:123)(cid:122) (cid:124) The edge sum with coordination cost function is then defined as w−(e) + κ (C) (cid:88) (cid:124) (cid:20)(cid:88) otherwise. e∈edges(C) e∈edges(C) w+(e) + V (CS) = (cid:21) (cid:125) . (cid:21) (cid:125) (cid:123)(cid:122) C∈CS C∈CS 0, V +(CS) V −(CS) PROPOSITION 6.2. The edge sum with coordination cost function is m + a. 7Such penalising factor makes the edge sum with coordination cost function to violate the IDM property (cf. Section 2.1.4), therefore the approach proposed by Voice et al. [2012a] cannot be used. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:15 SKETCH OF PROOF. It is easy to verify that V + (CS), i.e., the sum of all positive edges, is superadditive, while the sum of the negative ones, i.e., V − (CS), is subaddi- tive. Full proof is provided in the Online Appendix. 6.1.3. Coalition size with distance cost. The coalition size with distance cost can be mod- elled evaluating each coalition C with the function (7) where α ≥ 1, and d : A × A → R+ is a function that measures the distance between the opinions of agent i and agent j. From Equation 7 it follows that the input of our problem has size N 2, where N is the total number of agents, since we must know the distances between each pair or agents. The coalition size with distance cost function of a coalition structure CS is then defined as (i,j)∈C×C d (i, j) , v (C) = Cα −(cid:88) (cid:88) (cid:124) (cid:123)(cid:122) C∈CS V +(CS) Cα + (cid:125) (cid:88) (cid:124) (cid:20) −(cid:88) (cid:123)(cid:122) V −(CS) V (CS) = C∈CS (i,j)∈C×C d (i, j) . (cid:21) (cid:125) PROPOSITION 6.3. The coalition size with distance cost function is m + a. PROOF. On the one hand, it is easy to verify that v+(C) = Cα is a superadditive (i,j)∈C×C d (i, j) is sub- function, assuming that α ≥ 1. On the other hand, v−(C) = −(cid:80) additive, since v−(C1∪C2) = v−(C1)+v−(C2)−(cid:80) d (i, j) ≤ v−(C1)+v−(C2). i∈C1,j∈C2 These functions will be used in our experimental evaluation in the next section. 7. EMPIRICAL EVALUATION The main goals of our empirical evaluation of CFSS are: (1) To evaluate its runtime performance with respect to DyCE considering a variety of graphs, both realistic (i.e., subgraphs of the Twitter network) and synthetic (i.e., scale-free networks). Additional experiments on community networks and a de- tailed discussion on these network topologies are in the Online Appendix. (2) To evaluate the effectiveness of our bounding technique. (3) To evaluate the anytime performance and guarantees that our approach can pro- vide when scaling to very large numbers of agents (i.e., more than 2700). (4) To compare the quality of our approximate solutions with the ones computed by C-Link [Farinelli et al. 2013] on large-scale instances. (5) To evaluate the speed-up that can be obtained by using multi-core machines. (6) To evaluate the speed-up produced by our edge ordering heuristic. Following Voice et al. [2012b], we consider scale-free networks generated with the Barab´asi-Albert model with m ∈ {1, 2, 3}. This parameter determines the sparsity of the graph, as every newly added node is connected, on average, to m existing nodes. It is easy to verify that the average degree of a scale-free network is ∼ 2 · m. We compare our approach with DyCE in our three reference domains, measuring the runtime in seconds. In our characteristic functions we use the following parameters: - Following Farinelli et al. [2013], in the collective energy purchasing function we set pS=−80 and pF =−70. The consumption data is provided by a realistic dataset, com- prising the measurements collected over a month from 2732 households in the UK. - In the edge sum with coordination cost function we assigned a uniformly distributed random weight within [−10, 10] to each edge. - Following Farinelli et al. [2013], in both the above scenarios we considered γ = 1.3. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:16 Filippo Bistaffa et al. - In the coalition size with distance cost function we assigned a uniformly distributed random value within [0, 100] to each distance between a pair of different agents (with d(i, i) = 0), and we considered α = 2.2, motivated by the remarks in Section 7.4. We conducted an additional set of experiments in which the graph G is a subgraph of a large crawl of the Twitter social graph. Specifically, such dataset is a graph with 41.6 million nodes and 1.4 billion edges published as part of the work by Kwak et al. [2010]. We obtain G by means of a standard algorithm [Russell 2013] to extract a subgraph from a larger graph, i.e., a breadth-first traversal starting from a random node of the whole graph, adding each node and the corresponding arcs to G, until the desired number of nodes is reached. Moreover, we implemented a multi-threaded version of CFSS, namely P-CFSS (i.e., Parallel CFSS), and we analysed the speed-up of P-CFSS using Amdahl's law [Amdahl 1967], as it provides the maximum theoretical speed-up that can be achieved. All our results refer to the average value over 20 repetitions for each experiment. CFSS8 and C-Link are implemented in C, while we used the DyCE implementation provided by its authors. We run our tests on a machine with a 3.40GHz CPU and 32 GB of memory. 7.1. DyCE vs CFSS: runtime comparison In our experiments using scale-free networks, CFSS outperforms DyCE when coalition values are shaped by the above-described benchmark functions (as shown in Figures 6a, 6b and 6c). Specifically, for the edge sum with coordination cost function, CFSS outperforms DyCE by 4 orders of magnitude on networks with average connectivity (i.e., for m = 2), and by 3 orders of magnitude on networks with higher connectivity (i.e., for m = 3). Most probably this is due to the fact that the upper bound we adopt in this case closely resembles the function, allowing us to prune significant portions of the search space (see Section 7.2 for a more detailed discussion). In the collective energy purchasing scenario with 30 agents and m = 2, CFSS is 4.7 times faster than DyCE, and it is at least 2 orders of magnitude faster for m = 1. However, DyCE is significantly faster (44 times) than CFSS for m = 3. The adoption of the coalition size with distance cost function produces a similar behaviour, with a performance improvement for our method. In fact, CFSS is 17 times faster than DyCE for m = 2, and only 3 times slower for m = 3. On the other hand, the runtime of DyCE equals the previous case, since this approach is not sensitive to the values of the characteristic function. In our tests using subgraphs of the Twitter network, CFSS is at least four orders of magnitude faster than DyCE when solving instances with 30 agents (the biggest instances that DyCE can solve), and it can scale up to 45 agents. These results confirm the very good performance of CFSS when considering sparse networks. In fact, the average degree of these subgraphs is comparable with the one of a scale-free network with 1 < m < 2. In all our tests, we increased the number of agents until the execution time reached 105 seconds. Notice that, in general, DyCE cannot scale over 30 agents (due to its exponential memory requirements), while CFSS does not have such limitation, hence it is possible to reach instances with thousands of agents, as shown in Section 7.4. CFSS (m = 1) DyCE (m = 1) CFSS (m = 2) DyCE (m = 2) CFSS (m = 3) DyCE (m = 3) Fig. 5. Legend for scale-free networks. 8Our implementation of CFSS is publicly available at https://github.com/filippobistaffa/CFSS. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:17 Fig. 6. Runtime to compute the optimal solution. 7.2. Bounding technique effectiveness Here we compare the number of configurations explored by CFSS w.r.t. the entire search space, i.e., the one explored by Algorithm 1, to measure of the number of search nodes pruned by our bounding technique. We consider n = 30, adopting scale-free net- works with m = 2. When the coalitional values are provided by the collective energy purchasing function, CFSS can compute the optimal solution exploring a number of configurations which is, on average, 0.32% of the entire search space. We measured a similar value in the coalition size with distance cost scenario (i.e., 0.28%). In the edge sum with coordination cost scenario (which allows a more precise upper bound, as explained in Remark 5.2), only 0.0045% of the entire search space is explored. 7.3. Edge ordering heuristic The above table shows the speed-up obtained by using the ordering heuristic described in Section 5.1 and considering the collective energy purchasing and the coalition size with distance cost functions. Even though our heuristic is applicable also in the edge sum with coordination cost scenario, such function has not been included in this anal- ysis since, as stated in Remark 5.2, it allows an ad-hoc bounding method that is more effective than the general one. Our experiments show a clear benefit in the adoption of such a heuristic, producing a maximum performance gain of 843% in the first scenario and 338% in the second one. Across all experimental scenarios, such a heuristic allows an average speed-up of 295% considering both domains. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. AlgorithmsforGraph-ConstrainedCoalitionFormationintheRealWorldA:1720253035404550556010−310−210−1100101102103104105106105slimitDyCElimitNumberofagentsExecutiontime(s)(a)Edgesumwithcoordinationcost,scale-freenetworks.152025303540455010−310−210−1100101102103104105106105slimitDyCElimitNumberofagentsExecutiontime(s)(b)Collectiveenergypurchasing,scale-freenetworks.152025303540455010−310−210−1100101102103104105106105slimitDyCElimitNumberofagentsExecutiontime(s)(c)Coalitionsizewithdistancecost,scale-freenetworks.1520253035404510−310−210−1100101102103104105106105slimitDyCElimitNumberofagentsExecutiontime(s)CFSSDyCE(d)Collectiveenergypurchasing,Twittersubgraphs.Fig.6:Runtimetocomputetheoptimalsolution.7.2.BoundingtechniqueeffectivenessHerewecomparethenumberofconfigurationsexploredbyCFSSw.r.t.theentiresearchspace,i.e.,theoneexploredbyAlgorithm1,tomeasureofthenumberofsearchnodesprunedbyourboundingtechnique.Weconsidern=30,adoptingscale-freenet-workswithm=2.Whenthecoalitionalvaluesareprovidedbythecollectiveenergypurchasingfunction,CFSScancomputetheoptimalsolutionexploringanumberofconfigurationswhichis,onaverage,0.32%oftheentiresearchspace.Wemeasuredasimilarvalueinthecoalitionsizewithdistancecostscenario(i.e.,0.28%).Intheedgesumwithcoordinationcostscenario(whichallowsamorepreciseupperbound,asexplainedinRemark5.2),only0.0045%oftheentiresearchspaceisexplored.7.3.EdgeorderingheuristicTheabovetableshowsthespeed-upobtainedbyusingtheorderingheuristicdescribedinSection5.1andconsideringthecollectiveenergypurchasingandthecoalitionsizewithdistancecostfunctions.Eventhoughourheuristicisapplicablealsointheedgesumwithcoordinationcostscenario,suchfunctionhasnotbeenincludedinthisanal-ysissince,asstatedinRemark5.2,itallowsanad-hocboundingmethodthatismoreeffectivethanthegeneralone.Ourexperimentsshowaclearbenefitintheadoptionofsuchaheuristic,producingamaximumperformancegainof843%inthefirstscenarioand338%inthesecondone.Acrossallexperimentalscenarios,suchaheuristicallowsanaveragespeed-upof295%consideringbothdomains.ACMTransactionsonIntelligentSystemsandTechnology,Vol.V,No.N,ArticleA,Publicationdate:JanuaryYYYY.A:18 Filippo Bistaffa et al. Characteristic function Collective energy purchasing Coalition size with distance cost Minimum Average Maximum 176% 136% 367% 222% 843% 338% 7.4. Anytime approximate performance We evaluate the performance of the approximate version of CFSS on instances with thousands of agents considering the Performance Ratio (PR) [Ausiello et al. 2012], a standard measure to evaluate approximate algorithms defined as the ratio between the approximate solution and the optimal one on a given instance I. As computing the optimal solution for such large instances is not possible, we define the Maximum Performance Ratio (MPR) as the ratio between the approximate solution and the upper bound on the optimal solution defined in Equation 3. DEFINITION 8. Given an instance I, an approximate solution Approx(I) and an upper bound on the optimal solution as Bound(I), we define the Maximum Performance Ratio M P R(I) = max . (cid:17) (cid:16) Approx(I) Bound(I) , Bound(I) Approx(I) M P R(I) represents an upper bound of the PR on the instance I. The MPR provides an important quality guarantee on the approximate solution Approx(I), since Approx(I) cannot be worse than by a factor of M P R(I) w.r.t. the optimal solution. 7.4.1. Collective energy purchasing. Figure 7a shows the value of the MPR in the col- lective energy purchasing scenario, using n ∈ {100, 500, 1000, 1500, 2000, 2732}, adopting scale-free networks with m = 4 and Twitter subgraphs as network topologies, and con- sidering a time budget of 100 seconds. Other values for m show a similar behaviour (not reported here). We plot the average and the standard error of the mean over 20 repetitions. It is clear that the network topology does not impact the quality guarantees of our approach, hence we only adopt scale-free networks in the following experiments. In contrast, the MPR is heavily influenced by the nature of the characteristic function, as clarified later in this section. In addition, the results show that, for 100 agents, the provided bound is only 4.7% higher than the solution found within the time limit, reaching a maximum of +11.65% when the entire dataset is considered, i.e., with 2732 agents. Such small decrease is due to the fact that, for bigger instances, it is possible to explore a smaller part of the search space in the considered time budget, leaving a bigger portion to the estimation of the bound. Nonetheless, in this experiment CFSS provides a MPR of at most 1.12 and thus solutions that are at least 88% of the optimal. This confirms the effectiveness of this bounding technique when applied to the energy domain, which allows us to provide solutions and quality guarantees for problems in- volving a very large number of agents. In our tests, the bound is assessed at the root, without any frontier expansion, so it can be computed almost instantly, thus devoting all the available runtime to the search for a solution. This choice is further motivated by the fact that, in this scenario, the bound improves of a negligible value in the first levels of the search tree, due to the particular definition of the characteristic function. More precisely, if we consider a frontier formed by the children of the root, in each of them the bound of V −(·) will improve by a factor of 2γ − 2 ≈ 1.5 (i.e., the difference between the coalition management cost of the new coalition and the ones of the two merged singletons). On the other hand, the bound of V +(·) will remain constant: in fact, since we are taking the maximum (i.e., the worst) bound at the frontier (as shown in Equation 3), the result of this maximisation will still be equal to v+(A), because in at least one of the children nodes the computation of CS will result in joining all the agents together. In this case, it is not worth to expand the frontier from the root, since the gain would be insignificant w.r.t. the additional computational cost. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:19 Fig. 7. Maximum Performance Ratio (MPR) in the considered domains. 7.4.2. Edge sum with coordination cost. We further evaluate the scalability of our ap- proach by considering Twitter subgraphs as network topologies, and the edge sum with coordination cost function, which allows to generate coalitional values for instances with any number of agents. Such a function can be either positive or negative (in con- trast with the collective energy purchasing one, which is always negative to represent its nature of cost). Hence, it is possible that Approx(I) is negative and Bound(I) is pos- itive, resulting in a negative MPR. In order to avoid this unreasonable behaviour, here we consider M P R(I) = Bound(I)−LB(I) Approx(I)−LB(I), where LB(I) is a lower bound on the charac- teristic function considering the instance I. Notice that it is always possible to compute LB(I) for the edge sum with coordination cost function as LB(I) = V −(A). Figure 7b shows that, on our machine, CFSS can scale up to instances with 30000 agents, providing solutions with a MPR of 1.127 (at least 89% of the optimal). C∈CS (cid:80) additive component (i.e., −(cid:80) (i.e.,(cid:80) 7.4.3. Coalition size with distance cost. The MPR exhibits a different behaviour when con- sidering the coalition size with distance cost function, being heavily influenced by the value of the α exponent. Figure 7c shows how the MPR varies significantly with respect to α ∈ [2, 3], growing up to 41825.6 for α = 2.4 and then falling down to 1.13 for α = 2.7, with a tendency to 1 when increasing this exponent. This behaviour can be explained by reasoning about the structure of the characteristic function. Up to α = 2.4, the sub- (i,j)∈C×C d (i, j)) dominates the superadditive one C∈CS Cα), hence the search for a solution is not able to find any coalition struc- ture better than the initial one (i.e., the coalition structure with all singletons, which is probably the optimal one). Nonetheless, the MPR keeps growing when we increase N = N α−1, i.e., the bound computed at the root (i.e., V +(A) = N α) α, since it equals N α divided by the value of the initial solution (i.e., N). On the other hand, when α is sufficiently large (i.e., for α = 2.5), this behaviour is inverted, because V +(·) has a greater impact and the entire characteristic function tends to become superadditive. Thus, coalition structures closer to the grand coalition represent good solutions, which explains why the MPR tends to 1 when we increase α. These remarks motivate us to study the impact of α also on the optimal algorithm. Figure 8 displays the runtime needed to find the optimal solution on random instances with 25 agents on scale-free networks with m = 2, showing that the performance of CFSS decreases when we in- crease α from 2 to 3. The value of the bound provided by Equation 2 is larger when α grows, hence its quality decreases, producing a less effective bounding technique and, thus, a higher runtime. To summarise, the adoption of a bigger α in the coalition size with distance cost function negatively impacts the performance of our approach when computing optimal solutions, while improving approximate solutions as α grows. This ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. AlgorithmsforGraph-ConstrainedCoalitionFormationintheRealWorldA:1910050010001500200027321.051.061.071.081.091.101.111.12NumberofagentsMPRScale-freenetworksTwittersubgraphs(a)Collectiveenergypurchasing.500010000150002000025000300001.1211.1221.1231.1241.1251.1261.1271.1281.1291.130NumberofagentsMPR(b)Edgesumwithcoordinationcost.10050010001500200022.22.42.62.83100101102103104105NumberofagentsαMPR100101102103104105(c)Coalitionsizewithdistancecost.Fig.7:MaximumPerformanceRatio(MPR)intheconsidereddomains.7.4.2.Edgesumwithcoordinationcost.Wefurtherevaluatethescalabilityofourap-proachbyconsideringTwittersubgraphsasnetworktopologies,andtheedgesumwithcoordinationcostfunction,whichallowstogeneratecoalitionalvaluesforinstanceswithanynumberofagents.Suchafunctioncanbeeitherpositiveornegative(incon-trastwiththecollectiveenergypurchasingone,whichisalwaysnegativetorepresentitsnatureofcost).Hence,itispossiblethatApprox(I)isnegativeandBound(I)ispos-itive,resultinginanegativeMPR.Inordertoavoidthisunreasonablebehaviour,hereweconsiderMPR(I)=Bound(I)−LB(I)Approx(I)−LB(I),whereLB(I)isalowerboundonthecharac-teristicfunctionconsideringtheinstanceI.NoticethatitisalwayspossibletocomputeLB(I)fortheedgesumwithcoordinationcostfunctionasLB(I)=V−(A).Figure7bshowsthat,onourmachine,CFSScanscaleuptoinstanceswith30000agents,providingsolutionswithaMPRof1.127(atleast89%oftheoptimal).7.4.3.Coalitionsizewithdistancecost.TheMPRexhibitsadifferentbehaviourwhencon-sideringthecoalitionsizewithdistancecostfunction,beingheavilyinfluencedbythevalueoftheαexponent.Figure7cshowshowtheMPRvariessignificantlywithrespecttoα∈[2,3],growingupto41825.6forα=2.4andthenfallingdownto1.13forα=2.7,withatendencyto1whenincreasingthisexponent.Thisbehaviourcanbeexplainedbyreasoningaboutthestructureofthecharacteristicfunction.Uptoα=2.4,thesub-additivecomponent(i.e.,−(cid:80)C∈CS(cid:80)(i,j)∈C×Cd(i,j))dominatesthesuperadditiveone(i.e.,(cid:80)C∈CSCα),hencethesearchforasolutionisnotabletofindanycoalitionstruc-turebetterthantheinitialone(i.e.,thecoalitionstructurewithallsingletons,whichisprobablytheoptimalone).Nonetheless,theMPRkeepsgrowingwhenweincreaseα,sinceitequalsNαN=Nα−1,i.e.,theboundcomputedattheroot(i.e.,V+(A)=Nα)dividedbythevalueoftheinitialsolution(i.e.,N).Ontheotherhand,whenαissufficientlylarge(i.e.,forα=2.5),thisbehaviourisinverted,becauseV+(·)hasagreaterimpactandtheentirecharacteristicfunctiontendstobecomesuperadditive.Thus,coalitionstructuresclosertothegrandcoalitionrepresentgoodsolutions,whichexplainswhytheMPRtendsto1whenweincreaseα.Theseremarksmotivateustostudytheimpactofαalsoontheoptimalalgorithm.Figure8displaystheruntimeneededtofindtheoptimalsolutiononrandominstanceswith25agentsonscale-freenetworkswithm=2,showingthattheperformanceofCFSSdecreaseswhenwein-creaseαfrom2to3.ThevalueoftheboundprovidedbyEquation2islargerwhenαgrows,henceitsqualitydecreases,producingalesseffectiveboundingtechniqueand,thus,ahigherruntime.Tosummarise,theadoptionofabiggerαinthecoalitionsizewithdistancecostfunctionnegativelyimpactstheperformanceofourapproachwhencomputingoptimalsolutions,whileimprovingapproximatesolutionsasαgrows.ThisACMTransactionsonIntelligentSystemsandTechnology,Vol.V,No.N,ArticleA,Publicationdate:JanuaryYYYY.A:20 ) s ( e m i t n o i t u c e x E 104 103 102 101 100 n o i t u l o s S S F C / n o i t u l o s k n i L C - 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 NαN 1.10 1.09 1.08 1.07 1.06 1.05 1.04 1.03 Filippo Bistaffa et al. Amdahl's Law (94%) CFSS 12 11 10 9 8 7 6 5 4 3 2 1 p u - d e e p S 100 500 1500 1000 2000 Number of agents 2732 4 6 8 10 12 14 16 18 20 22 24 Number of threads Fig. 8. Runtime w.r.t. α. Fig. 9. C-Link vs. CFSS. Fig. 10. Parallel speed-up. motivates our choice of defining α = 2.2 in the previous experiments, as it represents a good value to benchmark CFSS. In fact, it is big enough to avoid excessively low run- times in the optimal version, but it does not exceed the 2.4 boundary, beyond which the quality guarantees it provides are extremely good (i.e., the MPR tends to 1). 7.5. CFSS vs C-Link: solution quality comparison We further evaluate the approximate performance of CFSS by comparing it against C-Link [Farinelli et al. 2013], an heuristic approach to solve CSG based on hier- archical clustering. We chose C-Link among the other approaches discussed in Sec- tion 2.1.2 because it is the most recent one and it has also been tested using the collec- tive energy purchasing function by its authors. Here we adopt the same experimental setting discussed in the previous section, i.e., we consider scale-free networks with n∈{100, 500, 1000, 1500, 2000, 2732} and m=4 (generating 20 random repetitions of each experiment), and we adopt the collective energy purchasing characteristic function. We solve each instance with C-Link (adopting the best heuristic proposed by Farinelli et al. [2013], i.e., Gain-Link) and then we run CFSS on the same instance with a time budget equal to C-Link's runtime. Figure 9 shows the average and the standard error of the mean of the ratio between the value of the solution computed by C-Link and the one computed by CFSS. Since we consider solutions with negative values, when such ratio is > 1 the solution computed by C-Link is better (i.e., has a lower cost) than the one computed by CFSS. Our results show that, even though C-Link can compute bet- ter solutions, the quality of our solutions is worse only by 3% for 100 agents. When we consider the entire dataset (i.e., with 2732 agents) the quality of our solutions is still within the 9% w.r.t. the counterpart. Notice that C-Link slightly outperforms CFSS. This comes as no surprise since the fundamental difference between C-Link and CFSS is that C-Link does a backtrack-free visit of the search graph adopting a greedy heuris- tic to determine the choice at each step. In other words, C-Link explores only one path of the search graph. On the other hand, CFSS does not employ any heuristic as it is designed to execute a systematic visit of the search graph with backtracking. Notice that we can easily include the C-Link's greedy heuristic into CFSS to guide the visit of the children nodes in the search. With C-Link's heuristic, CFSS first explores the same path explored by C-Link, and then, if given more time, continues the visit of the rest of the search space by backtracking. Since we provide CFSS with a time budget equal to C-Link's runtime, if we employ C-Link's heuristic then CFSS effectively becomes the same algorithm as C-Link, and hence returns solutions of the same quality. 7.6. P-CFSS Here we detail the parallelisation approach of the multi-threaded version of CFSS, analysing the speed-up with respect to its serial version. Following Bader et al. [2005], parallelisation is achieved by having different threads searching different branches ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:21 rightmost subtrees are computed by a team of ta −(cid:80)i of the search tree. The only required synchronisation point is the computation of the current best solution that must be read and updated by every thread. In particular, the distribution of the computational burden among the ta available threads is done by considering the first i subtrees rooted in every node of the first generation (starting from the left) and assigning each of them to tj threads (1 ≤ j ≤ i). The remaining j=1 tj threads using a dynamic schedule.9 Parameters i and tj are arbitrarily set, since it is assumed (and verified by an empirical analysis) that the distribution of the nodes over the search tree does not significantly vary among different instances. More advanced techniques, such as estimating the number of nodes in the search tree as suggested by Lelis et al. [2013], will be considered in the future. We run P-CFSS on random instances with 27 agents on scale-free networks with m = 2, using a machine with 2 Intel R(cid:13) Xeon R(cid:13) E5-2420 processors. The speed-up measured during these tests has been compared with the maximum theoretical one provided by Amdahl's Law, considering an estimated non- parallelisable part of 6%, due to memory allocation and thread initialisation. As can be seen in Figure 10, the actual speed-up follows the theoretical one up to 12 threads, the number of physical cores. After that, hyper-threading still provides some improvement, reaching a final speed-up of 9.44 with all 24 threads active. 8. CONCLUSIONS In this paper we considered the GCCF problem and proposed a branch and bound solu- tion (the CFSS algorithm) that can be applied to a general class of functions (i.e., m + a functions). Our empirical evaluation shows that CFSS outperforms DyCE, the state of the art algorithm, when applied to three characteristic functions. Specifically, CFSS is at least 3 orders of magnitude faster than DyCE in the first scenario, while solving bigger instances for the remaining two. Moreover, the adoption of our edge ordering heuristic provides a further speed-up of 296%. P-CFSS, the parallel version of CFSS, achieves a speed-up of 944% on a 12-core machine, close to the maximum theoreti- cal speed-up. Finally, our algorithm provides approximate solutions with good quality guarantees (i.e., with a MPR of 1.12 in the worst case) for systems of unprecedented scale (i.e., more than 2700 agents). Overall, our work is the first to show how coalition formation techniques can start coping with real-world scenarios, opening the possi- bility of employing coalition formation on practical applications, rather than purely synthetic, small-scale environments. Future work will look at applying our approach to other realistic scenarios (e.g., the formation of team of experts connected by a social network [Lappas et al. 2009]) and focusing on different multi-threading models (e.g., GPUs). ELECTRONIC APPENDIX The electronic appendix for this article can be accessed in the ACM Digital Library. ACKNOWLEDGMENTS COR (TIN 2012-38876-C02-01), Collectiveware TIN 2015-66863-C2-1-R (MINECO/FEDER), and the Gener- alitat of Catalunya 2014-SGR-118 funded Cerquides and Rodr´ıguez-Aguilar. This work was also supported by the EPSRC-Funded ORCHID Project EP/I011587/1. REFERENCES Gene M. Amdahl. 1967. Validity of the single processor approach to achieving large scale computing capa- bilities. In American Federation of Information Processing Societies. 483–485. 9Once a thread has completed the computation of one subtree, it starts with one of the remaining ones. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. A:22 Filippo Bistaffa et al. Giorgio Ausiello, Pierluigi Crescenzi, Giorgio Gambosi, Viggo Kann, Alberto Marchetti-Spaccamela, and Marco Protasi. 2012. Complexity and approximation: Combinatorial optimization problems and their approximability properties. Springer. David A. Bader, William E. Hart, and Cynthia A. Phillips. 2005. Parallel Algorithm Design for Branch and Bound. In Emerging Methodologies and Applications in Operations Research. Springer, 5–44. Daniel Berend and Tamir Tassa. 2010. Improved bounds on Bell numbers and on moments of sums of ran- dom variables. Probability and Mathematical Statistics 30 (2010), 185–205. Filippo Bistaffa, Alessandro Farinelli, Jes ´us Cerquides, Juan Rodr´ıguez-Aguilar, and Sarvapali D. Ram- churn. 2014a. Anytime Coalition Structure Generation on Scale-Free and Community Networks. In International Joint Workshop on Optimisation in Multi-Agent Systems and Distributed Constraint Rea- soning. Filippo Bistaffa, Alessandro Farinelli, Jes ´us Cerquides, Juan Rodr´ıguez-Aguilar, and Sarvapali D. Ram- churn. 2014b. Anytime Coalition Structure Generation on Synergy Graphs. In International Conference on Autonomous Agents and Multi-Agent Systems. 13–20. Filippo Bistaffa, Alessandro Farinelli, and Sarvapali D. Ramchurn. 2015. Sharing Rides with Friends: a Coalition Formation Algorithm for Ridesharing. In AAAI Conference on Artificial Intelligence. 608–614. Viet Dung Dang and Nicholas R. Jennings. 2004. Generating coalition structures with finite bound from the optimal guarantees. In International Conference on Autonomous Agents and Multi-Agent Systems. 564–571. Prithviraj Dasgupta, Vladimir Ufimtsev, Carl Nelson, and S. G. M. Hossain. 2012. Dynamic reconfiguration in modular robots using graph partitioning-based coalitions. In International Conference on Autonomous Agents and Multi-Agent Systems. 121–128. Gabrielle Demange. 2004. On Group Stability in Hierarchies and Networks. Political Economy 112 (2004), 754–778. Xiaotie Deng and Christos H. Papadimitriou. 1994. On the Complexity of Cooperative Solution Concepts. Mathematics of Operations Research 19 (1994), 257–266. Nicola Di Mauro, Teresa M A. Basile, Stefano Ferilli, and Floriana Esposito. 2010. Coalition Structure Gen- eration with GRASP. In International Conference on Artificial Intelligence: Methodology, Systems, Ap- plications. 111–120. Daniela Dos Santos and Ana Bazzan. 2012. Distributed clustering for group formation and task allocation in Multi-Agent systems: A swarm intelligence approach. Applied Soft Computing 12 (2012), 2123–2131. Alessandro Farinelli, Manuele Bicego, Sarvapali Ramchurn, and Mauro Zucchelli. 2013. C-link: A Hier- archical Clustering Approach to Large-scale Near-optimal Coalition Formation. In International Joint Conference on Artificial Intelligence. 106–112. Michael R. Garey and David S. Johnson. 1990. Computers and Intractability: A Guide to the Theory of NP- Completeness. W. H. Freeman & Co. Matthew E. Gaston and Marie desJardins. 2005. Agent-organized Networks for Dynamic Team Formation. In International Conference on Autonomous Agents and Multi-Agent Systems. 230–237. Hampshire County Council. 2014. Switch Hampshire. (2014). http://www3.hants.gov.uk/switch Atsushi Iwasaki, Suguru Ueda, Naoyuki Hashimoto, and Makoto Yokoo. 2015. Finding core for coalition structure utilizing dual solution. Artificial Intelligence 222 (2015), 49–66. David R. Karger. 1993. Global Min-cuts in RNC, and Other Ramifications of a Simple Min-out Algorithm. In ACM-SIAM Symposium on Discrete Algorithms. 21–30. George Karypis and Vipin Kumar. 1998. A Fast and High Quality Multilevel Scheme for Partitioning Irreg- ular Graphs. SIAM Journal on Scientific Computing 20 (1998), 359–392. Helena Keinanen. 2009. Simulated Annealing for Multi-Agent Coalition Formation. In Agent and Multi- Agent Systems: Technologies and Applications. 30–39. Haewoon Kwak, Changhyun Lee, Hosung Park, and Sue Moon. 2010. What is Twitter, a Social Network or a News Media?. In World Wide Web. 591–600. Theodoros Lappas, Kun Liu, and Evimaria Terzi. 2009. Finding a team of experts in social networks. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 467–476. Levi H. S. Lelis, Lars Otten, and Rina Dechter. 2013. Predicting the Size of Depth-First Branch and Bound Search Trees. In International Joint Conference on Artificial Intelligence. 594–600. Somchaya Liemhetcharat and Manuela Veloso. 2014. Weighted synergy graphs for effective team formation with heterogeneous ad hoc agents. Artificial Intelligence 208 (2014), 41–65. Shuo Ma, Yu Zheng, and Ouri Wolfson. 2013. T-share: A large-scale dynamic taxi ridesharing service. In International Conference on Data Engineering. 410–421. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY. Algorithms for Graph-Constrained Coalition Formation in the Real World A:23 Leandro Soriano Marcolino, Albert Xin Jiang, and Milind Tambe. 2013. Multi-agent Team Formation: Di- versity Beats Strength?. In International Joint Conference on Artificial Intelligence. 279–285. Reshef Meir, Yair Zick, and Jeffrey S Rosenschein. 2012. Optimization and stability in games with restricted interactions. In Workshop on Cooperative Games in Multi-Agent Systems. Roger B. Myerson. 1977. Graphs and Cooperation in Games. Mathematics of Operations Research 2 (1977), 225–229. Hariharan Narayanan. 1997. Submodular functions and electrical networks. Elsevier. George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. 1978. An analysis of approximations for maximizing submodular set functions. Mathematical Programming 14 (1978), 265–294. Naoki Ohta, Vincent Conitzer, Ryo Ichimura, Yuko Sakurai, Atsushi Iwasaki, and Makoto Yokoo. 2009. Coalition structure generation utilizing compact characteristic function representations. In Principles and Practice of Constraint Programming. 623–638. Guillermo Owen. 1995. Game Theory. Academic Press. Talal Rahwan and Nicholas. R. Jennings. 2008a. Coalition Structure Generation: Dynamic Programming Meets Anytime Optimisation. In AAAI Conference on Artificial Intelligence. 156–161. Talal Rahwan and Nicholas R. Jennings. 2008b. An improved dynamic programming algorithm for coali- tion structure generation. In International Conference on Autonomous Agents and Multi-Agent Systems. 1417–1420. Talal Rahwan, Tomasz P. Michalak, Edith Elkind, Piotr Faliszewski, Jacek Sroka, Michael Wooldridge, and Nicholas R Jennings. 2011. Constrained Coalition Formation. In AAAI Conference on Artificial Intelli- gence. 719–725. Talal Rahwan, Tomasz P. Michalak, and Nicholas R. Jennings. 2012. A hybrid algorithm for coalition struc- ture generation. In AAAI Conference on Artificial Intelligence. 1443–1449. Talal Rahwan, Tomasz P Michalak, Michael Wooldridge, and Nicholas R Jennings. 2015. Coalition structure generation: A survey. Artificial Intelligence 229 (2015), 139–174. Talal Rahwan, Sarvapali Ramchurn, Nicholas Jennings, and Andrea Giovannucci. 2009. An anytime algo- rithm for optimal coalition structure generation. Journal of Artificial Intelligence Research 34 (2009), 521–567. Matthew A. Russell. 2013. Mining the Social Web. O'Reilly Media. Tuomas Sandholm, Kate Larson, Martin Andersson, Onn Shehory, and Fernando Tohm´e. 1999. Coalition structure generation with worst case guarantees. Artificial Intelligence 111 (1999), 209–238. Alexander Schrijver. 2003. Combinatorial optimization: polyhedra and efficiency. Springer. Sandip Sen and Partha S. Dutta. 2000. Searching for optimal coalition structures. In International Confer- ence on Autonomous Agents and Multi-Agent Systems. 287–292. Onn Shehory and Sarit Kraus. 1998. Methods for task allocation via agent coalition formation. Artificial Intelligence 101 (1998), 165–200. Alexander Shekhovtsov. 2006. Supermodular decomposition of structural labeling problem. Control systems and computers 1 (2006), 20. Alexander Shekhovtsov, Vladimir Kolmogorov, Pushmeet Kohli, V´aclav Hlav´ac, Carsten Rother, and Philip Torr. 2008. LP-relaxation of binarized energy minimization. Technical Report. Czech Tech. University. Oskar Skibski, Tomasz P. Michalak, Talal Rahwan, and Michael Wooldridge. 2014. Algorithms for the Shap- ley and Myerson Values in Graph-restricted Games. In International Conference on Autonomous Agents and Multi-Agent Systems. 197–204. Long Tran-Thanh, Tri-Dung Nguyen, Talal Rahwan, Alex Rogers, and Nicholas R Jennings. 2013. An effi- cient vector-based representation for coalitional games. In International Joint Conference on Artificial Intelligence. 383–389. Suguru Ueda, Makoto Kitaki, Atsushi Iwasaki, and Makoto Yokoo. 2011. Concise characteristic function representations in coalitional games based on agent types. In International Conference on Autonomous Agents and Multi-Agent Systems. 1271–1272. Meritxell Vinyals, Filippo Bistaffa, Alessandro Farinelli, and Alex Rogers. 2012. Coalitional energy purchas- ing in the smart grid. In IEEE International Energy Conference. 848–853. Thomas Voice, Maria Polukarov, and Nicholas R. Jennings. 2012a. Coalition structure generation over graphs. Journal of Artificial Intelligence Research 45 (2012), 165–196. Thomas Voice, Sarvapali D. Ramchurn, and Nicholas R. Jennings. 2012b. On coalition formation with sparse synergies. In International Conference on Autonomous Agents and Multi-Agent Systems. 223–230. D. Yun Yeh. 1986. A dynamic programming approach to the complete set partitioning problem. BIT Numer- ical Mathematics 26 (1986), 467–474. ACM Transactions on Intelligent Systems and Technology, Vol. V, No. N, Article A, Publication date: January YYYY.
1903.09259	1	1903	2019-03-21T22:14:46	Local Interactions for Cohesive Flexible Swarms	[ "cs.MA" ]	Distributed gathering algorithms aim to achieve complete visibility graphs via a "never lose a neighbour" policy. We suggest a method to maintain connected graph topologies, while reducing the number of effective edges in the graph to order n. This allows to achieve different goals and swarming behaviours: the system remains connected but flexible, hence can maneuver in environments that are replete with obstacles and narrow passages, etc.	cs.MA	cs	Local Interactions for Cohesive Flexible Swarms Rotem Manor1, Ariel Barel2, and Alfred M. Bruckstein3 Center for Intelligent Systems (CIS) Computer Science Department Technion, Haifa 32000, Israel. 9 1 0 2 r a M 1 2 ] A M . s c [ 1 v 9 5 2 9 0 . 3 0 9 1 : v i X r a Abstract -- Distributed gathering algorithms aim to achieve complete visibility graphs via a "never lose a neighbour" policy. We suggest a method to maintain connected graph topologies, while reducing the number of effective edges in the graph to order n. This allows to achieve different goals and swarming behaviours: the system remains connected but flexible, hence can maneuver in environments that are replete with obstacles and narrow passages, etc. INTRODUCTION We present a novel method for distributed control of multi-agent systems, which ensures maintaining flexible and connected spatial constellations. Such a requirement is needed in the context of many tasks that swarms of agents must carry out, such as mapping unknown environments, and deploying a connected network of sensors to cover unknown regions. The agents are assumed to be identical, anonymous (i.e. indistinguishable) and simple in the sense of having little or no memory (i.e. oblivious), with limited computation and sensing capabilities. In this paper we assume limited sensing range, denoted by V . The swarm remains connected, never splits into disjoint groups, hence remains able to perform tasks cohesively. We consider systems comprising mobile agents that interact solely by adjusting their motion according to the relative location of their neighbours. The agents are assumed capable of sensing the presence of other agents within the given sensing range. The agents then implement rules of motion based on information on the geometric constellation of their neighbours. The motion of agents is designed to ensure global connectivity without attempting to maintain all existing visibility connections between agents. Therefore, we obtain swarms that are cohesive but not rigid, allowing the agents to move more freely and assume various desirable formations, for example in order to pass through narrow passages between obstacles, and reduce communication loads when agents must act as a backbone network for communication, etc. Our algorithm is truly distributed, hence insensitive to the size of the swarm, since each agent carries out simple calculations and makes local individual decisions, rendering it suitable for swarms with very large numbers of agents. set of edges (representing connections between the agents). The neighbourhood set of a vertex is the set of vertices Interactions between agents in multi-agent systems are often mathematically described using a graph, commonly labelled as G(V,E), where V = {ν1, ν2, ..., νn} is the set of vertices (representing the agents), and E ⊆V×V is the connected to it, i.e. Niá{νj∈V{νi, νj}∈E}. the plane R2), i.e. on the set{piá(xi, yi)}i=1,2,3,.... In the V , and the interconnection graph G(V,E) is the visibility G(V,E), the setE is defined as follows: ei,j∈E ⇐⇒ pi− pj≤ V The connection between the agents is either fixed and known in advance, or it is defined based on geometric relationships between the agents' positions (in space, say in limited visibility case, agents are visible to each other if their mutual distance does not exceed the visibility range limit graph corresponding to the agents' locations. In this case, in A fundamental issue in a multi-agent system is to maintain cohesiveness, i.e. the interconnection graph corresponding to the system configuration, must stay connected, otherwise the system splits into disjoint parts. For a graph to be connected, there must always be a path between each pair of nodes in the graph. In a recent survey on multi-agent geometric consensus [1], different types of distributed gathering algorithms are presented in detail, where gathering agents with limited vis- ibility is addressed, as discussed in [2], [3], [4], [5], [6], [7], [8], [9]. Other recent articles e.g. [10], [11], [12], [13], [14] also deal with the consensus problem in distributed multi- agent systems. A common principle in the distributed control algorithms dealing with gathering, rendezvous, clustering, and aggregation of multi-agent systems is the "never lose a neighbour" policy, requiring: ∀i and ∆t≥ 0 ∶ j∈ Ni(t)⇒ j∈ Ni(t+ ∆t) that is, once agents become neighbours, they must remain neighbours forever, hence the number of edges in the interconnection graph never decreases. Local behaviours for the agents that enforce this policy, with additional bias aiming to gradually add more edges to the graph, are often sufficient for the graph to become complete in finite expected time. In the limited visibility case, such systems between agents i and j. Coordination between the agents is already a feature built into this method because the decision of two agents whose connection is not necessary to maintain connectivity is symmetrical, that is, if a "bridging" agent k exists, we have that j∉ N e i ⇐⇒ i∉ N e j . is an agent k inside the intersection lens of these circles, it is necessarily Fig. 1. Two circles of radiuspi− pj< V centered at pi and pj. If there closer to i than j and closer to j than i. Hence in this case j ∈ Ni but i , and i∈ Nj but i∉ N e j∉ N e Define Ge as G(V,E e), i.e. Ge is a subgraph of the visibility graph G whose set of edgesE e is defined by (1). j . An example of a neighbourhood graph vs. effective neigh- bourhood graph is shown in Figure 2, where both solid and dashed lines belong to the neighbourhood graph, but only the solid lines belong to the effective neighbourhood graph. In this example only half of the edges remain after reducing unnecessary edges due to rule (1). necessarily gather to a region with "diameter" less than V . Hence, the "never lose a neighbour" policy is optimally suited for the purpose of gathering, since increasing the number of connected agents leads to clustering of agents to a small, visibility-horizon-defined area. However, for connectivity maintenance only, the "never lose a neighbour" policy is certainly too conservative and stringent: maintaining connectivity only requires a connected neighbourhood graph, rather than a complete graph, and this goal may be achieved if some existing mutually visible agents cease to be defined as neighbours. The conservative policy is quite popular in the multi- agent distributed control field since the lack of coordination between agents' decisions, may, in general, cause the graph to become disconnected. We here propose a local method for maintaining a connected graph in distributed multi-agent systems, which ensures, along with preserving the graph connectivity, that the number of edges does not exceed 3n, where n=V is the number of agents in the system. The proposed method maintains graph connectivity withE≤ 3n as compared toE= n(n−1) in the complete graph. 2 Why can such a generic method be useful? A swarm in which the constraints that apply to agents are fewer is more flexible, so it can be steered through regions full of obstacles and narrow passages. A swarm that is connected but the number of neighbourhood defining edges in the graph is small, may self-organise in more interesting and useful constellations. A swarm in which the connections are determined according to the proposed algorithm can be spread and stretched so that the distance between its agents can attain distance of the order of (n− 1)V , i.e. in some cases it can even be arranged to form a line. Hence we envisage that the proposed method can be a useful basic process in designing multi-agent swarming algorithms that are usually called upon to ensure cohesiveness in various cases of controlled swarm actions, where, for example, the swarm is guided by leaders or by some exogenous broadcast controls. Following THE COHESION ENSURING PROCESS on Toussaint's work relative neighbourhood graphs (RNG) the [15], we define N e i , "effective neighbourhood" of agent i, as a subset of Ni: (1980) ∀j∈ Ni i ⇐⇒ ࢜k∶ ∶ j∈ N e pi− pk<pi− pj &pj− pk<pi− pj Fig. 2. In this example, the original neighbourhood graph on 20 nodes, contain 42 edges between every pair of agents distanced V or less (both solid and dashed lines). In the effective neighbourhood graph, 21 edges that are unnecessary for connectivity maintenance due to rule (1) are shown dashed, and only 21 edges, shown in solid line, remain. (1) i.e. if agents i and j are neighbours, and there is an agent k whose distance from agent i and from agent j is smaller than the distance between agents i and j (see Figure 1), then the edge eij is not necessary to maintain connectivity Lemma 1: If G is connected, then Ge is connected. Proof: We prove this Lemma by contradiction. Assume G is connected and Ge is disconnected to two disjoint connected setsV1 andV2. Hence, by assumption, ∀(i∈V1, j∈V2)∶∃k∶ pi− pk<pi− pj &pj− pk<pi− pj without loss of generality, assume i∈V1 and j∈V2 are the closest agents between the two groupsV1 andV2. Since G is connected by assumption, we have that pi− pj ≤ V , i∉ N e j and j∉ N e being the closest pair. By assumption these two agents are not effective neighbours, hence are not connected in Ge, i.e. i . But by (1) there is necessarily an agent k that links them, see (2). Hence i and j can not be the closest agents. (2) We note that the RNG has been proposed in multi- agent robotics in conjunction with multi-agent gathering and/or deployment algorithms, e.g. [16], [17], [18], [19]. The novelty in our work is (1) proposing the use of the operations, and (2) defining and using, along with various RNG and some extensions for maximally flexible swarm goal oriented constraints, the largest local allowable regions for displacement that ensure maintenance of connectivity, in the spirit of [4]. We note that methods to reduce the number of edges in geometric graphs while maintaining connectivity are also used for purposes that do not concern dynamics of agents, but, for example, for saving energy in ad-hoc communication networks, see e.g. [20]. THE MOTION LAWS In order to allow the agents free movement that does not violate the preservation of connectivity, we define for each agent an "Allowable Region". If all agents move into their allowable regions, preservation of connectivity is satisfied. Recalling Ando et. al. [2] for a pair of agents, the Allowable 2 centered at the Region of each agent is a disc of radius V midpoint position between the two agents mij (see figure 3). ARij= ARji= D V 2 pi+ pj 2 If an agent has more than one neighbour, its allowable region is defined by the union of is allowable regions given by (3), see Figure 4: ARi= j∈Ni ARij (4) Fig. 4. regions, illustrated here by the dashed area ARi. The allowable region of an agent is the union of its allowable clearly if an agent moves to the union of its allowable regions defined by (4), it is guaranteed that the distance from all its neighbours will not exceed V . Similar to (4),we define an "Effective Allowable Region": i= j∈N e i ARe ARij (5) the effective allowable region of agent the i.e. intersection of all its allowable regions relative to all the agents in its effective neighbourhood, defined by Ge. is i Hence, the motion law will be: pi(t) -- → pi(t+ 1)∈ ARe i(t) (3) i.e. the next position of agent i is any point inside its current effective allowable region (5). Since ARi(t)⊆ ARe i(t), the restriction of movement in (5) is less constrained than that of (4). (6) The freedom to move to any location inside allowable regions rather to a specific point, provides flexibility and allows other considerations and optimisations to be carried out by the agents. maintains connectivity. Theorem 1: A system of agents obeying dynamic law (6) ated with the system is connected by assumption. We have Proof: The initial visibility graph G(V,E)(0) associ- by Lemma 1 that Ge(V,E)(t) is connected. If all agents obey motion law (6), we have that E e(t) ⊆ E(t+ 1), i.e. all edges of Ge(V,E) are preserved. Therefore G(V,E)(t+ 1) is connected. Since G(V,E)(t+ 1) is connected, also Ge(V,E)(t+ 1) will be connected. Fig. 3. The Allowable Region ARij of two agents i and j is a disc of radius V 2 centered at their meadpoint position mij. If two agents whose mutual distance is smaller or equal V move into their allowable region defined in (3) that they both can calculate, they keep their mutual distance smaller than V , hence the agents will stay connected. APPLICATIONS AND SIMULATION RESULTS The distributed meta-algorithm presented so far opens up many possibilities of applications for distributed control of swarms. A number of interesting research directions that we consider relevant are algorithms with the ability to move a swarm of agents through narrow passages and obstacles, algorithms for mapping unknown areas, and reducing communication load when agents constitute a relay station for transmitting messages, as well as other topics requiring geometrically flexible swarms. We discuss below several examples. Fig. 6. (a) Edges exist between every pair of mutually visible agents at a distance V or less (b) Evolution of the network: only the necessary edges are kept (c) Final constellation: a required structured is achieved. This simulation result is based on the "RNG Plus" approach described at the end of this article with the value m= 1. Example 1: Traversal of narrow passages Example 3: Following a leader In this example (Figure 5), which will be discussed in detail in a subsequent article, the swarm is required to pass through a narrow passage. we implemented a rule of motion that prevents two agents from approaching each other closer 10 V . Then the agents movement is also governed by than 1 the cohesion rule (6) with two additional restrictions on the agents' allowable regions: (1) they exclude the obstacle areas (1) they ensure line-of-sight preservation between effective neighbours. One designated "leader" agent pulls the rest of the agents through the narrow passage. Because the swarm is not rigid, this passage is possible. By comparison, if the "never lose a neighbour" policy would have been applied together with the minimum distance limit ( 1 10 V ), the swarm would have been too rigid to pass through the narrow passage. Fig. 5. Passing through a narrow passage. (a) Initial constellation (b) The leader pulls the rest of the agents while only necessary edges are preserved (c) The swarm successfully completed the passage. This simulation result is based on the "RNG Plus" approach described at the end of this article with the value m= 1. Example 2: Formation This example presents simulations of convergence to quasi regular formation (Figure 6). In a topology where each agent relates to all other agents (i.e. in complete graph), by geometry it is impossible to create a structure in which all the lengths of the edges are identical. In this simulation, each agent worked according to rule (1) in order to choose only its effective neighbours, and only in relation to these (effective) neighbours he aspires to maintain a 10 V . The system "anneals" to a nice quasi fixed distance of 1 regular formation enabling nice and uniform area coverage as presented in the network evolution below. In this example (Figure 7), one random agent in the group becomes a designated leader which will seek to move along a predetermined path. A process is shown in which the number of edges in the neighbourhood graph decreases, while the graph remains connected, until a linear constellation of agents is formed, and the linear "snake-like" swarm follows the leader. Fig. 7. Following the leader. The randomly designated leader agent is shown as a large red circle, and the evolution of forming a line is presented from left to right. (a) Edges exist between every pair of agents distanced V or less, and a random leader is selected (b) gradually unnecessary edges disappear while the graph remains connected (c) Final constellation: the required line structured is achieved while all the agents follow the leader agent. Example 4: Ad-hoc networking In this example, each node describes an agent with limited communication and visibility range V , and each solid-line edge describe an "active" communication channel. In order to receive information from all the agents, a connected graph is required, but since the operation of transmission and reception consumes energy, there is motivation to reduce the number of necessary communication channels. In this simulation (see Figure 2), all lines (both solid and dashed) are shorter than V , but not all the lines are necessary for maintaining connectivity, i.e. communication is guaranteed using the solid lines only. The sub graph shown in solid lines was distributedly established by the agents themselves, each agent having information only about neighboring agents in the range V . The upper limit of number of edges is n(n−1) (i.e. clique). Under the geometric law (1), in the 2d case, each agent can have up to 6 effective agents, and therefore the upper limit of effective edges in this case will be 3n. 2 [11] Dimos V Dimarogonas and Kostas J Kyriakopoulos. On the ren- dezvous problem for multiple nonholonomic agents. IEEE Transac- tions on automatic control, 52(5):916 -- 922, 2007. [12] Housheng Su, Xiaofan Wang, and Guanrong Chen. Rendezvous of multiple mobile agents with preserved network connectivity. Systems & Control Letters, 59(5):313 -- 322, 2010. [13] Ariel Barel, Rotem Manor, and Alfred M Bruckstein. Proba- arXiv preprint bilistic gathering of agents with simple sensors. arXiv:1902.00294, 2019. [14] Rotem Manor and Alfred M Bruckstein. Chase your farthest neigh- bour: A simple gathering algorithm for anonymous, oblivious and non-communicating agents. In Proc. 13th Int. Symp. Distrib. Auton. Robotic Syst, 2016. [15] Godfried T Toussaint. The relative neighbourhood graph of a finite planar set. Pattern recognition, 12(4):261 -- 268, 1980. [16] Shouwei Li, Christine Markarian, Friedhelm Meyer auf der Heide, and Pavel Podlipyan. A continuous strategy for collisionless gathering. In International Symposium on Algorithms and Experiments for Sensor Systems, Wireless Networks and Distributed Robotics, pages 182 -- 197. Springer, 2017. [17] Anurag Ganguli, Sara Susca, Sonia Mart´ınez, Francesco Bullo, and Jorge Cortes. On collective motion in sensor networks: sample problems and distributed algorithms. In Proceedings of the 44th IEEE Conference on Decision and Control, pages 4239 -- 4244. IEEE, 2005. [18] Jorge Cort´es, Sonia Mart´ınez, and Francesco Bullo. Robust rendezvous for mobile autonomous agents via proximity graphs in arbitrary dimensions. Automatic Control, IEEE Transactions on, 51(8):1289 -- 1298, 2006. [19] Sonia Mart´ınez, Jorge Cortes, and Francesco Bullo. Motion coordina- tion with distributed information. Control Systems, IEEE, 27(4):75 -- 88, 2007. [20] Farinaz Koushanfar, Abhijit Davare, David T Nguyen, Alberto Sangiovanni-Vincentelli, and Miodrag Potkonjak. Techniques for maintaining connectivity in wireless ad-hoc networks under energy constraints. ACM Transactions on Embedded Computing Systems (TECS), 6(3):16, 2007. Toussaint [15] proved that finite planar set RNG, similar to ours effective graphs, contains at most 3n−6 edges. It is clear that using the method reduces overall energy consumption associated with communication [20]. SUMMARY A general method was introduced for the distributed control of multi-agent systems, which maintains connected but flexible swarms. An algorithm was introduced, allowing to locally and in a distributed manner, to remove edges in an interconnection graph between agents, without risking disconnecting it, which reduces the number of edges in the graph over time from order n2 to order n. DISCUSSION AND FURTHER INVESTIGATION The examples presented by simulation, did not include formal proofs. In a forthcoming article, we deal with formal definitions of motion laws in the presence of obstacles, and a formal analysis that the system indeed converges in finite expected time to the desired constellation. An interesting extension of the ideas presented above is the design of local rules based on allowing "at most m" agents of the swarm to be present in the intersection between the visibility discs of neighbours, in order to maintain the connection between them. In our previous analysis on RNG we took m = 0. This extension named "RNG Plus" will ensure more reliable connectivity at the expense of slightly less flexibility. This tradeoff too will be the subject of further investigation. A video of some of the simulations shown above is avail- able on-line at https://youtu.be/A3jnYpY15DY. REFERENCES [1] Ariel Barel, Rotem Manor, and Alfred M Bruckstein. Come together: multi-agent geometric consensus (gathering, rendezvous, clustering, aggregation). arXiv preprint arXiv:1902.01455, 2019. [2] Hideki Ando, Yoshinobu Oasa, Ichiro Suzuki, and Masafumi Ya- mashita. Distributed memoryless point convergence algorithm for mobile robots with limited visibility. Robotics and Automation, IEEE Transactions on, 15(5):818 -- 828, 1999. [3] Noam Gordon, Israel A Wagner, and Alfred M Bruckstein. Gathering In Ant multiple robotic a (ge) nts with limited sensing capabilities. Colony Optimization and Swarm Intelligence, volume 3172 of Lecture Notes in Computer Science, pages 142 -- 153. Springer, 2004. [4] Noam Gordon, Israel A Wagner, and Alfred M Bruckstein. A ran- domized gathering algorithm for multiple robots with limited sensing capabilities. In Proc. of MARS 2005 workshop at ICINCO Barcelona, 2005. [5] Veysel Gazi and Kevin M. Passino. Stability analysis of swarms. IEEE Transactions on Automatic Control, 48:692 -- 697, 2003. [6] Veysel Gazi and Kevin M Passino. Stability analysis of social foraging swarms. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, 34(1):539 -- 557, 2004. [7] Reza Olfati-Saber, J Alex Fax, and Richard M Murray. Consensus and cooperation in networked multi-agent systems. Proceedings of the IEEE, 95(1):215 -- 233, 2007. [8] Luc Moreau. Stability of continuous-time distributed consensus algo- rithms. In Decision and Control, 2004. CDC. 43rd IEEE Conference on, volume 4, pages 3998 -- 4003. IEEE, 2004. [9] Ali Jadbabaie, Jie Lin, and A. Stephen Morse. Coordination of groups of mobile autonomous agents using nearest neighbor rules. Automatic Control, IEEE Transactions on, 48(6):988 -- 1001, 2003. Ozdemir, Melvin Gauci, Salom´e Bonnet, and Roderich Gross. Finding consensus without computation. IEEE Robotics and Automa- tion Letters, 3(3):1346 -- 1353, 2018. [10] Anıl
1007.0803	1	1007	2010-07-06T04:00:10	Soft Control on Collective Behavior of a Group of Autonomous Agents by a Shill Agent	[ "cs.MA" ]	This paper asks a new question: how can we control the collective behavior of self-organized multi-agent systems? We try to answer the question by proposing a new notion called 'Soft Control', which keeps the local rule of the existing agents in the system. We show the feasibility of soft control by a case study. Consider the simple but typical distributed multi-agent model proposed by Vicsek et al. for flocking of birds: each agent moves with the same speed but with different headings which are updated using a local rule based on the average of its own heading and the headings of its neighbors. Most studies of this model are about the self-organized collective behavior, such as synchronization of headings. We want to intervene in the collective behavior (headings) of the group by soft control. A specified method is to add a special agent, called a 'Shill', which can be controlled by us but is treated as an ordinary agent by other agents. We construct a control law for the shill so that it can synchronize the whole group to an objective heading. This control law is proved to be effective analytically and numerically. Note that soft control is different from the approach of distributed control. It is a natural way to intervene in the distributed systems. It may bring out many interesting issues and challenges on the control of complex systems.	cs.MA	cs	PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 1 Soft Control on Collective Behavior of a Group of Autonomous Agents by a Shill Agent Jing Han, Ming Li and Lei Guo Abstract This paper asks a new question: how can we control the collective behavior of self-organized multi- agent systems? We try to answer the question by proposing a new notion called 'Soft Control', which keeps the local rule of the existing agents in the system. We show the feasibility of soft control by a case study. Consider the simple but typical distributed multi-agent model proposed by Vicsek et al. for flocking of birds: each agent moves with the same speed but with different headings which are updated using a local rule based on the average of its own heading and the headings of its neighbors. Most studies of this model are about the self-organized collective behavior, such as synchronization of headings. We want to intervene in the collective behavior (headings) of the group by soft control. A specified method is to add a special agent, called a 'Shill', which can be controlled by us but is treated as an ordinary agent by other agents. We construct a control law for the shill so that it can synchronize the whole group to an objective heading. This control law is proved to be effective analytically and numerically. Note that soft control is different from the approach of distributed control. It is a natural way to intervene in the distributed systems. It may bring out many interesting issues and challenges on the control of complex systems. Index Terms Collective Behavior, Multi-agent System, Soft Control, Boid Model, Shill Agent This work was supported by the National Natural Science Foundation of China. Jing Han and Lei Guo are with the Institute of Systems Science, AMSS, Chinese Academy of Sciences, Beijing, 100080, China. Ming Li is with the Institute of Theoretical Physics, Chinese Academy of Sciences. Corresponding author: [email protected]. 0 1 0 2 l u J 6 ] A M . s c [ 1 v 3 0 8 0 . 7 0 0 1 : v i X r a PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 2 I. INTRODUCTION Collective behavior is the high level (macroscopic) property of a self-organized system which consists of a large number of (microscopic) individuals (agents). Examples are synchronization, aggregation, phase transition, pattern formation, swarm intelligence, fashion, etc. People found this kind of phenomena in many systems, such as flocking of birds, schools of fish, cooperation in ant colonies, panic of crowds [10], norms in economic systems [9], etc. Without any question, collective behavior is one of the fundamental and difficult topics of the study of complex systems. We classify the research on collective behavior into three categories. (I) Given the local rules of agents, what is the collective behavior of the overall system? Many people have been working on this category which is about the mechanism of how collective behavior emerges from multi-agent systems. "More is different"[1]. The physicists have applied theory of statistical physics to explored some simple models, from the Ideal Gas model, Spin Glasses, to the panic model and network dynamics. (II) Given the desired collective behavior, what are the local rules for agents? Some people work on this category. One typical example is Swarm Intelligence. Since the high level function of the overall system can be more than the sum of all individuals, how do we construct robust intelligence by a large number of locally interacting simple agents? An ant is simple and often moves randomly. But a colony of ants can efficiently find the shortest path from their nest to a food source. This natural phenomenon inspired the Ant Colony Optimization Algorithm[2]. (III) Given the local rules of agents, how we control the collective behavior? In some real applications, it is very difficult or even impossible to change the local rules of agents, such as the behavioral rules of people in panic and the flying strategies of birds. Yet we need to control the system to avoid danger or improve efficiency. Then what is the feasible way to intervene in the collective behavior? There should be a way especially for the multi-agent system that utilizes the property of collective behavior. This question is not well noticed in the control literature. In this paper, we propose a new control notion, called 'soft control'1, which keeps the basic local rule of the existing agents in the system. There are no global parameters can be adjusted 1The idea of soft control in this paper was inspired by discussions at the "Systems, Control and the Complexity Science" Xiangshan Science Conference (May, 2004, Beijing) and the preliminary result was presented at the 2nd Chinese-Swedish Conference on Control [13]. PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 3 to achieve the control purpose, such as adjusting the temperature to change water from liquid phase to gas/solid phase or broadcasting orders directly to all agents by a controller. This feature makes a difference between soft control and the traditional control. These three categories are tightly related. The first category (I) is about how a distributed system behaves and what emerges from the system. Both of (II) and (III), however, require the system to behave at the high level as what we expected, i.e., we know what collective behavior should be emerged from the system. But (II) focuses on how to design a distributed system by constructing(changing) the local rules for interaction among agents, and (III) focuses on how to intervene in the existing system by a nondestructive way. Since collective behavior is a kind of macroscopic feature of a multi-agent system, usually adding one or a few more agents will not affect it. But if those agents are special ones which can be controlled, even though only a small number (in some cases, only one), it can dramatically change the collective behavior as will be shown in Section 3. So in this paper, a way to control the system without changing the local rules of the existing agents is to add one (or a few) special agent which can be controlled. Therefore its behavior does not necessarily obey the local rules as the ordinary agents do. However, the existing agents still treat the special agent as an ordinary agent, so the special agent only affects the local area with limited power. The special agent is the only controlled part of the system and it indeed changes the collective behavior of the system by 'cheating' and 'seducing' its neighboring ordinary agents. So we do not call the special agent a leader. Instead, we call it a 'Shill'. Soft control is different from distributed control. Distributed control recently gets more and more attention because many real-world applications are distributed systems, such as power networks and traffic systems. The system consists of many interacting subsystems that all have their own controllers with local feedback. The study of the relationship about performance between the overall system and the subsystems falls into category (I). On the other hand, designing a practical distributed system for resource sharing and cooperation is a demanding and complicated task. This research is of the category (II). So we can see that both soft control and distributed control concern with the macroscopic collective behavior (such as synchronization) of the self-organized multi-agent system with local rules. But in distributed control, every agent is treated as a control system and has its control law (which is the set of local rules); while in soft control framework, all these agents are treated as one system and the control law is for PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 4 the shill. So soft control can be regarded as a way of intervention in the distributed systems. With the growing literature on complex systems, the challenge of controlling complex systems is likely to become a key problem for control scientists. This paper is going to explore the idea of soft control by a case study, where we will show how a shill can be used to control the headings of a group of mobile agents demonstrated based on the modified Boid model [4] proposed by Vicsek et al [3]. The current control theory cannot be applied directly to this problem because our soft control is a nondestructive intervention in a locally interacting autonomous multi-agent system, which is not allowed to change the local rules of the existing agents. The formation control [5] mainly focuses on how to design the local rules of a team of robots to maintain a geometric configuration during movement, which is a very special kind of collective behavior, and its approach belongs to distributed control. The pinning control [12] is especially for stabilizing dynamical networks (with special topology, such as random networks, small-world and scale-free networks) by imposing controllers on a small fraction of selected nodes in the network. No one appear to have directly studied the above mentioned soft control problem in the literature. Recently, Jadbabie et al [6] investigated a multi- agent model which is slightly different from the one proposed by Vicsek et al, they showed that under some a priori connectivity conditions, the group will eventually move in a same direction even without centralized control. This is a problem of the category (I) of collective behavior study. An idea in [6] worthy of mention is the Leader Following model, which has a leader agent with fixed heading. But it did not tell how to implement synchronization by a leader. The so-called 'virtual leader'[7] is not treated as an agent but part of the potential, which requires to introduce new rules for the ordinary agents to recognize it. The rest of this paper is organized as follows: The modified BOID model proposed by Vicsek et al and the notion of soft control are set up in Section 2. Section 3 is a case study for soft control. We will solve a simple but non-trivial problem by constructing an effective control law for the shill. Both theoretical analysis and computer simulation show that one shill is possible to turn the direction of the whole group. Some concluding remarks are given in Section 4. II. SOFT CONTROL OF THE MODIFIED BOID MODEL Eighteen years ago, Renolds proposed a simultaneous discrete-time multi-agent model, which is called Boid[4], to catch the natural phenomenon of flocking of birds and fish and make PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 5 computer animation. In this model, each bird decides its flying direction only by looking at its current neighboring birds and using three rules (Alignment, Separation and Cohesion) based on the status of its neighbors. These rules are local and simple but the overall system exhibits the flocking behavior. The Boid model became very popular in complex systems studies. Unfortunately, this simple model is not simple at all for theoretical analysis. In 1995, to investigate the emergence of self-ordered motion, Vicsek et al [3] proposed a model which is actually a modified version of Boid, by only keeping the Alignment rule– steer towards the average heading of neighbors, which still can catch the flocking behavior. There are n agents (xi(·), θi(·)) for n birds, labelled from 1 through n, all moving simultaneously in the two-dimensional space. The velocity of agent i at time t is defined by vi(t) = (v cos(θi(t)), v sin(θi(t))) (1) which is constructed to have an absolute value v and a heading given by the angle θi(t) ∈ [0, 2π). The position of agent i at time t is denoted as xi(t). The neighborhood of agent i at time t is defined as Ni(t) = {jkxj(t) − xi(t)k ≤ r, j = 1, 2, . . . , n}, where r is the radius of the neighborhood circle. For any agent i, its heading and position are updated by (2) and (3) below if ignoring noise: θi(t + 1) = hθi(t)ir xi(t + 1) = xi(t) + vi(t + 1) (2) (3) where vi(t + 1) is defined by (1), and hθi(t)ir is the angle of the sum of the velocity vectors of neighbors of agent i: Pj∈Ni(t) vj(t). In other words, hθi(t)ir can be obtained by arctan( X j∈Ni(t) sin(θj(t))/ X j∈Ni(t) cos(θj(t))) (4) with some necessary regulations reflecting angles of [0, 2π). Some recent results of Jadbabaie et al [6] have shown that under some connectivity conditions, the group will eventually move in a same direction. However, the model of [6] calculates hθi(t)ir by simply taking the average of angles of all neighbors. This brings convenience for mathematical analysis, but the system will sometimes exhibit counter-intuitive phenomena as they claimed in PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 6 their paper2. So this paper keeps the updating rule of the model proposed by Vicsek et al, even though the analysis result about synchronization of the model in [6] cannot be utilized3. One can think of many related questions: what is the way to help the group to synchronize if the group will not synchronize by self-organization? What if the group self-organize to a synchronized direction α which is not what we want? What if we want the group to fly to a desired destination? And so on. In this paper, we consider the case of adding one shill to control the collective behavior with the shill denoted as (x0, θ0). The ordinary agent (i = 1, . . . , n) still keep the local rule as formula (2)-(3) to update its heading and position. The only difference is that the neighborhood Ni(t) for agent i at time t will consider the shill: Ni(t) = {jkxj(t) − xi(t)k ≤ r, j = 0, 1, 2, . . . , n}. And the shill does not exactly obey the local rules as the ordinary agents do, its position and heading is decided by the control law u: (x0(t), θ0(t)) = u(x1(t), . . . , xn(t), θ1(t), . . . , θn(t), t) Note that (x0, θ0) may be subject to some constraints. For example, the constraint of x0(t+1) = x0(t) + v0(t + 1) will make the shill behave like an ordinary agent except the way it decides its own heading. Without any constraint, the shill can fly to anywhere with a 'deceptive' heading. The control law u will be different for different control purposes/tasks, such as to synchronize headings of the group (n agents), to keep the group connected or to dissolve a group, to turn 2 This is because the zero angle is special than other angles in (0, 2π). The angles do not symmetrically behave. 3 The result about synchronization for the model of [6] is not true for the model of Vicsek et al. Here is a counterexample: when v = 0, six agents are regularly put on the edge of a circle with radius of r, which form an ordinary hexagon. So each agent has three neighbors (because itself is also counted). The headings are 0, π/3, 2π/3, π, 4π/3, 5π/3. This case will synchronize for the model in [6]. But it will not for the model we consider here, i.e. the Vicsek et al model without noise. PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 7 headings of the group (minimal circling), to lead the group to a destination (in a shortest time), to avoid hitting an object, etc. There are many problems to be considered in terms of control: under what conditions we can use the shill to control the state of the multi-agent system (the controllability problem); which kind of state information can be used for control (the observation problem), e.g., the shill can only see some nearest agents but not all other agents; how can one design the control law when the local rule and other information is not clear (the learning and adaptation problem), etc. We do not intend to answer all these questions in this paper. As a beginning, we will just show how it works for a specific case, which is to synchronize and turn the headings of a group. III. A CASE STUDY Flocking of birds in airports is dangerous. How we drive them away? In this case, obviously we can not use centralized control, such as broadcasting orders to the birds to change to the desired flying direction. The current solution is to use cannon to shoot them away. But can we drive them away by soft control? If the modified Boid model proposed by Vicsek et al works for real birds flocking, and if we can make a controllable powerful robot bird, can we use this robot bird as the shill and guide flight of the birds? In this section, we will study a simple but non-trivial related problem 4. The problem we consider here is: for a group of n agents with initial heading of θi(0) ∈ [0, π), 1 ≤ i ≤ n, what is the control law for the shill, so that all agents will move to the direction of π eventually? Suppose the local rule about the ordinary agents is known. Suppose also that the position x0(t) and heading θ0(t) of the shill can be controlled at any time step t. Suppose further that the state information (headings and positions) of all ordinary agents are observable at any time step. Now we propose an effective control law uβ which is defined as (see Fig. 2): 4In flocking behavior of real animals, more factors such as different sensory systems should be considered. The Boid model is an abstract model which successfully simulates the flocking behavior. But no one has proved it to be true for real birds. So in this paper we just want to use a simple model to demonstrate how soft control works for self-organized multi-agent systems in general, but not to solve the specified airport birds problem. PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 8 x0(t) = xs(t)(t) θs(t)(t) + β π if if θs(t) ≤ π − β; θs(t) > π − β; (5) (6) θ0(t) =   where β ∈ (0, π) is a constant, and s(t) is defined as the 'worst' agent (or one of the 'worst' agents): s(t) = argmin {θi(t)}; 1≤i≤n The intuition behinds uβ is to put the shill to the position of the 'worst' agent s(t) at each time step t and try to 'pull' it to the desired direction π. This greatly reduces the search space of positions and simplifies the control strategy. Note that we can not use θ0(t) = π for all the time because: (A) in the case where θs(t)(t) = 0 and the all the neighboring ordinary agents of the shill have headings of zero degree, the shill with θ0(t) = π can not change θs(t)(t) according to the update rule (2); (B) to avoid the ill-case where both of the numerator and denominator in (4) are zero5. But in the final stages we do need θ0 = π to help the system to synchronize. In the sequel, we use ∆(t) = π − θs(t)(t) to denote the distance between the angel of the 'worst' agent and the objective angel π at time t. We now have the following main technical result of the paper: sin(θj(t)) = 0 implies θj(t) = 0 or π for all j ∈ Ni(t) because sin α > 0 for α ∈ (o, π). cos(θj (t)) = 0 we further conclude that half of the headings should be zero and the other half should 5In the ill-case, Pj∈Ni(t) Consequently, by Pj∈Ni(t) be π. So there must be at least one agent j 6= 0 such that θj(t) = 0 (since θ0(t) > 0 for t > 0 by the control law), which means θs(t)(t) = 0, and so we must have θ0(t) = β ∈ (0, π) if 0 ∈ Ni(t). However, this will contradict with the fact that θj(t) = 0 or π for all j ∈ Ni(t). On the other hand, if 0 /∈ Ni(t), Pj∈Ni(t) sin(θj(t)) = 0 only holds when θj(t) = 0 for all j ∈ Ni(t) since obviously θj (t) < π for all j = 1, . . . , n and t ≥ 0, but in this case we have Pj∈Ni(t) ill-case will never happen in (4). cos(θj (t)) > 0. So the PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 9 Theorem 1. For any n ≥ 2, r ≥ 0, v ≥ 0, and any initial headings and positions θi(0) ∈ [0, π), xi(0) ∈ R2, (1 ≤ i ≤ n), the update rule (2)-(3) and the soft control law uβ defined by (5)-(6) will lead to the asymptotic synchronization of the group, i.e., limt→∞ ∆(t) = 0. Proof. Because π ≥ θ0(t) > θs(t)(t) ≥ 0, and π > θi(t) ≥ θs(t)(t) for 1 ≤ i ≤ n, it is obvious by the update rule (2) that π > θi(t + 1) ≥ θs(t)(t) for 1 ≤ i ≤ n which implies that θs(t)(t) ≤ θs(t+1)(t + 1). Hence by definition of ∆(t), we have ∆(t) ≥ ∆(t + 1) ≥ 0, so ∆(t) will converge to some limit µ ≥ 0. If µ > 0, then for any 0 < ε < µ, we know that ∆(t) ≥ ε > 0 holds for all large t.Consequently, by letting t → ∞ in the following Lemma 2, we have µ ≤ µ − δ which contradicts with δ >0.So we can conclude µ = 0. This completes the proof of Theorem 1. The main task now is to prove the following lemma. Lemma 2. If for some t > 0 and ε > 0, ∆(t) ≥ ε > 0, then there exists a constant δ ∈ (0, ε) which does not depend on t, such that ∆(t + n) ≤ ∆(t) − δ. Proof. At any time step t, the shill is affecting the agent s(t) with heading θ0(t). We now proceed to analyze the lower bound to the angle change θs(t)(t + k) − θs(t)(t) for 1 ≤ k ≤ n, which is denoted by η(k). For k = 1, the shill is affecting the agent s(t), so the worst case where η(1) has the smallest value is that the agent s(t) is surrounded by n − 1 agents with heading θs(t)(t) (see Fig.3-(a)). For k ≥ 2, however, the shill may not affect the agent s(t), so the worst case where η(k) has the smallest value is that the agent s(t) is surrounded by n − 1 agents with heading θs(t)(t) again, but without the shill in its neighborhood (see Fig.3-(b)). Now by the definition of η(k) we have θs(t)(t + k) ≥ θs(t)(t) + η(k), and a simple calculation will give η(k) =   min{arctan(sin β/(n + cos β)), arctan(sin ε/(n + cos ε))}, k = 1; arctan(sin(η(k − 1))/(n − 1 + cos(η(k − 1))), 2 ≤ k ≤ n; (7) (8) First, note that ∆(t) ≥ ε > 0 and the function ψ(α) = arctan(sin α/(m + cos α)), (m > 0) is monotonically increasing when α ∈ [0, arccos(−1/m)] and monotonically decreasing when α ∈ [arccos(−1/m), π]. PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 Second, to explain the definition of η(1) in (8) we need to show that 0 < η(1) ≤ arctan(sin(θ0(t) − θs(t)(t))/(n + cos(θ0(t) − θs(t)(t)))). 10 (9) In fact, by the control law (6) and the expression of η(1) in (8), the above inequality (9) is trivial in the case where θs(t) ≤ π − β. So we only need to consider the case where θs(t) > π − β. In this case the control law is θ0(t) = π. We consider two subcases as follows: (i) θ0(t)−θs(t)(t) ≥ arccos(− 1 n ). Since θ0(t)−θs(t)(t) ≤ π −(π −β) = β, by the monotonicity of the function ψ(·), we know that the claim (9) is true. (ii) θ0(t) − θs(t)(t) < arccos(− 1 n ). Since θ0(t) − θs(t)(t) = ∆(t) ≥ ε, by the monotonicity of the function ψ(·) again, we know that the claim (9) is also true. Hence, (9) holds in any case. Now, note that 0 < η(1) < π/2, and again by the monotonicity of the function ψ(·), we have 0 < η(n) < η(n − 1) < . . . < η(1) < π/2. Let us take δ = η(n) which is a finite positive value depending on n, β and ε only. Obviously, we have θs(t)(t + n) − θs(t)(t) ≥ δ. Next, we consider the other agents: i = 1, 2, . . . , n, i 6= s(t). Let Λ(k) denote the set of agents whose angles are inside (θs(t)(t), θs(t)(t) + δ) at time step t + k, i.e., Λ(k) = {iθs(t)(t) < θi(t + k) < θs(t)(t) + δ, i = 1, 2, . . . , n}. We now proceed to show that we must have Λ(n) = ∅. First of all, as shown above, s(t) does not belong to Λ(k) for 1 ≤ k ≤ n because θs(t)(t+k) ≥ θs(t)(t) + η(k) ≥ θs(t)(t) + δ. So we have Λ(k) ≤ n − 1 for 1 ≤ k ≤ n. Therefore, at time step t + 1, there are at most n − 1 agents with angles which are less than θs(t)(t) + δ. If Λ(1) 6= ∅, s(t + 1) will be picked up from Λ(1). The shill will change the heading of agent s(t + 1). Because θs(t+1)(t + 1) ≥ θs(t)(t) and θs(t)(t + k) ≥ θs(t)(t) + η(k) for any t, we have for k ≥ 2, θs(t+1)(t + k) ≥ θs(t+1)(t + 1) + η(k − 1) ≥ θs(t+1)(t + 1) + δ ≥ θs(t)(t) + δ. So s(t + 1) will not belong to Λ(k) for any k ∈ [2, n], we have Λ(k) ≤ n − 2 for 2 ≤ k ≤ n. By repeating this argument, we get for 1 ≤ d ≤ k ≤ n, θs(t+d)(t + k) ≥ θs(t)(t) + δ and PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 11 Λ(k) ≤ n − d. So by taking d = k = n, we have Λ(n) = 0. This means that after at most n steps, Λ will be empty. Consequently, we know that θi(t + n) ≥ θs(t)(t) + δ for all i. Then we get θs(t+n)(t + n) ≥ θs(t)(t) + δ ⇒ ∆(t + n) ≤ ∆(t) − δ. This complete the proof of Lemma 2. The snapshots of the relating computer simulation are showed in Figure 4. The video of demo and the simulation program can be downloaded from the project website [14]. Although our case study is a simple starting point, it shows that it is possible to change the collective behavior (headings) of a group by soft control. Based on this work, we are going to explore more complicated control problems, such as what if the energy of the shill is concerned (the energy of the shill is limited so that it can not 'jump' as in (5)), what if the shill only can see locally, and what if the shill also needs to keep the group connected while turning the group, etc. To support this research, we have put a computer demo for soft control of the modified Boid model which is like a computer game on the project homepage [14]. After initially putting n agents on the two-dimensional space, the user can manually control the position and heading of the shill by keyboard and mouse. It is an open competition for the best control strategy. IV. CONCLUDING REMARKS The collective behavior of complex systems is the macroscopic feature that we concerned and studied in this paper. How we control or intervene in a multi-agent system without changing PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 12 the local rule of the existing agents is an important issue but has not been well recognized yet. The idea of soft control that we introduced and explored here appear to be an initial attempt to tackle this issue. We believe that our general soft control framework reflects a large class of control problems in complex systems. In this paper, we have restricted ourselves on the study of the modified Boid model which does catch certain key properties of many real world complex systems (see, e.g.,[3][4][6]). Without changing the local rule of the existing agents, we have added a shill in the distributed system and designed the control law for the shill carefully for the control purpose. Although the shill does not behave as the ordinary agents do, it is not recognized by the ordinary agents and is therefore still treated as an ordinary agent by them. This soft control idea works well as we have shown in the case study. We would like to emphasis that the idea of adding a shill is just one of the methods of soft control, and for different systems, they would be certainly different. There are many potential applications of soft control. Currently a project called Claytronics [11], a new form of programmable matter, is to study how to form interesting dynamic shapes and configurations by simple and local interactions of a billion micro-scale units. This research is in the category (II) of collective behavior. Using soft control to help to form shapes may release us from designing subtle local rules. Another possible case is to use soft control in the language diffusion model to guide evolution of language, and it might be able to save dying language [8]. A recent study in economics [9] shows that a small proportion of special agents who have preferences can dramatically alter the expected waiting time between norm transitions in a population of agents who repeatedly play the Nash demand game. If those special agents have feedbacks, how does the soft control help changing the transition of norms more efficiently ? How we use soft control to avoid and handle panic of crowd [10] ... In conclusion, we need a theory of soft control to efficiently intervene in many complex systems. The first author would like to thank Prof. John Holland, Dr. Liu Zhixin and Prof. Eric Smith ACKNOWLEDGMENT for helpful discussions. [1] P.W.Anderson. More is Different. Science 177, 393-396, 1972. REFERENCES PUBLISHED IN JOURNAL OF SYSTEMS SCIENCE AND COMPLEXITY, 2006(19):54-62 13 [2] E. Bonabeau, M. Dorigo, G. Theraulaz. Swarm intelligence: from natural to artificial systems. Oxford University Press, New York, 1999. [3] T. Vicsek, A. Czirok, E.B. Jacob, I. Cohen, and O. Schochet. Novel type of phase transitions in a system of self-driven particles. Physical Review Letters, 75:1226–1229, 1995. [4] C. Reynolds. Flocks, birds, and schools: a distributed behavioral model. Computer Graphics, 21:25–34, 1987. The demo can be downloaded from http://www.red3d.com/cwr/boids/applet/. [5] J. Desai, J. Otrowski, and V. Kumar. Modeling and control of formations of nonholonomic robots. IEEE Trans. on Robotics and Automation, 17(6):905–908, 2001. [6] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Trans. Automatic Control, vol. 48, pp. 988–1001, June 2003. [7] N.E. Leonard and E.Fiorelli. Virtual Leaders, Artificial Potentials and Coordinated Control of Groups. In Proceedings of the 40th IEEE Conference on Decision and Control. Orlando, FL, December 2001, 2968–2973. [8] W.S-Y.Wang, J.Ke and J.W.Minett. Computational studies of language evolution. In Computational Linguistics and Beyond,R. Huang and W. Lenders, eds. Nankang, 65-106, 2004. [9] R.L. Axtell, S. Chakravarty. Radicals, Revolutionaries and Reactionaries in a Multi-Agent Model of Class Norms. [10] D. Helbing, I. Farkas and T. Vicsek. Simulating dynamical features of escape panic. Nature vol.407, 487-490, 2000. [11] Synthetic Reality project at Carnegie Mellon. http://www-2.cs.cmu.edu/∼claytronics/. [12] X.F. Wang, G. Chen. Pinning control of scale-free dynamical networks. Physica A, Vol. 310, pp. 521-531, July 2002. [13] J. Han, L. Guo, M. Li. Guiding a Group of Locally Interacting Autonomous Mobile Agents. The 2nd Chinese-Swedish Conference on Control, October 16-17, 2004, Beijing. [14] Project of soft control, 2004-. http://complex.amss.ac.cn/hanjing/softcontrol/. Videos and demos can be downloaded from http://complex.amss.ac.cn/hanjing/softcontrol/demo.html.
1902.06228	1	1902	2019-02-17T09:08:52	Optimizing Online Matching for Ride-Sourcing Services with Multi-Agent Deep Reinforcement Learning	[ "cs.MA" ]	Ride-sourcing services are now reshaping the way people travel by effectively connecting drivers and passengers through mobile internets. Online matching between idle drivers and waiting passengers is one of the most key components in a ride-sourcing system. The average pickup distance or time is an important measurement of system efficiency since it affects both passengers' waiting time and drivers' utilization rate. It is naturally expected that a more effective bipartite matching (with smaller average pickup time) can be implemented if the platform accumulates more idle drivers and waiting passengers in the matching pool. A specific passenger request can also benefit from a delayed matching since he/she may be matched with closer idle drivers after waiting for a few seconds. Motivated by the potential benefits of delayed matching, this paper establishes a two-stage framework which incorporates a combinatorial optimization and multi-agent deep reinforcement learning methods. The multi-agent reinforcement learning methods are used to dynamically determine the delayed time for each passenger request (or the time at which each request enters the matching pool), while the combinatorial optimization conducts an optimal bipartite matching between idle drivers and waiting passengers in the matching pool. Two reinforcement learning methods, spatio-temporal multi-agent deep Q learning (ST-M-DQN) and spatio-temporal multi-agent actor-critic (ST-M-A2C) are developed. Through extensive empirical experiments with a well-designed simulator, we show that the proposed framework is able to remarkably improve system performances.	cs.MA	cs	1 Optimizing Online Matching for Ride-Sourcing Services with Multi-Agent Deep Reinforcement Learning Jintao Ke, Feng Xiao, Hai Yang, and Jieping Ye, Senior Member, IEEE Abstract -- Ride-sourcing services are now reshaping the way people travel by effectively connecting drivers and passengers through mobile internets. Online matching between idle drivers and waiting passengers is one of the most key components in a ride-sourcing system. The average pickup distance or time is an important measurement of system efficiency since it affects both passengers' waiting time and drivers' utilization rate. It is naturally expected that a more effective bipartite matching (with smaller average pickup time) can be implemented if the platform accumulates more idle drivers and waiting passengers in the matching pool. A specific passenger request can also benefit from a delayed matching since he/she may be matched with closer idle drivers after waiting for a few seconds. Motivated by the potential benefits of delayed matching, this paper establishes a two-stage framework which incorporates a combinatorial optimization and multi-agent deep reinforcement learning methods. The multi-agent reinforcement learning methods are used to dynamically determine the delayed time for each passenger request (or the time at which each request enters the matching pool), while the combinatorial optimization conducts an optimal bipartite matching between idle drivers and waiting passengers in the matching pool. Two reinforcement learning methods, spatio-temporal multi-agent deep Q learning (ST-M-DQN) and spatio-temporal multi-agent actor-critic (ST-M-A2C) are developed. Through extensive empirical experiments with a well-designed simulator, we show that the proposed framework is able to remarkably improve system performances. Index Terms -- ride-sourcing, dispatching, delayed matching, multi-agent reinforcement learning. ! 1 INTRODUCTION W ITH the rapid development of mobile internet based technologies and the popularity of smart phones, ride- sourcing services, such as Uber, Lyft, and DiDi, has been reshaping the way we travel. In contrast to traditional taxi markets that rely on the physical meeting between drivers and passengers, the ride-sourcing platforms implement an on-line matching process that can match unserved passengers and idle drivers more efficiently. Besides, the tremendous amount of information collected through the APPs, such as the real-time locations, trajectories and travel patterns of passengers and drivers, can further enhance the on- line matching procedure by reducing the search frictions between passengers and drivers. There are two main modes of matching process [1], [2]. One is the broadcast mode, in which the ride-sourcing platforms collect requests from passengers and broadcasts them to nearby idle drivers, each of whom opts for one of the requests by considering his/her individual benefits (such as trip fare) and costs (such as the distance to pick-up the passenger). Another is the dispatch mode, in which the ride-sourcing platforms collect • • • J. Ke and H. Yang are with the Department of Civil and Environmental Engineering, Hong Kong University of Science and Technology, Hong Kong, China. E-mail: [email protected], [email protected]. F. Xiao (corresponding author) is with School of Business Administration, Southwestern University of Finance and Economics, Chengdu, China. E-mail: [email protected] J. Ye . is with AI Labs, Didi Chuxing, Beijing, China E-mail: [email protected] Manuscript submitted March 10, 2019 the information of idle drivers and requests of passengers on the fly and implement a centralized algorithm to match these drivers and passengers pair by pair. Among these two modes, the dispatch mode is thought to be more efficient and thus widely adopted in various ride-sourcing platforms. For example, DiDi's statistics showed that a significant improvement on system efficiency - the request completion rate being improved by more than 10% - was observed when DiDi switched from broadcast mode to dispatch mode [1]. In dispatch mode, during each short match time interval (such as one or two seconds), the ride-sourcing platforms collect all idle drivers and unserved passenger requests, and then execute matching between drivers and passengers with combinatorial optimization approaches. These approaches aim to improve the overall system performance by increasing the matching rate (the number of matching driver-passenger pairs per unit time), decreasing the passengers' waiting time (from announcing a request to being matched by the on- line platform), and reducing the passengers' pick-up time (from being matched to dropping on the corresponding vehicle). Of particular importance here is the trade-off among pick-up time, waiting time, and success rate of matching. When the platforms wait for a few match time intervals to accumulate sufficient pairs of idle drivers and unserved passengers, instead of immediately matching them, a better matching are expected, and as a result, the average pick- up time will be reduced. As shown in Fig. 1, if a delayed matching is used, more drivers and passengers will occur in the matching pool, and thus a more efficient matching between drivers and passengers with shorter average pickup time can be implemented. As for a single passenger, he/she 2 be executed with a combinatorial optimization algorithm. Clearly, under this setting, the action space of each agent reduces to a binary decision (delayed or not) in each match time interval, which makes the originally intractable problem available to be solved. The problem setting of this paper is different from the recent related studies in this domain, such as [1], [4], [8], [9], [10], in the following aspects. These studies majorly focused on employing reinforcement learning methods to control the dispatching from orders to drivers and drivers' movements. However, none of these studies considered the potential benefits of delayed matching or investigated the dynamic control of matching time intervals. Moreover, most of these studies treated drivers as agents and tried to optimize the long-term revenue of drivers, whereas our paper focus more on the passenger side, and try to minimize passengers' pickup time through a dynamic control of delayed matching. The main contributions of this paper are listed below: • We propose a two-stage framework that first determines the delayed time of each passenger request through multi-agent reinforcement learning methods and then executes optimal matching between idle drivers and unserved passengers in the matching pool with a combinatorial optimization model. • We formulate the delayed matching problem as a MMDP by a proper design of agent, state, action, and reward. Two types of rewards, global reward and individual reward, are used and then tested. • We propose two learning methods, named spatio- temporal multi-agent deep Q-learning (ST-M-DQN) and spatio-temporal multi-agent actor-critic (ST-M-A2C), which observe the spatio-temporal patterns of supply and demand and find optimal policies that make a sequence of delayed-or-not decisions for each agent. • We build a simulator that characterizes passengers' and drivers' activities and is calibrated with both synthetic datasets and actual mobility datasets provided by DiDi. Experiment results show that the proposed framework can significantly improve the system efficiency. The rest of the paper is organized as follows. Section 2 describes the research problem, highlights the potential benefits of delayed matching, and points out the main challenges. Section 3 presents the proposed multi-agent reinforcement learning framework and algorithms. The sim- ulator settings, experimental results and sensitivity analyses are demonstrated in Section 4, while the related work is presented in Section 5. Section 6 summarizes the study and outlook future research. 2 RESEARCH PROBLEM This section presents the relevant background knowledge and formulates a combinatorial optimization problem for the matching between unserved passengers and idle drivers. In a ride-sourcing system, when a passenger raises an on- demand ride request, he or she will be placed in a queue (or matching pool) and wait for being matched with a driver through the platform. The time from raising a request to being matched is referred to as the waiting time. After the passenger is matched with a driver, the driver should drive to the location of the passenger and pick him/her up (the time Fig. 1. Potential benefits of delayed matching may be matched with a much closer idle driver (much lower pick-up distance) if his/her request is delayed for a few match time intervals by the platform. However, holding passengers and drivers for too many match time intervals brings a long passengers' waiting time, under which some impatient passengers may cancel their requests. With no doubt, determining the delayed time (or the number of match time intervals to wait) for each passenger request is essential and has potentials to well balance these trade-offs. In reality, with the help of high-efficient computing resources, it is applicable for the ride-sourcing platforms to dynamically determine the delayed time of each request in response to the real-time supply-demand conditions. Nevertheless, in a highly stochastic and dynamic system, in which passengers' requests and idle drivers pop up and disappear at any time, to simultaneously determine an optimal delayed time for each passenger request is non- trivial. Although rich historical and real-time information is available, the decisions on the delayed time of each request will impact the future supply-demand conditions (such as the number of unserved requests left from the previous match time interval), which in turn affects the decisions in the following match time intervals. The dynamic interactions between the decisions and the environment (i.e. the supply- demand conditions) cannot be directly characterized by the widely used supervised learning methods. One solution to this problem is the reinforcement learning (RL) [3] that learns a policy of sequential decision makings by interacting with a complex environment. Recent years have witnessed tremendous success of RL in many transportation problems, such as taxi fleet management [4] [5], signal control [6], and adaptive routing problems [7]. However, with the traditional RL approaches that use a centralized agent, the action space that includes the decisions for the delayed time of all passengers will become extremely large. To tackle these issues, we model the dynamics of the dispatch system as a multi-agent Markov decision process (MMDP), in which each passenger request is viewed as an agent. During each match time interval, each agent decides to be delayed or not to be delayed towards the next time interval. If it is delayed, it will not join the matching pool during this match time interval (simultaneously, one match time interval is added to its delayed time); otherwise, it will join the matching pool, then a bipartite matching between unserved requests and idle drivers in the matching pool will wasted in this process is denoted as pick-up time). When a passenger drops-on a vehicle and reaches the destination, or cancels his/her request halfway (during the waiting time or the pick-up time), the life cycle of his/her request is viewed as terminated. On the supply side, at each instant, each registered ride-sourcing driver has three status: off-line, occupied, and available (idle). Off-line status means that a driver closes his/her APP and is not willing to be dispatched by the platform, occupied status indicates that a driver has been already dispatched to a passenger (on the road to pick- up the passenger or to take the passenger to the destination), while the available status implies that a driver waits for being dispatched. With no doubt, the major concerns of the online matching are the unserved passengers, and available drivers. The goal of the online matching can be formulated as a bipartite matching problem between the unserved passengers and available drivers in the matching pool at each match time interval. To be specific, during each time interval, given I unserved passengers (i = 1, 2, . . . , I) and J idle drivers (j = 1, 2, . . . , J), the objective function and constraints for optimizing the online matching can be formulated as: min aij s.t. J(cid:88) i=1 I(cid:88) I(cid:88) J(cid:88) i=1 j=1 [c(i, j) − B]aij j=1 aij ≤ 1, j = 1, 2, . . . , J (1) aij ≤ 1, i = 1, 2, . . . , I where c(i, j) is the cost of assigning driver j to passenger i, which is equal to the pick-up time from driver j to passenger i in this study. aij are the index variables to be optimized, with the following definitions: (cid:40) aij = 1, 0, if passenger i is matched with driver j if passenger i is not matched with driver j And, B is a large number which ensures at least one of the two constraints in Eq. 1 is binding. It implies that: if the number of available drivers is greater than the number of unserved passengers, then every unserved passenger will be matched (the first constraint is binding while the second constraint is non-binding); otherwise, then every available driver will be matched (the first constraint is non-binding while the second constraint is binding). As aforementioned, a passenger request may experience a lower cost if its matching is delayed for a few match time intervals. Unlike most of the previous studies that focused on developing efficient bipartite graph matching algorithms or using reinforcement learning to control drivers' decisions, the main goal of this study is to propose a method that automatically determine the delayed time of each unserved passenger request, for better balancing the trade-offs among pick-up time, waiting time and matching success rate. Our proposed method is established on top of the bipartite graph matching and determines the time when each unserved request enters the matching pool to be matched with an idle driver. Fig. 2 illustrates the online matching procedure of 3 Fig. 2. Online matching of a ride-sourcing system. a ride-sourcing system, which first determines the delayed time for each passenger and then implements the bipartite graph matching. 3 A MULTI-AGENT LEARNING FRAMEWORK This section first builds an MMDP in which each passenger request represents an agent, and the action of each agent during each match time interval is simplified as a binary choice - delayed or not delayed. Then the two reinforcement learning methods for optimal policy learning are presented. 3.1 Problem formulation The problem of determining each passenger's delayed time is modelled as an MMDP Game, with the agents, states, actions, state transitions, and rewards defined as follows. Agent: we regard each passenger request as an agent, who emerges when it is raised and terminates when it is successfully matched with a driver or cancelled by the passenger. Clearly, the number of agents in each match time interval, denoted as Nt, may vary across different match time intervals t. This multi-agent based definition can greatly reduce the magnitude of the action space in comparison to the way that defines the centralized platform as an agent. t,(G), and the local-view state si State: states in each match time interval are spatio- temporal patterns of supply and demand that the agents observe. To be specific, the state for agent i in match time interval t, si t, consists of two components: the global-view state si t,(L). Suppose the whole examined area is partitioned into M small grids (such as squares and hexagons), then global-view state vector si t,(G) includes four types of spatio-temporal supply-demand features: 1) the number of remaining unserved passengers left from the previous match time interval in each grid; 2) the number of remaining idle drivers left from the previous match time interval in each grid; 3) the expected arrival rate of new unserved passengers in each grid; 4) the expected arrival rate of new idle drivers in each grid; The latter two types of features can be predicted with historical and real-time data. Clearly, global-view state vector t,(G) ∈ RM×4. Evidently, the is of M × 4 dimensions, i.e. si global-view state is shared by all agents in match time interval t. Note that more relevant spatio-temporal features can be incorporated into the global-view state vector. The local-view state vector si t,(L) includes three components: 1) the location of agent i that is represented with one- hot encoding; 2) agent i's cumulative waiting time; 3) the expected distance from agent i to its matched driver T i t , which is estimated according to a combinatorial optimization program shown in Eq. 1. Note that here the optimization program is run virtually for getting the features, i.e. T i t for all agents, rather than for actually matching unserved passengers and idle drivers. . t , . . . , sNt t Clearly, the local-view state is different across agents and has a dimensionality of M + 2. Therefore, the state for agent i during match time interval t can be written as si agents in match time interval t can be represented st = (cid:3) ∈ RM×5+2, then the joint state of all t = (cid:2)st,(G), st,(L) (cid:111) (cid:110) t ∈ {0, 1}, where ai t , s2 s1 Action: agent i in match time interval t has two available actions, ai t = 1 denotes that the agent enters the matching pool where a bipartite matching is implemented, while ai t = 0 indicates that the agent choose to not enter the matching pool and be delayed to the next match time interval. Then the joint action of all agents in match time interval t can be written as at = t , a2 a1 t , . . . , aNt t (cid:110) (cid:111) . State Transition: When time moves from time interval t to time interval t + 1, the state of each agent gets updated according to the interactions between the agents' actions and environment. First, the agents at time interval t with action equal to 1 are assigned to idle drivers based on a combinatorial optimization algorithm. Clearly, some old agents are matched and leave the system. Second, the environment changes according to the simulator (which will be presented in the next section), and particularly, some new agents (passenger requests) may enter the system. Third, the states of the agents at time interval t + 1 (including those unmatched agents and new agents) are calculated based on the current environments. It is worthy mentioned that the dimensionality of st is not necessarily equal to the dimensionality of s(t + 1), in other words, the number of agents in match time interval t is not necessarily identical to that in match time interval t + 1. This is because the number of matched agents at time interval t is not necessarily equal to the number of newly coming agents at time interval t + 1. Reward: to examine the trade-offs on cooperation and competition among agents, two types of reward, i.e. global reward and individual reward, are considered. The individ- ual reward of agent i in match time interval t, denoted as ri t, is designed as follows. (cid:40) ri t = 0, V − c(i, j), if t < Ti if t = Ti (2) where Ti is the last match time interval of agent i (the life of an agent is terminated when it is matched or cancelled by the passenger). V refers to the positive reward (benefit) for successfully matching one pair of passenger and driver. Although this positive reward can depend on the characteristics of the request, such as the trip fare and expected trip time, we do not consider request discrimination and set a constant value V for all successfully matched agents. To encourage agents to cooperate with each other, we further introduce a notion of global reward, which is defined as the average reward of all agents. Formally, the global reward assigned to agent i in match time interval t, denoted as ¯ri t, equals to: 4 (cid:40) 0, 1 N ¯ri t = (cid:80)N (cid:80)N i [V − c(i, j)] , if t < Ti if t = Ti (3) i where N refers to the number of agents that have appeared in the whole epoch. Notice that the average reward [V − c(i, j)], is calculated at the of all agents, i.e. 1 N end of the epoch (when the life of some agents has already terminated) and then assigned to the terminal match time interval Ti for each agent. The final reward assigned to agent i in match time interval t can be written as: t + (1 − ρ) · ¯ri (4) where ρ ∈ [0, 1] is a weighted factor that balance the tradeoff between the individual reward ri t and global reward ¯ri t. Clearly, ρ = 1 indicates that each agent aims to optimize its own reward instead of considering other agents' reward, while ρ = 0 implies that each agent tries to maximize the total reward of all agents. The goal of each agent i is to maximize t=0 γtri t its own total expected discounted return Ri = (cid:80)Ti t = ρ · ri ri t where γt is a discount factor. 3.2 Simulator One of the most important component for applying deep reinforcement learning methods is the learning environment, with which the agents interact. For training multi-agent reinforcement learning algorithms, many hand-crafted envi- ronments are designed [11] [12], such as cooperative naviga- tion where agents cooperate to reach various landmarks while avoiding colliding with each other, and predator- prey where some slower agents cooperate to chase the faster adversary. Multi-agent reinforcement learning methods are shown to achieve great performance in these human- designed games. However, implementation of these methods on real-world applications is much more difficult, due to the high dimensionality of action and state space, non-stationary environment, and improper reward design. The real-world environment is of high stochasticity and randomness, which makes the training and evaluation for the reinforcement learning algorithms difficult. Moreover, the training of reinforcement learning algorithms requires a large number of epochs, which may not be supported by limited historical data. One common solution is to build simulators which are calibrated with real historical data. In this section, we design a simulator that explicitly describes the dynamics in an online matching system. As shown in Algorithm 1, each match time interval includes the following steps: policy implementation of DRL that separates agents into two groups (entering the matching pool or being delayed to the next interval), bipartite matching between unserved passengers and idle drivers in the matching pool, update of drivers' status due to completed trip and online/offline activities, and generation of new requests, etc. In essence, the simulator iteratively updates the waiting list of unserved requests and the status of drivers ("idle", "occupied" or "offline"). When a new request is generated, it is appended to the waiting list; when an unserved request is matched with an idle driver, it is removed from the waiting list. When a driver gets online (or offline), his/her status changes from "offline" to "idle" (or from "idle" to "offline"), respectively. When a driver is matched with a request or completes his/her trip, his/her status changes from "idle" to "occupied" (or from "occupied" to "idle"). Note that the number of agents N t may change with the match time interval since new agents join and some old agents may leave the waiting list of unserved requests. This setting is different from most of the previous studies related to multi-agent reinforcement learning, in which the number of agents is unchanged. It makes the design of the multi-agent learning algorithms more intractable. It is also worthy mentioned that the bipartite matching is executed virtually for estimating the expected pickup distance for agent i while preparing the joint states for the next match time interval. Algorithm 1 Simulator for an online matching system 1: Input: information of the historical passenger requests and distributions of drivers' online/offline time. 2: Initialize joint states st. 3: for every match time interval of the online dispatching system (t = 0 to T ) do 4: 5: 6: 7: 8: 9: 10: Implement policies of DRL: Execute the joint actions at = π(st) according to the policies of multi-agent DRL algorithms, where at determines whether each agent i should enter the matching pool or be delayed to t + 1. Bipartite matching: Collect unserved passenger re- quests with action equal to 1 (i.e. entering the matching pool) and idle drivers, and execute bipartite matching according to Eq. 1. Update matching outcomes: The matched requests are removed from the waiting list of unserved passenger requests; the status of the matched drivers are updated from "idle" to "occupied". Completed trips update: Update the status of drivers who completes their trips from "occupied" to "idle". Request generation: new passenger requests are boot- strapped from historical requests occurred in the same time period. The new requests are appended to the waiting list of unserved requests. Note that the number of agents Nt changes after this step. Online/offline status update: update the status of drivers who become online (available to serve) or offline (unavailable to serve) in the current interval, according to real historical distributions. Generate next states: collect the spatio-temporal demand-supply conditions and update the state for each agent, outputting st+1. Particularly, for estimating the expected distance from each agent i to its matched driver, we execute Eq. 1 with the input of all unserved agents and idle drivers. 11: end for 3.3 Reinforcement learning methods The objective of reinforcement learning is to identify a policy π that achieves optimal accumulated reward for one agent or multiple agents. For single-agent applications, 5 Q-learning is one widely used method. It estimates the expected total discounted rewards of state-action pairs, which can be approximated by a Q-function table that is solved with Bellman equation. A recent extension of Q- learning is Deep Q-network (DQN) which uses a neural network for approximating the Q-function table. DQN tries to minimize the difference between predicted Q-values and target Q-values that combines the current reward and estimated value of next state. The adaptation of Q-learning to multi-agent settings has been examined in the literature [13], however, it is still an open question and the theoretical guarantees for multi-agent DQN remain unsolved. There are two main solutions: decentralized and centralized methods. In decentralized methods, each agent is characterized with separate neural network, and the only interaction among agents is the environment. For example, [14] established a decentralized DQN in Pong game environment where two (players) agents try to beat each other (in competition setting) or keep the ball bouncing between them (in cooperative setting). However, the decentralized methods are not well adapted to our problem for the two reasons below. First, the number of agents change with intervals, thus it is hard to identify the number of neural networks to be constructed ex ante. Second, there are numerous requests popping up and disap- pearing dynamically, indicating that the number of agents is extremely large. Modelling each agent with one neural net- work must require high amount of computational resources. Therefore, in this paper, we resort to centralized methods, which use one neural network to model the behaviors of all agents, in other words, the weights of the centralized neural network is shared by all agents. The parameter of the centralized network θ is updated by minimizing the total difference between the predicted value of the Q-networks and the target Q values, as shown in Eq. 5, in terms of the transitions of all agents (cid:0)si (cid:1). As shown t+1, di t in Eq. 6, the target Q values are estimated by the obtained reward if the current state is the terminated state, and by the sum of the obtained reward and the discounted estimated value of next state otherwise. t, ai t, ri t, si (cid:104) t θ) − y)2(cid:105) L(θ) = Esi t,ai t,ri t,si t+1 (Q(si t, ai (cid:40) y = ri t, ri t + γ maxai t+1 Q(si t+1, ai t+1), if di if di t = 1 t = 0 The details of the algorithm of the proposed ST-M-DQN are shown in Algorithm 2. To avoid instable training, the replay memory is used. In each match time interval, we first sample action for each agent with -greedy policy based on the centralized Q network, then execute the joint actions in the simulator. The simulator will re-compute the spatio- temporal patterns of supply and demand and then returns rewards and next states for all agents, then the transitions of (cid:1) are stored in the reply memory. each agent(cid:0)si t+1, di t For training the centralized Q-network, we randomly sample a mini-batch of transitions from the reply memory each time, and update the network parameter θ by minimizing the loss function in Eq. 5. t, ai t, ri t, si (5) (6) Algorithm 2 ST-M-DQN 1: Initialize replay memory D; 2: Use random weights θ to initialize the action-value function. 3: for k = 1 to number of epochs do 4: Reset the environment and obtain the initial joint 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: for every match time interval (t = 0 to T ) do states s0. for i = 1 to Nt do Sample action of each agent, ai t, according to -greedy policy: ai t, ai t) with probability 1 − otherwise choose a random action. t = arg maxai Q(si t end for Execute the simulator (Algorithm 1) with the input of the joint actions at, then observe joint reward rt and next joint state st+1. For agent i, we denote di t = 1 if its life is terminated, i.e. t = Ti, and di t = 0 otherwise. t, si ∀i = 1, . . . , Nt in the replay memory D. Store transitions of all agents(cid:0)si (cid:0)si (cid:1), ∀i = 1, . . . , Nt transitions from mini-batch a t, ri t, si Sample t+1, di t (cid:1), t, ai t, ri of end for for m = 1 to M1 do t, ai t+1, di t replay memory D. Calculate target value according to Eq. 6. Update the parameters of Q-network by: θ ← θ + β1∇θL(θ)2. end for 16: 17: end for The second reinforcement learning method is the spatio- temporal multi-agent actor-critic (ST-M-A2C). Actor-critic (A2C) is a popular policy gradient method for reinforcement learning tasks. A2C establishes two networks, one policy network (known as critic and used to output policy) and one value network (known as actor and used to evaluate the performance of the policy network). It updates the parameters of the policy network θp, and that of the value network θv iteratively. Similar to the DQN, due to the large and dynamically changing number of agents, we design a centralized multi-agent A2C. The weights of the centralized value network θv are shared across all agents, and the update of θv can be achieved by minimizing a loss function presented in Eq. 7, where Vθv (Si t) is the predicted value of value network and Vtarget(Si v, π) is the target value constituted of immediate reward and discounted estimated value of next state, as shown in Eq. 8. t+1; θ(cid:48) (cid:104) L(θv) = Vθv (Si (cid:105) t+1; θ(cid:48) v, π) Vtarget(si t+1; θ(cid:48) v, π) = ri t + γVθ(cid:48) v(si t+1) ai t With the parameters of the value network θv, ST-M- A2C updates the parameters of the policy network θp by a gradient descent rule θp ← θp + β2∇θp J(θp), where β2 is the learning rate for actor, and the gradient ∇θp J(θp) is given by Eq. 9. To reduce high variability of value functions, t) − Vtarget(Si (cid:104) (cid:88) t si t) π(ai (cid:105) (7) (8) an advantage function estimated by the TD-error defined in Eq. 10 rather is used for calculating the policy gradient. 6 ∇θp J(θp) = ∇θp log πθp (ai t ai t)A(si t, ai t) A(si t, ai t) = ai t+1 + γVθ(cid:48) v t+1) − Vθv (si (si t) (9) (10) The details of the ST-M-A2C is illustrated in Algorithm 3. Algorithm 3 ST-M-A2C 1: Initialize replay memory D; 2: Use random weights θv to initialize the value network. 3: for k = 1 to number of epochs do 4: Reset the environment and obtain the initial joint 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: for every match time interval (t = 0 to T ) do states s0. for i = 1 to Nt do Sample action of each agent, ai t based on the action probability P (si t). end for Execute the simulator (Algorithm 1) with the input of the joint actions at, then observe joint reward rt and next joint state st+1. t, si t, ai 1, . . . , Nt in the replay memory D. Store transitions of all agents(cid:0)si Sample a mini-batch of transitions(cid:0)si (cid:1), ∀i = (cid:1), end for for m = 1 to M2 do t+1 t, ai t, si t+1 from the replay memory D. Update the parameters of value network θv by maximizing Eq. 7. Compute the advantage A(si t) by Eq. 10 and update the parameters of policy network θp by θp ← θp + β2∇θp L(θp), where ∇θp L(θp) is calculated with Eq. 9. t, ai end for 16: 17: end for 4 EXPERIMENTS In this section, extensive experiments and sensitivity analyses are conducted to evaluate the performance of the proposed methods and investigate the impacts of the key parameters. 4.1 Experiments on a customized environment Here we first design a customized environment for illustrat- ing the effectiveness of the proposed methods. We consider an 4 km × 4 km area with 30 match time intervals (each interval equals 1 s), and a simplified scenario the waiting passengers arrival at a rate of qd (in unit/s), while the idle drivers arrival at a rate of qs (in unit/s). Meanwhile, new arrival waiting passengers' and idle drivers' are generated based on a two-component mixture of Gaussians in each match time interval. The mean and standard deviation of arrival locations of waiting passengers are (1.2 km, 1.2 km) and (0.8 km, 0.8 km), while the mean and standard deviation of arrival locations of idle drivers are (2.8 km, 2.8 km) and (0.8 km, 0.8 km). Without considering the traffic congestions, we assume a constant vehicle speed as 25 km/h. The road distance between any two points is estimated by the Manhattan distance, and given the constant vehicle speed, then the pickup time between each pair of waiting passenger and idle driver can be identified. The whole examined area is uniformly partitioned into 10 × 10 zones. Then the spatio- temporal features or states can be extracted by counting the number of idle drivers and waiting passengers in each zone during each time interval. Note that the partitioned zones are only used for perceiving the spatio-temporal features. Other features, such as the agent's expected pickup distance, are calculated based on a continuous space. The two proposed approaches, ST-M-DQN and ST-M- A2C, are compared with two benchmarks: pure optimization strategy and tabular Q-learning. Pure optimization strategy matches idle drivers and waiting passengers immediately without considering any delayed matching with a combi- natorial optimization depicted in Eq. 1. In this simplified customized environment, we do not consider the request cancelling behaviors of passengers, and thus the answer rate (proportion of the successfully matched passengers) is always equal to 1 if supply (idle drivers) is larger than or equal to demand (waiting passengers). The tabular Q-learning learns a Q-table that maps the states to action, with -greedy policy. In tabular Q-learning, the state only includes the location and time of the agents, and thus the Q-table has a dimension of T × N × 2, where N is the total number of separated zones. Three metrics are used for comparing the effectiveness of the proposed models and benchmarks: mean reward of each agent (Ri for agent i), answer rate and mean pickup time. The reward function is specified as follows. Let the positive reward (benefit) for successfully matching one passenger- driver pair, V = 800s. It is noteworthy that V is a decision variable determined by the platform and reflects the trade- offs between different objectives. A large V indicates that the platform focuses more on successfully matching drivers and passengers with less considerations on the cost of pickup time, and vice versa. Therefore, the decision variable V can be adjusted according to the objectives of different platforms. To compare the effectiveness of the models under different environments, three environmental settings are implemented: qd=qs=1 unit/s, qd=qs=2 unit/s, and qd=qs=3 unit/s. Both the value function approximation networks and policy networks are three-layer networks, with 512, 256, and 128 neurons from the first to last hidden layer. The activations of all hidden units are ReLu, while output layers of the value function approximation networks and policy networks use Linear and Softmax activations, respectively. All the experiments are repeated by 5 times to ensure the robustness of the results. Fig. 3 shows the resulting mean reward of each agent, answer rate, and mean pickup time of Pure optimization, Q- learning, ST-M-DQN, ST-M-A2C on the three environments. The results show that the mean reward of each agent increases with the arrival rate of drivers and passengers. This is intuitive since a better bipartite matching (with lower mean pickup time) becomes available as the increase of the density (dominated by the arrival rates) of drivers and passengers. In addition, it can be observed that the two proposed reinforcement learning methods significantly outperform the pure optimization and the tabular Q-learning methods in terms of the mean reward of each agent (which is the objective of the algorithms). The ST-M-A2C achieves 7 the best performance in all environmental settings, which demonstrates its robustness. From Fig. 3 (b)-(c), we find that the delayed matching controlling by the two deep multi-agent reinforcement learning methods can significantly reduce the mean pickup time of passenger-driver pairs with little loss on the answer rate. For example, when qd=qs=1 unit/s, the implementation of ST-M-A2C reduced the mean pickup time by 14.3% while decreasing the mean answer rate by 3.8%, resulting in 19.1% of increase in agents' mean reward. Fig. 4 shows the trends of average reward of agents with respect to training epochs of the three examined reinforce- ment learning models under different environment settings. Results show that ST-M-A2C and ST-M-DQN significantly outperform Tabular Q-learning in terms of converging speed, converged agents' reward, and robustness. In general, ST-M- A2C shows more robust training curves and achieves higher converged agents' reward than ST-M-DQN. t and total reward ¯ri 4.2 Sensitivity analysis in terms of reward weighted factor Next we conduct sensitivity analysis to look at the effects of reward weighted factor ρ for the training of the pro- posed deep multi-agent reinforcement learning methods. As aforementioned, ρ measures the tradeoff between individual reward ri ti. The higher the ρ, the more emphasis that each agent pays to its individual reward versus total reward of all agents. It is naturally expected that using the total reward as signals for agents has potential to achieve better overall system performance. However, as [15] mentioned, the use of total reward generates weak incentives to each agent, and may make some agents "lazy" to obtain rewards. Consequently, the resulting total rewards of all agents can be even worse than the algorithms using individ- ual reward for each agent. Here, we train and test ST-M-A2C under the environmental settings of qd = qs = 1 unit/s, given five groups of reward weighted factor ρ (from 0.0 to 1.0 with a step of 0.25). Table 1 demonstrates the answer rate, average pickup time and average agents' reward of ST-M-A2C given different reward weighted factor ρ. Fig. 5 further plots the training trends of the agents' average reward with episodes under different reward weighted factors. The results show that the effectiveness and robustness of ST-M-A2C increases with the reward weighted factor ρ, indicating that the multi-agent framework behaves better when they are rewarded by their individual rewards in our problem. The possible reason is that the global reward (the sum of rewards of all agents) may be not strong enough to guide each agent and may encourage the "lazy" behaviors of agents. Those "lazy" agents do not learn from trial and error but simply select action 1 (not- delayed action) when they interact with any environment states. This observation is also consistent to the previous studies in the domain of taxi order dispatching or fleet management [1] [4] which normally use individual rewards as incentives for each agent in the multi-agent environments. 4.3 Model performances on a real environment Apart from the customized environment, we further evaluate the performances of the proposed methods based on a 8 (a) Mean reward of agent (b) Mean pickup time (c) Answer rate Fig. 3. Model comparisons on a customized environment. (a) Arrival rates qd = qs = 1 unit/s (b) Arrival rates qd = qs = 2 unit/s (c) Arrival rates qd = qs = 3 unit/s Fig. 4. Convergence curves of different models. Model performances under different reward weighted factors TABLE 1 Model ST-M-A2C ST-M-A2C ST-M-A2C ST-M-A2C ST-M-A2C Pure Optimization Weighted factor ρ Answer rate Mean pickup time (s) Average reward (s) 0 0.25 0.5 0.75 1 - 0.997 1.000 0.986 0.986 0.944 1.000 494.68 474.34 439.11 436.37 405.65 495.56 302.36 323.88 352.70 355.70 371.13 302.17 Fig. 5. Evolution of episode mean reward under different reward weighted factor ρ simulator calibrated with real mobility data. The dataset, provided by the largest ride-sourcing platform in China, Didi Chuxing, includes four-week orders and drivers' information in the downtown area (20km × 20km) of a city of China. As shown in Algorithm 1, the input of the simulator includes a table of passenger requests and a table of drivers' status. The former table contains the following trip-based attributes: the time and location (longitude and latitude) at which the passenger requests her order, the destination location, and the trip duration (from the time at which the passenger gets aboard to the time at which she drops off), and passenger ID. These attributes are extracted from the real-happened trips. The later table records the following driver-based attributes: the time and location at which the driver gets online and starts listening to orders, the time at which she gets offline and terminates her work, and driver ID. These attributes of drivers are also retrieved from the actual ride-sourcing drivers who register with the platform. With these two calibrated tables as input, the simulator with Algorithm 1 can be executed. The performances in terms of average pickup time, answer rate and reward are evaluated on the four models, including the two proposed models, ST-M-DQN, ST- M-A2C, and two baselines, Pure optimization, Q-learning. In terms of spatio-temporal features, we partition the examined 7;07,3/!,88030777;,7,908:398 0,3584/0#0,7/147,.039!:704592,943" 0,733$% "$% 7;07,3/!,88030777;,7,908:398 0,3!.:5%20!:704592,943" 0,733$% "$% 7;07,3/!,88030777;,7,908:398 0,33807#,90!:704592,943" 0,733$% "$% 54.08 0,370,7/41,0398%,-:,7" 0,733$% "$% 54.08 0,370,7/41,0398%,-:,7" 0,733$% "$% 54.08 0,370,7/41,0398%,-:,7" 0,733$% "$% 584/08 584/00,3#0,7/= = = = = Effectiveness of models in comparison to pure optimization TABLE 2 Agent Increases in answer rate Increases in mean pickup time Increases in average reward Tabular Q-learning ST-M-DQN ST-M-A2C -0.78% -0.508% -0.468% -1.408% -1.318% -12.938% 0.098% 0.108% 6.028% Fig. 6. Evolution of episode scaled reward in real environment area into 20 × 20 zones, each of which occupies an area of 1 square kilometer. At each step, the number of idle drivers and the number of waiting passengers within each zone are counted and fed into the deep reinforcement models as spatio-temporal features. Table 2 shows the performance measures (answer rate, mean pickup time and mean reward of agents) of the three reinforcement learning models compared to pure optimization. Fig. 6 shows the evolution of the mean reward of agents (which is scaled to a range from 0 to 1 due to privacy concerns) with episodes. It can be observed ST-M- A2C has the best performance; it significantly reduces the mean pickup time by 12.93% with minor decrease on answer rate (by 0.46%), and as a result achieves 6.02% higher mean reward of agents than pure optimization. However, neither of ST-M-DQN and Tabular Q-learning has a good performance in an environment calibrated by real historical data. From Fig. 6, we can also observe that the training of ST-M-A2C is more robust and effective than ST-M-DQN and Tabular Q-learning. As a supplement to the experiments on a customized environment, this experiment on a simulator calibrated by real mobility data further evaluates the effectiveness of the proposed methods. 5 RELATED WORK Taxi dispatch: Taxi dispatch, or driver dispatch, is a term usually referring to the process of matching vacant drivers with passengers' requests using some algorithms to maxi- mize the system's performance. Traditional dispatch systems maximize the driver acceptance rate for each individual order by sequentially dispatching taxis to riders. [16] proposed to 9 dispatch taxis to server multiple bookings at the same time thus to maximize the global success rate. [17] considered the individual participant's benefit and proposed a notion of a stable match. [18] constructed an end-to-end framework to predict the future supply and demand in order to optimally schedule the drivers in advance. [16] investigated the pre- ferred service and proposed a recommendation system to enhance the prediction accuracy and reduce the user's effort in finding the desired service. [1] proposed an order dispatch algorithm that combined an offline learning step and an online planning step. The offline learning step estimated the value functions for the spatio-temporal patterns of passenger demand and taxi supply, then the online planning step executed an optimal matching between drivers and orders with the learned value functions. Their main focus was to enhance the traditional combinatorial optimization algorithm, such as Kuhn-Munkres (KM) algorithm, by adding the action- space values to the objective function. [4] proposed a multi- agent reinforcement learning framework to tackle the fleet management problem that relocated the idle taxis to improve taxi utilization rate. [9] proposed a Mean Field Multi-Agent Reinforcement Learning for order dispatching in Didi's platform. They found that the mean field approximation was able to globally capture the demand and supply dynamics by propagating many local interactions between agents and the environment. [10] combined a deep Q-network with transfer learning techniques in a large-scale online order dispatching system. They showed that the strategies learned by knowl- edge transfer from sources city to target cities remarkably improved the system efficiencies, compared to strategies without transfer learning. However, as aforementioned, none of these studies have investigated the impacts of matching time intervals as well as the potential benefits of delayed matching. Deep Reinforcement Learning: Deep Reinforcement Learning (DRL) is a rapidly developed field which has attracted great attention especially since the emergence of AlfaGo [19], [20]. DRL combines the advantages of both deep learning (DL) and reinforcement learning (RL), where DL learns abstract or hidden features from large-scale data with the capability of using multiple processing layers [21] and RL enables agent learning by interacting with its environment. Regarding the application of DRL in the transportation system, it is a multi-agent problem in the high-dimensional and non-stationary space. The powerful DRL provides a possible way to solve the multi-agent system problem, which is conventionally encountered in a variety of domains including robotics, control and communication [22]. [14] evaluated the cooperation and competition between different agents when they share same environment. [23] investigated the wireless sensor networks using a multi-agent framework and found that cooperative neighbors effectively help the sensor to relay packets. 6 CONCLUSIONS This study proposes a two-stage framework for online matching in ride-sourcing systems. The lower part contains a traditional convex combinatorial optimization algorithm that matches idle drivers and waiting passengers in the matching 54.08 $.,0/70,7/%,-:,7" 0,733$% "$% 10 [12] I. Mordatch, and P. Abbeel. "Emergence of grounded composi- tional language in multi-agent populations". In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. [13] H. M. Schwartz. "Multi-agent machine learning: A reinforcement approach". John Wiley & Sons, 2014. [14] A. Tampuu, T. Matiisen, D. Kodelja, I. Kuzovkin, K. Korjus, J. Aru, ... and R. Vicente. "Multiagent cooperation and competition with deep reinforcement learning". PloS one, 12(4), e0172395, 2017. [15] P. Sunehag, G. Lever, A. Gruslys, W. M. Czarnecki, V. Zambaldi, M. Jaderberg, ... and T. Graepel. "Value-decomposition networks for cooperative multi-agent learning". arXiv preprint arXiv:1706.05296, 2017. [16] L. Zhang, T. Hu, Y. Min, G. Wu, J. Zhang, P. Feng, ... and J. Ye. "A taxi order dispatch model based on combinatorial optimization". In the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017. [17] X. Wang, N. Agatz, and A. Erera. "Stable matching for dynamic ride-sharing systems". Transportation Science, 2017. [18] D. Wang, W. Cao, J. Li, and J. Ye. "DeepSD: supply-demand predic- tion for online car-hailing services using deep neural networks". In 2017 IEEE 33rd International Conference on Data Engineering (ICDE), 2017. [19] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, ... and S. Dieleman. "Mastering the game of Go with deep neural networks and tree search". nature, 529(7587), 484, 2016. [20] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton and Y. Chen. "Mastering the game of Go without human knowledge." Nature, 550(7676), 354, 2017. [21] Y. LeCun, Y. Bengio, and G. Hinton. "Deep learning". nature, 521(7553), 436, 2015. [22] L. Busoniu, R. Babuska, and B. De Schutter. "A comprehensive survey of multiagent reinforcement learning". IEEE Transactions on Systems, Man, And Cybernetics-Part C: Applications and Reviews, 38 (2), 2008. [23] D. Ye, M. Zhang, and Y. Yang. "A multi-agent framework for packet routing in wireless sensor networks". sensors, 15(5), 10026-10047, 2015. pool with minimum cost measured by pickup time. The upper part establishes two multi-agent deep reinforcement learning models, named ST-M-DQN and ST-M-A2C, which serve as multiple gates to control whether an agent should enter the matching pool in each matching time interval. These reinforcement learning models essentially determine the delayed time of each agent and help improve the effec- tiveness of the matching process, while the potential benefits of delayed matching are discussed in the literature. Through extensive experiments, we show that the proposed ST-M- DQN and ST-M-A2C well balance the trade-off between pickup time and matching rate and significantly improve the matching effectiveness in comparison to pure optimization and other benchmarks. It shows that the delayed matching controlled by the reinforcement learning methods can indeed remarkably reduces the average pickup time by incurring little loss on answer rate. This paper provides a novel frame- work that combines reinforcement learning and bipartite matching and evaluate its effectiveness in a ride-sourcing online matching system. ACKNOWLEDGMENTS The work described in this paper was supported by Hong Kong Research Grants Council under projects HKUST16222916, NHKUST627/18 and the National Nat- ural Science Foundation of China under projects 71622007, 7181101024. REFERENCES [1] Z. Xu, Z. Li, Q. Guan, D. Zhang, Q. Li, J. Nan, ... and J. Ye. "Large-Scale Order Dispatch in On-Demand Ride-Hailing Platforms: A Learning and Planning Approach". In the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018. [2] H. Yang, J. Ke, and J. Ye. "A universal distribution law of network detour ratios". Transportation Research Part C: Emerging Technologies, 96, 22-37, 2018. [3] R. S. Sutton, and A. G. Barto. "Introduction to reinforcement learning". Cambridge: MIT press, 1998. [4] K. Lin, R. Zhao, Z. Xu, and J. Zhou. "Efficient Large-Scale Fleet Management via Multi-Agent Deep Reinforcement Learning". In the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2018. [5] T. Oda, and C. Joe-Wong. "MOVI: A Model-Free Approach to Dynamic Fleet Management". arXiv preprint arXiv:1804.04758, 2018. [6] H. Wei, G. Zheng, H. Yao, and Z. Li. "Intellilight: A reinforcement learning approach for intelligent traffic light control". In the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018. [7] C. Mao and Z. Shen. "A reinforcement learning framework for the adaptive routing problem in stochastic time-dependent network". Transportation Research Part C: Emerging Technologies, 93, 179-197, 2018. [8] I. Jindal, Z. T. Qin, X. Chen, M. Nokleby, and J. Ye. "Optimizing Taxi Carpool Policies via Reinforcement Learning and Spatio-Temporal Mining". In 2018 IEEE International Conference on Big Data (Big Data) (pp. 1417-1426). IEEE, 2018. [9] M. Li, Y. Jiao, Y. Yang, Z. Gong, J. Wang, C. Wang, ... and J. Ye. "Efficient Ridesharing Order Dispatching with Mean Field Multi- Agent Reinforcement Learning". In ACM WWW (Web Conference 2019), May 2019, San Francisco, USA, 2019. [10] Z. Wang, Z. Qin, X. Tang, J. Ye, and H. Zhu. "Deep Reinforcement Learning with Knowledge Transfer for Online Rides Order Dispatch- ing". In 2018 IEEE International Conference on Data Mining (ICDM) (pp. 617-626). IEEE, 2018. [11] R. Lowe, Y. Wu, A. Tamar, J. Harb, O. P. Abbeel, and I. Mordatch. "Multi-agent actor-critic for mixed cooperative-competitive environ- ments". In Advances in Neural Information Processing Systems, 2017.
0911.3723	1	0911	2009-11-19T08:38:30	Applications of the Dynamic Distance Potential Field Method	[ "cs.MA" ]	Recently the dynamic distance potential field (DDPF) was introduced as a computationally efficient method to make agents in a simulation of pedestrians move rather on the quickest path than the shortest. It can be considered to be an estimated-remaining-journey-time-based one-shot dynamic assignment method for pedestrian route choice on the operational level of dynamics. In this contribution the method is shortly introduced and the effect of the method on RiMEA's test case 11 is investigated.	cs.MA	cs	Applications of the Dynamic Distance Potential Field Method Tobias Kretz PTV Planung Transport Verkehr AG Stumpfstrasse 1 D-76131 Karlsruhe [email protected] June 13, 2021 Abstract Recently the dynamic distance potential field (DDPF) was intro- duced as a computationally efficient method to make agents in a simu- lation of pedestrians move rather on the quickest path than the short- est. It can be considered to be an estimated-remaining-journey-time- based one-shot dynamic assignment method for pedestrian route choice on the operational level of dynamics. In this contribution the method is shortly introduced and the effect of the method on RiMEA's test case 11 is investigated. 1 Introduction In traffic science a lot of effort is put into work discussing dynamic assign- ment methods: trying to find proofs on the existence of an equilibrium and/or implementing it in an efficient manner for the simulation of real- world networks. This effort is justified by the conviction that Wardrop's first principle "The journey times in all routes actually used are equal and less than those which would be experienced by a single vehicle on any un- used route." [1] is at least a good approximation for real world traffic flow distribution. Compared to this strong emphasize of (estimated remaining) journey time in vehicular traffic science as a determinant of route choice, explicit discussion of the influence on the dynamics of pedestrians [2, 3] is surpris- ingly scarce. Although evidently in case of emergency it's their egress time that occupants desire to minimize and not their egress path length, models of pedestrian evacuation dynamics are often constructed such that various 1 influences change the basic direction of desired velocity, which is calculated as a gradient of the field of distances toward the exit and not a field of es- timated remaining journey time. While this is not to say that pedestrian traffic always moves on quickest paths -- there are probably more occasions than in vehicular traffic when it deviates from it -- there are at least numer- ous situations, where it is close to travel time referred user-equilibrium or at least travel time is relevant for route choice. Recently introduced dynamic distance potential fields [4, 5] do not hold information on remaining journey time either, but -- on any spatial scale -- on the load of the remaining path and implicitly the possibility to detour jams. With this information it is possible to shift the agents' walking behavior toward user-equilibrium. In the two subsequent sections of this paper at first the method of dy- namic distance potential fields is shortly sketched and then applied in a simulation of RiMEA's test case 11 [6]. 2 Dynamic Distance Potential Field Let's begin with a summary of the method: a dynamic distance potential field can be imagined as a "shadow" (as shown in figure 1) of agents cast "upwards" the static potential, i.e. away from the destination. One can think of a destination area as a source of light sending light rays that travel under consideration of obstacles on approximate geodisces away from the destination area. When an agent or a group of agents throws a shadow, successive agents will try to avoid the shadow, as it is a hint for upcoming delays. A shadow is the longer the larger and denser a group of agents (i.e. a jam) is. This functionality is achieved using flood fills, where a value of 1 is added, if a cell is unoccupied and some larger value sadd, if it is occupied by an agent. But the two simple flood fill methods (over common edges or common edges and corners) result in Manhattan or Chebyshev but not the desired Euclidean metric. In this sense these two methods produce large errors and unwanted artifacts in the movement. However, the errors can be reduced for once by combining the two methods and second by using as heuristic not the time dependent potential, but its difference to the potential calculated on the empty (unoccupied) geometry. For two mutually visible coordinates (x0, y0) and (x1, y1), Manhattan (1) (2) distance (i.e. a distance according to Manhattan metric) is dM = x1 − x0 + y1 − y0 = ∆x + ∆y and Chebyshev distance is dC = max(∆x,∆y) 2 Figure 1: A visualization of the dynamic distance potential field. The green line is the exit, agents are located on the bright white spots, which by that have a large value of the dynamic distance potential field. The darker a spot is, the smaller is its dynamic potential field value. The gradient of the dynamic distance potential field enters into the calculation of the direction of desired motion. 3 From this the "minimum distance" follows: and therefore the Euclidean distance is dm = dM − dC = min(∆x,∆y) (cid:113) (cid:113)∆x2 + ∆y2 dE = (dC)2 + (dm)2 = (3) (4) Combining the whole potentials cell by cell in this manner -- even if cells are not mutually visible -- will now be called "method V1". All mentioned metrices and norms are depicted in figure 2. A discussion of the resulting errors of this method compared to Euclidean distance is given in [7]. Let's call the potential calculated on the empty (no agents) geometry V 1 and the potential after a certain time step SV 1(t). Then the heuristic S0 influencing the motion of pedestrians is Sdyn(t) = SV 1(t) − S0 V 1. This has been used and coupled to the F.A.S.T. model [8 -- 12] as an additional partial probability pdyn = e−kSdynSdyn(t), (5) multiplied to the original probability that a cell is selected as destination cell and normalized accordingly p = N · p · pdyn. (6) Figure 2: F.l.t.r.: Manhattan metric, Chebyshev metric, Minimum norm, and Variant 1 (which here is identical to Euclidean metric). 3 RiMEA Test Case 11 In RiMEA's test case 11 a group of 1000 agents is meant to leave a room using two doors, which have different distance to the agents' starting area. The geometry of test case 11 is shown in figure 3. RiMEA prescribes an exit width of 1 m for the simulations. As space is discretized into cells of width 40 cm in the F.A.S.T. model, the exit width has been set to 80 cm. The parameters of the F.A.S.T. model were set to kS = 1.0 and any other k = 0. The speeds were normal distributed (3.5 cells/s, ± 1.0 cell/s) with cut-offs at 1 and 4 cells/s. All results given are averages of 100 simulation runs. In the description of the test case the shape and position of the actual destination area, i.e. the area from which agents are taken out of the sim- ulation upon arrival, is not defined. Therefore three variants were tested, which are shown in figure 4. 4 Figure 3: Test case 11 of RiMEA. Figure 4: Three variants of geometric exit modelling. Agents are taken out of the simulation as soon as they reach the green area, obstacles are grey, free space black. Note that in any case there is only one destination area with one static floor field, such that the exit choice method proposed in [10] is not made use of in the simulations for this contribution. 5 Exit geometry 1 2 3 588.8 447.9 539.5 474.9 289.1 400.4 T w/o DDPF T with DDPF Ti w/o DDPF Ti with DDPF 270.4 212.6 248.0 229.1 143.5 191.3 Table 1: Total times T and average individual egress times Ti without and with dynamic distance potential field. In the latter case kSdyn = 1.0 and sadd = 10. Exit geometry w/o DDPF with DDPF 1 2 3 407.0 198.2 338.9 495.7 483.4 492.3 Table 2: Number of agents that leave through the right exit. Tables 1 and 2 show in numbers the effect the dynamic distance potential field has. Simulated evacuation times shrink as the agents distribute more equally on the two exits. As a by-observation the effect of the exit geometry becomes obvious. Ammending table 2 figure 5 shows for exit geometry 2 the dependence of the number of agents leaving through the right exit in dependence on parameter sadd. Figure 5: Number of agents leaving through the right exit in dependence on parameter sadd (exit geometry 2). The difference becomes even more obvious, if one looks at the situation after 100 seconds in figure 6. One gets the impression that without dynamic distance potential field the agents are either only in small numbers "aware" of the existence of the right exit or are reluctant to walk five extra meters. This is totally different with dynamic distance potential field. Here it's almost difficult to distinguish the shape of the crowd from a geometry, where 6 there is only one exit placed between the two actually existing ones. Figure 6: Situation after 100 seconds without (left) and with (right) dynamic distance potential field. 4 Summary and Conclusions The dynamic distance potential field method has shortly been described. It is a computationally efficient heuristic method to mimic the effects of esti- mated remaining journey time on route choice on the operational level of pedestrian dynamics. For fundamental reasons, but also to keep computa- tion times small, deviations from real remaining journey time and limited errors in the metric of the field are accepted. The method has been applicated in a simulation of RiMEA test case 11. This led to the agents utilizing the exit 5 m further away almost as well as the nearby exit, which could best be seen in a snapshot from the simulation's animation. As long as there is no experiment for test case 11 or an observation that resembles test case 11, it's not safe to say that the dynamic distance potential field makes the simulation in this case more realistic. But with the two parameters that control the method a planner has the flexibility to model almost any point between strong preference for shorter walking paths and strong preference for shorter walking times, respectively shorter evacuation times. References [1] J.G. Wardrop, "Some Theoretical Aspects of Road Traffic Research. Proceedings", Institution of Civil Engineers 11 (1952) no. 1, 325 -- 378. [2] A. Schadschneider, W.W.F. Klingsch, H. Klupfel, T. Kretz, C. Rogsch, and A. Seyfried, "Evacuation Dynamics: Empirical Results, Modeling and Applications", in Meyers [18], p. 3142. arXiv:0802.1620 [physics.soc-ph]. ISBN:978-0-387-75888-6. 7 [3] A. Schadschneider, H. Klupfel, T. Kretz, C. Rogsch, and A. Seyfried, "Fundamentals of Pedestrian and Evacuation Dynamics", in Bazzan and Klugl [17], ch. VI, pp. 124 -- 154. ISBN:978-1-60566-226-8. [4] T. Kretz, "Pedestrian Traffic: on the Quickest Path", Journal of Statistical Mechanics: Theory and Experiment P03012 (2009) , arXiv:0901.0170 [physics.soc-ph]. [5] T. Kretz, "The use of dynamic distance potential fields for pedestrian flow around corners", in First International Conference on Evacuation Modeling and Management. TU Delft, 2009. arXiv:0804.4336 [cs.MA]. contribution to ICEM09 conference, eprint. [6] U. Brunner, H. Kirchberger, C. Lebeda, M. Oswald, R. Konnecke, M. Kraft, A. Thoss, L. Mulli, A. Seyfried, C. Hartnack, S. Wader, G. Spennes, and T. Kretz, RiMEA -- Richtlinie fur Mikroskopische Entfluchtungs-Analysen. Initiatoren des RiMEA-Projekts: M. Schwendimann, N. Waldau, P. Gattermann, C. Moroge, T. Meyer-Konig, and M. Schreckenberg, 2.2.0 ed., 2009. (eprint from http://www.rimea.de/). (in German). [7] T. Kretz, C. Bonisch, and P. Vortisch, "Comparison of Various Methods for the Calculation of the Distance Potential Field", in Rogsch et. al. [16], pp. 335 -- 346. arXiv:0804.3868 [physics.comp-ph]. ISBN: 978-3-642-04503-5. [8] T. Kretz and M. Schreckenberg, "The F.A.S.T.-Model", in El Yacoubi et. al. [13], pp. 712 -- 715. arXiv:0804.1893 [cs.MA]. ISBN:3-540-40929-7. [9] T. Kretz and M. Schreckenberg, "F.A.S.T. -- Floor field- and Agent-based Simulation Tool", in Transport simulation: Beyond traditional approaches, E. Chung and A.-G. Dumont, eds., ch. 8, pp. 125 -- 135. EPFL press, Lausanne, CH, April, 2009. arXiv:physics/0609097. ISBN:978-1420095098. [10] T. Kretz, Pedestrian Traffic -- Simulation and Experiments. PhD thesis, Universitat Duisburg-Essen, February, 2007. urn:nbn:de:hbz:464-20070302-120944-7. [11] T. Kretz, M. Kaufman, and M. Schreckenberg, "Counterflow Extension for the F.A.S.T.-Model", in Umeo et. al. [14], pp. 555 -- 558. arXiv:0804.4336 [cs.MA]. ISBN:978-3-540-79991-7. [12] T. Kretz, "Computation Speed of the F.A.S.T. Model", in Traffic and Granular Flow '09, S. Dai et al., ed. 2010. arXiv:0911.2900 [cs.MA]. (to appear in TGF09 proceedings). 8 [13] S. El Yacoubi, B. Chopard, and S. Bandini, eds., Cellular Automata - 7th International Conference on Cellular Automata for Research and Industry, ACRI 2006. Springer-Verlag Berlin Heidelberg, Perpignan, France, September, 2006. ISBN:3-540-40929-7. [14] H. Umeo, S. Morishita, K. Nishinari, T. Komatsuzaki, and S. Bandini, eds., Cellular Automata -- 8th International Conference on Cellular Automata for Research and Industry, ACRI 2008, Proceedings. Springer-Verlag Berlin Heidelberg, Yokohama, Japan, September, 2008. ISBN:978-3-540-79991-7. [15] N. Waldau, P. Gattermann, H. Knoflacher, and M. Schreckenberg, eds., Pedestrian and Evacuation Dynamics 2005. Springer Berlin Heidelberg New York, Vienna, December, 2006. ISBN:978-3-540-47062-5. [16] C. Rogsch, A. Schadschneider, W.W.F. Klingsch, and M. Schreckenberg, eds., Pedestrian and Evacuation Dynamics 2008. Springer Berlin Heidelberg New York, Wuppertal, 2009. ISBN: 978-3-642-04503-5. [17] A. Bazzan and F. Klugl, eds., Multi-Agent Systems for Traffic and Transportation Engineering. Information Science Reference, Hershey, PA, USA, April, 2009. ISBN:978-1-60566-226-8. [18] R.A. Meyers, ed., Encyclopedia of Complexity and Systems Science. Springer, Berlin Heidelberg New York, 2009. ISBN:978-0-387-75888-6. 9
1805.09090	1	1805	2018-05-23T12:30:23	Volunteers in the Smart City: Comparison of Contribution Strategies on Human-Centered Measures	[ "cs.MA" ]	Several smart city services rely on users contribution, e.g., data, which can be costly for the users in terms of privacy. High costs lead to reduced user participation, which undermine the success of smart city technologies. This work develops a scenario-independent design principle, based on public good theory, for resource management in smart city applications, where provision of a service depends on contributors and free-riders, which benefit from the service without contributing own resources. Following this design principle, different classes of algorithms for resource management are evaluated with respect to human-centered measures, i.e., privacy, fairness and social welfare. Trade-offs that characterize algorithms are discussed across two smart city application scenarios. These results might help Smart City application designers to choose a suitable algorithm given a scenario-specific set of requirements, and users to choose a service based on an algorithm that matches their preferences.	cs.MA	cs	Volunteers in the Smart City: Comparison of Contribution Strategies on Human-Centered Measures Stefano Bennati, Ivana Dusparic, Rhythima Shinde, Catholijn M. Jonker Keywords: Participatory Sensing , Smart Cities , Public Good , Privacy , Fairness This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Abstract Several smart city services rely on users contribution, e.g., data, which can be costly for the users in terms of privacy. High costs lead to reduced user participation, which undermine the success of smart city technologies. This work develops a scenario-independent design principle, based on public good theory, for resource management in smart city applications, where provision of a service depends on contributors and free-riders, which benefit from the service without contributing own resources. Following this design principle, different classes of algorithms for re- source management are evaluated with respect to human-centered mea- sures, i.e., privacy, fairness and social welfare. Trade-offs that character- ize algorithms are discussed across two smart city application scenarios. These results might help Smart City application designers to choose a suitable algorithm given a scenario-specific set of requirements, and users to choose a service based on an algorithm that matches their preferences. 1 Introduction Recent years have seen a substantial increase in active user participation in the smart city, and with it an increase in the resources contributed by the citizens. Sensor data is one type of user-contributed resource that is at the base of many smart city applications [1]. Collecting data in a smart city allows to predict the needs of the citizens [2], thus enabling the creation of more advanced and more efficient services with a high potential for innovation [3]. User participation is crucial for the success of several smart city applications, but it entails costs that disincentivize users. For example, transmitting privacy-sensitive data in a participatory sensing scenario increases the risk of disclosure and misuse of private information, e.g., discrimination. Similarly, increasing energy availability in the smart grid scenario, e.g., by postponing the use of appliances, comes with a risk of unfair treatment and disproportinatelly low access to the resource [4]. In order for the 1 smart city applications to be successful, these costs have to be reduced. Examples of existing solutions for reducing contribution costs are privacy-enhancing technologies [5] that reduce the risks associated with disclosing information [6], and fair resource- management technologies [7], which improve the perceived fairness of the system. Comparing different implementations of these mechanisms is difficult because they are tied to specific assumptions about the scenario and the resource. To enable a comparison of different scenarios, resources and contribution strategies, this paper introduces a design principle for developing smart city algorithms, which allows the evaluation of privacy-enhancing technologies in a scenario-independent approach. Sev- eral smart city services are produced from user-contributed data. Proposed approach is based on the theory of public goods and voluntary contribution games [8], as it is an approach well suited for modeling scenarios where a common resource, i.e. the smart-city service, depends on the action of a population [9]. The orchestration of de- mand and offer of some resource is driven by contribution strategies, algorithms that determine which users should contribute to the system at each point in time, based on the state of the system and a set of system requirements. By acting on the decision of whether to participate, a contribution strategy is independent of the characteristics of the resource, hence it can be combined with existing mechanism that work at the content level, e.g., privacy-preserving algorithms. Additionally, it makes data shar- ing across smart city applications easier, thanks to the uniform way of dealing with different kinds of data. Following this design principle, a simulation framework is developed [10] that allows a comparison of multiple algorithms for resource contribution, to address the research question: how do different contribution strategies compare in terms of privacy, fairness and social welfare? The first contribution of this paper is to introduce the design principle and verify its applicability to two different application scenarios, i.e., traffic congestion information and charging of electric vehicles, by means of real-world datasets. Similar validation in different scenarios can provide useful insights for the design of other resource-based smart city applications. The second contribution is to showcase the generality of the framework by imple- menting two classes of algorithms -- centralized optimization and localized reinforce- ment learning in two flavors, with and without contextual awareness -- in these two smart city scenarios and evaluating them in terms of trade-offs between system effi- ciency, user privacy or resource distribution fairness, as opposed to optimizing them towards a single measure. The results presented show that the same trade-offs between algorithms repeat across the application scenarios. Specifically, a localized solution, which works only on local data, is found to deliver higher privacy and equality than a centralized solution. Centralized optimization offers instead higher efficiency thanks to the global knowledge of data. Localized algorithms with and without contextual knowledge are evaluated, and those considering the context in the decision-making are found to deliver higher fairness. This work can be of interest to designers of smart city applications and services that look for a guideline for the choice of algorithm, given specific scenario priorities in terms of different measures. Another target audience is that of citizens that are concerned with the risk of abuse of their data, and particularly privacy-aware prosumers. The results help quantify different service implementations along measures such as privacy and fairness, thus help the citizens choose a service provider that best matches their preferences, reducing the concerns about the system and fostering user participation. 2 The rest of the paper is organized as follows: Section 2 introduces the design principle and how it can be applied to two smart city scenarios, Section 3 describes and discusses the main results obtained from simulating voluntary contributions in different scenarios, and Section 4 concludes the paper discussing possible avenues for future work and giving general recommendations about the choice of contribution strategies. 2 Model This section introduces the design principle, based on the theory of public goods and Voluntary Contribution Games [8], and it describes how it can be applied to the smart city scenarios of participatory sensing, with the example of traffic congestion [11], and smart grids, with the example of electric vehicle charging [12]. Voluntary contribution games involve the need by a number of users to coordinate the provision of some public good, e.g., financing a playground with private donations. Table 1 illustrates the notation of the model, symbols are listed in order of appearance. Description The service provider Math symbol S V = {1, . . . , n} The set of n users T ∈ N>0 The number of rounds in the simulation rt The resource produced by user i at time t i ∈ R i vt The value associated to disclosing rt i ∈ R i The cost associated to transmitting rt ct i ∈ R i pt The privacy leaked when disclosing rt A = {D, C} i The action set i ∈ A at The action of user i at time t + ⊆ V At The set of contributors at time t qt ∈ R The service quality at time t f : V → R The quality function τ t ∈ R The quality requirement for time t S : R × R → R The success function Gt ∈ R G : R × R → R Bt ∈ R B : R × R → R U t i Q The payoff for a successful round at time t The payoff function for successful rounds The payoff for an unsuccessful round at time t The payoff function for unsuccessful rounds The utility of user i at time t The total quality of service after T rounds Scenario: smart grid The energy production of household i at time t The baseline consumption of household i at time t The energy surplus at time t i ∈ R πt i ∈ R βt σt ∈ R Table 1: Mathematical notation, in order of appearance. V = {1, . . . , n} is the set of users and S is a service provider. In each round t ≤ T each user produces a resource rt i . Users perform an action i ∈ A = {C, D}: C corresponds to contributing the resource and D to opt out from at i = C}. contribution. The set of contributors at time t is denoted as At i that depends on the char- acteristics of resource and communication medium. Contribution might also entail a privacy cost pt i, which models the risks of revealing private information to third parties. Contribution is costly, thus contributors pay a cost ct + = {i ∈ V : at i with value vt 3 Q = max(cid:80)T t=1 f t(At +). +) =(cid:80) i∈At The service provider determines the service quality qt = f (At satisfy the requirement τ t ≤ (cid:80) vt i based on all contributions received. A quality requirement τ t, either global or per user, is generated at every timestep. Not all users are required to contribute in order to i . A round is successful, i.e. the service can be provided, if the quality is higher than the threshold: the success function is defined as S(qt, τ t) = {Gt if qt ≥ τ t else Bt}. Each user gets the same positive payoff Gt = G(τ t, qt) from accessing the service, including those who did not contribute. If the quality threshold is not met, the service cannot be provided and every agent receives a large negative payoff B(τ t, qt) > 0. Given that payoffs are distributed equally, there is no incentive for users to contribute unilaterally: the public goods theory predicts the existence of an equilibrium where nobody contributes. i vt + The individual goal of the users is to maximize their individual utility over T rounds (see Table 2a): U t , where 1X is the goal the indicator function (equals 1 if event X is true). The social goal, i.e. of the service provider, is to maximize the total quality of the service over T rounds i = G(τ t, qt)1qt≥τ t − B(τ t, qt)1qt<τ t − ct i 1 at i The model relies on the following assumptions: • Users contribute their resources to a central entity, which uses them to provide a service. • Contributions cannot be doctored, e.g., to reduce the cost. • Users are not allowed to communicate with each other; this is generally the case for simple devices such as smart meters [13]. • The system implements a privacy-preserving resource-management algorithm that optimizes the use of the resource. 2.1 Applicability to real-world scenarios This section describes how the scenarios of smart and participatory sensing can be modelled using the proposed design principle. 2.1.1 Smart grid value is computed from the current production of renewable energy πt =(cid:80) the current network load βt =(cid:80) Electric vehicles (EVs), associated to households connected to the smart grid, are required to be periodically recharged. The public good consists of the total energy surplus σt = πt − βt, which is available to all households for charging the EVs, its i , and i , obtained from EirGrid data [14]. Contribution to the public good is then defined as opting out from consumption, i.e. renouncing to charge the EV of a charge vt i that might depend on contextual variables such as the current charge level, the availability of a charging station, and the current energy surplus. Contribution entails a comfort cost that is assumed to be proportional to the corresponding value, as the utility of an EV depends on it being charged. Values are determined in the experiments by sampling a uniform probability distribution, and costs are determined by sampling a normal distribution, centered around the respective value. i βt i πt 4 qt−i qt−i + vt i < τ t outcome i (ai = C) −B(τ t, qt) − ct U t −B(τ t, qt) U t i (ai = D) (a) Definition of utilities for agent i. Let qt−i =(cid:80) failure i τ t − vt i ≤ qt−i < τ t depends on ai G(τ t, qt) − ct −B(τ t, qt) i τ t ≤ qt−i success G(τ t, qt) − ct i G(τ t, qt) i be the total contributions, excluding agent i, and τ t be the global requirement. This game qualifies as a threshold public goods game if G(τ t, qt) + B(τ t, qt) > ct i, which is always verified for large enough values of G or B. j∈At +\{i} vt j = qt − vt Social welfare Measure Success a) b) Efficiency c) d) Privacy Definition Σt = min(1, qt/τ t) Et = τ t/qt if τ t ≤ qt else 0 W t = i P t = 1 − At T F t = 1 n Fairness, over time G = 1 n (cid:80)T (cid:16) t=1 U t +/n i (cid:16) n + 1 − 2 n + 1 − 2 (cid:16)(cid:80)n (cid:16)(cid:80)n (cid:80)n i=1(n+1−i)yi (cid:80)n i=1(n+1−i)yi Fairness e) f) i=1 yi i=1 yi (cid:17)(cid:17) (cid:17)(cid:17) where yi =(cid:80)T where yi = vt i t=1 vt i (b) Measures used during evaluation. Table 2: Analytical definition of criteria of performance. 2.1.2 Participatory sensing Traffic congestion is computed by aggregating speeds and locations reported by cars using a participatory sensing approach [15]. The value vt i of individual measurements reflects the novelty of the information, which might depend on the actions of other users [16], e.g., duplicated information due to local correlation in measurements. In the specific case, measurements cannot be linked with one another as the dataset lacks GPS coordinates [17], therefore the value of novelty is approximated with the change of speed, as a sudden change in speed is considered more informative than keeping a constant speed. Contributing data comes at a transmission cost ct i -- that might depend on the characteristics of the message, e.g., size, and of the medium, e.g., congestion, power consumption -- and a privacy cost pt i -- that might depend on disclosing private information, such as location and speed [18]. Costs are determined as the distance to the only points of interests known in the data: the origin and destination of a trip. The public good is created if the sum of individual contributions is higher than a certain threshold τ , which is set to 80% of the size of the population, which means that the service is successfully generated if at least 80% of the users contribute a value of 1. 2.2 Measures The performance of different contribution strategies are quantified with the following measures (see Table 2b): a) Success rate: The fraction of the threshold that has been covered by contribu- tions, or 1 if the total contribution is higher than the threshold. b) Efficiency: The ratio between the requirement and the sum of contributions. Efficiency is 1 if the sum of contributions is equivalent to the threshold e.g., all 5 agents contribute 1, it is lower than 1 if the sum of contributions is larger than the threshold. c) Social welfare: The average reward, sum of a constant benefit and a constant negative cost (only for contributors). d) Privacy: the fraction of agents that did not disclose private information during the current timestep. e) Fairness: The Gini coefficient represents the fairness of the current round of contributions (0 is total equality). It is computed for each timestep, as the Gini coefficient of the values that agents contributed in that timestep, and aggregated across time and repetitions. f) Fairness over time: It represents the fairness of the history of contributions. It measures the Gini coefficient computed across agents on the individual histories of contribution, from the start of the simulation to the current timestep. The definition of privacy can be made scenario-specific by adopting an appropriate privacy measure, e.g., K-anonymity, differential privacy [19]. 2.3 Algorithms This section describes the contribution strategy algorithms chosen for the analysis in our framework. The choice favored well-established and general-purpose algorithms as opposed to algorithms with state-of-the-art efficiency, as the goal is to highlight trade-offs between measures over different scenarios. The review and comparison of scenario-specific algorithms is not aligned to the goals of this paper and is therefore out of scope. 2.3.1 Centralized algorithms Centralized or top-down algorithms rely on a central optimizer that satisfies the public good while minimizing the cost of contribution. This problem can be modeled as the well-know Knapsack problem: i > τ t , where the weight of items is given by the contribution value and the value of each item is defined as the inverse of the cost. We chose a customized "fully polynomial time approximation scheme" (FPTAS) that reaches the knapsack constraints from above, instead of from below, such that the threshold can be met. (cid:40) minimize (cid:80)n subject to (cid:80)n i ∗ at i ∗ at i i=1 ct i=1 vt 2.3.2 Localized algorithms Decentralized algorithms distribute decision making at the local level and allow com- munication between agents for coordination [20] or learning [21]. In this paper we focus on localized algorithms, a type of decentralized algorithms that operate only on local knowledge, without assuming the availability of special hardware for communi- cation [13]. Aspiration learning is a learning algorithm that is specifically tailored for coordination games [22]; agents contribute based on their "satisfaction value", which depends on their previous experiences, but does not consider the current context, e.g., current value or cost. Q-Learning is a model-free unsupervised reinforcement learning approach that considers both the history and the current context in the decision [23]. A disadvantage of reinforcement learning is its sensitivity to initial conditions, for example a multi-agent learning process might converge to an inefficient equilibrium 6 where nobody contributes. In order to make this outcome less likely, agents are pre- trained to prefer contribution in order to bias the initial exploration period. Pre- training is a reasonable solution as it can be performed during device manufacturing and its effect on the behavior of agents fades off quickly as agents start learning. 3 Results and Analysis This section presents the results of computational experiments. Trade-offs are evalu- ated between the three contribution strategies presented in section 2.3 and two base- lines: a baseline in which users contribute "fully", i.e. always, and a "random" baseline in which users have 50% probability of contributing at each round 1. All experiments are performed in a simulation framework developed in Python and run on ETHs cluster Euler. Results presented in this paper represent the average state of 20 simulations after 5000 timesteps, and error bars represent the confidence intervals at 95%. The contribution strategies are tested with real-world datasets in two smart city applications, using cost, value and public good values as described in Section 2.1. The traffic dataset contains mobility traces of private cars obtained from U.S. Depart- ment of Energy National Renewable Energy Laboratory[17]. The electric vehicle (EV) dataset contains residential consumption data obtained from Irish Smart Meter trial and renewable energy production data from Irish elecriticy grid operator EirGrid [14]. Full results are shown in Figure 1 and present the comparisons of the contribution strategies by the six measures discussed in Section 2.2. Most measures do not show a dependency on the size of the population, i.e., the number of sensors, because the value of the public good threshold is chosen to be proportional to the size of the population. If the threshold would be constant, an increase in the number of sensors, i.e., potential contributors, would make the creation of the public good easier, hence affect all measures. In terms of success rate (Figure 1a), full contribution and centralized optimization always succeed,i.e., are always able to provide the services, while Q-Learning does not guarantee success and fails in about 2-3% of the cases. In terms of efficiency (Figure 1b), centralized optimization is the most efficient solution, as it finds the subset of users whose contribution satisfies the requirement 2 at the lowest cost 3. Efficiency measures how close the total contribution approaches the needs of the system, hence baseline in which all agents contribute will have the lowest possible efficiency. Efficiency increases with the size of the population as a higher number of possible solutions -- combinations of individual contributions -- makes it more likely to find an efficient solution. Similarly, in terms of social welfare (see Figure 1c), optimization scores the highest, while localized strategies reach a welfare around 30% lower. Social welfare is the difference between the rewards from the public good and the costs of contributions, so a negative value indicates that costs are higher than gains. Differently from the previous result, the performance of aspiration learning is equivalent to that of Q- Learning, as opposed to that of optimization. 150% chance of contribution does not imply that 50% chance of success, because each contribution has an average value greater than 1. 2Optimization is not successful if the total requirement is larger than the sum of contribu- tions from all agents, but the experiments are generated to be successful if all users contribute. 3The approximation algorithm used to solve the optimization problem does not guarantee to find the global minimum. 7 In terms of privacy (Figure 1d), centralized optimization achieves no privacy, like full contribution, as it requires knowledge about the state of all users. Random con- tribution offers the highest privacy -- around 50% of users -- at the expenses of other measures, e.g., success rate and efficiency. Localized contribution strategies allow a fraction of the user to keep their data private. This fraction increases with the pop- ulation size for aspiration learning, while Q-Learning trades a lower privacy off for a higher fairness. Fairness is measured in two ways: "fairness of contributions" compares the actions of all agents at the current time t (see Figure 1e), while "fairness of contribution over time" considers the histories of contribution up to time t (see Figure 1f). Full contribution requires all users to contribute, thus it trades perfect fairness off for other measures such as efficiency. Optimization offers low fairness because it considers only the current state, and users in certain states, e.g., with high values, are more likely to contribute than others. Conversely, optimization offers high fairness over time because agents are randomly assigned to states, hence the chance of being in any state is over time the same. This result might not hold if states are not randomly assigned, e.g., some users are more likely to obtain high values/costs than others. Aspiration learning bases decisions only on the history of decisions, this leads to higher fairness, as contributions are independent of the state, and to lower fairness over time, caused by individual differences in training that accumulate over time. These values decrease with the population size, while other contribution strategies are not affected by this parameter. Q-Learning scores high values in both measures as it considers both the current context and the history of actions, this allows agents to learn similar behaviors by interacting with one another. Algorithm Pros Full Random Type Baseline Baseline Centralized Knapsack Aspiration Localized Localized Q-Learning Success Privacy Success, Efficiency Privacy Fairness (over time) Efficiency Cons Efficiency Success Privacy, Fairness F. over time Table 3: Comparison of contribution strategies. Summary of trade-offs is presented in Table 3. Centralized optimization assures the success of the service and high efficiency, it is hence appropriate for mission critical services for which computation and network constraints are not an issue, e.g., measur- ing current load on the smart grid to prevent outages. Localized strategies are instead best suited for applications where privacy concerns might reduce user adoption, e.g., participatory sensing. Finally, Q-Learning offers the highest fairness, hence it is ideal for applications where fair access to the service is desirable, e.g., charging of electric vehicles. 4 Conclusions and Future Work This paper provides a scenario-independent design principle and simulation framework for smart city applications that rely on voluntary user contribution. Voluntary contri- butions empower users to control the ownership of their resource, e.g., by contributing 8 (a) Ratio of success in ser- vice creation. (b) Efficiency of total contri- butions. (c) Social welfare of the pop- ulation. (d) Average privacy of users. (e) Average fairness of con- tributions. (f) Average fairness over time. Figure 1: Comparison of contribution strategies, the x axis represents the population size. Dashed lines represent baselines, solid lines represent contribution strategies. The plots show trade-offs between centralized opti- mization, aspiration learning and Q-Learning in terms of efficiency, privacy and fairness. Optimization offers the highest success rate and efficiency, while as- piration learning and Q-Learning offers higher privacy. Q-Learning is the only contribution strategy to offer high fairness on both measures. 9 AspirationQ-LearningOptimizationRandomFull contrib.Legend:10203040500.750.800.850.900.951.00Success rateTraffic data10203040500.860.880.900.920.940.960.981.00Success rateEV data10203040500.30.40.50.60.70.80.9EfficiencyTraffic data10203040500.30.40.50.60.70.80.9EfficiencyEV data10203040504202468Social welfareTraffic data102030405012345678Social welfareEV data10203040500.00.10.20.30.40.5PrivacyTraffic data10203040500.00.10.20.30.40.5PrivacyEV data10203040500.50.60.70.80.91.0Fairness of contributionsTraffic data10203040500.50.60.70.80.91.0Fairness of contributionsEV data10203040500.750.800.850.900.951.00Fairness of contributions, over timeTraffic data10203040500.600.650.700.750.800.850.900.951.00Fairness of contributions, over timeEV datadata towards a service, independently of the type of resource and its use. The applica- bility of this framework is verified using real-world data from the application scenarios of traffic congestion monitoring and electric vehicle charging. Results quantify trade-offs produced by different contribution strategies along mea- sures such as efficiency, privacy and fairness. The trade-offs identified hold in both scenarios, suggesting that they depend on characteristics of the algorithms and not on characteristics of the scenarios. Therefore the results can be used as implementa- tion recommendations to service providers and system designers about the choice of contribution strategies. Modeling and application in other scenarios with different cost and value character- istics is left to future work, for example "negative" contribution in traffic congestion, where users contribute to the public good by choosing a longer route instead of the shortest but congested route [24]. Incentives for contributions are important in scenar- ios that rely on user participation [25]. However, incentives mechanisms would require quantifying privacy [26], which could be addressed in future work. Another limitation of the current work is that only localized algorithms are considered. Relaxing this as- sumption is a worthy avenue for future work. Communication between agents would introduce privacy concerns during communication between agents, which would need to be measured, but would would allow analysis of new classes of algorithms, such as decentralized optimization. 5 Acknowledgements Stefano Bennati acknowledges support by the European Commission through the ERC Advanced Investigator Grant Momentum [Grant No. 324247] References [1] A. Gaur, B. Scotney, G. Parr, and S. McClean, "Smart city architecture and its applications based on iot," Procedia computer science, vol. 52, pp. 1089 -- 1094, 2015. [2] E. Al Nuaimi, H. Al Neyadi, N. Mohamed, and J. Al-Jaroodi, "Applications of big data to smart cities," Journal of Internet Services and Applications, vol. 6, no. 1, p. 25, 2015. [3] I. A. T. Hashem, V. Chang, N. B. Anuar, K. Adewole, I. Yaqoob, A. Gani, E. Ahmed, and H. Chiroma, "The role of big data in smart city," International Journal of Information Management, vol. 36, no. 5, pp. 748 -- 758, 2016. [4] S. Bennati and E. Pournaras, "Privacy-enhancing Aggregation of Internet of Things Data via Sensors Grouping," Sustainable Cities and Society, 2018. [5] D. Eckhoff and I. Wagner, "Privacy in the smart city -- applications, technologies, challenges and solutions," IEEE Communications Surveys & Tutorials, 2017. [6] S. Finster and I. Baumgart, "Privacy-Aware Smart Metering: A Survey," IEEE Communications Surveys & Tutorials, vol. 17, no. 2, pp. 1088 -- 1101, 2015. [7] C. K. Tham and T. Luo, "Fairness and social welfare in service allocation schemes for participatory sensing," Computer Networks, vol. 73, pp. 58 -- 71, nov 2014. 10 [8] M. Bagnoli and B. L. Lipman, "Provision of Public Goods: Fully Implement- ing the Core through Private Contributions," The Review of Economic Studies, vol. 56, no. 4, pp. 583 -- 601, 1989. [9] M. Wolsink, "The research agenda on social acceptance of distributed generation in smart grids: Renewable as common pool resources," Renewable and Sustainable Energy Reviews, vol. 16, no. 1, pp. 822 -- 835, jan 2012. [10] S. Bennati, "The simulation framework," 2018. [Online]. Available: https: //github.com/bennati/EnergyVCG [11] L. Duan, T. Kubo, K. Sugiyama, J. Huang, T. Hasegawa, and J. Walrand, "Moti- vating smartphone collaboration in data acquisition and distributed computing," IEEE Transactions on Mobile Computing, vol. 13, no. 10, pp. 2320 -- 2333, 2014. [12] X. Wang, C. Shao, X. Wang, and C. Du, "Survey of electric vehicle charging load and dispatch control strategies," Proceedings of the CSEE, vol. 33, no. 1, pp. 1 -- 10, 2013. [13] Z. M. Fadlullah, M. M. Fouda, N. Kato, A. Takeuchi, N. Iwasaki, and Y. Nozaki, "Toward intelligent machine-to-machine communications in smart grid," IEEE Communications Magazine, vol. 49, no. 4, 2011. [14] EirGrid group. (2018) Smart grid dashboard. [Online]. Available: http: //smartgriddashboard.eirgrid.com/ [15] J. W. S. Brown, O. Ohrimenko, and R. Tamassia, "Haze," in Proceedings of the 21st ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems - SIGSPATIAL'13. New York, New York, USA: ACM Press, 2013, pp. 530 -- 533. [16] M. Vuran, O. Akan, and I. Akyildiz, "Spatio-temporal correlation: theory and applications for wireless sensor networks," Computer Networks, vol. 45, pp. 245 -- 259, 2004. [17] NREL. (2015) Transportation secure data center. [Online]. Available: www.nrel. gov/tsdc [18] G. Tsoukaneri, G. Theodorakopoulos, H. Leather, and M. K. Marina, "On the inference of user paths from anonymized mobility data," Proceedings - 2016 IEEE European Symposium on Security and Privacy, EURO S and P 2016, pp. 199 -- 213, 2016. [19] M. L. Damiani, "Location privacy models in mobile applications: conceptual view and research directions," GeoInformatica, vol. 18, no. 4, pp. 819 -- 842, 2014. [20] J. Kennedy and R. Eberhart2, "Particle Swarm Optimization," IEEE Interna- tional Conference on Neural Network, vol. 4, pp. 1942 -- 1948, 1995. [21] I. Dusparic, A. Taylor, A. Marinescu, F. Golpayegani, and S. Clarke, "Residen- tial demand response: Experimental evaluation and comparison of self-organizing techniques," Renewable and Sustainable Energy Reviews, vol. 80, pp. 1528 -- 1536, dec 2017. [22] G. C. Chasparis, J. S. Shamma, and A. Arapostathis, "Aspiration learning in co- ordination games," Proceedings of the IEEE Conference on Decision and Control, pp. 5756 -- 5761, 2010. [23] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. MIT press Cambridge, 1998, vol. 1, no. 1. 11 [24] B. A. Huberman, "Social Dilemmas and Internet Congestion," Science, vol. 277, no. 5325, pp. 535 -- 537, 1997. [25] G. Radanovic and B. Faltings, "Incentive Schemes for Participatory Sensing." Aamas, pp. 1081 -- 1089, 2015. [26] P. A. Norberg, D. R. Horne, and D. A. Horne, "The privacy paradox: Personal information disclosure intentions versus behaviors," Journal of Consumer Affairs, vol. 41, no. 1, pp. 100 -- 126, 2007. 12
1902.07781	1	1902	2019-02-20T21:12:44	Empathic Autonomous Agents	[ "cs.MA" ]	Identifying and resolving conflicts of interests is a key challenge when designing autonomous agents. For example, such conflicts often occur when complex information systems interact persuasively with humans and are in the future likely to arise in non-human agent-to-agent interaction. We introduce a theoretical framework for an empathic autonomous agent that proactively identifies potential conflicts of interests in interactions with other agents (and humans) by considering their utility functions and comparing them with its own preferences using a system of shared values to find a solution all agents consider acceptable. To illustrate how empathic autonomous agents work, we provide running examples and a simple prototype implementation in a general-purpose programing language. To give a high-level overview of our work, we propose a reasoning-loop architecture for our empathic agent.	cs.MA	cs	Empathic Autonomous Agents Timotheus Kampik[0000−0002−6458−2252], Juan Carlos Nieves[0000−0003−4072−8795], and Helena Lindgren[0000−0002−8430−4241] Umea University, 901 87, Umea, Sweden {tkampik,jcnieves,helena}@cs.umu.se Abstract. Identifying and resolving conflicts of interests is a key challenge when designing autonomous agents. For example, such conflicts often occur when com- plex information systems interact persuasively with humans and are in the future likely to arise in non-human agent-to-agent interaction. We introduce a theoretical framework for an empathic autonomous agent that proactively identifies potential conflicts of interests in interactions with other agents (and humans) by consider- ing their utility functions and comparing them with its own preferences using a system of shared values to find a solution all agents consider acceptable. To illus- trate how empathic autonomous agents work, we provide running examples and a simple prototype implementation in a general-purpose programing language. To give a high-level overview of our work, we propose a reasoning-loop architecture for our empathic agent. Keywords: Multi-agent systems · Utility theory · Conflicts of interests 1 Background and Problem Description In modern information technologies, conflicts of interests between users and informa- tion systems that operate with a high degree of autonomy (autonomous agents) are of increasing prevalence. For example, complex web applications persuade end-users, possibly against the interests of the persuaded individuals1. Given the prevalence of autonomous systems will increase, conflicts between autonomous agents and humans (or between different autonomous agent instances and types) can be expected to occur more frequently in the future, e.g. in interactions with or among autonomous vehicles in scenarios that cannot be completely solved by applying static traffic rules. Conse- quently, one can argue for the need to develop empathic intelligent agents that consider the preferences or utility functions of others, as well as ethics rules and social norms when interacting with their environment to avoid severe conflicts of interests. As a sim- ple example, take two vehicles (A and B) that are about to enter a bottleneck. Assume they cannot enter the bottleneck at the same time. A and B can either wait or drive. Considering only its own utility function, A might determine that driving is the best action to execute, given that B will likely stop and wait to avoid a crash. However, A should ideally assess both its own and B's utility function and act accordingly. If B's 1 E.g., research provides evidence that contextual advertisement influences how users process online news [25]; social network applications have effectively been employed for political persuasion (see for an example: [4]. 2 T. Kampik et al. utility for driving is considered higher than A's, A can then come to the conclusion that waiting is the best action. As A does not only consider its own goals, but also the ones of B, one can regard A as empathic, following Coplan's definition of empathy, as "a process through which an observer simulates another's situated psychological states, while maintaining clear self -- other differentiation" [12]. While existing literature covers conflict resolution in multi-agent systems from a broad range of perspectives (see for a partial overview: [2]), devising a theoretical framework for autonomous agents that consider the utility functions (or preferences) of agents in their environment and use a combined utilitarian/rule-based approach to identify and resolve conflicts of interests can be considered a novel idea. However, existing multi-agent systems research can be leveraged to implement core components of such a framework, as is discussed later. In this chapter, we provide the following research contributions: 1. We create a theoretical framework for an empathic agent that uses a combination of utility-based and rule-based concepts to compromise with other agents in its environment when deciding upon how to act. 2. We provide a set of running examples that illustrate how the empathic agent works and show how the examples can be implemented in a general-purpose programing language. 3. We propose a reasoning-loop architecture for a generic empathic agent. The rest of this chapter is organized as follows: in Section 2, we present a theoreti- cal framework for the problem in focus. Then, we illustrate the concepts with the help of different running examples and describe the example implementation in a general- purpose programing language in Section 3. Next, we outline a basic reasoning-loop architecture for the empathic agent in Section 4. In Section 5, we analyze how the architecture aligns with the belief-desire-intention approach and propose an implemen- tation using the Jason multi-agent development framework. Finally, we discuss how our empathic agent concepts relate to existing work, propose potential use cases, highlight a set of limitations, and outline future work in Section 6, before we conclude the chapter in Section 7. 2 Empathic Agent Core Concepts In this section, we describe the core concepts of the empathic agent. To allow for a precise description, we assume the following scenario2: -- The scenario describes the interaction between a set of empathic agents {A0, ..., An}. -- Each interaction scenario takes place at one specific point in time, at which all agents execute their actions simultaneously. -- At this point in time, each agent Ai(0 ≤ i ≤ n) has a finite set of possible ac- tions Actsi := {Act0 i }, resulting in an overall set of action sets Acts := {Acts0, ..., Actsn}. Each agent can execute an action tuple that contains one or 2 As we will explain later, the scenario and the resulting specification can be gradually extended to allow for better real-world applicability. i , ..., Actm Empathic Autonomous Agents 3 multiple actions. In each interaction scenario, all agents execute their actions si- multaneously and receive their utility as a numeric reward based on the actions that have been executed. -- The utility of an agent Ai is determined by a function ui of the actions of all agents. The utility function returns a numerical value or null3: ui := Acts0 × ... × Actsn → {null,−∞, R,∞} The goal of the empathic agent is to maximize its own utility as long as no conflicts with other agents arise. We define a conflict of interests between several agents as any interaction scenario in which there is no tuple of possible actions that maximizes the 4. utility functions of all agents. I.e., we need to compare arg max uA0 , ..., arg max uAn Note that arg max uAi returns a set of tuples (that contains all action tuples that yield the maximal utility for agent Ai). For this, we create a boolean function c that the empathic agent uses to determine conflicts between itself and other agents, based on the utility functions of all agents: c(uA0, ..., uAn ) := arg max uA0 ∩ ... ∩ arg max uAn (cid:54)= {};  true, if : f alse, otherwise. Considering the incomparability property of the von Neumann-Morgenstern utility the- orem [24], such a conflict can be solved only if a system of values exists that is shared between the agents and used to determine comparable individual utility values. Hence, we introduce such a shared value system. To provide a possible structure for this system, we deconstruct the utility functions into two parts: -- An actions-to-consequences mapping (a function a2ci that takes the actions the agents potentially decide to execute and returns a set of consequences (proposi- tional atoms) Consqs := {Consq0 i , ..., Consqn i }): a2ci := Actsi × ... × Actsn → 2Consqs -- A consequences-to-utility mapping (utility quantification function uq). Note that the actions-to-consequences mapping is agent-specific, while the utility quantifica- tion function is generically provided by the shared value system5: uq := 2Consqs → {null,−∞, R,∞} 3 We allow for utility functions to return a null value for action tuples that are considered im- possible, e.g. in case some actions are mutually exclusive. While we concede that the elegance of this approach is up for debate, we opted for it because of its simplicity. 4 The arg max operator takes the function it precedes and returns all argument tuples that max- imize the function. 5 I.e., for the same actions, an agent should only receive a different utility outcome than another agent if the impact on the two is distinguishable in its consequences. We again allow for null values to be returned in case of impossible action tuples. 4 T. Kampik et al. Then, agents can agree on the utility value of a given tuple of actions, as long as the quality of the consequence is observable to all agents in the same way. In addition, the value system can introduce generally applicable rules, e.g. to hard-code a prioritization of individual freedom into an agent. With help of the value system, we create a prag- matic definition of a conflict of interests as any situation, in which there is no tuple of actions that is regarded as acceptable by all agents when considering the shared set of values, given each agent executes the actions that maximize their individual utility function. To support the notion of acceptability, we introduce a set of agent-specific acceptability functions accs := {accA0 , ..., accAn}. The acceptability functions are de- rived from the corresponding utility functions and the shared system of values and take a set of actions as their inputs. Acceptability functions are domain-specific and there is no generic logic to be described in this context: accAi := ActsA0 × ... × ActsAn → {null, true, f alse} The notion of acceptability rules adds a normative aspect to the otherwise consequen- tialist empathic agent framework. Without this notion, our definition of a conflict of interests would cover many scenarios that most human societies regard as not conflict- worthy, e.g. when one agent would need to accept large utility losses to optimize its own actions towards improving another agents' utility. Considering the acceptabil- ity functions, we can now determine whether a conflict of interests in terms of the pragmatic definition approach exists for an agent Ai by using the following func- tion cp that takes the utility function ui of agent Ai and the acceptability functions Accs := {accA0, ..., accAn} as input arguments: cp(ui, Accs) :=  true, if : (cid:64)acts ∈ arg max ui ∧ ∀acc∈Accs : acc(acts) = true f alse, otherwise. We define an empathic agent Ai as an agent that, when determining the actions it exe- cutes, considers the utility functions of the agents it could potentially affect and max- imizes its own utility only if doing so does not violate the acceptability function of any other agent; otherwise it acts to maximize the shared utility of all agents (while also considering the acceptability functions)6. Algorithm 1 specifies an initial, naive approach towards the empathic agent core algorithm. The empathic agent core algo- rithm of an agent Ai in its simplest form can be defined as a function that takes the utility functions {u0, ..., un} of the different agents, the set of all acceptability func- tions Accs := {acc0, ..., accl}, and all possible actions Actsi of agent Ai and returns the tuple of actions Ai should execute7. 6 As different aggregation approaches are possible (for example: sum, product) to deter- mine the maximal shared utility, we introduce the not further specified aggregation function aggregate(u0, ..., un). In our running examples (see Section 3), we use the product of the in- dividual utility function outcomes to introduce some notion of fairness; inequality should not be in the interest of the empathic agent. However, the design choice for this implementation detail can be discussed. 7 To facilitate readability, we switch to a pseudo-code notation for the following algorithms. Algorithm 1 Naive empathic agent algorithm: D A N (determine actions naive) 1: procedure D A Ni({u0, ..., un}, Accs, Actsi) (cid:46) Utility & acceptability functions of all Empathic Autonomous Agents 5 : ∀acc∈Accsacc(acts) = true if ∃acts ∈ arg max ui ∧ ∀acc∈Accs : acc(acts) = true then agents, actions of Ai(0 ≤ i ≤ n) best acceptable acts ← return Actsi ∩ f irst(actsk ∈ best acceptable acts) return Actsi ∩ f irst(arg max(aggregate(u(cid:48) 0, ..., u(cid:48) acts∈arg max ui (cid:83) else n)) 2: 3: 4: 5: 6: 7: 8: end procedure end if Note that in the context of the empathic agent algorithms, the function f irst(set) turns the provided set of tuples into a sequence of tuples by sorting the elements in decreasing alphanumerical order and then returns the first element of the sequence. This enables a deterministic action tuple selection. Moreover, we construct a set of new utility func- n} that assign all not acceptable action tuples a utility of null (Algo- tions {u(cid:48) rithm 2)8: 0, ..., u(cid:48) i(ui,{acts0, ..., actsn}, accs) is acceptable ← ∀ acc ∈ accs : acc(actsi) = true if is acceptable then Algorithm 2 Helper function: new utility function based on ui; all not acceptable action tuples yield utility of null. 1: procedure u(cid:48) 2: 3: 4: 5: 6: 7: 8: end procedure return ui(actsi, ..., actsn) return null end if else In Algorithm 1, we specify that the agent picks the first item in the sequence of determined action tuples if it finds multiple optimal tuples of actions. Alternatively, the agent could employ one of the following approaches to select between the optimal action tuples: -- Random. The agent picks a random action tuple from the list of the tuples it deter- mined as optimal. This would require empathic agents to use an additional protocol to agree on the action tuple that should be executed. -- Utilitarian. Among the action tuples that were determined as optimal, the agent picks the one that provides maximal combined utility for all agents and falls back to a random or first-in-sequence selection between action tuples if several of such tuples exist. 8 We already use null to denote impossible action tuples. This implies an acceptable action tuple should always exists. To achieve a distinction, a value of −∞ could be assigned. 6 T. Kampik et al. Still, the algorithm is somewhat naive, as agents that implement it will decide to execute suboptimal activities if the following conditions apply: -- Multiple agents find that the actions that optimize their individual utility are incon- sistent with the actions that are optimal for at least one of the other agents. -- Multiple agents find that executing these conflicting actions is considered accept- able. -- Executing these acceptable actions generates a lower utility for both agents than optimizing the shared utility would. Hence, we extend the algorithm so that the agent selects the tuple of actions that maxi- mizes its own utility, but falls back to maximize shared utility if the utility-maximizing action tuple is either not acceptable, or would lead to a lower utility outcome than maximizing the shared utility, considering the other agent follows the same approach (Algorithm 3): Algorithm 3 Lazy empathic agent algorithm: D A L (determine actions lazy) 1: procedure D A Li({u0, ..., un}, Accs) (cid:46) Utility & acceptability functions of all agents, actions of all agents {A0, ..., An} {acts max0, ..., acts maxn} ← DET ERM IN E ACT M AX(ui, Accs) {good acts max0, ..., good acts maxn} ← { DET ERM IN E GOOD ACT S M AX(u0, Accs, acts max0), ..., DET ERM IN E GOOD ACT S M AX(un, Accs, acts maxn), } if good acts max0 ∩ ... ∩ good acts maxn (cid:54)= {} then return Actsi ∩ f irst(good acts max) return Actsi ∩ f irst(arg max(aggregate(u(cid:48) 0, ..., u(cid:48) n))) 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: end procedure end if else Algorithm 3 calls two helper functions. Algorithm 4 determines acceptable action tu- ples that maximize a provided utility function ui: Algorithm 4 Helper function: determine acceptable action tuples that maximize utility function ui 1: procedure DETERMINE ACT MAX(ui, Accs) 2: 3: end procedure : ∀acc∈Accsacc(acts) = true acts∈arg max ui return (cid:83) Algorithm 5 determines all action tuples that would maximize an agent's (Ai's) utility if this agent could dictate the actions of all other agents, given the action tuples provide Empathic Autonomous Agents 7 a better utility for this agent than the action tuples that maximize all agents' combined utility, given all agents execute an action tuple that maximizes their own utility if they could dictate the other agents' actions. Note that Algorithm 5 makes use of the previ- ously introduced algorithm (Algorithm 1): Algorithm 5 Helper function: determines all maximizing action tuples that would still yield a good utility result for agent Ai(0 ≤ i ≤ n), given all other agents also pick an action tuple that would maximize their own utility, if all other agents "played along". 1: procedure DETERMINE GOOD ACTS MAX(ui, Accs) 2: : ∀acc∈Accs : acc(acts) = true ∧ return (cid:83) acts∈arg max ui n(cid:83) ui( ≥ ui(acts max) k=0 D A Nk({u0, ..., un}, Accs, acts)) 3: 4: 5: end procedure However, this algorithm only considers two types of action tuples for execution: action tuples that provide the maximal individual utility for the agent and action tuples that provide the maximal combined utility for all agents. Action tuples that do not maximize the agent's individual utility, but are still preferable over the action tuples that maximize the combined utility, remain unconsidered. Consequently, we call an agent that imple- ments such an algorithm a lazy empathic agent. We extend the algorithm to also consider all action tuples that could possibly be relevant. I.e., if an action tuple is not considered acceptable, or if the tuple is considered acceptable but the agent chooses to not execute it, the agent falls back to the tuple of actions that provides the next best individual util- ity. We construct a function ne that returns the Nash equilibria based on the updated n}, considering we have a strategic game (cid:104)N, (Ai) (cid:37)i(cid:105), with utility functions {u(cid:48) N := {A0, ...An}, Ai := ActsAi, and acts (cid:37)i acts(cid:48) i(acts(cid:48))9. Then, we create the full empathic agent core algorithm D A Fi for an agent Ai that n} and all agents' possible actions as in- takes the updated utility functions {u(cid:48) puts {Acts0, ..., Actsn}. The algorithm determines the (first of) the Nash equilibria that provide the highest shared utility and, if no Nash equilibrium exists, chooses the first tuple of actions that maximizes shared utility: i(acts) ≥ u(cid:48) 0, ..., u(cid:48) 0, ..., u(cid:48) := u(cid:48) 9 See the Nash equilibrium definition provided by Osborne and Rubinstein [19, p. 11 et sqq.]. 8 T. Kampik et al. 0, ..., u(cid:48) 0, ..., u(cid:48) ∀acts ∈ equilibria : n},{Acts0, ..., Actsn}) n},{Acts0, ..., Actsn}) shared max equilibria ← acts∗ ∈ equilibria : Algorithm 6 Full empathic agent algorithm: D A F (determine actions full) 1: procedure D A Fi({u(cid:48) equilibria ← ne({u(cid:48) 2: if equilibria (cid:54)= {} then 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: end procedure return Actsi ∩ f irst(shared max equilibria) return Actsi ∩ f irst(arg max(aggregate(u0, ..., un)) end if (u(cid:48) 0(acts∗) × ... × u(cid:48) n(acts∗)) ≥ (u(cid:48) 0(acts) × ... × u(cid:48) n(acts)) else Going back to the selection between several action tuples that might be determined as optimal, it is now clear that a deterministic approach for selecting a final action tuple is preferable for both lazy and full empathic agents, as it avoids agents deciding upon executing action tuples that are not aligned with one another and lead to an unnecessary low utility outcome. Hence, we propose using a utilitarian approach with a first-in- sequence selection if the utilitarian approach is inconclusive10. The proposed agent can be considered a rational agent following the definition by Russel and Norvig in that it "acts so as to achieve the best outcome or, when there is uncertainty, the best expected outcome" [22, p. 4-5] and an artificially socially intelli- gent agent as defined by Dautenhahn as it instantiates "human-style social intelligence" in that it "manage[s] the individual's [its own] interests in relationship to the interests of the social system of the next higher level" [13]. 3 Running Examples In this section, we present two simple running examples of empathic agents and de- scribe the implementation of the examples in a general-purpose programming language (JavaScript). 3.1 Example 1: Vehicles We provide a running example for the "vehicle/bottleneck" scenario introduced above. Consequently, we have a two-agent scenario {A, B}. Each agent has a utility function uA,B := ActsA × ActsB → {−∞, R,∞}. ActsA and ActsB are the possible actions A and B, respectively, can execute. To fully specify the utility functions, we follow the approach outlined above and first construct the actions-to-consequences mappings a2cA and a2cB for both agents. The possible actions are ActsA = {driveA, waitA} and ActsB = {driveB, waitB}. I.e., Acts = {driveA, waitA, driveB, waitB}. To assess 10 As state above, we assume that the f irst function sorts the action tuples in a deterministic order before returning the first element. the consequences that include waiting, we assume B is twice as fast as A (without waiting, A needs 20 time units to pass the bottleneck while B needs 10)11: Empathic Autonomous Agents 9 We construct the following utility quantification functions and subtract an amount pro- portional to the waiting time from the utility value 1 of wait 0: a2cA(acts) := a2cB(acts) := crash, if : acts = (driveA, driveB); wait 0, if : acts = (driveA, waitB); wait ∞, if : acts = (waitA, waitB); wait 10, if : acts = (waitA, driveB); null, otherwise. crash, if : acts = (driveA, driveB); wait 20, if : acts = (driveA, waitB); wait ∞, if : acts = (waitA, waitB); wait 0, if : acts = (waitA, driveB); null, otherwise.   −∞, if : consqs = {crash}; if : consqs = {wait 20}; 0.9, if : consqs = {wait ∞}; 0 if : consqs = {wait 0}, 1, null, otherwise. −∞, if : consqs = {crash}; if : consqs = {wait 20}; 0.8, if : consqs = {wait ∞}; 0, if : consqs = {wait 0}; 1, null, otherwise. u2cA(consqs) := u2cB(consqs) :=     Actions-to-consequences mappings and utility quantification functions can then be com- bined to utility functions: uA(acts) = uB(acts) = 1, if : acts = (driveA, waitB); 0.9, if : acts = (waitA, driveB); 0, if : acts = (waitA, waitB); −∞, if : acts = (driveA, driveB); null, otherwise. if : acts = (driveA, waitB); 0.8, if : acts = (waitA, driveB); 1, if : acts = (waitA, waitB); 0, −∞, if : acts = (driveA, driveB); null, otherwise. 11 driveA ∧ waitA and driveB ∧ waitB, respectively, are mutually exclusive ({driveA ⊕ waitA, driveB ⊕ waitB}, with A ⊕ B := (A ∨ B) ∧ ¬(A ∧ B)). I.e., the functions re- turn null if driveA ∧ waitA ∨ driveB ∧ waitB. 10 T. Kampik et al. accA,B(acts) := We assume scenarios where both agents are driving or both agents are waiting are not acceptable by either agents and introduce the corresponding acceptability rules:  f alse, if : acts = (driveA, driveB) ∨ (waitA ∧ waitB); null, if : (driveA ∈ acts ∧ waitA ∈ acts) ∨ (driveB ∈ acts ∧ waitB ∈ acts); true, otherwise. Based on the utility functions (uA, uB), we create new utility functions (u(cid:48) consider the acceptability rules: A, u(cid:48) B) that 2 , 1 3 ,  1  1, 1 3 , u(cid:48) A(acts) := u(cid:48) B(acts) := if : acts = (driveA, waitB); if : acts = (waitA, driveB); null, otherwise. if : acts = (waitA, driveB); if : acts = (driveA, waitB); null, otherwise. Finally, we apply the empathic agent algorithms to our scenario. Using the naive al- gorithm, the agents apply the acceptability rules, but do not consider the other agent's strategy. Hence, both agents decide to drive, (and consequently crash). B,},{accA, accB}, ActsA) = driveA B,},{accA, accB}, ActsB) = driveB D A NA({u(cid:48) D A NB({u(cid:48) A, u(cid:48) A, u(cid:48) The resulting utility is −∞ for both agents. None of the two other algorithms (lazy, full) allows any agent to decide to execute an action tuple that does not optimize shared utility. I.e., both algorithms yield the same result: D A LA({u(cid:48) D A LB({u(cid:48) A, u(cid:48) A, u(cid:48) B,},{accA, accB}, ActsA) = waitA B,},{accA, accB}, ActsB) = driveB D A FA({u(cid:48) D A FB({u(cid:48) A, u(cid:48) A, u(cid:48) B,},{accA, accB}, ActsA) = waitA B,},{accA, accB}, ActsB) = driveB The resulting utility is 0.9 for agent A and 1 for agent B. As can be seen, the differ- ence between agent types is not always relevant. The following scenario will provide a distinctive outcome for all three agent variants. 3.2 Example 2: Concert As a second example, we introduce the following scenario12. Two empathic agents {A, B} plan to attend a concert of music by either Bach, Stravinsky, or Mozart (Acts := 12 The scenario is an adjusted and extended version of the "Bach or Stravinsky? (BoS)" example presented by Osborne and Rubinstein [19, p. 15 -- 16] Empathic Autonomous Agents 11 {BachA, StravinskyA, M ozartA, BachB, StravinskyB, M ozartB}). A considers the Bach and Mozart concerts of much greater pleasure when attended in company of B (utility of 6, respectively 3) and not alone (either concert: 1). In contrast, the Stravinsky concert yields good utility, even if A attends it alone (4). Attending it in company of B merely gives a utility bonus of 1 (total: 5). B prefers concerts in company of A as well (2 for Stravinsky and 4 for Mozart), but gains little additional utility from attending a Bach concert with A (1.1 with A versus 1 alone) because they dislike listening to A's Bach appraisals. Attending any concert alone yields a utility of 1 for B. As the utility is in this scenario largely derived from the subjective musical taste and social preferences of the agents and to keep the example concise, we skip the actions-to-consequences mapping and construct the utility functions right away13: null, 6, 5, 4, 3, 1, null, if : length(acts) (cid:54)= 2 ∨ length(set(BachA, StravinskyA, M ozartA) ∩ set(acts)) (cid:54)= 1∨ length(set(BachB, StravinskyB, M ozartB) ∩ set(acts)) (cid:54)= 1; else if : acts = (BachA, BachB); else if : acts = (StravinskyA, StravinskyB); else if : StravinskyA ∈ acts ∧ StravinskyB /∈ acts; else if : acts = (M ozartA, M ozartB); otherwise. if : length(acts) (cid:54)= 2 ∨ length(set(BachA, StravinskyA, M ozartA) ∩ set(acts)) (cid:54)= 1∨ length(set(BachB, StravinskyB, M ozartB) ∩ set(acts)) (cid:54)= 1; 1.1, else if : acts = (BachA, BachB); 2, 4, 1, else if : acts = (StravinskyA, StravinskyB); else if : acts = (M ozartA, M ozartB); otherwise.   uA(acts) = uB(acts) = We introduce the following acceptability function that applies to both agents (although it is of primary importance for agent A). As agent A is banned from the venue that hosts the Stravinsky concert, the action StravinskyA is not acceptable: accA,B(acts) := (cid:26) f alse, if : acts = StravinskyA ∈ acts; true, otherwise. 13 Note that the if-condition that triggers the return of a null value simply defines that BachA, StravinskyA, and M ozartA are mutually exclusive, as are BachB, StravinskyB, and M ozartB. 12 T. Kampik et al. Considering the acceptability function, we create the following updated utility func- tions: null, 6, 4, 3, 1, null,   u(cid:48) A(acts) = u(cid:48) B(acts) = if : length(acts) (cid:54)= 2 ∨ length(set(BachA, StravinskyA, M ozartA) ∩ set(acts)) (cid:54)= 1∨ length(set(BachB, StravinskyB, M ozartB) ∩ set(acts)) (cid:54)= 1∨ StravinskyB ∈ acts; else if : acts = (BachA, BachB); else if : StravinskyA ∈ acts; else if : acts = (M ozartA, M ozartB); otherwise. if : length(acts) (cid:54)= 2 ∨ length(set(BachA, StravinskyA, M ozartA) ∩ set(acts)) (cid:54)= 1∨ length(set(BachB, StravinskyB, M ozartB) ∩ set(acts)) (cid:54)= 1∨ StravinskyB ∈ acts; 1.1, else if : acts = (BachA, BachB); 4, 1, else if : acts = (M ozartA, M ozartB); otherwise. Now, we can run the empathic agent algorithms. The naive algorithm returns Bach for agent A and M ozart for agent B: B,},{accA, accB}, ActsA) = BachA A, u(cid:48) B,},{accA, accB}, ActsB) = M ozartB D A NA({u(cid:48) A, u(cid:48) D A NB({u(cid:48) The resulting utility is 1 for both agents. The lazy algorithm returns M ozart for both agents: D A LA({u(cid:48) D A LB({u(cid:48) A, u(cid:48) A, u(cid:48) B,},{accA, accB}, ActsA) = M ozartA B,},{accA, accB}, ActsB) = M ozartB The resulting utility is 3 for agent A and 4 for agent B. The full algorithm returns Bach for both agents: D A FA({u(cid:48) D A FB({u(cid:48) A, u(cid:48) A, u(cid:48) B,},{accA, accB}, ActsA) = BachA B,},{accA, accB}, ActsB) = BachB The resulting utility is 6 for agent A and 1.1 for agent B. JavaScript Implementation 3.3 We implemented the running examples in JavaScript14. As a basis for the implementa- tion, we created a simple framework that consists of the following components: -- Web socket server: environment and communications manager. The environ- ment and communications interface is implemented by a web socket server that consists of the following components: 14 The code, as well as documentation and tests, are available at http://s.cs.umu.se/qxgbfi. Empathic Autonomous Agents 13 • Environment and communications manager. The web server provides a generic environment and communications manager that relays messages between agents and provides the shared value system of acceptability rules. • Environment specification. The environment specification contains scenario- specific information and enables the server to determine and propagate the util- ity rewards to the agents. -- Web socket clients: empathic agents. The empathic agents are implemented as web socket clients that interact via the server described above. Each agent consists of the following two components: • Generic empathic agent library. The generic empathic agent library provides a function to create an empathic agent object with the properties ID, utilityMap- pings, acceptabilityRules, and type (naive, lazy, or full). The empathic agent object is then equipped with an action determination function that implements the empathic agent algorithm as described above. • Agent specifications. The agent specification consists of the scenario-specific information of all agents in the environment, as well as of the current agents' identifier and type (naive, lazy, or full) and is used to instantiate a specific em- pathic agent. Note that in the implementation, we construct the utility functions right away and do not use actions-to-consequences mappings. The implementation assumes that the specifications provided to both agents agents and to the server is consistent. Fig. 1 depicts the architecture of the empathic agent JavaScript implementation for the vehicle scenario. We chose JavaScript as the lan- Fig. 1: Empathic intelligent: architecture guage for implementing the scenario to show how to implement basic empathic agents using a popular general-purpose programing language, but concede that a more power- ful implementation in the context of MAS frameworks like Jason is of value. 4 Reasoning-loop Architecture We create a reasoning-loop architecture for the empathic agent and again assume a two- agent scenario to simplify the description. The architecture consists of the following components: 14 T. Kampik et al. -- Empathic agent (EA). The empathic agent is the system's top-level component. It has three generic components (observer, negotiator, and interactor) and five dy- namically generated functions/objects (utility function and acceptability function of both agents, as well as a formalized model of the shared system of values). -- Target agent (TA). In the simplest scenario, the empathic agent interacts with ex- actly one other agent (the target agent), which is modeled as a black box. Pre- existing knowledge about the target agent can be part of the models the empathic agent has of the target agent's utility and acceptability functions. -- Shared system of values. The shared system of values allows comparing the utility functions of the agents and creating their acceptability functions, as well as their actions-to-consequences mappings and utility quantification functions, from which the utility functions are derived. -- Utility function. Based on the actions-to-consequences mappings and utility quan- tification functions, each empathic agent maintains its own utility function, as well as models of the utility function of the agent it is interacting with. -- Acceptability function. Based on the shared system of values, the agent derives the acceptability functions (as described above) to then incorporate them into updated utility functions, which it feeds into the empathic agent algorithm to determine the best possible tuple of actions. -- Observer. The observer component scans the environment, registers other agents, receives their utility functions, and also keeps the agent's own functions updated. To construct and update the utility and acceptability functions without explicitly re- ceiving them, the observer could make use of inverse reinforcement learning meth- ods, as for example described by [10]. -- Negotiator. The negotiator identifies and resolves conflicts of interests using the ac- ceptability function models and instructs the interactor to engage with other agents if necessary, in particular, to propose a solution for a conflict of interest, or to re- solve the conflict immediately (depending on the level of confidence that the solu- tion is indeed acceptable). The negotiator could make use of argument-based nego- tiation (see e.g.: [3]). -- Interactor. The interactor component interacts with the agent's environment and in particular with the target agent to work towards the conflict resolution. The means of communication is domain-specific and not covered by the generic architecture. Fig. 2 presents a simple graphical model of the empathic agent's reasoning loop archi- tecture. 5 Alignment with BDI Architecture and Possible Implementation with Jason Our architecture reflects the common belief-desire-intention (BDI) model as based on [7] to some extent: -- If a priori available to both agents in the forms of rules or norms, beliefs, and belief sets are part of the shared value system. Otherwise, they qualify the agents' utility Empathic Autonomous Agents 15 Fig. 2: Empathic intelligent: architecture and acceptability functions directly. In contrast, desires define the objective(s) to- wards which an agent's utility function is optimized and are -- while depending on beliefs -- not directly mutable through persuasive argumentation between the agents. -- Intentions are the tuples of actions the agents choose to execute. -- As it strives for simplicity, our architecture does for now not distinguish between desires and goals, and intentions and plans, respectively. We expect to improve the alignment of our framework with the BDI architecture to facilitate the integration with existing BDI-based theories and implementation using BDI frameworks. The Jason platform for multi-agent system development [6] can serve as the basis for implementing the empathic agent. While simplified running examples of our architecture can be implemented with Jason, extending the platform to provide an empathic agent-specific abstraction layer would better support complex scenarios. 6 Discussion In this section, we place our empathic agent concepts into the context of existing work, highlight potential applications, analyze limitations, and outline future work. 6.1 Similar Conflict Resolution Approaches Our empathic agent can be considered a generic and basic agent model that can draw upon a large body of existing research on multi-agent learning and negotiation tech- niques for possible extensions. A survey of research on agents that model other agents 16 T. Kampik et al. is provided by Albrecht and Stone [1]. The idea of combining a utility-based approach with acceptability rules to emulate empathic behavior is to our knowledge novel. How- ever, a somewhat similar concept is presented by Black and Atkinson, who propose an argumentation-based approach for an agent that can find agreement with one other agent on acceptable actions and can develop a model of the other agent's preferences over time [5]. While Black's and Atkinson's approach is similar in that it reflects Coplan's definition of empathy (it maintains "a process through which [it] simulates another's situated psychological states, while maintaining clear self -- other differentiation" [12]) to some extent we identify the following key differences: -- The approach is limited to a two-agent scenario. -- The agent model is preference-based and not utility-based. While this has the ad- vantage that it does not require reducing complex preferences to a simple numeric value, it makes it harder to combine with existing learning concepts (see below). -- The agent has the ability to learn another agent's preferences over time. However, the learning concept is -- according to Black and Atkinson -- "not intended to be com- plete" [5]. We suggest that while our empathic agent does not provide learning capabilities by default, it has the advantage that its utility-based concept allows for integration with established inverse reinforcement learning algorithms (see: Sub- section 6.4). -- The agent Black and Atkinson introduce is not empathic in that it tries to compro- mise with the other agent, but rather uses its ability to model the agent's preferences to improve its persuasive capabilities by tailoring the arguments it provides to this agent. 6.2 Potential Real-World Use Cases In this chapter, we exemplified the empathic agent with two simple scenarios, with the primary purpose of better explaining our agent's core concepts. These scenarios do not fully reflect real-world use cases. However, the core concepts of the agent can form the basis of solutions for real-world applications. Below, we provide a non-exhaustive list of use case types empathic agents could potentially address: -- Handling aspects of traffic navigation scenarios that cannot be covered by static rules. Besides adjusting the assertiveness levels to the preferences of their drivers, as suggested by Sikkenk and Terken [23], and Yusof et al. [26], autonomous vehicles could consider the driving style of other human- or agent-controlled ve- hicles to improve traffic flow, for example by adjusting speed or lane-changing behavior according to the (perceived) utility functions of all traffic participants or to resolve unexpected incidents (in particular emergencies). -- Mitigating negative effects of large-scale web applications on their users. Ev- idence exists that suggests the well-being of passive (mainly content-consuming) users of social media is frequently negatively impacted by technology, while the well-being of at least some users, who actively engage with others through the tech- nology, improves [20]. To facilitate social media use that is positive for the users' well-being, an empathic agent could serve as a mediator between user needs (social Empathic Autonomous Agents 17 inclusion) and the business goals of the technology provider (often: maximization of advertisement revenue). -- Decreasing the negotiation overhead for agent-based manufacturing systems. Autonomous agent-based manufacturing systems are an emerging alternative to traditional, hierarchically managed control architectures [16]. While agent-based systems are considered to increase the agility of manufacturing processes, one dis- advantage of agent-based manufacturing systems is the need for negotiation be- tween agents and the resulting overhead (see for example: Bruccoleri et al. [8]). Employing empathic agents in agent-based manufacturing scenarios can possibly help solve conflicts of interests efficiently. -- Improving persuasive healthcare technology. Persuasive technology -- "comput- erized software or information system designed to reinforce, change or shape atti- tudes or behaviours or both without using coercion or deception" [18] -- is frequently applied in healthcare scenarios [11], in particular, to facilitate behavior change. Per- suasive functionality is typically implemented using recommender systems [14], which in general struggle to compromise between system provider and end-user needs [21]. This can be considered as a severe limitation in healthcare scenarios, where trade-offs between serving public health needs (optimizing for a low burden on the healthcare system) and empowering patients (allowing for a subjective as- sessment of health impact, as well as for unhealthy choices to support individual freedom) need to be made. Hence, employing the empathic agent concepts in this context can be considered a promising endeavor. 6.3 Limitations The purpose of this chapter is to introduce empathic agents as a general concept. When working towards a practically applicable empathic agent, the following limitations of our work need to be taken into account: -- The agent is designed to act in a fully observable world, which is an unrealistic assumption for real-world use cases. For better applicability, the agent needs to support probabilistic models of the environment, the other agents, and the shared value system. -- Our formal empathic agent description is logic-based. Integrating it with Markov decision process-based inverse reinforcement learning approaches is a non-trivial endeavor, although certainly possible. -- In the example scenarios we provided, all agents are identically implemented em- pathic agents. An empathic agent that interacts with non-empathic agents will need to take into account further game-theoretic considerations and to have negotiation capabilities. -- The presented empathic agent concepts use a simple numeric value to represent the utility an agent receives as a consequence of the execution of an action tuple. While this approach is commonly employed when designing utility-based autonomous agents, it is an oversimplification that can potentially limit the applicability of the agent. 18 T. Kampik et al. -- Software engineering and technological aspects of empathic agents need to be fur- ther investigated. In particular, the implementation of an empathic agent library using a higher-level framework for multi-agent system development, as we discuss in Section 5 could provide a more powerful engineering framework for empathic agents. 6.4 Future Work We suggest the following research to address the limitations presented in Subsection 6.3: -- So far, we have chosen a logic-based approach to the problem in focus to allow for a minimalistic problem description with low complexity. Alternatively, the prob- lem could be approached from a reinforcement learning perspective (see for an overview of multi-agent reinforcement learning: [9]). Using (partially observable) Markov decision processes, one can introduce a well-established temporal and probabilistic perspective15. A key capability our empathic agent needs to have is the ability to learn the utility function of other agents. A comprehensive body of research on enabling this ability by applying inverse reinforcement learning exists (for example: [10] and [17]). Hence, creating a Markovian perspective on the em- pathic agent to enable the application of reinforcement learning methods for the observational learning of the utility functions of other agents can be considered relevant future work. -- To better assess the applicability of the empathic agent algorithms, it is important to analyze its computational complexity in general, as well as to evaluate it in the context of specific use cases that might allow for performance-improving adjust- ments. -- To enable empathic agents to reach consensus in case of inconsistent beliefs argumentation-based negotiation approaches can be applied that consider uncer- tainty and subjectivity (e.g. [15]) for creating solvers for finding compromises be- tween utility/acceptability functions. Similar approaches can be used to enhance utility quantification capabilities by considering preferences and probabilistic be- liefs. -- The design intention of the architectural framework we present in Section 4 is to form a high-level abstraction of an empathic agent that is to some extent agnos- tic of the concepts the different components implement. We are confident that the framework can be applied in combination with existing technologies to create a real-world applicable empathic agent framework, at least for use cases that allow making some assumptions regarding the interaction context and protocol. -- The ultimate goal of this research is to apply the concept in a real-world scenario and evaluate to what extent the application of empathic agents provides practically relevant benefits. 15 However, the same can be achieved with temporal and probabilistic logic. Empathic Autonomous Agents 19 7 Conclusion In this chapter, we introduced the concept of an empathic agent that proactively iden- tifies potential conflicts of interests in interactions with other agents and uses a mixed utility-based/rule-based approach to find a mutually acceptable solution. The theoretical framework can serve as a general purpose model, from which advanced implementa- tions can be derived to develop socially intelligent systems that consider other agents' (and ultimately humans') welfare when interacting with their environment. The exam- ple implementation, the reasoning-loop architecture we introduced for our empathic agent, and the discussion of how the agent can be implemented with a belief-desire- intention approach provide first insights into how a more generally capable empathic agent can be constructed. As the most important future research steps to advance the empathic agent, we regard the conceptualization and implementation of an empathic agent with learning capabilities, as well as the development of a first simple empathic agent that solves a particular real-world problem. Acknowledgements We thank the anonymous reviewers for their constructive criti- cal feedback. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. References 1. Albrecht, S.V., Stone, P.: Autonomous agents modelling other agents: A comprehensive sur- vey and open problems. Artificial Intelligence 258, 66 -- 95 (May 2018) 2. Alshabi, W., Ramaswamy, S., Itmi, M., Abdulrab, H.: Coordination, cooperation and conflict resolution in multi-agent systems. In: Sobh, T. (ed.) Innovations and Advanced Techniques in Computer and Information Sciences and Engineering. pp. 495 -- 500. Springer Netherlands, Dordrecht (2007) 3. Amgoud, L., Dimopoulos, Y., Moraitis, P.: A unified and general framework for argumentation-based negotiation. In: Proceedings of the 6th International Joint Conference on Autonomous Agents and Multiagent Systems. pp. 158:1 -- 158:8. AAMAS '07, ACM, New York, NY, USA (2007) 4. Berinsky, A.J.: Rumors and health care reform: experiments in political misinformation. British Journal of Political Science 47(2), 241 -- 262 (2017) 5. Black, E., Atkinson, K.: Choosing persuasive arguments for action. In: The 10th Interna- tional Conference on Autonomous Agents and Multiagent Systems-Volume 3. pp. 905 -- 912. International Foundation for Autonomous Agents and Multiagent Systems (2011) 6. Bordini, R.H., Hubner, J.F.: BDI agent programming in AgentSpeak using Jason. In: Inter- national Workshop on Computational Logic in Multi-Agent Systems. pp. 143 -- 164. Springer (2005) 7. Bratman, M.: Intention, Plans, and Practical Reason. Center for the Study of Language and Information (1987) 8. Bruccoleri, M., Nigro, G.L., Perrone, G., Renna, P., Diega, S.N.L.: Production planning in reconfigurable enterprises and reconfigurable production systems. CIRP Annals 54(1), 433 -- 436 (2005) 20 T. Kampik et al. 9. Busoniu, L., Babuska, R., De Schutter, B.: A comprehensive survey of multiagent reinforce- ment learning. IEEE Trans. Systems, Man, and Cybernetics, Part C 38(2), 156 -- 172 (2008) 10. Chajewska, U., Koller, D., Ormoneit, D.: Learning an agent's utility function by observing behavior. In: ICML. pp. 35 -- 42 (2001) 11. Conroy, D.E., Yang, C.H., Maher, J.P.: Behavior change techniques in top-ranked mobile apps for physical activity. American journal of preventive medicine 46(6), 649 -- 652 (2014) 12. Coplan, A.: Will the real empathy please stand up? a case for a narrow conceptualization. The Southern Journal of Philosophy 49(s1), 40 -- 65 (2011) 13. Dautenhahn, K.: The art of designing socially intelligent agents: Science, fiction, and the human in the loop. Applied artificial intelligence 12(7-8), 573 -- 617 (1998) 14. Hors-Fraile, S., Rivera-Romero, O., Schneider, F., Fernandez-Luque, L., Luna-Perejon, F., Civit-Balcells, A., de Vries, H.: Analyzing recommender systems for health promotion using a multidisciplinary taxonomy: A scoping review. International journal of medical informatics 114, 143 -- 155 (2018) 15. Marey, O., Bentahar, J., Khosrowshahi-Asl, E., Sultan, K., Dssouli, R.: Decision making under subjective uncertainty in argumentation-based agent negotiation. Journal of Ambient Intelligence and Humanized Computing 6(3), 307 -- 323 (Jun 2015) 16. Monostori, L., V´ancza, J., Kumara, S.: Agent-based systems for manufacturing. CIRP An- nals 55(2), 697 -- 720 (2006) 17. Ng, A.Y., Russell, S.J., et al.: Algorithms for inverse reinforcement learning. In: Icml. pp. 663 -- 670 (2000) 18. Oinas-Kukkonen, H., Harjumaa, M.: Towards deeper understanding of persuasion in soft- ware and information systems. In: Advances in computer-human interaction, 2008 first in- ternational conference on. pp. 200 -- 205. Ieee (2008) 19. Osborne, M.J., Rubinstein, A.: A course in game theory. MIT press (1994) 20. Philippe, V., Oscar, Y., Maxime, R., John, J., Ethan, K.: Do social network sites enhance or undermine subjective well-being? a critical review. Social Issues and Policy Review 11(1), 274 -- 302 21. Ricci, F., Rokach, L., Shapira, B.: Recommender systems: introduction and challenges. In: Recommender systems handbook, pp. 1 -- 34. Springer (2015) 22. Russell, S.J., Norvig, P.: Artificial intelligence: a modern approach. Malaysia; Pearson Edu- cation Limited, (2016) 23. Sikkenk, M., Terken, J.: Rules of conduct for autonomous vehicles. In: Proceedings of the 7th International Conference on Automotive User Interfaces and Interactive Vehicular Ap- plications. pp. 19 -- 22. AutomotiveUI '15, ACM, New York, NY, USA (2015) 24. Von Neumann, J., Morgenstern, O.: Theory of games and economic behavior. Bull. Amer. Math. Soc 51(7), 498 -- 504 (1945) 25. Wojdynski, B.W., Bang, H.: Distraction effects of contextual advertising on online news pro- cessing: an eye-tracking study. Behaviour & Information Technology 35(8), 654 -- 664 (2016) 26. Yusof, N.M., Karjanto, J., Terken, J., Delbressine, F., Hassan, M.Z., Rauterberg, M.: The exploration of autonomous vehicle driving styles: Preferred longitudinal, lateral, and vertical accelerations. In: Proceedings of the 8th International Conference on Automotive User Inter- faces and Interactive Vehicular Applications. pp. 245 -- 252. Automotive'UI 16, ACM, New York, NY, USA (2016)
1904.02895	1	1904	2019-04-05T06:59:52	Sensory Regimes of Effective Distributed Searching without Leaders	[ "cs.MA" ]	Collective animal movement fascinates children and scientists alike. One of the most commonly given explanations for collective animal movement is improved foraging. Animals are hypothesized to gain from searching for food in groups. Here, we use a computer simulation to analyze how moving in a group assists searching for food. We use a well-established collective movement model that only assumes local interactions between individuals without any leadership, in order to examine the benefits of group searching. We focus on how the sensory abilities of the simulated individuals, and specifically their ability to detect food and to follow neighbours, influence searching dynamics and searching performance. We show that local interactions between neighbors are sufficient for the formation of groups, which search more efficiently than independently moving individuals. Once a member of a group finds food, this information diffuses through the group and results in a convergence of up to 85\% of group members on the food. Interestingly, this convergence behavior can emerge from the local interactions between group members without a need to explicitly define it. In order to understand the principles underlying the group's performance, we perturb many of the model's basic parameters, including its social, environmental and sensory parameters. We test a wide range of biological-plausible sensory regimes, relevant to different species and different sensory modalities and examine how they effect group-foraging performance. This thorough analysis of model parameters allows for the generalization of our results to a wide range of organisms, which rely on different sensory modalities, explaining why they move and forage in groups.	cs.MA	cs	SENSORY REGIMES OF EFFECTIVE DISTRIBUTED SEARCHING WITHOUT LEADERS A PREPRINT Ravid Cohen School of Computer Science, Tel-Aviv University, Tel-Aviv 69978, Israel Yossi Yovel Sagol School of Neuroscience, Tel Aviv University, Tel Aviv 39040, Israel Dan Halperin School of Computer Science, Tel-Aviv University, Tel-Aviv 69978, Israel April 8, 2019 ABSTRACT Collective animal movement fascinates children and scientists alike. One of the most commonly given explanations for collective animal movement is improved foraging. Animals are hypothesized to gain from searching for food in groups. Here, we use a computer simulation to analyze how moving in a group assists searching for food. We use a well-established collective movement model that only assumes local interactions between individuals without any leadership, in order to examine the benefits of group searching. We focus on how the sensory abilities of the simulated individuals, and specifically their ability to detect food and to follow neighbours, influence searching dynamics and searching performance. We show that local interactions between neighbors are sufficient for the formation of groups, which search more efficiently than independently moving individuals. Once a member of a group finds food, this information diffuses through the group and results in a convergence of up to 85% of group members on the food. Interestingly, this convergence behavior can emerge from the local interactions between group members without a need to explicitly define it. In order to understand the principles underlying the group's performance, we perturb many of the model's basic parameters, including its social, environmental and sensory parameters. We test a wide range of biological-plausible sensory regimes, relevant to different species and different sensory modalities and examine how they effect group-foraging performance. This thorough analysis of model parameters allows for the generalization of our results to a wide range of organisms, which rely on different sensory modalities, explaining why they move and forage in groups. Introduction Why do so many animals move in groups? One of the most accepted benefits of moving in a group is hypothesized to be improved foraging, which is achieved via enhanced sensing. Specifically, searching for food in a group, has been hypothesized to improve searching by allowing individuals to glean information from conspecifics [1, 2, 3, 4, 5]. Group foraging has been suggested to be especially important when searching for an ephemeral resource. This has also been backed by several mathematical frameworks [6, 7, 8]. In parallel to these analytic models of group foraging, many models, mostly numeric (but not exclusively), try to explain how animals maintain a group while moving [9, 10, 11, 12, 13, 14, 15, 16]. Most of these models assume that collective movement emerges from the behavior of individuals and their local interactions with neighbors (sometimes including distant neighbors). The group does not need leaders, and there is no necessity for individual recognition, or signaling to achieve such coordinated behavior. However, only a few attempts have been made to use such collective movement simulations to test the improved searching hypothesis, that is to use a movement model to simulate collective movement and to examine if and how the groups that are formed by this model benefit from enhanced sensing while searching for food. Two examples include A PREPRINT - APRIL 8, 2019 [17, 18]) but they mostly focus on how group foraging improves tracking of the food landscape. Several other models that suggest how group searching may assist sensing (e.g [1, 19]) focus on the situation where the target generates a (noisy) gradient that can be sensed and utilized from a distance. This scenario is relevant for olfactory based searching or for light and shade patterns, but with most sensory modalities (e.g., vision, echolocation, tactile sensing) searching is a binary problem with food being either detectable or not. Some additional attempts to model collective searching have been made in collective robotics (e.g., [19, 20, 21, 22]) but these models often rely on non-biological assumptions, such as active transmission of location-information between individuals (agents) or division of labor (but see [23] for a biological approach). In this work, we use a simple well-studied movement model [24] to examine how moving in a group improves searching. We focus on the sensory aspects of a commonly used agent-based movement model. We test a wide range of biological-plausible sensing regimes, relevant to different species relying on different sensory modalities and we examine how they effect group-foraging performance. We consider a scenario of n agents moving in a two-dimensional region while searching for a sparse resource. To simplify the problem, we simulated a single target (representing food) with an unknown location at each simulation. In the simulations, all agents start moving from the center of the region, each in a random direction. At every time step, each agent adjusts its movement according to its current direction and according to the agents around it (see below). The agents' steps are thus characterized by a constant step-length (i.e., speed is constant) and a continuously changing direction. To simplify the sensory model, we assumed that the agents sense equally in all directions, and can be sensed from all directions. We define two sensing radii: rt is the detection radius of the target, and rs is the detection radius of neighboring individuals. In the case of visual animals, these radii represent the visual range for detecting food and a conspecific respectively, and in the case of echolocating bats or dolphins, where sensing is based on sound, they would represent the echo detection range of a food item, and the eavesdropping range on conspecifics respectively. In reality, the values of these two radii will depend on the sensory modality and the characteristics of the system (e.g., the size of the food and the behavior of the conspecific). We tested several combinations of biologically-reasonable radii and analyzed how they influence group foraging. At each time step an agent advances according to its velocity and direction, where pt agent i at time t. The agent's step magnitude is defined by, i is the location of agent i at time t, x is the agent's step magnitude, and dt pt+1 i = pt i + xdt i, (1) i is the unit direction vector of an x = v cf , (2) where v is the agent's velocity and cf is the sensory direction update rate, i.e., how often is sensory information collected and the movement updated accordingly. This parameter represents an actual biological feature that might be adapted through evolution and we therefore examined how it influences group foraging performance. While searching, the agent's movement direction is determined based on its previous direction and on the location and movement direction of its neighbors. When moving socially, the agent is influenced by neighboring agents as follows: , (3) −→ di , as already mentioned, is the direction of an agent i, −−−→ where social are the individual and social effects on di agent i respectively. The parameter ρ determines the relative weight of the self and the social effects. Note that when ρ is set to zero, the agents are not affected by their neighbors and move independently. When moving as independent individuals, each of our agents applied a correlated walk with Gaussian noise represented −−→ by a distribution of turning angles whose width (σ) we manipulated. The individual movement direction, self, can be di described as follow, −→ di = (cid:13)(cid:13)(cid:13)(1 − ρ) −−−→ −−→ (1 − ρ) di di self + ρ −−−→ −−→ social di di self + ρ social −−→ self and di (cid:13)(cid:13)(cid:13)2 (cid:26)α, di(t) self = t = 0 di(t−1) + β t > 0, α ∼ U[0, 360), β ∼ N (0, σ). −−→ self and di 2 where di and σ represents the width of the distribution of turning angles. At the first step, di self and di are the angles (in degrees) that the vector −→ di form with the positive x-axis respectively, self is drawn from a uniform angle (4) (5) (6) A PREPRINT - APRIL 8, 2019 distribution and afterwards it is determined according to the direction of the agent in the previous step with the addition of Gaussian noise. −−−−→ social, is equal and does not depend on its distance. di −−−→ The social direction of agent i, social, is calculated according to the direction and location of its neighbors. We adopt the di commonly used 3-radii model [24] to govern the interactions between agents. In brief, each agent has three concentric zones in which: repulsion, alignment and attraction govern the interactions respectively (Fig. 1b). The agent will try to avoid collision with neighbors that are located in the repulsion zone by moving away from them. If there are no neighbors in the repulsion zone, the agent will align with neighbors that are located in the alignment zone and it will be attracted to neighbors that are located in the attraction zone. Note that the three interaction zones are limited by the social detection radius rs. That is, only neighbors within a distance rs from the agent are sensed and interacted with. Essentially, the attraction zone was defined by rs while the two internal zones (repulsion and alignment) were smaller than rs and were determined experimentally (Methods). To simplify the model, we assumed that the weight of the contribution of each neighbor to The agents could be in one of three behavioral states: Search, Lock and Find (Fig. 1a). Initially, all the agents start at the Search state. An agent that arrives within the sensing range of the target switches to the Find state and moves directly to the target, while an agent that detects another agent in the Find state switches to the Lock state (we assume that individuals can sense when another individual within rs is feeding). An agent in the Lock state homes in on the location of the agent that found the target. Importantly, agents in the search mode do not lock on agents in the lock mode. In other words, an agent only knows when another agent is feeding but not when another agent detected a feeding agent. All agents that arrive at the target remain there until the end of the simulation (we assume that food patches are very sparse but that they contain plenty of food so that there is no competition).Importantly, we also test the special condition where rs=rt. In this condition, agents can only be in the Search or Find states (but never in the Lock state) and thus, agents only know the positions of their conspecifics, but not their state. This is because an agent will always find the food at the same instance as it finds an individual that found the food. This is a very important condition, because if group foraging is beneficial under this condition, this implies that animals only need to know the positions of their neighbors and not their state in order to gain from group foraging, making it easier to evolve. Figure 1c and 1d and movies S1-2 demonstrate our model's behavior for ρ = 0 and ρ = 0.6 respectively, we highly recommend viewing the movies. Because in nature mostly rt < rs, finding the target by the first agent results in an effective increase of the target- detection radius from rt to rs, as from this point onward, the agents can detect the feeding individuals instead of the food itself (except for the special case described above where rt = rs). Such an increase occurs in nature, for example, when a bird of prey finds a carcass and circles it in the air or when a marine bird detects a school of fish and dives towards it repeatedly. Other birds can now home in on the carcass or the school from much larger distances by detecting the circling or the diving bird [25, 26]. Similarly, when a bat finds a patch of insects, other bats can home in on this patch from larger distances by eavesdropping on the echolocation attack signals emitted by the first bat rather than by finding the insect patch [3]. Interestingly, our model generated a convergence effect on-to the food patch even though we did not explicitly define it. When one individual found the food and was moving towards it, other individuals followed this individuals attracting farther individuals long before they could directly sense the food or the finding individual. Therefore, the information about the location of food diffuses through the group attracting more and more individuals (from farther circles) long before they can sense the food itself or any of the individuals that already found it (Fig. 1e and movie S1). This convergence occurred even in the special case where rt = rs in which individuals never sense that others have found food (they only sense the food directly) proving that it is an emergent property of the movement. We next discuss the effect of different sensory and social model parameters on group searching performance. Importantly, we measured performance from an individual's point of view as the average time to finding food. Results To delineate the sensory regimes where grouping improves searching, we first set out to find the best non-social foraging model, that is, the model that minimizes the time for finding food when the individuals move without interacting. To this end, we set ρ to 0 and we varied σ (the width of turning angle distribution) until we found the value that optimizes searching as individuals (minimized the mean searching time). The best σ was small (in the range 0 < σ ≤ 3 degrees, Fig. S1) meaning that the agents moved almost in straight lines (we only allowed Gaussian noise in the turning angle distribution, so other movement distribution, such as Levy walks were not possible). Note that we only varied σ because in the non-social model, σ is the only meaningful parameter while all other parameters are either irrelevant (because they describe interactions) or were set to be fixed in both the social and the non-social models (see Methods). In the process of finding the best non-social model (and everywhere else, unless stated otherwise), we simulated 50 agents searching 3 A PREPRINT - APRIL 8, 2019 (a) Behavioral state diagram. The distance to the target is less than rt Search Lock F ind The distance to an agent that found the target is less than rs The distance to the target is less than rt (b) (c) (e) (d) Figure 1: The principles of the system. (a) The behavioral state diagram. rt is the target detection radius and rs is the social detection radius. When rs > rt only the first agent will advance from the Search state straight to the Find state while the following agents will detect the first one and thus switch to the Lock state. (b) The social interaction model includes three concentric zones: Attraction, Alignment and Repulsion. The dashed arrows represent the effect of each agent on the social direction of agent i. The black solid arrow is the average direction of arrows 2 and 3. di social is equal to arrow 1 if the repulsion zone is non-empty (i.e. if agent 1 exists) and equal to arrow 4 otherwise. Note that the attraction range is defined by rs. Agents whose distance is larger than rs have no effect on di social (such as agent 4). (c) A snapshot of the simulation where ρ = 0. When ρ was set to 0, and hence individuals were moving independently, the only social interactions occurred when an agent homed in on a neighbor that already found the target (only when rs > rt). The gray lines represent the agents' paths after ∼ 400 steps, the black circles represent the social detection region in which conspecifics can be sensed, the black lines (inside the circles) depict the heading of the agents and the dark-gray disc represents the food detection region. In panels c-e we simulated n = 50 agents. In both (c) and (d) an agent has already detected the food and this is why the target's disc (dark gray) is defined by rs. Note that the agents' circles have a radius of rs 2 (and not rs) because two agents must be at most rs apart in order to sense each other. In these simulations, the size of the search area was set to 4km2 instead of 400km2 for better visualization.(d) A snapshot of the simulation for the case where ρ = 0.25. All symbols are the same as in c. Notice how groups are now formed. (e) A zoom-in on the target's area demonstrates the convergence of the group on-to the target after one of the agents found it, for ρ = 0.6. The white circle in the center depicts the detection range of the food with several individuals who found it. Individuals in the dark grey circle are locked on the individuals that have already found the food. All other individuals (the great majority) are moving in the direction of the food even though they have not detected it and even though they have not detected individual that already found food. 4 A PREPRINT - APRIL 8, 2019 rt = 150m for a single target (food patch) in a 20km · 20km two-dimensional area with a detection ratio of 15 (specifically, we mostly used rs 10m ). This ratio represents a common situation in which an animal can detect its prey from a shorter distance than it can detect its peers. This is, for example, the case for grazing sheep [27] and for scavenging birds of prey [25, 26], and it is also the case for small birds that are searching for seeds on the ground and can see that another bird is pecking from much longer distances than they can detect the seeds themselves. The exact radius values that we started off with (i.e., rt = 10m and rs = 150m) are typical for echolocating bats, which can detect prey from much shorter distance (∼ 10m) than they can detect a neighbor (∼ 150m). Bats can realize when a neighbor found prey based on its echolocation attack signals [3]. Below, we discuss the effect of varying this ratio and varying the specific radii. Note that even in the non-social model, individuals could recognize when an agent within rs from them has found the target and they moved towards it. (a) (b) (c) Figure 2: Social searching improves individual gain. The searching time in all panels was normalized by dividing the results by the searching time of the model where ρ = 0 (each line in the graph was normalized separately) and the results are shown as an average over 1,000 simulations. The searching time was defined as the average time it took an agent to reach the target. In all panels, unless stated otherwise, the number of agents was 50, the target detection radius was set to 10m and the social detection radius to 150m. In all panels, for each ρ we plot the best result depending on σ (deviation in the Gaussian noise of the individual direction). The black line is the same in all panels allowing comparison. (a) The mean searching time as a function of ρ for different numbers of searching agents. The insert shows the results for ρ = 0.6 without normalization. Note the reduction in both the mean searching time and the standard deviation. (b) The mean searching time depending on ρ for different ratio between the social radius and the food detection radius. (c) The mean searching time as a function of ρ for different values of cf (the movement direction update rate). (d) The mean searching time as a function of ρ for different numbers of targets. Insert shows the results for ρ = 0.6 without normalization to allow comparing the actual time to finding a target. (d) After finding the best non-social model, we used the same setup (e.g., same: area, number of agents and sensing radii), while varying the social weight ρ, simulating different degrees of social foraging (we tested the range: 0 ≤ ρ ≤ 0.9). We assumed that finding a social model in which individuals perform better than in the best non-social model would demonstrate that social foraging is beneficial for searching. Indeed, increasing the social weight (ρ), improved searching by a factor of 1.6 in comparison to the best non-social model - the average time to find the food was 1.6 times faster (Fig. 2a-2d, there was a significant Pearson negative correlation between ρ and the mean searching time, P = 0.005. The difference between searching times at ρ = 0.6 and ρ = 0 was significant, P < 10−5, two-sample t-test). The best searching performance was observed when ρ was between 0.6 and 0.9 (the exact value depended on the setup, see below). Social foraging improved searching in all group-sizes we examined (from 10 to 500 individuals), but it improved more in larger groups, reaching an improvement factor of 1.9 in a group of 500 agents (Fig. 2a). Another advantage of searching in a larger group was a reduction in the searching time variability in accordance with the predictions of analytic models [6, 7]. The variance was estimated over multiple simulations and could be thought of as 5 A PREPRINT - APRIL 8, 2019 the variability in searching time over multiple days or nights of foraging. We took the overall variance of all individuals in all simulations with specific parameters (e.g. 50,000 individuals for n = 50 individuals per simulation and 1, 000 simulations). The variability was smaller in larger groups (see insert in Fig. 2a and Fig. S2). Note that the agents did not necessarily form one large searching group that included all individuals, but that they typically split into multiple groups, which were disconnected from each other. In this analysis of group size, and in all further analyses, we always compared our social model to the best non-social model that had the same parameters. For example, if we changed the number of agents in the simulation; we also searched for the σ that optimized the non-social model (ρ = 0) with this new number of agents. To further generalize our findings, we tested the effect of the sensing ratio (rs/rt) on searching in a group. Social searching was always better than the best non-social model, but the best social model (i.e., the best ρ) differed depending on this ratio (best ρ was always between 0.5-0.9 Fig. 2b). The advantage of social foraging decreased as the detection ratio increased, i.e., as the social detection range (of other individuals) increased relative to the target detection range (rs >> rt). The reason for this is that when individuals can detect other individuals from very large distances, they can home in on the food from very large distances once it is found by the first individual, so there is less need for searching in a group. Moreover, when (rs >> rt), moving independently actually has an advantage because the target is usually found faster by the first agent. This is why when the social detection range is very large (e.g., rs = 500m, grey line in Fig. 2b), highly social models (large ρ values) are detrimental; because large groups are formed and the target is found for the first time rather late. Notably, social searching was better than searching independently even in the special case in which the ratio was 1 (rs = rt), when agents never know when a neighbor finds food (because they arrive at the food and at the feeding individuals at the same time). Next, we tested how the optimal social model is influenced by another sensory parameter -- the sensory update rate -- the rate of updating the movement direction based on the positions of the neighbors (cf , see Equation 5). So far, we assumed that agents update their movement direction once a second. Increasing this rate to 10Hz shifted the best ρ from above 0.6 to around 0.3 (Fig. 2c), and improved the performance of social foraging even further in comparison with the non-social model up to factor of 1.9 (so far, the best 50-agent system achieved an improvement of 1.6-fold). The reason for this improvement was that larger and more spread-out groups were formed when a higher update rate was used (Movie S3). The range of the update-rates that we tested (1-10Hz) is typical for biological systems (e.g., [28]). Finally, we tested how the presence of multiple targets affects the results by introducing more targets in the area. We found that the advantage of group searching improved further when more than one target was present, reaching an improvement factor of 4.5 with 100 targets in the area (Fig. 2d). This result proves that our findings hold also in the general case of many food sources. An ideal searching system would be composed of independently moving agents with perfect communication between everyone, such that all individuals are informed when one finds food (assuming that competition over the food is not detrimental). Such a setup is non-realistic for most biological systems, and surely non-realistic for animals that must search large areas for food due to the limited range of social communication and eavesdropping (the range from which an individual can follow another one). In a realistic scenario, such as the one we modeled, two competing goals must be achieved for optimizing group searching. On the one hand, the group should spread as much as possible in order to minimize the searching overlap between its agents. On the other hand, individuals should maintain a connection to other group members so that they converge on-to the target when a member of the group finds it. Being part of a group does not help if the information about finding food by a group member does not reach other individuals. In practice, these two factors typically negatively correlate, meaning that a group that spreads out (i.e., with less searching-overlap) often has poor connection between its members. We found that the number of individuals per group increased rapidly with ρ, reaching a range of 10-16 individuals per group when ρ was set to 0.2 or more (in a system with a total of 50 individuals). Larger groups did not form in our setup probably due to the set of interaction rules that we used. The increase in group size, was accompanied by a decrease in the proportion of converging individuals (Fig. 3a). Convergence is less efficient in larger groups because each individual is influenced by more individuals some of which have not detected the food and have not detected individuals that already found the food. In other words, the movement is more noisy (large groups often split when arriving at the food). The best models showed a convergence proportion of 65-85%, that is, when the target was found by a group member 65-85% of the other members of the same group found it as well (see also Fig. S3). Together, these effects determined the best social model (i.e., the best ρ). The importance of the convergence effect for improved searching can be learned by observing the dynamics of finding food in a single simulation (Fig. 3b). When ρ is set to 0 (blue line) agents find food individually as can be learned from the monotonous increase of the proportion of agents who found the food. In comparison, when ρ is set to 0.6 (red line), every event of finding of the food by an individual, is followed by many others converging on it as can be learned from the staircase shape of the graph. Note that when operating as individuals (blue line), the food is found faster for the first time (because individuals spread over the entire area rapidly), but very few other agents join (See Fig. S4 for averages). 6 A PREPRINT - APRIL 8, 2019 (a) (c) (b) Figure 3: The mechanisms underlying improved social searching. In all panels, the number of agents was 50, the food detection radius was set to 10m and the social detection radius to 150m. Results are an average over 1, 000 simulations. (a) Group size and convergence proportion as a function of ρ. The convergence proportion was defined as the proportion of individuals that converged on-to the target when it was found by a member of the group. Note that the peaks in group size at ρ = 0.25 and ρ = 0.6 are not measurement errors. We ran this analysis with a resolution of 0.02 to confirm this. (b) Searching performance in a single simulation. The number of agents arriving at the target are shown as a function of time for ρ = 0 (blue) and 0.6 (red). When ρ = 0, agents continuously reach the target as individuals. When ρ = 0.6, agents reach the target in groups (as can be learned from the stair case shape of the graph). (c) The normalized cover time of 50% of the area for the average group that is created at each ρ. For example, at ρ = 0.6, the average groups size was 16 agents, so we estimated the time it would take a group of 16 agents to cover 50% of the area. In both the Group Size line of panel a, and in panel c, we exclude the food from the searching area to examine searching dynamics without a reduction in the number of searching agents, which occurs when they reach the target (see Methods for more details). In addition to these two criteria (i.e., good convergence and little overlap), the group must also move efficiently in order to improve searching -- a well spread group that hardly moves will not perform efficient searching. To quantify the overall searching performance of the model we estimated the time it takes groups of different size (equivalent to different values of ρ) to cover 50% of the area (see Methods). This analysis (Fig. 3c) shows that covering is better for social models (ρ > 0). In fact, the best performance reaches a plateau from values of ρ > 0.2 but because groups size is slightly smaller and so is the convergence rate (Fig. 3a) the overall best performance is achieved only at ρ = 0.6 (for a system of 50 agents, see also Fig. S5). Discussion By using a simple collective movement model we show that a group, with no leaders and without any direct commu- nication between its individuals is more efficient in searching for food than individual agents. We tested a realistic biological model using biological plausible sensing ranges. We could have, for example, forced the agents to spread out in an optimally stretched line (with minimal overlap between them) and we could have imposed perfect communication between agents such that even when the most extreme agent in the line found food, all group members converge on it. This would have obviously improved searching, but such a behavior would require defining global rules. Instead, we used a model that only assumes simple local interactions between individuals. Still, even this simple model shows how improved searching can emerge spontaneously when animals move together. Our general finding, that social searching improves searching was robust to various sensing and social model-parameters that we perturbed. We find that better group performance positively correlates with group size, with fast coverage of the areas and with good convergence once food is found (Fig. 3). However, we also show that there is a trade-off between group size and convergence -- as the group increases, less of its members will converge on-to the food when one of the members finds it. The actual improvement achieved by the group and the searching dynamics varied depending on the system's sensory, social and environmental parameters. It is likely that an organism that evolved to perform group searching, has evolved a different set of parameters and rules than those we used; that would further improve searching. For example, such an organism could have evolved a higher sensory update rate, which we find advantageous under certain conditions. Additional sensory abilities that could be tuned by evolution to improve social foraging include an increase in the social detection range (of a conspecific) which could for example be achieved by improved sensing directionality or, in the case of sound emitting animals, by adapting the signal design. Importantly, these adaptations should result in a gain for the individuals in the group as our framework indeed suggests. In reality, the sensing radii and especially the food detection range rt might change seasonally and spatially depending on the available prey. Our prediction would thus be that the degree of social foraging within a species changes seasonally and spatially according to the season and region specific rt . For group searching to improve searching, we only had to make two assumption: (1) Food is distributed in patches, that is, many group members can enjoy a food-patch when it is found (without competition), and (2) A neighbor's location and movement direction can be tracked to some degree. Interestingly, it was not essential to assume that an agent can detect when another agent found the food. We originally assumed this, but found that it is not obligatory. This can 7 A PREPRINT - APRIL 8, 2019 be learned from the results of the model in the special case when rs = rt (purple line in Fig. 2b). When rs = rt an individual must be within the target's detection range to sense that another individual found it, so it will have to reach the food itself and will never detect an individual that has already found the food. Even under this condition, the social model performed better than the non-social model. This condition also clarifies that the detection range of another agent (rs) must not be larger than the target's detection range (rt) for social foraging to be beneficial. Our original intuition was that if the two sensory ranges are the same (as in this case) there is no benefit to the group, but our results show otherwise. The reason that group foraging is beneficial even in this condition, is the convergence effect, which occurs when someone finds the target and occurred even when rs = rt. Namely, even if the target is only sensed when the agent reaches it (i.e., is within rt from it), when a group member reaches the target, its movement will attract others in its direction, leading to a convergence on-to the target. In fact, social searching was more important in the situation when rs = rt than when rs >> rt (compare purple and gray lines in Fig. 2b). In contrast, and slightly counter-intuitive at first, the only situation where the benefit of social foraging was very small, based on our model, was when rs was very large, that is when individuals can sense the location and state of other individuals from a large distance (e.g., grey line in fig. 2b). In this situation, agents can home in on a food item that was found by another individual from a very large distance, so there is less advantage in searching together. This might be the case for vultures that can detect individuals circling a carcass from very large distances. Importantly, the basic interaction model that we used, was sufficient to generate a diffusion of information about the location of the target and a convergence on-to it once it was found by one of the group members. We did not have to explicitly model this behavior, that is, the same interaction rules that were applied during search also resulted in this convergence. This can be learned, once again, from the special case where rs = rt because in this condition individuals never get the chance to know when another individual found food (if they are close enough to a feeding individual, they will already detect the food itself). Moreover, our results suggest that the internal zone of the 3-radii model -- the repulsion zone -- which has been originally designed to prevent collisions between agents, also assists social searching, because it forces the group to spread out and cover more area while searching. As much as we increased ρ, the typical group size did not exceed 17 individuals (larger groups split into several smaller groups). When we tuned other model parameters to force the formation of larger groups, their movement was often not efficient and they tended to get stuck in place. Indeed, in nature, foraging animals are often seen in groups of several individuals to dozens of individuals while swarms of thousands or more (which are often mathematically modeled) are more commonly observed in sleeping sites or in stationary situations when the group remains in place aiming to avoid predation [29]. Mathematical models of group size (e.g., [30]) rarely consider group-movement constraint which seem to play a role based on our model. But, even though the size of a typical group did not grow, when there were more agents in the area, searching became more efficient for the average individual. Food was found faster and the inter-simulation variability decreased dramatically (see insert in Fig. 2a). This was mainly a result of the increase in the target's detection range from rt to rs once it was found by the first agent. This improvement with number of agents was thus not relevant for systems in which rs = rt. All of the system parameters that we tested influenced the performance of the model, but in all cases, there was always a social model that was better than the best comparable individual model. In reality, animals that evolved to rely on social searching might be flexible, adjusting how they weigh personal and social information according to the specific scenario, i.e., the type of food that is available, the sensory modality they are using to search for it and the sensory modality they are using to follow neighbors. Our agent based model thus suggests how we can use simulations to gain intuition about social foraging and more generally, how we can gain insight about the reasons underlying field observations on animal behavior. Methods Experimental setup. The model was defined by a set of environmental parameters and agent parameters. The environmental parameters included: S -- the size of the searching area, n -- the number of agents and t -- the number of targets (whose positions were sampled from a uniform distribution at each simulation). The agent parameters included: v -- the agent's velocity, cf - the movement direction update rate, rt -- the target detection range, rs -- the social detection range, ρ - the weight of the neighbors' influence on the agent's direction, σ - the width of the distribution of turning angles of the Self Direction and r1, r2 - the radii of the concentric (repulsion and alignment) zones (where r3 = rs). We refer to a model with a specific set of parameters as an instance of the model. In our experiments, we fixed some of the parameters based on a simplified but realistic scenario. The fixed parameters were: v = 10m/s and S = 20km · 20km. We simulated a two dimensional world with toroidal boundaries, meaning that when an agent reached a border of the searching area it reappeared on the opposite side. This methodology is commonly used [23] to avoid the edge case when the agents reach the borders. In most simulations we also fixed the following parameters: n = 50, t = 1 ,cf = 1Hz, rt = 10m and rs = 150m, but unlike v and S, these parameters were varied in specific analyses to test 8 A PREPRINT - APRIL 8, 2019 10 rs and r2 was set to 7 their influence on performance . The main aim of this study was to find out whether social behavior improves searching. To that end, we started our experiments by searching for the best instance of the model where the agents search as individuals, i.e. ρ = 0 (see below the definition of searching performance). Note that when ρ = 0 the model depends only on the distribution of turning angles, determined by σ (r1 and r2 are meaningless in this case). We therefore searched over different values of σ = {0 − 10, 15, 20, 30, 50} running 1, 000 simulation for each instance. Then we searched for a social model where ρ > 0 improves searching performance. In order to reduce the number of model parameters we first searched for concentric radii values that showed good performance. We did this by searching over four parameters simultaneously: ρ = {0.1 : 0.1 : 0.9}, σ = {1, 3, 5, 10, 15, 20, 30, 50}, and r1 = {0 : 0.1 ∗ rs : rs} and r2 = {0 : 0.1 ∗ rs : rs} (r1 ≤ r2), total of ∼ 4, 500 instances (4-parameters sets); each instance was tested over 50 simulations. According to the results, we fixed the concentric radii, r1 and r2 that achieved on average the minimal searching time: r1 was set to 3 10 rs. Later, we isolated some of the additional agent parameters and tested their effect on the performance of the model while focusing on changing the social weight ρ = {0 : 0.1 : 0.9} (we always tested different values of σ = {3, 10, 20, 30, 50} for each model instance Fig. S6). In these experiments (after r1 and r2 were fixed) each instance was tested over 1, 000 simulations. We tested the model for different numbers of agents, n = {2, 5, 10, 50, 100, 200, 500}, different ratios between the social detection radius to the target detection rt = {1, 15, 100}, different rates of updating the movement direction, cf = {10Hz, 5Hz, 2Hz, 1Hz} and radius, rs different numbers of targets, t = {1, 2, 5, 10, 30, 50, 100}. Metrics. In order to evaluate the model and explain its behavior we defined the following metrics (see below for formal definitions): Savg - the mean searching time to find a target (Fig. 2a-2d); Cprop - the convergences proportion of a subgroup on-to a target (Fig. 3a); Gsize - the average group size (Fig. 3a); f k ac(x) - the time it takes a subgroup of k agents to cover x percent of the searching area (Fig. 3c). Note that in order to estimate the last three metrics, we excluded all targets from the search area and ran the simulation for a constant time of 500K − 5, 000K time steps (depending on cf ). We used this approach when we did not want to be affected by the convergence of individuals on-to the target, which decreased the number of individuals over time. The metrics were evaluated in each simulation and the reported results are average across all simulations. Next, we formally define the metrics. The mean searching time to find a target, Savg, is defined as follows, where n is the number of agents and si is the time it took agent i to find the target. To define Cprop and f k ac(x) we will first define the average number of connected components, C∗ comp, where n is the number of agents, T is the number of iterations of the simulation, Ccomp(Gt) is the number of connected components in the graph Gt. The Graph Gt, is defined as follows, Gt(V, Et) = (cid:13)(cid:13)2 ≤ rs, (9) i − pt j where n is the number of agents, pt i is the position of an agent i at time t and rs is the social radius. We match a graph for each time step. The vertices of the graph represent the agents, that is vi represent agent i. The edges of the graph represent which agents are in range of detection from each other. That is, we connect two vertices by an edge if the distance between their corresponding agents is less or equal to rs (the social detection radius). For each such graph we calculated the number of connected components and C∗ comp is the average number of connected components over all of the graphs in a single simulation. The convergence proportion of a subgroup to a target, Cprop, is an estimation of the number of agents that will reach the target immediately after it is found by one of the subgroup members (it is possible that some of the agents will miss the target). Cprop is defined as follows, where C∗ directly (i.e., not because an agent followed another agent to the target). As F# is closer to C∗ comp is the average number of connected components and F# is the number of times that the target is found comp the convergence Cprop = , (10) C∗ comp F# 9 n(cid:88) i=1 si, Savg = 1 n Ccomp(Gt), t=0 1 T C∗ comp = T(cid:88) V = {v1, ..., vn} < vi, vj >∈ Et iff(cid:13)(cid:13)pt Et = {< vi, vj >}, (7) (8) A PREPRINT - APRIL 8, 2019 n(cid:88) i=1 effect is larger, i.e. more agents of a subgroup will reach the target when it is found by one of the subgroup's agents. F# is defined as follow, F# = xi (11) where xi indicates (with high probability) if agent i found the target by himself, i.e. not because it followed another agent. xi is defined as follow, (cid:40) 0,∃j (cid:54)= i s.t. 0 < ti − tj ≤ rs∗cf 1, otherwise, v xi = , (12) , which means that it will take an agent rs∗cf where ti is the searching time of agent i, rs is the social detection radius, cf is the movement direction update rate and v is the agent's velocity. Agents can effect each other only if the distance between them is less than rs. At each step an agent advances v time steps to advance a distance of rs. If the cf reaching time between agent i and all the other agents (that reach the target before him) is greater than rs∗cf then we can conclude that agent i found the target by himself. On the other hand, if there is an agent j, that reached the target before agent i and ti − tj ≤ rs∗cf v we can conclude with high probability that agent i followed agent j to the target. We can not be sure that agent i followed agent j because it is possible that they reached the target from different directions. The average group size, Gsize, is defined as follows, v v Gsize = n C∗ comp (13) (14) where n is the number of agents and C∗ To estimate the value of f k ,we first calculated the accumulated cover time of the complete group, f n grid, each cell of size rt stays that way until the end of the simulation. f n ac(x), namely the time it takes a subgroup of k agents to cover x percent of the searching area, ac(x). To do so, we divided S in to an m × m 2 × rt 2 . Each cell that is visited by an agent is considered covered. Once a cell is covered, it comp is the average number of connected components. ac(x) is calculated as follows, }, ac(x) ∼ min{t ≥ 0 : f n mt M ≥ x S where mt is the number of covered cells at time t and M is the total number of cells in the grid. f k follows, ac(x) is calculated as ac(x) ∼ n f k Gsize f n ac(x), (15) ac(x) is the accumulated where n is the number of agents, Gsize is the average number of agents in a subgroup and f n cover time of the complete group. To estimate f k ac(x) we measure the accumulated cover time of the complete group and then we normalized it by the number of agents divided by the average group size. Another approach to measure ac(x) is to run a simulation with k agents, 1 ≤ k ≤ n. In addition to requiring more simulations this latter approach is f k also not accurate. This is because a simulation with a group of k agents will split into subgroups which will reduce the overlap between agents and increase the cover rate efficiency. ACKNOWLEDGMENTS. We thank H. Balaban and N Gov for commenting on the manuscript. This project was partially funded by ONRG grant no. N62909-13-1-N066. Work by D.H. and R.C. has been supported in part by the Israel Science Foundation (grant no. 825/15), by the Blavatnik Computer Science Research Fund, by the Blavatnik Interdisciplinary Cyber Research Center at Tel Aviv University, and by grants from Yandex and from Facebook. References [1] A Berdahl, C J Torney, C C Ioannou, J Faria, and I D Couzin. Emergent sensing of complex environments by animal groups. Science, 339(6119):574 -- 576, 2013. [2] S Creel, P Schuette, and D Christianson. Effects of predation risk on group size, vigilance, and foraging behavior in an African ungulate community. Behavioral Ecology, 25:773 -- 784, 2014. [3] Noam Cvikel, Katya Egert Berg, Eran Levin, Edward Hurme, Ivailo Borissov, Arjan Boonman, Eran Amichai, and Yossi Yovel. Bats aggregate to improve prey search but might be impaired when their density becomes too high. Current Biology, 25(2):206 -- 211, 2015. 10 A PREPRINT - APRIL 8, 2019 [4] Deborah M. Gordon. The ecology of collective behavior. PLoS Biology, 12(3), 2014. [5] E. Danchin, L.A. Giraldeau, T.J. Valone, and R.H. Wagner. Public information: from nosy neighbours to cultural evolution. American Association for the Advancement of Science, 305(5683):487 -- 491, 2004. [6] Colin W. Clark and Marc Mangel. The evolutionary advantages of group foraging. Theoretical Population Biology, 30(1):45 -- 75, 1986. [7] H. Ronald Pulliam and George C. Millikan. Social organization in the nonreproductive season. Avian Biology, pages 169 -- 197, jan 1982. [8] Ehud D. Karpas, Adi Shklarsh, and Elad Schneidman. Information socialtaxis and efficient collective be- havior emerging in groups of information-seeking agents. Proceedings of the National Academy of Sciences, 114(22):5589 -- 5594, 2017. [9] Alessandro Attanasi, Andrea Cavagna, Lorenzo Del Castello, Irene Giardina, Stefania Melillo, Leonardo Parisi, Oliver Pohl, Bruno Rossaro, Edward Shen, Edmondo Silvestri, and Massimiliano Viale. Collective behaviour without collective order in wild swarms of midges. PLoS Computational Biology, 10(7), 2014. [10] Iain D. Couzin, Jens Krause, Nigel R. Franks, and Simon A. Levin. Effective leadership and decision-making in animal groups on the move. Nature, 433(7025):513 -- 516, 2005. [11] Jacques Gautrais, Francesco Ginelli, Richard Fournier, Stéphane Blanco, Marc Soria, Hugues Chaté, and Guy Theraulaz. Deciphering interactions in moving animal groups. PLoS Computational Biology, 8(9), 2012. [12] Máté Nagy, Zsuzsa Ákos, Dora Biro, and Tamás Vicsek. Hierarchical group dynamics in pigeon flocks. Nature, 464(7290):890 -- 893, 2010. [13] Nicholas T. Ouellette. Empirical questions for collective-behaviour modelling. Pramana - Journal of Physics, 84(3):353 -- 363, 2015. [14] Julia K. Parrish and Leah Edelstein-Keshet. Complexity, pattern, and evolutionary trade-offs in animal aggregation. Science, 284(5411):99 -- 101, 1999. [15] David J T Sumpter, Jens Krause, Richard James, Iain D. Couzin, and Ashley J W Ward. Consensus decision making by fish. Current Biology, 18(22):1773 -- 1777, 2008. [16] W. Bialek, A. Cavagna, I. Giardina, T. Mora, E. Silvestri, M. Viale, and A. M. Walczak. Statistical mechanics for natural flocks of birds. Proceedings of the National Academy of Sciences, 109(13):4786 -- 4791, 2012. [17] Mathieu Lihoreau, Michael A. Charleston, Alistair M. Senior, Fiona J. Clissold, David Raubenheimer, Stephen J. Simpson, and Jerome Buhl. Collective foraging in spatially complex nutritional environments. Philosophical Transactions of the Royal Society B: Biological Sciences, 372(1727), 2017. [18] Colin J. Torney, Andrew Berdahl, and Iain D. Couzin. Signalling and the evolution of cooperative foraging in dynamic environments. PLoS Computational Biology, 7(9), 2011. [19] M. Yamashita and I. Suzuki. Agreement on a common X-Y coordinate system by a group of mobile robots. Intelligent Robots: Sensing, Modeling and Planning [Dagstuhl Workshop, September 1-6, 1996], pages 305 -- 321, 1996. [20] Noa Agmon, Noam Hazon, and Gal A. Kaminka. The giving tree: Constructing trees for efficient offline and online multi-robot coverage. Annals of Mathematics and Artificial Intelligence, 52(2-4):143 -- 168, 2008. [21] Sze Kong Chan, Ai Peng New, and Ioannis Rekleitis. Distributed coverage with multi-robot system. In Proceedings - IEEE International Conference on Robotics and Automation, volume 2006, pages 2423 -- 2429, 2006. [22] Sasanka Nagavalli, Andrew Lybarger, Lingzhi Luo, Nilanjan Chakraborty, and Katia Sycara. Aligning coordinate frames in multi-robot systems with relative sensing information. In IEEE International Conference on Intelligent Robots and Systems, pages 388 -- 395, 2014. [23] Luca Giuggioli, Idan Arye, Alexandro Heiblum Robles, and Gal Kaminka. From Ants to Birds: A Novel Bio-Inspired Approach to Online Area Coverage. Springer Tracts in Advanced Robotics. Springer, 9 2016. [24] Iain D. Couzin, Jens Krause, Richard James, Graeme D. Ruxton, and Nigel R. Franks. Collective memory and spatial sorting in animal groups. Journal of Theoretical Biology, 218(1):1 -- 11, 2002. [25] Roi Harel, Orr Spiegel, Wayne M. Getz, and Ran Nathan. Social foraging and individual consistency in following behaviour: testing the information centre hypothesis in free-ranging vultures. Proceedings of the Royal Society B: Biological Sciences, 284(1852):20162654, 2017. [26] A. L. Jackson, G. D. Ruxton, and D. C. Houston. The effect of social facilitation on foraging success in vultures: a modelling study. Biology Letters, 4(3):311 -- 313, 2008. 11 A PREPRINT - APRIL 8, 2019 [27] Francesco Ginelli, Fernando Peruani, Marie-Helène Pillot, Hugues Chaté, Guy Theraulaz, and Richard Bon. Intermittent collective dynamics emerge from conflicting imperatives in sheep herds. Proceedings of the National Academy of Sciences, 112(41):12729 -- 12734, 2015. [28] Nadav S. Bar, Sigurd Skogestad, Jose M. Marçal, Nachum Ulanovsky, and Yossi Yovel. A sensory-motor control model of animal flight explains why bats fly differently in light versus dark. PLoS Biology, 13(1), 2015. [29] D. J. Hoare, I. D. Couzin, J. G.J. Godin, and J. Krause. Context-dependent group size choice in fish. Animal Behaviour, 67(1):155 -- 164, 2004. [30] H. R. Pulliam and T. Caraco. Living in groups: is there an optimal group size? In Behavioural ecology: an evolutionary approach, pages 122 -- 147. 1984. 12 Supplementary A PREPRINT - APRIL 8, 2019 Figure 1: The mean searching time for the non-social model as a function of the width of the turning angle distribution (σ). 13 A PREPRINT - APRIL 8, 2019 Figure 2: The probability density function of the mean searching time for different values of ρ and n. Note how the mean of the distribution decreases for ρ > 0 and how the width decreases with n. 14 A PREPRINT - APRIL 8, 2019 Figure 3: The number of groups and the number of convergence events as a function of ρ. When the two numbers are the same this implies that all group members converged when one of them found the target. The larger the difference between the two - the poorer the convergence. 15 A PREPRINT - APRIL 8, 2019 Figure 4: Number of agents reaching the target as a function of time for different values of ρ. For each ρ we present the results of the best σ. The social graphs (ρ > 0) do not show steps (like in Fig. 3b) because they are an average of 1, 000 simulations. In these simulation the following parameters were used: n = 50, rs rt = 15, σ = 3, t = 1. 16 A PREPRINT - APRIL 8, 2019 Figure 5: Snapshots of the covered area for different models. We exclude the target from the area. (a) n = 1, rt = 30m, ρ = 0, σ = 3, T = 500, 000 time steps (∼ 140 hours). (b) same as a. but with σ = 30. (c) n = 10, rt = 30m, ρ = 0, σ = 3, T = 50, 000 (∼ 14 hours). (d) same as c. but with ρ = 0.6. 17 A PREPRINT - APRIL 8, 2019 Figure 6: Mean searching time for different combinations of ρ and σ. In these simulation the following parameters were used: n = 50, rs/rt = 15, t = 1. 18 A PREPRINT - APRIL 8, 2019 Movie S1. Demonstration of individual search. Movie S2. Demonstration of social search. Movie S3. Demonstration of group motion for different ρ and σ. 19
1604.04737	1	1604	2016-04-16T12:13:02	Studying the impact of negotiation environments on negotiation teams' performance	[ "cs.MA", "cs.AI" ]	In this article we study the impact of the negotiation environment on the performance of several intra-team strategies (team dynamics) for agent-based negotiation teams that negotiate with an opponent. An agent-based negotiation team is a group of agents that joins together as a party because they share common interests in the negotiation at hand. It is experimentally shown how negotiation environment conditions like the deadline of both parties, the concession speed of the opponent, similarity among team members, and team size affect performance metrics like the minimum utility of team members, the average utility of team members, and the number of negotiation rounds. Our goal is identifying which intra-team strategies work better in different environmental conditions in order to provide useful knowledge for team members to select appropriate intra-team strategies according to environmental conditions.	cs.MA	cs	Studying the Impact of Negotiation Environments on Negotiation Teams' Performance V´ıctor S´anchez-Anguix, Vicente Juli´an, Vicente Botti, Ana Garc´ıa-Fornes Universidad Polit´ecnica de Valencia, Departamento de Sistemas Inform´aticos y Computaci´on, Cam´ı de Vera s/n, 46022, Valencia, Spain, {sanguix,vinglada,vbotti,agarcia}@dsic.upv.es Abstract In this article we study the impact of the negotiation environment on the perfor- mance of several intra-team strategies (team dynamics) for agent-based negoti- ation teams that negotiate with an opponent. An agent-based negotiation team is a group of agents that joins together as a party because they share common interests in the negotiation at hand. It is experimentally shown how negotiation environment conditions like the deadline of both parties, the concession speed of the opponent, similarity among team members, and team size affect perfor- mance metrics like the minimum utility of team members, the average utility of team members, and the number of negotiation rounds. Our goal is identifying which intra-team strategies work better in different environmental conditions in order to provide useful knowledge for team members to select appropriate intra-team strategies according to environmental conditions. Keywords: Negotiation Teams, Agreement Technologies, Automated Negotiation, Collective Decision Making, Multi-agent Systems 1. Introduction Agreement technologies [26, 37] conform an emergent research area among scholars in artificial intelligence and autonomous agent systems. Autonomous software agents act reactively and proactively with the objective of maximizing their human users' goals. Nevertheless, as systems tend to be more complex, so do agents' goals, and agents cannot achieve their goals without the cooperation of other agents. Given the open nature of many multi-agent systems, con- flict may be inherent among agents. Hence, distributed mechanisms that allow agents to solve conflict and cooperate are a necessity. Agreement technologies have been actively researched bearing in mind the aforementioned necessity. Automated negotiation [22, 18, 25] is one of the core topics in agreement technologies. Basically, agents in conflict engage in an automatic offer exchange process which gradually leads towards a final solution, or agreement, that solves conflict and makes cooperation among agents possible. The most common use Preprint submitted to Elsevier October 26, 2018 for automated negotiation has been electronic commerce [24], but it should be highlighted that the applicability of this technology has been demonstrated in other domains like collaborative design [21], labor management disputes [42], and mediation between human negotiation parties [3]. Despite being widely studied by scholars from different disciplines like arti- ficial intelligence, game theory, and social sciences, studies have largely focused on processes whose parties (bilateral, or multiparty) are formed by single indi- viduals [10, 8, 11, 16, 6, 12, 1, 14, 13]. However, some real world scenarios bring about negotiation parties that are formed by more than a single individual. For instance, when an organization negotiates with another organization the selling of a product line, it is usual for organizations to send a group of representatives to negotiate with the other organization. Another example, probably a more quotidian example, involves a married couple that negotiates the purchase of a house with a seller. In this case, the married couple is actually a negotiation party which is formed by two individuals instead than a single individual party. To conclude with the list of real examples, the reader could also think of a group of friends that want to go on a holiday together. This party, conformed by all the friends, has to negotiate a deal with the travel agency if they want to achieve their desired goal. This kind of multi-individual party is known in the social sciences as a ne- gotiation team [43, 2, 1, 14]: a group of interdependent people that join and act together as a single negotiation party because of their shared interests, related to a negotiation. The rationale behind negotiation teams is mainly twofold. First, team members may have different expertise and negotiation skills that are needed to tackle the negotiation problem successfully. Second, the multi- individual entity that negotiates may be formed by multiple stakeholders with different sub-goals and preferences regarding the final negotiation outcome. We can imagine how an IT company may send a negotiation team formed by experts (different knowledge and skills) from the sales department, marketing depart- ment, and R&D department to successfully negotiate a new project with the local administration, how the wife and the husband may have different opinions with respect to house pricing, location, and facilities, and how each friend may have different interests regarding hotel location, number of days to spend, and pricing regarding their travel. Electronic applications, and consequently automated negotiation, are not alien to scenarios that may involve agent-based negotiation teams (ABNT). For instance, group travel e-markets, group buying in e-markets, electronic management of farming cooperatives, negotiation support systems for real hu- man teams, and agent-based simulation may be some of the applications where ABNT may be used. From our point of view, we are interested in ABNT whose members may have different preferences regarding the negotiation issues, and, more specifically, we are interested in models for electronic markets. In this paper, we present four intra-team strategies for an ABNT that ne- gotiates with a single opponent. Intra-team strategies, also known as team dynamics, govern which decisions are taken as a team, and how and when those decisions are taken [34]. The relationship between intra-team strategies and 2 team performance is direct. Hence, it became the focus of our current research. It has been documented that environment conditions such as the deadline, con- cession speed, and reservation utility may affect the impact of single-individual bilateral strategies [10]. However, in the team case, new conditions like the number of team members, team preferences' diversity, and the emergent ef- fect of aggregating team members' behaviors/actions may also end up affecting team performance. Prior to the negotiation process, negotiation teams face the challenge of selecting which intra-team strategy should be employed. If environ- mental conditions have an effect on the performance of the different intra-team strategies, the intra-team strategy for the negotiation at hand should be selected accordingly to the current environmental conditions inferred by team members. Our research goal is identifying which intra-team strategies perform better ac- cording to different negotiation environments under different team performance measures. The long term goal is employing the results of this article for helping team members to select the proper intra-team strategy. Hence, four intra-team strategies that guarantee four minimum levels of unanimity regarding team decisions are presented in this article: representa- tive (no unanimity guaranteed), Similarity Simple Voting (plurality/majority guaranteed), Similarity Borda Voting (semi-unanimity guaranteed), and Full Unanimity Mediated (unanimity guaranteed). Due to the large amount of vari- ables that may affect the negotiation, we employ an empirical approach to study the behavior of the four intra-team strategies. We study and identify which are the most appropriate strategies according to different environmental conditions and team performance measures. This article, is partially based on our previous work regarding intra-team strategies for negotiation teams [35, 33], where we presented initial results and simulations. In this article, we extend our empiri- cal experiments by incorporating new environmental conditions (i.e., team size, different deadlines), carrying out a more fine-grained analysis of previous envi- ronmental conditions (i.e., deadline, concession speeds), and presenting revised versions of the four intra-team strategies. The article is organized as follows. First, we describe the assumptions of our negotiation model (Section 2). After that, the details of the four intra- team strategies are thoroughly described in Section 3. Then, in Section 4 the article depicts which negotiation environments and team performance metrics have been studied, and it presents the results and analysis of our experiments. Afterwards, the present work is related to other works in the area of artificial intelligence and automated negotiation (Section 5). Finally, we briefly state the conclusions of our study and point out some future and interesting lines of work in Section 6. 2. General Model Description In this section, we describe the assumptions of our model. We have divided the assumptions in two different categories: general assumptions and opponent assumptions. The general assumptions directly affect the nature of the negoti- ation at hand and are shared between parties (e.g., protocol, number of parties, 3 attribute types, etc.), whereas opponent assumptions describe the strategy car- ried out by the opponent. 2.1. Negotiation Setting • The team A is formed by M different agents ai, 1 ≤ i ≤ M . It should be stated that team membership is considered static during the negotiation process. Dynamic ABNT are not considered in this work, and they are appointed as future work. • The common goal of the team A is negotiating a successful deal with the opponent op. Thus, in this case we assume an implicit representation of the teams' goal. • It is assumed that information is private, even among team members. Therefore, agents do not know other agents' utility functions, strate- gies, reservation utilities, or deadlines. On top of that, we also assume agents with bounded computational resources. Thus, we take a heuristic approach which seeks near optimal results while being computationally tractable. • It is assumed that the team A and the opponent op communicate following an alternating bilateral protocol [30]. One of the two parties acts as the initiator, and it is entitled to propose the first offer. The other party receives the offer and can respond with two different actions: accept the offer (successful negotiation), or propose a counteroffer. If a counteroffer is proposed, the initiator party receives the offer and it can either accept the counteroffer or propose another offer, starting a new negotiation round. Depending on the intra-team strategy, one of the team members or a team mediator is responsible of the communications with the opponent. In this setting, the fact that one of the parties is a team remains unknown to the other party. • Additionally, it is also assumed that the negotiation is time-bounded, and each party has a private deadline TA (team deadline), Top (opponent dead- line). When its deadline is achieved, the party leaves the negotiation and it is considered a failed negotiation. In the case of TA, it is considered a joint deadline for all of the team members, who have agreed upon this deadline prior to the negotiation at hand. • The mediator, if present, is never a perfect mediator that aggregates the utility functions of all the team members. This assumption is taken due to the fact that, depending on the application, some team members may not be completely trustable and may attempt to exaggerate/change their pref- erences to manipulate the negotiation process. This mischievous behavior is easily carried out when aggregating utility functions. • The negotiation domain is comprised of n real attributes whose domains can be scaled to [0, 1]. Thus, the possible number of offers is [0, 1]n. In this 4 domain, a complete offer is represented as X = {x1, x2, . . . , xn}, where xi is a specific instantiation of attribute i. Additionally, we use the notation X t i→j to denote that offer X was sent at round t from party i to party j. • Every agent i (team member or opponent) has its preferences represented by means of linear additive utility functions in the form: Ui(X) = wi,1 Vi,1(xi,1) + wi,2 Vi,2(x2) + ... + wi,n Vi,n(xn) (1) ized so that(cid:80)n where X is a complete offer, xj, is the value given to the j-th attribute, Vi,j(.) is the valuation function for attribute j used by agent i to normal- ize the attribute value to [0, 1], and wi,j is the weight/importance given by agent i to attribute j in the negotiation process. Several observations should be made regarding these utility functions: (i) weights are normal- j=1 wi,j = 1; (ii) attributes are assumed to be independent from each other. Thus, the valuation of one of the attributes does not alter the others attributes' valuation; (iii) it is assumed that valuation functions are either monotonically increasing or monotonically decreas- ing. Moreover, we assume that team members share the same type of monotonic function (i.e., increasing or decreasing) for each Vi,j(.). As for the opponent, it is assumed that the monotonic function for Vi,j(.) is the opposite type to that of team members. It is reasonable to assume this model for valuation functions in e-commerce scenarios. Buyers usually share the same type of valuation function for attributes such as the price (monotonically decreasing), product quality (monotonically increasing), and the dispatch time (monotonically decreasing), whereas sellers usually use the opposite type of monotonic functions (monotonically increasing for price, monotonically decreasing for product quality, and monotoni- cally increasing for dispatch time); (iv) attribute weights wi,j are different for team members. This way, we are able to represent the fact that some team members may be more interested in some attributes whereas other team members may be more interested in other attributes (e.g., some team members prefer price over quality, while others give a higher priority to the product quality). Obviously, the weights of the opponent's utility function may be different from those of team members; (v) since team members share the same type of monotonic function, if one of the team members increases its utility by increasing/decreasing one of the attribute values, the other team members will stay at the same utility level or they will also increase their utility. Thus, there is potential for cooperation among team members. • The opponent has a reservation utility RUop. Any offer whose utility is lower than RUop will be rejected. Each team member ai has a private reservation utility RUai . This individual reservation utility is not shared among teammates. Therefore, a team member ai will reject any offer whose value is under RUai . In this setting, reservation utilities represent the individual utility of each agent if the negotiation process fails. For the 5 experiments, the reservation utilities are drawn from uniform distributions RUop = U [0, 0.25] and RUai = U [0, 0.25]. 2.2. Opponent Model The opponent op is modeled as a single agent for the sake of simplicity. We acknowledge that the other party may be another team (e.g., organization vs organization), but the focus of our study in this article, and the kind of applications in our mind, involve negotiations between one team and one single agent. For the opponent model we used well-known strategies in the agent negotiation literature: • The opponent uses a time-based tactic during the negotiation process. In multi-attribute, time-bounded negotiations, assuming that agents con- cede gradually to reach an agreement is usual. This is especially important when agents do not know other agents' preferences and strategies, since an exploration of the agreement space is required for discovering possi- ble agreements. A concession strategy typically starts by demanding the maximum aspiration and, as the negotiation process advances, the aspi- ration demanded tends to be lowered. The speed at which the strategy concedes is regulated by the concession speed parameter. In this article, we employed a time-based concession tactic inspired in tactics used by other authors [10, 23]: sop(t) = 1 − (1 − RUop)( t Top 1 βop ) (2) where t is the current negotiation round and βop is a parameter that gov- erns the concession speed. On the one hand, when βop = 1 the concession is linear and each negotiation round the same amount of concession is performed, and when βop < 1 the concession is Boulware and very little is conceded at the start of the negotiation process but the agent concedes faster as the negotiation deadline approaches. On the other hand, when βop > 1 the tactic is conceder and the agents concede fast towards the reservation utility in the first rounds. • The opponent uses an offer acceptance criterion acop(.) during the nego- tiation process. It is formalized as follows: (cid:26) accept reject acop(X t A→op) = if sop(t + 1) ≤ Uop(X t otherwise A→op) (3) where t is the current round, X t A→op is the offer received from the team, Uop(.) is the utility function of the opponent, and sop(.) is the opponent concession strategy. Thus, an offer is accepted by op if it reports a utility that is equal to or greater than the utility of the offer that op would propose in the next round. 6 op→A whose utility is Uop(X t • When the opponent has to propose a counteroffer at negotiation round t, it proposes an offer X t op→A) = sop(t). However, depending on the negotiation domain, there may be an infinite number of offers that comply with the previous condition. Similarity heuristics are a largely used family of heuristics in agent-based negotiation literature [11, 23, 34]. Thus, we assumed that the opponent attempts to propose the offer that is the most similar to the last offer received from the team, and whose utility is Uop(X t op→A) = sop(t). The Euclidean distance was used as similarity function. 3. Intra-Team Strategies An intra-team strategy defines what decisions have to be taken by a negoti- ation team, how those decisions are taken, and when those decisions are taken. In a bilateral negotiation process between a team and an opponent, the deci- sions that must be taken (what) are which offers are sent to the opponent, and whether or not opponent offers are accepted. Given the fact that a negotiation team is formed by more than a single individual, decisions should take into account the interests of the team members. How decisions are taken will deter- mine the satisfaction level of the team with the final decision. In this article, most of the team interactions in intra-team strategies are carried out during the negotiation (when). Next, we describe the four intra-team strategies that we propose: Represen- tative (RE), Similarity Simple Voting (SSV), Similarity Borda Voting (SBV) and Full Unanimity Mediated (FUM). Each strategy is capable of guaranteeing a minimum level of unanimity regarding the offer sent to the opponent, and whether or not to accept the opponent's offer. Another difference between the four strategies is the presence of a mediator and its level of activity (none, coordi- nation tasks, very active participation). Table 1 captures the main qualitative differences between the four intra-team strategies according to the aforemen- tioned criteria. Intra-Team Strategy Unanimity level (How) Pre-negotiation? (When) Mediated? RE SSV SBV FUM Minimum Minimum Minimum Information sharing Unilateral Plurality/Majority Semi-Unanimity Unanimity No Only coordination Only coordination Very Active Table 1: A brief qualitative comparison among the four intra-team strategies presented in this article 3.1. Representative (RE) The Representative strategy (RE) is perhaps the simplest intra-team strat- egy. Basically, one of the team members is selected as representative are for the team during the negotiation. This agent will act on behalf of the team during the negotiation, making it responsible of selecting which offers are sent 7 to the opponent, and whether or not opponent's offers are accepted. The only communications are those carried out between the representative agent are and the opponent aop, and, therefore, this strategy is equivalent to a classic bilateral strategy. The representative agent negotiates according to its own utility function Uare(.) since it does not know the utility function of the other participants. The two decisions that have to be taken during the negotiation are which offers are sent to the opponent, and whether or not the opponent's offer is accepted. 3.1.1. Offer proposal Being a time-bounded negotiation, the representative employs a time-based concession tactic sare(.) to negotiate with the opponent. It is based on a team deadline TA and a concession speed βA, which have been agreed upon prior to the negotiation start: sare(t) = 1 − (1 − RUare)( 1 βA ) t TA (4) The concession strategy defines the aspiration level (utility demanded) by the agent at a specific round t. This utility is demanded from the point of view of the representative, and, so, any offer X t A→op proposed by are at round t will obey the following condition: Uare(X t A→op) = sare(t) (5) Since there is a large number of offers that may obey the equation above, we aimed to satisfy the opponent's preferences as much as possible. As in the case of the opponent strategy, the representative selects the offer that is the most similar to the previous offer received from the opponent using a similarity heuristic based on the Euclidean distance. X t A→op = 3.1.2. Offer acceptance max XUare (X)=sare (t) Sim(X, X t−1 op→A) (6) A common acceptance criterion in time-bounded negotiations is that an opponent's offer is accepted if it reports a utility which is higher than or equal to the utility that is to be demanded in the next negotiation step. In the case of the representative, it will accept the opponent's offer X t op→A at round t if it reports a utility Uare(X t op→A) greater than or equal to sare(t + 1). This can be formalized as follows: if sare(t + 1) ≤ Uare(X t otherwise op→A) (7) 8 (cid:26) accept reject acare(X t op→A) = 3.1.3. Unanimity Level It is clear that since the representative negotiates according to its own util- ity function and reservation utility, it cannot guarantee any kind of unanimity regarding team decisions. Decisions taken by the representative are acceptable to himself (1 agent), but nothing can be assured about the rest of team mem- bers. One could think that if no unanimity can be guaranteed, this strategy is not worth being used. However, when team members tend to be very similar this strategy is expected to yield acceptable results with communication costs equivalent to a bilateral negotiation process. 3.2. Similarity Simple Voting (SSV) The second intra-team strategy relies on a trusted mediator that helps team members to participate in the negotiation process. Its main tasks involve coordi- nation of voting processes and communications with the opponent. It should be highlighted that the mediator communicates team's decisions to the opponent, and broadcasts opponent's decisions among team members. Thus, the fact that every team member participates in the negotiation process remains unknown for the opponent. As for intra-team communications, it should be noted that team members do not communicate among them, but they only communicate anonymously with the mediator. The decision rule used for voting processes is plurality/majority. More specif- ically, a plurality rule is used in the voting process employed to decide which offer is sent to the opponent, and a majority rule is used in the voting pro- cess employed to decide opponent's offer acceptance. A detailed view of the intra-team strategy can be observed in Algorithm 1, which describes the whole process from the point of view of the mediator. 3.2.1. Offer proposal Whenever a new offer has to be proposed to the opponent at round t, the mediator opens a call for proposals among team members. Each team member ai is allowed to communicate anonymously one offer X t ai→A to be proposed to the opponent. Once every proposal has been gathered, the mediator opens a voting process where offers proposed XT t are made public among team members. Then, each agent ai anonymously sends a multi-vote V oteai to the mediator. A multi-vote gathers votes for every offer made public. We use the notation V oteai(j) to denote the vote given by agent ai to the offer j-th from XT t, and XT t(j) as the j-th offer in XT t. The votes can be either positive (1), if the offer j-th is acceptable for ai at round t, or negative (0), if the offer j-th is not acceptable for ai at round t. Once all votes have been gathered, the mediator sums up the number of positive votes and the most supported offer X t A→op is selected, made public among team members, and sent to the opponent. When a tie is produced, the tie-breaker rule consists in randomly selecting one of the most supported offers. The following Equation describes the selection rule of the previous mechanism: (cid:88) ai∈A 9 X t A→op = argmax Xj∈XT t V oteai (j) (8) The paragraph above describes the intra-team protocol followed by team mem- bers and mediator to determine which offer is sent to the opponent at round t. However, team members are faced with two decisions in this intra-team protocol: which offer should be proposed to the mediator during the call for proposals, and the acceptability of each offer proposed during the aforementioned process. We assume that, since the negotiation is time-bounded, team members follow a time-based concession tactic where the concession speed βA is common and agreed by teammates prior to the negotiation process: sai(t) = 1 − (1 − RUai)( 1 βA ) t TA (9) ai→A whose utility is equal to Uai(X t For the first decision, proposing an offer to team members, the agent ai proposes an offer X t ai→A) = sai(t). Since there may be more than a single offer with such utility, the agent has to choose one of ai→A to be accepted it should those offers. maximize the probability of it being the most supported by team members and the probability of it being accepted by the opponent: If the agent ai wants its offer X t X t ai→A = argmax XUai (X)=sai (t) pop(X) × pA(X) (10) where pop(X) is the probability for X to be accepted by the opponent, and pA(X) is the probability for X to be selected by team members. We incorporated agents with a similarity heuristic based on the Euclidean distance over attribute domains scaled to [0,1]. It takes into account the last offer proposed by the opponent X t−1 op→A and the offer sent by team members in the previous negotiation round X t−1 op→A, the more probable it is for the offer to be accepted by the opponent. Analogously, the most similar an offer is to X t−1 A→op, the more probable it is for the offer to be the most supported option in the voting process and, therefore, to be sent to the opponent. Thus, Equation 10 can be approximated by similarity heuristics as follows: A→op. The most similar an offer is to X t−1 X t ai→A = argmax XUai (X)=sai (t) XUai (X)=sai (t) argmax pop(X) × pA(X) ≈ Sim(X, X t−1 op→A) × Sim(X, X t−1 A→op) (11) Finally, for determining the acceptability of offers proposed by team members at round t, we used a rational criterion so that an agent ai emits a positive vote V oteai(j) = 1 for the j-th offer if it reports a utility that is greater or equal than the utility marked by the concession strategy sai(t). Otherwise, the offer is not supported and a negative vote is emitted. This process can be formalized as: (cid:26) 1 0 V oteai(j) = if Uai(XT t(j)) ≥ sai (t) otherwise (12) 3.2.2. Offer acceptance Whenever the mediator receives an offer X t op→A from the opponent at round t, it broadcasts the offer among team members. Then, the mediator opens up a 10 majority voting process where each agent ai states whether or not the opponent's offer is acceptable acai(X t op→A) (1 for accept, 0 for reject). The mediator counts A 2 ) the number of acceptances, and if the offers is supported by the majority (> then it is accepted by the team. Otherwise, the offer is rejected. If the number of team members is even and a tie has been produced, a random decision is taken by the mediator. This mechanism can be described as follows: acA(X t op→A) = accept reject acai(X t op→A) > acai(X t op→A) < A 2 A 2 (13)  if (cid:80) if (cid:80) ai∈A ai∈A random otherwise How team members ai decide the acceptability of the opponent's offer acai(X t follows the rational mechanism that we have employed so far. Basically, the offer is acceptable (1) if it yields a utility which is greater than or equal to the utility demanded by the concession strategy in the next negotiation round sai(t + 1). Otherwise, the offer is not considered acceptable. The following Equation for- malizes the acceptance criterion: op→A) (cid:26) 1 0 acai(X t op→A) = op→A) ≥ sai (t + 1) if Uai (X t otherwise (14) 11 t = 0; while t ≤ TA do Send (Call For Proposals −→ A); XT t = ∅; foreach ai ∈ A do Receive (X t XT t = XT t(cid:83) X t ai→A ←− ai); ai→A; end Send (Open Voting XT t −→ A); foreach ai ∈ A do Receive (V oteai ←− ai); end X t A→op = argmax Xj∈XT t (cid:80) A→op −→ op, A); op→A ←− op); ai∈A Send (X t Receive (X t if X t V oteai (j); op→A = Withdraw then Send (Opponent Withdraw −→ A); Return Failure; end else if X t op→A = Accept then Send (Offer Accepted −→ A); Return Success; end else Send (Open Voting X t foreach ai ∈ A do Receive (acai (X t op→A −→ A); op→A) ←− ai); end if acA(X t op→A)= accept then Send (Accept −→ op, A); Return Success; end else end Send (Opponent Offer Rejected −→ A); end t = t + 1; end Send (Withdraw −→ op, A); Return Failure; Algorithm 1: Pseudo-code algorithm for the mediator in the Similarity Sim- ple Voting intra-team strategy. Messages are represented as (Body direction agents). Therefore, (Accept −→ op) means that the agent sends an accept message to op, whereas (Reject ←− op) describes a message from op with the content "Reject". 3.2.3. Unanimity Level The proposed method is capable of guaranteeing team decisions that are sup- ported by a plurality/majority of the participants. More specifically, plurality is assured in case of the offer proposed to the opponent, and majority is assured when deciding opponent's offer acceptance. Exceptions for this minimum level of team unanimity are ties. For instance, the most extreme case is when team members propose offers to the team, but they only support their own offers. In that case, each proposal sums up exactly 1 positive vote and there is not a clear plurality winner. 12 3.3. Similarity Borda Voting (SBV) SSV is capable of assuring majority and plurality decisions within the team. However, some scenarios may need of intra-team strategies that ensure higher levels of unanimity. SBV and FUM (described later) are designed to solve this problem. The basic structure of SBV remains the same than in SSV, but the voting rules employed are different. More specifically, when each team member votes team proposals, borda count is employed to determine the winner, and a unanimity rule is used to determine opponent's offer acceptance. Next, we briefly describe the aspects which make SBV different to SSV. 3.3.1. Offer proposal As in SSV, when the team has to propose an offer to the opponent, the mediator opens a call for proposals where each team member can propose an offer to the mediator. Then, once every offer has been gathered, the mediator makes public the offers proposed to the team members and a voting process starts. The main difference between both intra-team strategies resides in the fact that team members vote according to a Borda count rule [27]. Basically, each team member ai ranks the proposals XT t in ascending order according to its own utility function Uai(.). We denote as rankai(XT t) the ascending rank according to ai's utility function, and P osition(X, rankai(XT t)) as the position (1 to XT t) that the offer X occupies in a ranked list. The vote emitted by ai for offer j-th in XT t is the position occupied by such offer in the ranked list minus one unit: V oteai(j) = P osition(XT t(j), rankai(XT t)) − 1 (15) Numerical votes for each offer are summed up by the mediator, who finally selects the offer that received the highest sum of scores from the team members (see Equation 8). It should be highlighted that the similarity heuristic employed by team members is the same than the one employed in SSV. 3.3.2. Offer acceptance As for the offer acceptance, the only difference remains in the rule used by the mediator. The opponent's offer is accepted only if it is acceptable for all the team members. The rationale acai(X t op→A) used by team members to determine if an offer is acceptable at round t is equivalent to the one used in SSV. Thus, the offer acceptance mechanism can be formalized as follows: acai(X t op→A) = A (16) (cid:40) accept if (cid:80) ai∈A reject otherwise acA(X t op→A) = 3.3.3. Unanimity Level When describing the minimum unanimity level guaranteed by SBV, we men- tioned the term semi-unanimity. It is clear that if an opponent offer is accepted by the team, it is acceptable for every team member due to the unanimity ruled 13 employed. However, such unanimity is not guaranteed regarding the team deci- sion on which offer is sent to the opponent. Borda count is generally referred as a method that selects broadly accepted options as winners instead of the ma- jority/plurality option (e.g., avoid the tyranny of the majority). In this sense, Borda count entails some degree of unanimity. Nevertheless, the specific de- gree of unanimity that Borda assures is difficult to determine in our negotiation scenario. 3.4. Full Unanimity Mediated (FUM) The last intra-team strategy, Full Unanimity Mediated (FUM), seeks to reach unanimity regarding all team decisions. In fact, every team decision taken (i.e., offer acceptance, offer proposal) following this intra-team strategy entails unanimity at each round t of the negotiation process. However, the type of mediator required for FUM is more sophisticated than in the rest of strate- gies presented in this article. It requires that the mediator participates in a pre-negotiation process where team members hand over decision rights over at- tributes that are not interesting for them. Additionally, the team mediator needs to be able to infer attributes' importance for the opponent. Finally, it also needs to coordinate unanimity voting processes, and an iterated building process that constructs the offers sent to the opponent. A complete view of the pseudo-algorithm carried out by the mediator can be observed in Algorithm 2. 3.4.1. Pre-negotiation: information sharing During the pre-negotiation, team members are allowed to hand over decision rights over some attributes that they do not consider interesting. The iterated offer building process relies on a mechanism which sets attributes' values one- per-one according to team members' will. When an agent hands over decision rights on an attribute, it does not participate in the setting of such attribute. All the communications in the pre-negotiation are private with the mediator, who asks each team member regarding the set of attributes which it is willing to hand over. The rationale behind the idea of handing over decision rights is that conflict may be reduced, and, so, the chances to build a more likeable offer for the opponent are increased while maintaining a good quality for one's own utility function. The fact that some attributes may yield little or no importance at all for some team members is also feasible in a team setting, since some of these attributes may have been introduced to satisfy the interests of a subgroup of team members. The pre-negotiation protocol goes as follows. First, the mediator opens a call for decision rights, where each team member ai is allowed to send (to the mediator) a set of negotiation attributes N Iai, whose decision rights are handed over by ai. Once all the responses have been gathered, the mediator keeps track of those attributes that are not interesting for each agent N Iai, and M(cid:84) i=1 those attributes that are not interesting for all team members N Iai. Once this process has finished, the team and the mediator are ready to start the negotiation process. 14 /Pre-negotiation: information sharing/; Send (Ask for N Iai −→ A); foreach ai ∈ A do Receive (N Iai ←− ai); t = 0; while t ≤ TA do /Offer proposal starts/; agenda = build agenda(); A(cid:48) = A; X /Attributes not interesting for every team member are maximized for the opponent/; (cid:48) t A→op = ∅; foreach j ∈ M(cid:84) N Iai do i=1 xj = maximize f or opponent(j); X (cid:48) t A→op = X (cid:48) t A→op (cid:83){xj}; end /The offer is built attribute per attribute/; foreach j ∈ agenda ∧ attribute not set(j) do /Ask about the value needed of attribute j/; Send (Needed value j, given X Receive (xai,j ←− aij /∈ N Iai ∧ ai ∈ A(cid:48)); if monotonically increasing(j) then xj = max ai∈A xai,j ; (cid:48) t A→op −→ aij /∈ N Iai ∧ ai ∈ A(cid:48)); else xj = min ai∈A (cid:48) t A→op (cid:48) t A→op = X xai,j ; (cid:83){xj}; X /Ask about the acceptability of the partial offer/; Send (Acceptable X foreach ai ∈ A(cid:48) do (X (cid:48) t A→op? −→ aiai ∈ A(cid:48)); (cid:48) t A→op) ←− ai); Receive (ac(cid:48) ai (cid:48)t A→op) = true then A(cid:48) = A(cid:48) − {ai}; if ac(cid:48) ai (X end /If no more active agents, we do not ask team members anymore/; if A(cid:48) = ∅ then break; end /Any attribute not set yet, is maximized for the opponent/; foreach j ∈ agenda ∧ attribute not set(j) do xj = maximize f or opponent(j); X (cid:48) t A→op = X (cid:48) t A→op (cid:83){xj}; A→op = X (cid:48) t A→op; end X t Send (X t /Opponent offer acceptance starts/; Receive (X t if X t A→op −→ op, A); op→A ←− op); op→A = Withdraw then Send (Opponent Withdraw −→ A); Return Failure; Send (Offer Accepted −→ A); Return Success; op→A = Accept then else if X t else op→A −→ A); Send (Open Voting X t foreach ai ∈ A do Receive (acai (X t if acA(X t op→A)= accept then Send (Accept −→ op, A); Return Success; Send (Opponent Offer Rejected −→ A); else end op→A) ←− ai); end t = t + 1; end Send (Withdraw −→ op, A); Return Failure; Algorithm 2: Pseudo-code algorithm for the mediator in the FUM intra-team strategy. 15 Of course, the set of attributes handed over by each team member is not controllable by the mediator. It depends on the behavior of each agent. In our model, the set of attributes handed over by each agent depends on a private parameter ai. The value of such parameter is related to the weight of the different negotiation attributes in one's own utility function. More precisely, if ai = 0, then the agent is only willing to hand over the decision rights over those attributes that are not interesting for himself (i.e., weight equal to zero in the utility function). When ai = 1, the agent is willing to hand over decision rights over every attribute in the negotiation. In general, the agent is willing to hand over decision rights over attributes whose sum of weight in the utility function is equal to or lower than ai: (cid:88) wai,j ≤ ai (17) j∈N Iai Given a certain ai, a reasonable heuristic is to assume that the agent is willing to concede as many decision rights as possible since this will enhance the possibility of finding an agreement with the opponent. Hence, each team member ai chooses the largest possible set N Iai that fulfills Eq. 17. A simple algorithm that solves this problem is ordering the negotiation attributes in as- cending order by weight in the utility function. The set N Iai starts empty, and, then, the array of ordered attributes is followed. If the attribute weight plus the weights of those attributes already in N Iai exceeds ai , then the search stops. Otherwise, the attributes is added to N Iai and the algorithm continues with the next attribute. Our initial experiments with FUM [33] suggested that team members should set its private ai to 0 and hand over decision rights only over those attributes that are not interesting at all. Therefore, in the experimental setting, we use ai = 0. 3.4.2. Negotiation: observing opponent's concessions and building an attribute agenda Once the negotiation starts, the mediator attempts to guess a ranking of attributes according to the opponent's preferences. This ranking is used to build an agenda of attributes, which is used in the iterated offer building process. The idea behind the agenda is attempting to satisfy team members as much as possible with those attributes that are less important for the opponent. This way, team members may reach their desired aspiration level with those attributes less interesting for the opponent, and use the rest of attributes to make the offer as satisfactory as possible for the opponent. The only information available for the mediator regarding the opponent's preferences are the offers received. Thus, the mediator has to infer a ranking of attributes according to that information. A possible heuristic is assuming that agents usually concede less in important attributes and greater concessions are performed in lesser important attributes at the first rounds of the negotiation. Our proposed heuristic assumes that the mediator observes opponent's offers In our experiments, the value of this parameter for the first k interactions. 16 4 (cid:99) [33]. Then, it calculates the concession performed in was set to k = (cid:98) TA each attribute. Since our model assumes that the opponent's utility function employs the opposite type of valuation function than team members for each attribute, it is relatively easy to calculate the amount of concession performed at each attribute. For instance, if the opponent is a seller, it is reasonable to assume that its valuation functions is monotonically increasing (e.g., higher prices report higher utilities) and, thus, any value below the maximum price can be considered a concession with respect to the maximum price. Therefore, the relative concession can be calculated in each attribute. For each attribute j, we calculate the total amount of relative concession Cj in the first k offers: k−1(cid:88) t=0 Cj = op→A − best value(j) X(j)t max value(j) − min value(j) (18) op→A it the value of attribute j in the offer X t where X(j)t op→A, best value(j) is the best possible value for the opponent in attribute j, and max value(j) and min value(j) are the maximum and minimum value of the attribute in the negotiation domain. The inner part of the summatory determines the relative concession on attribute j in the offer received at interaction/round t. So, the summatory counts the total relative concession for attribute j in the first k offers. The heuristic is that attributes that score lower in Equation 18 are usually those more important for the opponent, whilst those attributes scoring higher in Equation 18 are those less important for the opponent. Based on the available information (i.e., number of rounds up to k), the mediator builds an agenda of attributes according to the scores of Cj in descending order. This way, lesser important attributes for the opponent are first in the agenda. 3.4.3. Negotiation: Offer proposal In order to determine which offer is sent to opponent, the mediator governs an iterated building process. The aim of this iterated process is building an of- fer, attribute per attribute, so that the offer sent to the opponent is acceptable for every team member. The order in which the attributes are adjusted is deter- mined by the agenda built by the mediator. The first attribute in the agenda is the one considered less important for the opponent, the second attribute is the next lesser important attribute for the opponent, and so forth. Thus, the first attributes set are those less important for the opponent. The heuristic used by this iterated building process is attempting to satisfy team members' demands with those attributes that are less important for the opponent, and demand as less as possible from those attributes that are the most important for the opponent. Briefly, the iterated building process goes as follows. 1. The agenda of attributes agenda is built by the mediator according to the available information. The first attribute in the agenda is the one guessed as the less important attribute for the opponent. active member (ai ∈ A(cid:48)) in the construction process. 2. When the iterated process starts, every team member is considered an 17 3. The initial partial offer X not been set. (cid:48)t A→op starts as an offer whose attributes have M(cid:84) i=1 4. The mediator checks those attributes that are not interesting for every team member N Iai. These attributes are maximized according to the opponent's preferences (i.e., if the price was one of these attributes, it would be maximized for the opponent, thus, acquiring its minimum value). The partial offer X 5. The next attribute j in the agenda is selected. Those team members active in the construction process (ai ∈ A(cid:48)) and interested in j (j /∈ N Iai) are asked by the mediator to submit the value xai,j needed of attribute j to get as close as possible to their aspiration levels. (cid:48)t A→op is updated with the new attributes' values. 6. The values xai,j gathered from team members are aggregated. If the assumed valuation function is monotonically increasing, then the max op- erator is used to aggregate the values and obtain the final value for the attribute xj. Otherwise, if the assumed valuation function is monoton- ically decreasing, then the min operator is used to aggregate the values and obtain xj. (cid:48)t A→op and the new partial offer is broadcasted among team 7. xj is set in X members. Every team member that is active in the construction phase is asked if the current partial offer satisfies its current demands. 8. Every response is gathered by the mediator. Those agents that answered If there are still positively are removed from the list of active agents. active agents, the mediator goes back to 5. 9. When every team member has been satisfied by the partial offer X (cid:48)t A→op, if there are still attributes that have not been set, those attributes are maximized according to the opponent's preferences. Then, a final offer X t A→op is obtained, made public among team members, and sent to the opponent. In the protocol described above, team members are asked to submit a value for attributes in which they are interested, and to determine whether or not the partial offer satisfies their needs. In both cases, as in previous strategies, we have assumed that team members follow time-based concession tactics similar to the one described in Equation 9, where βA has been agreed upon by team members prior to the negotiation process. However, since team members may have handed over some decision rights, it is not possible for agents to demand the maximum utility. The value ai has to be subtracted from the maximum utility. Therefore, the concession strategy sai(t), which determines the level of demand at each negotiation round, can be formalized as: sai (t) = (1 − ai) − (1 − ai − RUai)( 1 βA ) t TA (19) When team members are asked about a value for j, each team member communicates anonymously the value xai,j. The value communicated is the 18 one that gets as close as possible to its desired aspiration level sai(t) at round t. Taking the linear additive utility function formula, this can be calculated as: xai,j = argmin x∈[0,1] (sai(t) − Uai (X A→op) − wai,jVai,j(x)) (cid:48)t (20) (cid:48)t where sai(t) is the utility demanded by the agent ai at round t, Uai(X A→op) is the utility reported by the current partial offer, and wai,jVai,j(x) is the weighted utility reported by the value demanded by the agent. Since the value demanded looks to be as close as possible to the utility necessary to get to the current aspi- ration, the function is minimized. However, the following constraint is fulfilled by team members in order to avoid surpassing the utility demanded: sai(t) − Uai(X A→op) − wai,jVai,j(xai,j) ≥ 0 (cid:48)t (21) As for determining when a partial offer is acceptable, team members follow a similar criterion to the method proposed in other intra-team strategies. Ba- sically, a partial offer is acceptable for an agent ai if it reports a utility that is greater than or equal to the aspiration level marked by its concession strategy: ac(cid:48) ai (cid:48)t A→op) = (X A→op) ≥ sai(t) (cid:48)t if Uai(X otherwise (22) (cid:26) true f alse where true indicates that the partial offer is acceptable at its current state for agent ai, and f alse indicates the opposite. 3.4.4. Negotiation: Offer acceptance Since this strategy looks for unanimity regarding team decisions, we em- ployed the same mechanism employed in SBV for determining whether or not an opponent offer is acceptable. When the mediator receives the opponent's offer X t op→A, the offer is publicly announced to all of the team members. Then, the mediator opens a private voting process where each team member ai should spec- ify whether or not it supports acceptance of the opponent's offer acai(X t op→A). The mediator counts the number of positive votes. The offer is accepted if the number of positive votes is equal to the number of team members. Otherwise, the offer is rejected. Similarly to SBV, an opponent offer is acceptable for a team member at round t if it reports a utility that is greater than or equal to the aspiration level marked by the concession strategy in the next round: acai(X t op→A) = if sai(t + 1) ≤ Uai (X t otherwise op→A) (23) (cid:26) true f alse where true means that the agent supports the opponent's offer, f alse has the opposite meaning, and sai(.) is the concession strategy employed by agent ai to calculate the aspiration level at each negotiation round t. 19 3.4.5. Unanimity Level As stated in the introduction of this section, this strategy is capable of guar- anteeing unanimity regarding team decisions. How unanimity is guaranteed in the offer acceptance phase is clear, since a voting process with unanimity rule is employed. In [33] we showed how unanimity is also guaranteed in the offer sent to the opponent. More specifically, the strategy is capable of guaranteeing a strict unanimity: for any team member ai, the offer sent to the opponent reports a utility that is greater than or equal to its aspiration level sai (t). This is possible thanks to the iterated building process and the assumptions in team members' utility functions. Since team members share the same type of monotonic valuation functions, the use of the max/min operator (max for monotonically increasing valuation functions, min for monotonically decreasing functions) ensures that for each attribute, each team member either gets exactly the value demanded for the attribute or it gets a value that reports a utility greater than or equal to the utility they demanded for the attribute. Hence, when team members demand the exact value needed to get as close as possible to their desired utility level, they will always get the same or greater utility than the one they actually demanded. Thus, in the end, the offer will yield a utility that is equal to or greater than their aspiration levels at round t. 4. Experimental Analysis In this section we study how the four intra-team strategies presented in this article perform under different environmental conditions. First, we introduce the negotiation case employed for our experiments. Then, the environmental conditions and performance measures studied are introduced and explained to the reader. Finally, we describe the experiments carried out, and we analyze the results provided by each intra-team strategy. 4.1. Negotiation Case: Group Booking The negotiation case employed for our experiments is based on a group book- ing negotiation with a hotel, which also illustrates the types of applications that can be built using the intra-team strategies proposed in this article. In this sce- nario, a group of friends who have decided to spend their holidays together has to book accommodation for their stay. Their destination is Rome, and they want to spend a whole week. Each friend is represented by his/her electronic agent, who acts semi-automatically on behalf of its user. This agent has pre- viously elicited the preferences of its user regarding booking conditions. Each group member has different preferences regarding possible booking conditions. Thus, the final agreement with the hotel should satisfy every friend as much as possible. The group of agents engages in a negotiation with a well-known ho- tel in their city of destination, which is also represented by an electronic agent. During the pre-negotiation, both parties have decided to negotiate the following issues: 20 • Price per person (pp): The price per person is the amount of money that each friend will pay to the hotel for the accommodation service. The issue domain goes from 210$, which is the minimum rate (30$ per night), to 700$, which is the maximum rate (100$ per night). A realistic assump- tion in the group of friends is that friends prefer to pay lower prices to higher prices (i.e., monotonically decreasing valuation function), whereas the seller prefers to charge higher prices to lower prices (i.e., monotonically increasing valuation function). • Cancellation fee per person (cf ): When a booking is cancelled, the hotel charges a fee to compensate for losses. The issue domain goes from 0$ (no cancellation fee) to 150$. A realistic assumption in the group of friends is that friends prefer to pay lower prices to higher prices (i.e., monotonically decreasing valuation function), whereas the seller prefers to charge higher prices to lower prices (i.e., monotonically increasing valuation function). • Full payment deadline (pd): The full payment deadline indicates when the group of friends has to pay the full price booking in order to confirm their reservation. The domain goes from "Today"=0 days (the date time when the final agreement has been signed) to "Departure Date"=30 days, which indicates that the team should only pay when leaving the hotel. A realistic assumption in the group of friends is that friends prefer to pay as late as possible (i.e., monotonically increasing valuation function), whereas the seller prefers to charge as soon as possible (i.e., monotonically decreasing valuation function). • Discount in bar (db): As a token of respect for good clients, the hotel offers nice discounts at the hotel bar. The issue domain goes from 0% (no discount) to 20%. A realistic assumption in the group of friends is that friends prefer higher discounts to lower discounts (i.e., monotonically increasing valuation function), whereas the seller prefers to offer lower dis- counts prices to higher discounts (i.e., monotonically decreasing valuation function). 4.2. Negotiation Environment Conditions & Team Performance We consider that the negotiation environment plays a very important part in team dynamics. It may not be the same using a representative approach in a setting where all of the team members' preferences are very similar than a setting where team members' preferences are exactly the opposite. Since condi- tions of the negotiation environment highly vary depending on the application domain, we decided to focus on those general conditions that are present in almost every negotiation scenario involving negotiation teams: opponent dead- line, team deadline, team members' preference similarity, opponent concession speed, and team size. Regarding team performance, it is also acknowledged that there are several well known social welfare measures to assess the quality of decisions in a society. A negotiation team can be considered a small society, and, thus, social welfare 21 measures can also be considered appropriate measures for measuring negotiation teams' performance. More specifically, we study the impact of the negotiation environment on the minimum utility of team members (i.e., egalitarian social welfare [4]), and the average utility of team members (i.e., a special case of ordered weighted averaging [4]). However, we do not only restrain our analysis to social welfare measures. Computational measures like the number of negotiation rounds are also analyzed for all of the intra-team strategies. 4.2.1. Environment Condition: Opponent Deadline Length One of the issues that can affect the negotiation process is the number of interactions that the opponent has until he decides that negotiating is no longer worthy, namely opponent deadline Top. We partitioned the opponent negoti- ation deadline in three different classes: short deadline Top = U [5, 10] = S1, medium deadline Top = U [11, 29] = M , and long deadline Top = U [30, 60] = L. 4.2.2. Environment Condition: Team Deadline Length Similarly, the maximum number of rounds that the team has to negotiate also may impact the performance of the different intra-team strategies. As in the case of the opponent deadline, we partitioned the team deadline in three different classes: short deadline TA = U [5, 10] = S, medium deadline TA = U [11, 29] = M , and long deadline TA = U [30, 60] = L. 4.2.3. Environment Condition: Team Similarity 25 different linear utility functions were randomly generated. These utility functions represented the preferences of potential team members. 25 linear utility functions were generated to represent the preferences of opponents. These utility functions were generated by taking potential teammates' utility functions and reversing the type of Vi(.). In order to determine the preference diversity in a team, we decided to compare team members' utility functions. We introduce a dissimilarity measure based on the utility difference between offers. The dissimilarity between two teammates can be measured as follows: (cid:80) ∀X∈[0,1]n Uai(X) − Uaj (X) X ∈ [0, 1]n D(Uai(.), Uaj (.)) = (24) If the dissimilarity between two team members is to be measured exactly, it needs to sample all of the possible offers. However, this is not feasible in the current domain where there is an infinite number of offers. Therefore, we limited the number of sampled offers to 1000 per dissimilarity measure. Due to the fact that a team is composed by more than two members, it is necessary to provide a team dissimilarity measure. We define the team dissimilarity measure as the average of the dissimilarity between all of the possible pairs of teammates. 1U[5,10] is a uniform distribution from 5 to 10 22 For all of the teams that had been generated, we measured their dissimilarity and calculated the dissimilarity mean ¯dt and standard deviation σ. We used this information to divide the spectrum of negotiation teams according to their diversity. Our design decision was to consider those teams whose dissimilarity was greater than, or equal to ¯dt+1.5σ as very dissimilar, and those teams whose dissimilarity was lower than, or equal to ¯dt− 1.5σ as very similar. In each case, 100 random negotiation teams were selected for the tests, that is, 100 teams were selected to represent the very similar team case, and 100 teams were selected to represent the very dissimilar team case. These teams participate in the different environmental scenarios, where they are confronted with one random half of all of the possible individual opponents. Therefore, each environmental scenario (complete instantiation of all the environmental conditions) consists of 100×12×4=4800 different negotiations (each negotiation is repeated 4 times to capture stochastic variations in the different intra-team strategies). 4.2.4. Environment Condition: Opponent Concession Speed The concession speed of the opponent during the negotiation process βop may determine the final quality of the agreement for team members. For instance, if the opponent concedes very quickly towards its reservation utility, better agree- ments for the team may come earlier in the negotiation process. In those cases, even intra-team strategies that guarantee less degree of unanimity may achieve good results. We divided the family of concession speeds based on the classic classification of time-tactics: we considered that when βop = U [0.1, 0.49] = V B the concession speed is very boulware, when βop = U [0.5, 0.99] = B the con- cession speed is boulware, when βop = U [1, 10] = C the concession speed is conceder, when βop = U [11, 40] = V C the concession speed is very conceder. Similarly, when we refer to βA (the team concession speed), we will also employ the same partition in boulware (B), very boulware (VB), conceder (C), and very conceder (VC). 4.2.5. Environment Condition: Number of Team Members We think that the number of team members may also influence the perfor- mance of the different intra-team strategies. Some of the strategies may become too demanding when the number of team members increases and it may result in more negotiations ending in failure. Therefore, we decided to study the effect of the team size on the performance of the different intra-team strategies. The number of team members A ranged from 4 to 8. This number of team members is motivated by the negotiation case employed in our experiments. We consider that groups of friends from 4 to 8 persons are reasonable in practice. 4.2.6. Team Performance: Number of Negotiation Rounds The number of negotiation rounds considers the number of interactions be- tween the team and the opponent. It is a measure employed to assess the negotiation time employed by the different negotiation strategies to reach a fi- nal agreement. In our study, every pair offer/counter-offer in the negotiation 23 thread is considered as a negotiation round. In equal conditions of utility per- formance, those intra-team strategies that spend less negotiation rounds are preferred since they employ less negotiation time to reach a final agreement. 4.2.7. Team Performance: Minimum Utility of Team Members The minimum utility of team members (Min.) in a negotiation represents the utility of the final agreement for the less benefited team member. If the final agreement is X and the team is composed of M different team members A = {a1.a2, ..., aM}, the minimum utility of team members can be calculated as: M in.(X) = min 1≤i≤M Uai(X) (25) In applications where there is a strong bond among team members (i.e., the group of travelling friends), team members may attempt to maximize the min- imum utility of team members in order to avoid extremely unsatisfied team members and a degradation of the relationship among team members. Even if a strong bond is not present among team members, an agent may attempt to maximize the minimum utility of team members if it thinks that its own utility is going to be the less favored utility by the final agreement. 4.2.8. Team Performance: Average Utility of Team Members If the final agreement is X and the team is composed of M different team members A = {a1.a2, ..., aM}, the average utility of team members can be cal- culated as: Ave.(X) = Uai(X) (26) (cid:88) 1 M 1≤i≤M A less conservative agent may attempt to maximize the average utility of team members if it thinks that its own utility is not going to be the less favored utility by the final agreement. 4.3. Results 4.3.1. Number of Negotiation Rounds Although we measured the number of negotiation rounds in each experiment, we found that a general pattern was found in almost every experiment. Thus, instead of commenting the results for the number of negotiation rounds in each experimental section, we decided to present the performance of the four intra- team strategies according to the number of negotiation rounds just once. As a sample for this behavior, we can observe the number of negotiation rounds spent by each intra-team strategy when team and opponent have a long deadline (Top = L and TA = L), the number of team A members is set to 4, and the opponent uses different concessions speeds βop in Table 2. As long as the concession speed of the four intra-team strategies can be categorized as the same type, RE is usually the fastest intra-team strategy in number of negotiation rounds, followed by SSV, then SBV, and finally FUM. Since less unanimity is guaranteed among team members, it is logical that there 24 Very Similar, Top = TA = L, M = 4 Very Dissimilar, Top = TA = L, M = 4 βop = V C βop = C βop = B βop = V B Rounds 2.01 RE βA = V C 2.02 SSV βA = V C SBV βA = V C 2.01 FUM βA = V C 2.11 2.39 RE βA = C 2.73 SSV βA = C 3.02 SBV βA = C 4.09 FUM βA = C 9.17 RE βA = B SSV βA = B 15.53 17.96 SBV βA = B 20.31 FUM βA = B 22.50 RE βA = V B 28.62 SSV βA = V B SBV βA = V B 31.50 FUM βA = V B 32.77 Rounds Rounds Rounds 2.28 2.41 2.70 2.63 3.77 5.17 6.18 6.23 13.63 19.99 22.40 23.25 25.47 31.44 33.21 33.67 7.25 8.35 10.48 10.31 11.07 13.17 15.55 14.01 22.48 26.97 28.88 25.59 31.59 35.27 36.24 33.97 19.57 22.08 24.83 24,10 23.47 25.33 27.32 26.45 30.73 32.95 34.21 33.09 35.51 37.22 37.80 37.15 βop = V C βop = C βop = B βop = V B Rounds 2.03 RE βA = V C 2.00 SSV βA = V C SBV βA = V C 2.05 FUM βA = V C 2.90 2.28 RE βA = C 2.45 SSV βA = C 2.99 SBV βA = C 6.54 FUM βA = C 6.57 RE βA = B SSV βA = B 12.09 16.50 SBV βA = B 25.93 FUM βA = B 17.22 RE βA = V B 25.44 SSV βA = V B SBV βA = V B 30.10 FUM βA = V B 35.00 Rounds Rounds Rounds 2.33 2.71 3.44 4.52 3.30 5.17 7.12 10.47 10.02 18.26 22.74 28.53 21.14 30.04 33.29 36.59 7.78 9.97 12.99 16.29 10.98 14.83 18.64 21.13 19.94 26.42 30.54 30.97 28.94 34.59 36.77 36.39 19.71 24.35 27.33 30.70 22.84 27.43 30.08 32.66 29.19 33.52 35.76 36.96 34.50 37.64 38.74 39.01 Table 2: The table depicts the comparison of the intra-team strategies when both parties have a long deadline. Results show the average number of rounds spent in the negotiation. may be less conflict with the opponent and, thus, agreements are found faster with low unanimity strategies like RE and SSV. The main exception for this rule is when team members are very similar and the opponent uses either boulware or very boulware concession speeds. In those cases, FUM is able to finalize negotiations successfully in fewer rounds than SBV (and sometimes SSV). The learning heuristic employed by FUM benefits from the fact that the opponent usually concedes more in those attributes that are less important and, thus, it is able to infer a proper agenda and propose better offers to the opponent (ending the negotiation faster). This pattern did not exist when team members are very dissimilar, since in that case, FUM also has to deal with more intra-team conflict. This results in more demanding offers to guarantee unanimity. Additionally, as expected, as the concession strategy of team members be- comes more conceder, the number of negotiation rounds spent is lower. Thus, RE using βA = V B is slower than RE using βA = B, which is slower than RE using βA = C, which is slower than RE using βA = V C. The number of negotiation rounds spent by each intra-team strategy is es- pecially interesting to select intra-team strategies when they perform equally in utility terms (minimum or average utility). For instance, if SBV and FUM tie in utility terms, a team is suggested to select SBV most of the times due to the fact that it usually requires less negotiation rounds, if SSV and SBV tie in utility terms, the team should select SSV since it usually requires less rounds than SBV, and so forth. 4.3.2. Same Type of Deadlines The next set of experiments that we conducted consisted in assessing which intra-team strategies work better when both parties have the same type of dead- line. More specifically, we chose those scenarios where both parties have short deadlines or long deadlines. Additionally, for each type of deadline, we simulated 25 βop = V C Min. 0.687 RE β = B 0.709 SSV β = B SBV β = B 0.719 FUM β = B 0.691 Ave. 0.794 0.797 0.796 0.772 Ro. 2.94 3.33 3.67 4.26 βop = V C Min. 0.758 RE β = B 0.757 SSV β = B 0.774 SBV β = B FUM β = B 0.752 Ave. 0.853 0.833 0.833 0.816 βop = V C Min. 0.264 RE β = B 0.503 SSV β = B SBV β = B 0.550 FUM β = B 0.554 Ave. 0.621 0.730 0.735 0.709 Ro. 9.18 15.54 17.97 20.32 Ro. 2.551 2.983 3.659 5.103 Very Similar, Top = TA = S, M = 4 βop = C Ave. 0.711 0.706 0.703 0.721 Ro. 3.67 4.35 4.73 4.88 Min. 0.590 0.610 0.620 0.622 βop = B Ave. 0.551 0.561 0.568 0.608 Ro. 5.08 5.66 5.93 5.93 Min. 0.415 0.459 0.477 0.483 Very Similar, Top = TA = L, M = 4 βop = C Ave. 0.779 0.765 0.765 0.770 Min. 0.667 0.671 0.698 0.695 Ro. 13.637 19.995 22.409 23.264 Min. 0.476 0.514 0.541 0.606 βop = B Ave. 0.617 0.629 0.628 0.725 Ro. 22.486 26.981 28.891 25.603 Very Dissimilar, Top = TA = S, M = 4 βop = C Ave. 0.549 0.662 0.653 0.657 Ro. 3.169 4.054 4.900 5.750 Min. 0.198 0.417 0.467 0.468 βop = B Ave. 0.421 0.526 0.526 0.584 Ro. 4.67 5.74 6.229 6.564 Min. 0.115 0.295 0.341 0.348 Very Dissimilar, Top = TA = L, M = 4 βop = V C Min. 0.383 RE β = B 0.571 SSV β = B SBV β = B 0.650 FUM β = B 0.627 Ave. 0.725 0.788 0.797 0.756 Ro. 6.583 12.103 16.508 25.938 Min. 0.270 0.451 0.551 0.564 βop = C Ave. 0.630 0.707 0.713 0.714 Ro. 10.030 18.625 22.746 28.536 Min. 0.118 0.283 0.373 0.489 βop = B Ave. 0.454 0.573 0.562 0.724 Ro. 19.945 26.430 30.552 30.976 βop = V B Ave. 0.408 0.423 0.427 0.475 Ro. 6.31 6.57 6.79 6.78 βop = V B Ave. 0.457 0.484 0.489 0.560 Ro. 30.735 32.960 34.220 33.100 βop = V B Ave. 0.303 0.371 0.368 0.445 Ro. 6.171 6.771 7.050 7.224 βop = V B Ave. 0.319 0.430 0.416 0.543 Ro. 29.198 33.525 35.766 36.968 Min. 0.292 0.335 0.348 0.356 Min. 0.327 0.372 0.402 0.442 Min. 0.072 0.182 0.213 0.236 Min. 0.064 0.180 0.242 0.318 Table 3: The table shows the comparison of the intra-team strategies when both parties have the same type of deadline. Results show the average for the minimum utility of team members (Min.), the average utility of team members (Ave.), and the number of rounds (Ro.). The results in bold font indicate those configurations that are statistically better and different (t-test α = 0.05) to the rest of configurations. scenarios where team members were either very dissimilar, or very similar, and gathered information about the minimum and average utility of team mem- bers regarding each possible strategy configuration (team concession speeds, intra-team strategies, opponent concession speeds, etc.). The number of team members remained static at A = 4. The results for this experiment can be found in Table 3. It shows the average minimum utility of team members (Min.), the average of the average utility of team members (Ave.), and the average number of rounds (Ro.). It only shows the results for intra-team strategies using a Boulware concession speed since we found that this concession speed worked better than the rest of concession speeds. When both parties have a short deadline (first and third sub-table in Table 3), independently of team similarity, SBV β = B and FUM β = B are usually the best options for the minimum utility. The unanimity and semi-unanimity rules employed by this strategy make possible for the worst affected team mem- ber to ensure that its situation is better than with other strategies. As for the average utility of team members, FUM β = B usually is the best option. The only exception for this pattern is when the opponent uses conceder strategies (βop = V C or βop = C). In that case, all of the strategies perform similarly, 26 especially when team members are very similar. For instance, we can observe that RE, SSV, SBV β = B are the best option for the average utility of team members when the deadline is short, team members are very similar, and the opponent uses a very conceder strategy. In the same setting, but with the op- ponent using a conceder strategy, FUM is statistically better but the differences are not very important (less than a 1.8%). However, when both parties have a long deadline to negotiate (subtables 2 and 4 in Table 3), FUM β = B becomes the best choice for the minimum and average utility of team members in almost every scenario. The only exceptions for this superiority are, again, scenarios where the opponent employs conceder strategies. For instance, when the deadline is long, team members are very dissimilar, and the opponent uses a very conceder strategy, SBV β = B is the best intra-team strategy for the minimum and average utility of team members. We can also observe that RE and SSV are specially affected by very dissimilar preferences' scenarios. When team members are very similar, both strategies are capable of being close to SBV and FUM in the minimum and average utility of team members as long as the opponent plays conceder strategies. However, both intra-team strategies' results get further from those of SBV and FUM when team members are very dissimilar. These intra-team strategies are not able to tackle situations where team members have very dissimilar preferences due to the type of decision rule applied, and their use in such situations is discouraged. The reason why several strategies perform similarly in utility terms when the opponent plays conceder strategies is simple: Since the opponent concedes very fast in the first rounds of the negotiation process, as long as the team does not concede very fast (i.e., boulware strategy), all of the strategies are capable of finding a reasonable good agreement in the first rounds by letting the opponent concede and then accepting the opponent's offer. However, there is an additional reading that explains why strategies like FUM, which guarantees unanimity regarding team decisions, does not perform so well when the opponent uses conceder strategies. FUM relies on the assumption that the opponent concedes very little in those attributes that are important for its interests at the first rounds. However, when the concession strategy carried out by the opponent is conceder or very conceder (a more acute effect) big concessions are usually carried out at the first rounds. Thus, FUM is not able to infer an appropriate agenda. In [33], it was shown that as the agenda gets further from the real ranking of opponent preferences, the more demanding becomes the strategy. This may have a negative effect in the negotiation, since more negotiations may end in failure due to the high demands of the team. In fact a slight effect is observed in the results: when the opponent uses a boulware strategy, the percentage of successful negotiations is 94.6% which is greater than the 92.6% obtained when the opponent uses a conceder strategy and the 93.1% obtained when the opponent uses a very conceder strategy. Another issue found in the results is the difference between FUM and other strategies when the deadline is long. FUM tends to obtain better results when the deadline is long for both parties. The differences with the other intra-team strategies become greater when compared with the short deadline scenario. The 27 reason for this phenomenon is similar to the reason mentioned in the paragraph above. FUM is a strategy that relies on the information gathered in the negoti- ation process. Thus, when interactions are lesser, like when deadlines are short, the agenda inferred by the trusted mediator is less close to the ideal agenda. When the agenda deviates from the ideal agenda, offers proposed by the team are more demanding and less probable to be accepted by the opponent. As a matter of fact, the reader can notice that the difference on average between FUM β = B in long deadline scenarios (aggregating all of the scenarios where the deadline is long) and the results obtained by FUM β = B in short deadline (aggregating the scenarios where the deadline is short) scenarios counterpart is 7.9%, whereas it is 5.3% for SBV, 5.4% for SSV and 5.8% for RE. Logically, every intra-team strategy benefits from having a longer deadline, but the results suggest that FUM benefits more than the rest of intra-team strategies due to its learning heuristic, which is based on the amount of information. 4.3.3. Different Types of Deadlines The next experiment consisted in studying the behavior of the different intra- team strategies when both parties have different types of deadline. Thus, in this case, one of the two parties has a deadline which is lower than the deadline of the other party. Clearly, the party with a lower deadline is at disadvantage with respect to the other party since it has fewer offers to send before ending the negotiation, and the pressure to accept the opponent's offers arises earlier. Short Team Deadline and Long Opponent Deadline. First, we start by analyzing the case where the deadline of the team is shorter (short deadline) than the deadline of the opponent party (long deadline). Hence, TA = U [5, 10] and Top = U [30, 60]. The results of this experiment can be found in Table 4. In this case, the team has a shorter deadline and, thus, it should be at disadvantage with respect to the opponent. However, we can observe that when the opponent uses a conceder or very conceder strategy, the results are similar to the analogous case where both parties had a short deadline. These results can be explained due to the fact that since the opponent concedes very quickly, a good deal can be found for the team in the first rounds of the negotiation process and the team is not affected by the fact that its deadline is shorter. Nevertheless, as the opponent moves towards Boulware strategies, there is a clear negative effect on the minimum and average utility of team members: all of the strategies are affected by the fact that the team has a shorter deadline. In the scenario where both parties have a short deadline (see Table 3), the average for the average utility of team members in conceder settings (aggregating those negotiations where βop = C or βop = V C) is 0.67, and the average for the average utility of team members in boulware settings (aggregating those negotiations where βop = B or βop = V B) is 0.45. Thus, the average utility for team members is reduced a 22%. In this experiment (see Table 4), the average of the average utility of team members in conceder settings is 0.63, whereas the average of the average utility of team members in boulware settings is 0.10. Therefore, the 28 average utility of team members is reduced a 53%, approximately doubling the difference found in the case where both parties had a short deadline. When team members are very similar (upper sub-table in Table 4), it can be observed that, as in the scenario where both parties have a short deadline and team members are very similar, several strategies perform very similarly. The main difference resides in the fact that the only strategy capable of reaching similar results to FUM β = B in the minimum and average utility is RE β = B. Differently to the case when team members are very similar and the deadline for both parties is short, the RE βA = B strategy is capable of achieving similar re- sults to the other intra-team strategies even in less conceding settings (βop = C, βop = B, and βop = V B). These results suggest that, despite not assuring any minimum level of unanimity, employing a representative with a reasonably slow concession (boulware) leads to good results compared with those obtained by other intra-team strategies. A closer look at the experiments threw some light over these results. For instance, when βop = B, the number of successful negotiations was 2695 for RE βA = B, 1925 for FUM βA = B, 1855 for UBS βA = B, and 2394 for SSV βA = B. The average utility for successful negotia- tions was 0.32 for RE βA = B, 0.34 for SSV βA = B, 0.39 for SBV βA = B, and 0.42 for FUM βA = B. Hence, despite obtaining less quality results in success- ful negotiations, the representative approach becomes a good option for these scenarios because it leads to a great number of negotiations ending in success where other intra-team strategies fail to succeed (utility=0). SSV, UBS, and FUM need more interactions to find a satisfactory deal, but when they find it, it is better in utility terms. However, in average, a representative approach may be more adequate for settings where the team has a shorter deadline than the opponent. As for the scenario where team members are very dissimilar (lower sub-table in Table 4), we can observe that the negative effect produced by having a shorter deadline is especially acute when the opponent uses boulware or very boulware concessions. The dissimilarities between team members, and the fact that there are very few interactions to find a deal that satisfies both team and opponent, contribute to a strong reduction in the minimum and the average utility of team members. In terms of the minimum utility of team members, F U M and SBV βA = B work better when the opponent uses conceder or very conceder concessions. However, almost every intra-team strategy performs equally bad in terms of the minimum utility of team members when the opponent moves to- wards boulware concessions (especially in the very boulware case). In this case, the representative approach can no longer compete with the rest of strategies in terms of utility in most scenarios. Nevertheless, despite team members being very dissimilar and RE not guaranteeing any unanimity regarding team deci- sions, RE performs slightly better than the rest in terms of the average utility of team members when the opponents concedes using boulware. The explanation to this phenomenon is similar to the case where team members were very simi- lar: a lesser number of negotiations end in failure (26% failures for RE, 33% for SSV, 48% for SBV, and 46% for FUM), which compensates for the dissimilarity between team members' preferences and the unanimity level guaranteed by RE. 29 Very Similar, Top = L, TA = S, M = 4 βop = V C βop = C βop = B βop = V B Min. 0.643 RE β = B 0.648 SSV β = B SBV β = B 0.656 FUM β = B 0.651 Ave. 0.756 0.748 0.743 0.747 Ro. 3.232 3.895 4.370 4.733 Min. 0.483 0.465 0.473 0.494 Ave. 0.242 0.227 0.199 0.222 Very Dissimilar, Top = L, TA = S, M = 4 Ave. 0.609 0.576 0.568 0.612 Ro. 4.542 5.468 5.995 5.931 Min. 0.152 0.145 0.139 0.150 βop = V C βop = C Min. 0.245 RE β = B 0.459 SSV β = B 0.511 SBV β = B FUM β = B 0.520 Ave. 0.596 0.703 0.706 0.704 Ro. 2.771 3.444 4.377 5.825 Min. 0.153 0.280 0.313 0.336 Ave. 0.482 0.540 0.496 0.545 Ro. 3.870 5.553 6.677 7.003 Min. 0.025 0.035 0.028 0.026 βop = B Ave. 0.170 0.156 0.082 0.084 Ro. 7.112 7.544 7.876 7.818 Ro. 7.111 7.919 8.282 8.333 Min. 0.027 0.024 0.019 0.029 Min. 0.002 0.002 0.001 0.001 Ave. 0.048 0.040 0.028 0.046 βop = V B Ave. 0.040 0.017 0.064 0.060 Ro. 8.284 8.396 8.457 8.424 Ro. 8.231 8.526 8.560 8.565 Table 4: Comparison of the intra-team strategies when the team has a short deadline and the opponent party has a long deadline. Results show the average for the minimum utility of team members (Min.), the average utility of team members (Ave.), and the number of rounds (Ro.). The results in bold font indicate those configurations that are statistically better and different (t-test α = 0.05) to the rest of configurations. In any case, the utility obtained for team members is so low in the average and minimum utility of team members that, in some cases, it may even be better not to negotiate with such kind of opponent and spend computational resources in looking for another alternative. Long Team Deadline and Short Opponent Deadline. In this case, the team has an advantage over the opponent since its maximum deadline is longer than the opponent's deadline. The goal of these experiments is to determine the combi- nation of intra-team strategies and negotiation parameters that maximize the different social welfare measures employed. Thus, if the team has a maximum deadline equal to the uniform distribution TA = U [30, 60], the team may decide to play (prior to the negotiation) a different class of deadline like a medium deadline (TA = U [11, 29]) or a short deadline (TA = U [5, 10]) if the results of the simulation suggest that better results are obtained by not playing the max- imum deadline. Thus, we also show the results for teams that play a medium deadline, and teams that play a short deadline. In this experiment, the oppo- nent plays a short deadline Top = U [5, 10]. The results of this experiment for the very similar scenario can be observed in Fig. 1, whereas the results for the very dissimilar scenario can be observed in Fig. 2. We start by analyzing the results for scenarios where team members are very similar (Fig. 1). We can observe that for situations where the opponent is very conceder, the team benefits from playing strategies with the same deadline. Since the opponent concedes very fast in the first negotiation rounds, the best deals for the team may be proposed in the first negotiation rounds. Playing a longer deadline may be risky since the team may have extremely high aspirations during the whole negotiation, which results in most offers being rejected and ending the negotiation in failure. As a matter of fact, the number of successful negotiations for intra-team strategies playing a short deadline and boulware con- cession was 95.1%, 68% for medium deadline and boulware concession, and 45% 30 Figure 1: Results for very similar team members when the team has a long deadline and the opponent has a short deadline. The dots indicate those configurations that perform statistically better than the rest (t-test α = 0.05) for long deadline and boulware concession, 29% for medium deadline and very boulware concession, and 14% for long deadline and very boulware concession. Other configurations may have a higher number of successful negotiations, but they are not able to retain as much utility as the boulware configuration. As the opponent starts to move towards strategies that concede more slowly, the best intra-team strategies for the team are those played with a medium deadline and 31 (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)BFigure 2: Results for very dissimilar team members when the team has a long deadline and the opponent has a short deadline. The dots indicate those configurations that perform statistically better than the rest (t-test α = 0.05) boulware strategy (RE, SBV and SSV β = B). In those cases, the opponent may not propose the best deals for the team until its last negotiation rounds. Thus, playing a slightly longer deadline with a boulware concession comes at an advantage for the team since the team does not fully concede in the whole negotiation and still accepts last opponent's offers. Some strategies played with a medium deadline like FUM β = B are still too demanding, end up in more 32 (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BDBA(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:9)FF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)(cid:6)(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)AB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)CD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:9)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAA(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BAB(cid:8)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)BCD(cid:6)(cid:3)(cid:4)(cid:5)E(cid:3)(cid:7)(cid:5)(cid:8)B(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BAA(cid:8)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BDBA(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BCD(cid:6)(cid:3)(cid:4)(cid:5)A(cid:3)(cid:7)(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)(cid:6)(cid:22)(cid:23)(cid:22)(cid:24)(cid:25)(cid:24)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)BFF(cid:16)(cid:17)F(cid:16)(cid:18)F(cid:16)(cid:19)F(cid:16)(cid:20)(cid:21)%&(cid:31)" '(cid:31)(cid:3)D(cid:26)(cid:22)(cid:27)(cid:22)(cid:26)(cid:28)(cid:3)(cid:29)(cid:30)(cid:3)(cid:4)(cid:31) (cid:24)(cid:3)(cid:6)(cid:31)(cid:24)!(cid:31)"#(cid:4)(cid:29)$(cid:5)A(cid:3)(cid:3)(cid:3)(cid:3)B(cid:29)$(cid:5)(cid:8)Bnegotiation failures, and have very little information to learn the opponents' preferences. The very dissimilar scenario (Fig. 2) is a little bit different. In this scenario, the team needs to deal with strong divergences in their preferences too. Thus, teams are prone to be more demanding in order to accommodate the preferences of as many team members as possible. We can observe that for cases where the opponent uses conceder strategies, the team should play boulware strategies with the same deadline. Similarly to the very similar scenario, playing a longer deadline is risky since it results in extremely high aspirations and most offers being rejected. However, in the very dissimilar scenario, the transition from selecting short deadline strategies to selecting medium deadline strategies does not appear until the opponent uses boulware strategies. This may be explained precisely due to the dissimilarity among team members, which requires stronger demands that are not met when playing medium deadline. As the opponent starts to concede using boulware strategies, the best intra-team strategies are usually found in the medium deadline, as in the very similar scenario case. In conclusion, in this experiment we have observed that, generally, even though the team is able to play a long deadline and the opponent plays a short deadline, the team would benefit more from playing the same type of deadline than the opponent or a slightly longer deadline. 4.3.4. Team size effect on intra-team strategies We also decided to analyze the effect of the team size on the performance of the different intra-team strategies. Thus, we repeated the conditions in 4.3.2 increasing the number of team members. However, we only analyzed intra- team strategies whose βA = B since they were those one that obtained better results in Table 3. We excluded the RE strategy from the analysis. Since team members do not interact in RE and no unanimity level is guaranteed, the inclusion of additional team members should not affect the way in which the strategy works. The results of this experiment can be found in Figure 3. It shows the average and minimum utility of team members for teams of size A = {4, 5, 6, 7, 8}. Generally, it can be observed in all of the graphics in Figure 3 that, as the number of team members increases, the quality of the results in terms of the minimum and the average utility is reduced. This behavior was expected since as the number of agents increases, the set of possible agreements is reduced and the conflict inside the team and with the opponent is increased. However, the reduction in utility terms can be appreciated more easily in the minimum utility of team members. The average for the average utility of team members when A = 4 is 0.70 (aggregating all other factors) and 0.67 for A = 8 (aggregating all other factors), whereas the average for the minimum utility of team members when A = 4 is 0.48 and 0.41 for A = 8. As the number of team members increases, the contribution of each team member to the average utility is lesser, and that is the reason why the negative effect of team size on utility measures can be observed more easily in the minimum utility of team members than in the average utility of team members. 33 We expected that as the number of team members increased, the perfor- mance of unanimity intra-team strategies like FUM would greatly decrease com- pared to the performance of SSV since more team members would increase the demands of the team and make offers less interesting for the opponent. How- ever, the difference in performance between the three strategies is approximately maintained in almost every graphic as the number of team members increases. Therefore, team size did not have a different effect on the performance of the three intra-team strategies, affecting all of intra-team strategies equally. The decision on which intra-team strategy should be chosen is not affected by team size. The only clear exceptions to this rule are scenarios where the opponent uses conceder strategies (βop = C and βop = V C) and team members' preferences are very dissimilar (first two graphics in rows 3 and 4, Figure 3). In these scenarios, we can observe that there is a special negative effect of team size on the performance (mininum utility and specially in the average utility) of FUM with respect to the other intra-team strategies, which results in FUM being one of the worst choices when the number of team members in large (e.g., second graphics in rows 3 and 4, Figure 3). As a numeric example of the reduction in the performance of FUM , the difference in the average utility between SBV and FUM goes from approximately a 2% (A = 4) to 10% (A = 8) when βop = V C and the deadline is short, from approximately a 0% (A = 4) to 5% (A = 8) when the deadline is short and βop = C, and from 3% (A = 4) to 8% (A = 8) when the deadline is long and βop = V C . This phenomenon has a reasonable explanation. When the opponent uses conceder strategies, FUM has greater difficulties to learn a proper attribute agenda. If the number of team members increases and they are very dissimilar, the demands of team members increase, which summed up to the fact that the agenda does not properly reflect the preferences of the opponent, results in demanding team proposals. 5. Related Work Multi-agent systems have gained a growing interest as the infrastructure necessary for the next generation of distributed systems. Due to the inherent conflict among agents, techniques that allow agents to solve their conflicts and cooperate are needed. This need is what has given birth to a group of tech- nologies which have recently been referred to as agreement technologies [26, 37]. Trust and reputation [31, 40, 41], norms [7, 5], agent organizations [15, 9, 38], argumentation [29, 28] and automated negotiation [18, 36] are part of the core that makes up this new family of technologies. Even though agreement technologies are a novel topic in the community of agent research, some of its core technologies like automated negotiation have been studied by scholars for a few years. In definition, automated negotiation is a process carried out between two or more parties in order to reach an agree- ment by means of exchange of proposals. Two different research trends can be distinguished in automated negotiation models. The first type of model aims to calculate the optimum strategy given certain information about the opponent 34 Figure 3: Size effect when team and opponent have the same type of deadline and the negotiation environment [16, 6, 12]. The second type of model encloses heuristics that do not calculate the optimum strategy but obtain results that aim to be as close to the optimum as possible [10, 19, 11, 23]. These models assume imperfect knowledge about the opponent and the environment, and aim to be computationally tractable while obtaining good results. This present work can be classified into the latter type of models. In multi-agent systems, most of the research has concentrated on bilateral models where each party is a single individual. The present article studies bilateral negotiations where at least one of the parties is a negotiation team, composed by more than a single individual. It should be noted that the problem of finding an agreement for a negotiation team is inherently complex since it 35 (cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)"(cid:8)(cid:9)"(cid:4)(cid:8)(cid:9)#(cid:8)(cid:9)#(cid:4)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)#(cid:8)(cid:9)#(cid:4)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)(cid:8)(cid:9)(cid:7)(cid:4)(cid:8)(cid:9)%ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)#(cid:8)(cid:9)#(cid:4)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)ABCDEF(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)#(cid:8)(cid:9)#(cid:4)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)((cid:8)(cid:9)((cid:4)(cid:8)(cid:9)"(cid:8)(cid:9)"(cid:4)(cid:8)(cid:9)#(cid:8)(cid:9)#(cid:4)(cid:8)(cid:9)(cid:3)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)#(cid:8)(cid:9)#(cid:4)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)FE(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)(cid:8)(cid:9)(cid:6)(cid:4)(cid:8)(cid:9)(cid:7)(cid:8)(cid:9)(cid:7)(cid:4)(cid:8)(cid:9)%ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:24)E(cid:23)(cid:21)(cid:22)(cid:2)(cid:24)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)((cid:5)(cid:8)(cid:9)"((cid:8)(cid:9)"(cid:5)(cid:8)(cid:9)#((cid:8)(cid:9)#(cid:5)(cid:8)(cid:9)(cid:3)((cid:8)(cid:9)(cid:3)(cid:5)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:27)(cid:16)(cid:28)(cid:16)(cid:17)(cid:29)(cid:17)E(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)D(cid:1)(cid:2)(cid:3)(cid:1)(cid:2)(cid:4)(cid:1)(cid:2)(cid:5)(cid:1)(cid:2)(cid:6)(cid:1)(cid:2)(cid:7)(cid:8)(cid:9)(cid:3)(cid:8)(cid:9)(cid:3)(cid:4)(cid:8)(cid:9)(cid:4)(cid:8)(cid:9)(cid:4)(cid:4)(cid:8)(cid:9)(cid:5)(cid:8)(cid:9)(cid:5)(cid:4)(cid:8)(cid:9)(cid:6)ABCDE&(cid:16)''(cid:16)(cid:17)(cid:16)(cid:18)(cid:19)CE(cid:20)(cid:1)(cid:2)(cid:20)(cid:21)(cid:22)(cid:2)$E(cid:23)(cid:21)(cid:22)(cid:2)(cid:23)E(cid:23)(cid:21)(cid:22)(cid:2)A(cid:23)FFAF(cid:23)A(cid:25)(cid:26)(cid:27)(cid:20)B(cid:19)(cid:17)EF(cid:16)(cid:31)B(cid:1) BC(cid:19)!BE(cid:26)(cid:30)(cid:16)(cid:18)(cid:16)(cid:30)Dnot only requires finding an agreement with the other party but it also entails reaching some type of unanimity within the team. Even though communications with the opponent party may be similar to classical bilateral models, negotiation teams may require an additional level of negotiation that involves team mem- bers. Thus, classical bilateral models cannot be applied directly if a certain level of unanimity regarding team decisions is necessary . As far as we know, our previous work [34, 35, 33, 32] is the only work that focuses on negotiation teams. In [34] we introduced the topic of negotiation teams in agent research from a descriptive perspective. We analyzed the different phases necessary for an agent- based negotiation team to face such negotiations with success. Apart from the phases that we identified, we also described the current technologies that may be appropriate for the development of such phases. Later, we introduced our first experimental study [35] comparing intra-team strategies in different negotiation environments. That paper should be considered the preliminary basis for our current analysis. We have introduced changes in the intra-team strategies, and our current study applies a more fine-grained analysis of the negotiation envi- ronment and its possible scenarios. Additionally, we also studied the properties of the Full Unanimity Mediated intra-team strategy in [33]. However, a thor- ough analysis of how environmental conditions affect team performance was not carried out. Finally, we should also highlight our work regarding the study of cultural factors in negotiation teams [32]. The setting is different to the current article. We attempted to propose a computational model for explaining how human cultural factors affect team dynamics in negotiation teams composed by humans. In this present article we do not consider humans but automated agents. Therefore, human factors are not relevant to the present study. Apart from agent-based negotiation teams, bilateral negotiation is perhaps the most similar topic to our current research. Hence, we describe some of the most important bilateral models that assume imperfect knowledge. Faratin et al. [10] propose a bilateral negotiation model for service negotiation where agents apply and mix different concession tactics (i.e., time-dependent, imita- tive and resource-dependent). In their work, they analyze the impact of the model's parameters and determine which configurations work better in different scenarios by means of experiments. Our proposed work also assumes the use of time-dependent concession tactics for the calculation of agents' aspirations at each negotiation round. Additionally, we also take an experimental approach to validate the impact of our model's parameters. Later, the authors proposed a bilateral negotiation model [11] whose main novelty was the use of trade-offs to improve agreements between two parties. A trade-off consists of reducing the utility obtained from some negotiation issues with the goal of obtaining the same exact utility from other negotiation issues. The rationale behind trade-offs is to make the offer more likable for the opponent while maintaining the same level of satisfaction for the proposing agent. For that purpose, the authors propose a fuzzy similarity heuristic that proposes the most similar offer to the last offer received from the opponent. Some of our intra-team strategies like Similarity Simple Voting and Similarity Borda Voting also employ similarity heuristics to 36 attempt to satisfy team members' preferences and the opponent's preferences. Jonker and Treur propose the Agent-Based Market Place (ABMP) model [19] where agents, engage in bilateral negotiations. ABMP is a negotiation model where proposed bids are concessions to previous bids. The amount of concession is regulated by the concession factor (i.e., reservation utility), the negotiation speed, the acceptable utility gap (maximal difference between the target utility and the utility of an offer that is acceptable), and the impatience factor (which governs the probability of the agent leaving the negotiation process). Lai et al. [23] propose a decentralized bilateral negotiation model where agents are allowed to propose up to k different offers at each negotiation round. Offers are proposed from the current iso-utility curve according to a similarity mechanism that selects the most similar offer to the last offer received from the opponent. The selected similarity heuristic is the Euclidean distance since it is general and does not require domain-specific knowledge and information regarding the opponent's utility function. Results showed that the strategy is capable of reaching agreements that are very close to the Pareto Frontier. Sanchez-Anguix et al. [36] proposed an enhancement for this strategy in envi- ronments where computational resources are very limited and utility functions are complex. It relies on genetic algorithms to sample offers that are interest- ing for the agent itself and creates new offers during the negotiation process that are interesting for both parties. Results showed that the model is capa- ble of obtaining statistically equivalent results to similar models that had the full iso-utility curve sampled, while being computationally more tractable. As commented above, some of our intra-team strategies use similarity heuristics to satisfy team members' preferences and the opponent's preferences. Another topic that resembles team negotiations are multi-party negotiations. Several works have been proposed in the literature along this line [8, 21, 13]. For instance, Ehtamo et al. [8] propose a mediated multi-party negotiation pro- tocol which looks for joint gains in an iterated way. The algorithm starts from a tentative agreement and moves in a direction according to what the agents prefer regarding some offers' comparison. Results showed that the algorithm converges quickly to Pareto optimal points. Klein et al. [21] propose a medi- ated negotiation model which can be extended to multiple parties. Their main goal is to provide solutions for negotiation processes that use complex utility functions to model agents' preferences. The negotiation attributes are not in- dependent. Therefore, preference spaces cannot be explored as easily as in the linear case. Later, Ito et al. [13] proposed different types of utility functions (cube and cone constraints) and multiparty negotiation models for such utility functions. The main difference between our work and multi-party negotiations lies in the nature of the conflict and how protocols are devised. Even though each team member could be viewed as a participant in a multi-party negotiation with the opponent, it is natural to think that team members' preferences are more similar (e.g., a team of buyers, a group of friends, etc.) and they trust other teammates more than the opponent (i.e., they may share more informa- tion). Furthermore, multi-party negotiation models may be unfair for agents that are alien to the team if the number of team members exceeds the num- 37 ber of other participants. In that case, multi-party models may be inclined to move the negotiation towards agreements that maximize the preferences of most participants (i.e., team members). Multi-agent teamwork is also a close research area. Agent teams have been proposed for a variety of tasks such as Robocup [39], rescue tasks [20], and transportation tasks [17]. However, as far as we know, there is no published work that considers teams of agents negotiating with an opponent. Most works in agent teamwork consider fully cooperative agents that work to maximize shared goals. The team negotiation setting is different since, even though team members share a common interest related to the negotiation, there may be competition among team members to maximize one's own preferences. 6. Conclusions and Future Work An agent-based negotiation team is a group of two or more interdependent agents that join together as a single negotiation party because they share some common interests in the negotiation at hand. Intra-team strategies govern which decisions are taken by the negotiation team, and how and when these decisions are taken. The goal of this article is studying how environmental conditions affect the performance of different intra-team strategies for a team negotiating with an opponent. We studied how the deadline of both parties, the concession speed of the opponent, similarity among team members' preferences and team size affect the performance of Representative (RE) intra-team strategy, Simi- larity Simple Voting (SSV) intra-team strategy, Similarity Borda Voting (SBV) intra-team strategy and Full Unanimity Mediated (FUM) intra-team strategy in terms of the minimum utility of team members, the average utility of team members and the number of negotiation rounds. The results suggest that de- pending on the environmental conditions and the team performance metric, team members should select different intra-team strategies, which confirms our initial hypothesis in this article. Next, we summarize some of the most important results found in this paper: • Generally, when the concession speed is the same for the different intra- team strategies, RE takes less numbers of negotiation rounds than SSV, which takes less number of rounds than SBV, which takes less number of rounds than FUM. The exception for this rule is when team members are very similar and the opponent uses boulware or very boulware strategies, which makes FUM usually faster than SBV. • FUM tends to clearly outperform the rest of intra-team strategies studied in utility terms (minimum and average utility of team members) when the deadline of both parties is long and the opponent uses either boulware of very boulware concession strategies. When the opponent uses conceder or very conceder strategies, different intra-team strategies tie in terms of the minimum and average utility of team members depending on the rest of environmental conditions. 38 • When the team deadline is way shorter than the opponent's deadline, all of the intra-team strategies are negatively affected in the results obtained in the minimum and average utility of team members. Additionally, if team members are very similar, RE becomes one of the best choices for the average utility of team members since it is capable of ending more negotiations successfully where other intra-team strategies fail. If team members are very dissimilar, FUM and SBV tend to work better in terms of utility (minimum and average). However, if the opponent uses boulware or very boulware concession strategies every intra-team strategy performs equally bad and team members are encouraged to look for other negotia- tion alternatives. • In situations where the team's maximum deadline is longer than the op- ponent's deadline, the team should not play intra-team strategies with the maximum deadline but intra-team strategies with the same type of deadline than the opponent or a slightly longer type of deadline. Other- wise, the team performance in utility terms is not maximized due to more negotiations ending in failure. • As the number of team members increases, the performance in utility terms of all of the intra-team strategies is negatively affected. However, in general, all of the intra-team strategies studied are equally affected by the increment in the number of team members. Thus, team size did not have an effect on the intra-team strategy that should be selected by team members to maximize the minimum or the average utility of team members. The field of negotiation teams is novel in the area of multi-agent systems. Therefore, there is much work to be done in order to advance the state-of-the- art. Current works in agent-based negotiation teams [34, 35, 33, 32] have focused on negotiation processes where the team has a strong potential for cooperation since team members share the same type of monotonic valuation function for negotiation issues. However, it is possible to assume that in some negotiation scenarios there is more conflict among team members since valuation functions may be of a different type of monotonic function among team members, or the valuation function itself is not predictable in the negotiation domain (e.g., colors, brands, etc.). Our current future work involves designing intra-team strategies that are able to tackle negotiation domains where attribute's valuation functions may be unpredictable. RE, SSV and SBV are able to handle such types of domains by definition. However, FUM, which is the strategy capable of guaranteeing unanimity regarding team decisions, is not capable of handling domains where attributes are unpredictable (due to the max/min aggregation operator). Hence, our future works consists in proposing an intra-team strategy capable of guaranteeing unanimity for negotiation domains where attributes may not be predictable. On the other hand, since the results of this present article have shown that environmental conditions do affect the performance of intra-team strategies, we 39 plan to propose a mechanism that allows team members to infer the most prob- able state of the negotiation environment, and according to that information, advise the use of an appropriate intra-team strategy. Finally, in our current work we assume that team members have the same knowledge about the negotiation domain and they have the same skills. It may be interesting to study scenarios where team members have different knowledge and skills. Acknowledgements This work is supported by TIN2011-27652-C03-01, TIN2009-13839-C03-01, CSD2007-00022 of the Spanish government, and FPU grant AP2008-00600 awarded to V´ıctor S´anchez-Anguix. We would also like to thank anonymous reviewers and assistants of AAMAS 2011 who helped us to improve our previous work, making this present work possible. References [1] K. Behfar, R.A. Friedman, J.M. Brett, The team negotiation challenge: Defining and managing the internal challenges of negotiating teams, in: Proceedings of the 21st Annual Conference for the International Association for Conflict Management. [2] S. Brodt, L. Thompson, Negotiation within and between groups in organi- zations: Levels of analysis, Group Dynamics (2001) 208 -- 219. [3] M. Chalamish, S. Kraus, Automed: an automated mediator for multi-issue bilateral negotiations, Autonomous Agents and Multi-Agent Systems (In Press) 1 -- 29. [4] Y. Chevaleyre, P.E. Dunne, U. Endriss, J. Lang, M. Lemaıtre, N. Maudet, J. Padget, S. Phelps, J.A. Rodr´ıguez-aguilar, P. Sousa, Issues in multiagent resource allocation, Informatica 30 (2006) 2006. [5] N. Criado, E. Argente, V. Botti, THOMAS: An Agent Platform For Sup- porting Normative Multi-Agent Systems, J. Logic Comput. (2011). [6] F. Di Giunta, N. Gatti, Bargaining over multiple issues in finite horizon alternating-offers protocol, Ann. Math. Artif. Intell. 47 (2006) 251 -- 271. [7] F. Dignum, Autonomous agents with norms, Artificial Intelligence and Law 7 (1999) 69 -- 79. [8] H. Ehtamo, E. Kettunen, R.P. Hamalainen, Searching for joint gains in multi-party negotiations, Eur. J. Oper. Res. 130 (2001) 54 -- 69. [9] S. Esparcia, E. Argente, Formalizing Virtual Organizations, in: 3rd Inter- national Conference on Agents and Artificial Intelligence, volume 2, pp. 84 -- 93. 40 [10] P. Faratin, C. Sierra, N.R. Jennings, Negotiation decision functions for autonomous agents, International Journal of Robotics and Autonomous Systems 24 (1998) 159 -- 182. [11] P. Faratin, C. Sierra, N.R. Jennings, Using similarity criteria to make issue trade-offs in automated negotiations, Artif. Intell. 142 (2002) 205 -- 237. [12] S.S. Fatima, M. Wooldridge, N.R. Jennings, Multi-issue negotiation with deadlines, J. Artif. Intell. Res. 27 (2006) 381 -- 417. [13] K. Fujita, T. Ito, M. Klein, Secure and efficient protocols for multiple interdependent issues negotiation, J. Intell. Fuzzy Syst. 21 (2010) 175 -- 185. [14] N. Halevy, Team negotiation: Social, epistemic, economic, and psycholog- ical consequences of subgroup conflict, Pers. Soc. Psychol. Bull. 34 (2008) 1687 -- 1702. [15] B. Horling, V. Lesser, A survey of multi-agent organizational paradigms, Knowl. Eng. Rev. 19 (2004) 281 -- 316. [16] Y. In, R. Serrano, Agenda restrictions in multi-issue bargaining (ii): unre- stricted agendas, Econ. Lett. 79 (2003) 325 -- 331. [17] N.R. Jennings, Controlling cooperative problem solving in industrial multi- agent systems using joint intentions, Artif. Intell. 75 (1995) 195 -- 240. [18] N.R. Jennings, P. Faratin, A.R. Lomuscio, S. Parsons, M.J. Wooldridge, C. Sierra, Automated negotiation: Prospects, methods and challenges, Group Decis. Negot. 10 (2001) 199 -- 215. [19] C.M. Jonker, J. Treur, An agent architecture for multi-attribute negotia- tion, in: 17th International Joint Conference on Artificial Intelligence, pp. 1195 -- 1201. [20] H. Kitano, S. Tadokoro, Robocup rescue: A grand challenge for multiagent and intelligent systems, AI Magazine 22 (2001) 39 -- 52. [21] M. Klein, P. Faratin, H. Sayama, Y. Bar-Yam, Negotiating complex con- tracts, Group Decis. Negot. 12 (2003) 111 -- 125. [22] S. Kraus, Negotiation and cooperation in multi-agent environments, Arti- ficial Intelligence 94 (1997) 79 -- 97. [23] G. Lai, K. Sycara, C. Li, A decentralized model for automated multi- attribute negotiations with incomplete information and general utility func- tions, Multiagent and Grid Systems 4 (2008) 45 -- 65. [24] A. Lomuscio, M. Wooldridge, N. Jennings, A classification scheme for nego- tiation in electronic commerce, Group Decision and Negotiation 12 (2003) 31 -- 56. 41 [25] F. Lopes, M. Wooldridge, A. Novais, Negotiation among autonomous com- putational agents: principles, analysis and challenges, Artificial Intelligence Review 29 (2008) 1 -- 44. [26] M. Luck, P. McBurney, Computing as Interaction: Agent and Agreement Technologies, in: IEEE Conference on Distributed Human-Machine Sys- tems, pp. 1 -- 6. [27] H. Nurmi, Voting systems for social choice, Handbook of Group Decision and Negotiation (2010) 167 -- 182. [28] P. Pardo, S. Pajares, E. Onaind´ıa, L. Godo, P. Dellunde, Multiagent Argu- mentation for Cooperative Planning in DeLP-POP, in: 10th International Conference on Autonomous Agents and Multiagent Systems, pp. 971 -- 978. [29] I. Rahwan, S. Ramchurn, N. Jennings, P. Mcburney, S. Parsons, L. So- nenberg, Argumentation-based negotiation, Knowl. Eng. Rev. 18 (2003) 343 -- 375. [30] A. Rubinstein, Perfect equilibrium in a bargaining model, Econometrica 50 (1982) 97 -- 109. [31] J. Sabater, C. Sierra, Review on computational trust and reputation mod- els, Artif. Intell. Rev. 24 (2005) 33 -- 60. [32] V. Sanchez-Anguix, T. Dai, Z. Semnani-Azad, K. Sycara, V. Botti, Mod- eling power distance and individualism/collectivism in negotiation team dynamics, Hawaii International Conference on System Sciences (2012) 628 -- 637. [33] V. Sanchez-Anguix, V. Julian, V. Botti, A. Garcia-Fornes, Reaching Unan- imous Agreements within Agent-Based Negotiation Teams with Linear and Monotonic Utility Functions, IEEE Transactions on Systems, Man and Cy- bernetics - Part B (2012). [34] V. Sanchez-Anguix, V. Julian, V. Botti, A. Garc´ıa-Fornes, Towards agent- based negotiation teams, in: Group Decision and Negotiation 2010, pp. 328 -- 331. [35] V. Sanchez-Anguix, V. Julian, V. Botti, A. Garc´ıa-Fornes, Analyzing Intra- Team Strategies for Agent-Based Negotiation Teams, in: 10th Interna- tional Conference on Autonomous Agents and Multiagent Systems, pp. 929 -- 936. [36] V. Sanchez-Anguix, S. Valero, V. Julian, V. Botti, A. Garc´ıa-Fornes, Evolutionary-aided negotiation model for bilateral bargaining in Ambient Intelligence domains with complex utility functions, Inf. Sci. (2011). [37] C. Sierra, V. Botti, S. Ossowski, Agreement computing, KI-Kunstliche In- telligenz (2011) 1 -- 5. 42 [38] D.C. Silva, R.A. Braga, L.P. Reis, E. Oliveira, Designing a meta-model for a generic robotic agent system using gaia methodology, Information Sciences 195 (2012) 190 -- 210. [39] P. Stone, M. Veloso, Task decomposition, dynamic role assignment, and low-bandwidth communication for real-time strategic teamwork, Artif. In- tell. 110 (1999) 241 -- 273. [40] J.M. Such, A. Espinosa, A. Garcia-Fornes, V. Botti, Partial identities as a foundation for trust and reputation, Eng. Appl. Artif. Intell. (2011). [41] J.M. Such, A. Espinosa, A. Garc´ıa-Fornes, C. Sierra, Self-disclosure decision making based on intimacy and privacy, Information Sciences 211 (2012) 93 -- 111. [42] K. Sycara, Machine learning for intelligent support of conflict resolution, Decision Support Systems 10 (1993) 121 -- 136. [43] L. Thompson, E. Peterson, S. Brodt, Team negotiation: An examination of integrative and distributive bargaining, J. Pers. Soc. Psychol. 70 (1996) 66 -- 78. 43
cs/0609041	1	0609	2006-09-08T10:12:35	Primitive operations for the construction and reorganization of minimally persistent formations	[ "cs.MA" ]	In this paper, we study the construction and transformation of two-dimensional persistent graphs. Persistence is a generalization to directed graphs of the undirected notion of rigidity. In the context of moving autonomous agent formations, persistence characterizes the efficacy of a directed structure of unilateral distances constraints seeking to preserve a formation shape. Analogously to the powerful results about Henneberg sequences in minimal rigidity theory, we propose different types of directed graph operations allowing one to sequentially build any minimally persistent graph (i.e. persistent graph with a minimal number of edges for a given number of vertices), each intermediate graph being also minimally persistent. We also consider the more generic problem of obtaining one minimally persistent graph from another, which corresponds to the on-line reorganization of an autonomous agent formation. We prove that we can obtain any minimally persistent formation from any other one by a sequence of elementary local operations such that minimal persistence is preserved throughout the reorganization process.	cs.MA	cs	Primitive operations for the construction and reorganization of minimally persistent formations Julien M. Hendrickx∗, Barı¸s Fidan†, Changbin Yu, Brian D.O. Anderson and Vincent D. Blondel June 13, 2021 Abstract In this paper, we study the construction and transformation of two- dimensional persistent graphs. Persistence is a generalization to directed graphs of the undirected notion of rigidity. In the context of moving autonomous agent formations, persistence characterizes the efficacy of a directed structure of unilateral distances constraints seeking to preserve a formation shape. Analogously to the powerful results about Henneberg sequences in minimal rigidity theory, we propose different types of di- rected graph operations allowing one to sequentially build any minimally persistent graph (i.e. persistent graph with a minimal number of edges for a given number of vertices), each intermediate graph being also mini- mally persistent. We also consider the more generic problem of obtaining one minimally persistent graph from another, which corresponds to the on-line reorganization of an autonomous agent formation. We prove that we can obtain any minimally persistent formation from any other one by a sequence of elementary local operations such that minimal persistence is preserved throughout the reorganization process. ∗J. M. Hendrickx (corresponding author) and V. Blondel are with Department of Mathe- matical Engineering, Universit´e catholique de Louvain, Avenue Georges Lemaitre 4, B-1348 Louvain-la-Neuve, Belgium; hendrickx,[email protected]. Their work is sup- ported by the Belgian Programme on Interuniversity Attraction Poles initiated by the Belgian Federal Science Policy Office, and The Concerted Research Action (ARC) "Large Graphs and Networks" of the French Community of Belgium. The scientific responsibility rests with its authors. Julien Hendrickx holds a FNRS fellowship (Belgian Fund for Scientific Research). This work was supported in part by DoD AFOSR URI for "Architectures for Secure and Robust Distributed Infrastructures," F49620-01-1-0365 (led by Stanford University). †B. Fidan, C. Yu and B. Anderson are with Australian National University and National ICT Australia , 216 Northbourne Ave, Canberra ACT 2601 Australia ; baris.fidan,brad.yu,[email protected]. Their work is supported by an Australian Research Council Discovery Project Grant and by National ICT Australia, which is funded by the Australian Government's Department of Communications, Information Tech- nology and the Arts and the Australian Research Council through the Backing Australia's Ability Initiative. Changbin Yu is an Australia-Asia Scholar supported by the Australian Gov- ernment's Department of Education, Science and Training through Endeavours Programs. 1 1 Introduction The recent progress in the field of autonomous agent systems has led to new problems in control theory [1,2] and graph theory [3,5,9]. By autonomous agent, we mean here any human controlled or unmanned vehicle that can move by it- self and has a local intelligence or computing capacity, such as ground robots, air vehicles or underwater vehicles. The results derived in this paper concern mostly autonomous agents evolving in a two dimensional space. Many applications require some inter-agent distances to be kept constant during a continuous move in order to preserve the shape of a multi-agent for- mation. In other words, some inter-agent distances are explicitly maintained constant so that all the inter-agent distances remain constant. The information structure arising from such a system can be efficiently modelled by a graph, where agents are abstracted by vertices and actively constrained inter-agent distances by edges. We assume here that those constraints are unilateral, i.e., that the responsibility for maintaining a distance is not shared by the two con- cerned agents but relies on only one of them while the other one is unaware of it. This asymmetry is modelled by the use of directed edges in the graph. The characterization of the directed information structures which can efficiently maintain the formation shape has begun to be studied under the name of "di- rected rigidity" [1, 3]. These works included several conjectures about minimal directed rigidity, i.e., directed rigidity with a minimal number of edges for a fixed number of vertices. In [5], Hendrickx et al. proposed a theoretical frame- work to analyze these issues, where the name of "persistence" was advanced in preference to "directed rigidity", since the latter notion does not correspond to the immediate transposition of the undirected notion of rigidity to directed graph. The intuitive definition of persistence is the following: An information structure is persistent if, provided that each agent is trying to satisfy all the dis- tance constraints for which it is responsible, all the inter-agent distances remain constant and as a result the formation shape is preserved. It is shown in [5] that persistence is actually the conjunction of two distinct notions: rigidity of the underlying undirected graph (i.e. the undirected graph obtained by ignor- ing the direction of the edges), and constraint consistence. Intuitively, rigidity means that, provided that all the prescribed distance constraints are satisfied during a continuous displacement, all the inter-agent distances remain constant, as shown in Figure 1. Constraint consistence of an information structure means that, provided that each agent is trying to satisfy all its distances constraints, all the agents actually succeed in doing so. In other words, no agent has an impossible task, as shown in the example in Figure 2. Observe that this last notion depends strongly on the directed structure of the graph, while rigidity only relies on its underlying undirected graph. An example of persistent graph is provided in Figure 3. Note that for agents evolving in a two-dimensional space, a purely combinatorial criterion to decide persistence is provided in [5]. In this paper, we focus on minimally persistent graphs, that are persistent graphs having a minimal number of edges (for a given number of vertices), and 2 (a) (b) Figure 1: Representation of (a) a non-rigid and (b) a rigid graph/formation. The solid structure in (a) can indeed be deformed to the dotted structure without breaking any distance constraint. 3 2 4 (a) 3 ?? 1 1 2 4 (b) Figure 2: Representation of (a) a constraint consistent and (b) a non-constraint consistent (in 2 dimension) graph/formation. One can indeed see in (b) that for almost any uncoordinated continuous displacement of the agents 2 and 4 (which are unconstrained), the agent 3 is unable to move in such a way that it maintains its distances to all of 1,2 and 4 constant. However, such a situation could not happen in graph (a). 3 2 4 1 Figure 3: Representation of a persistent graph, i.e., a rigid constraint consistent graph. 3 their connections with minimally rigid graphs. More particularly we analyze dif- ferent ways to sequentially build minimally persistent graphs, analogously to the Henneberg sequences for the minimally rigid graphs [7, 10]. It has indeed long been known that every minimally rigid graph can be obtained from the com- plete graph on two vertices by a sequence of two basic operations, as detailed in Section 2. The natural extension of these operations to directed graphs [3] does not allow one to build all minimally persistent graphs, as remarked in [5] and reviewed later in Section 2. We consider here different possible additional opera- tions that would help to achieve this purpose. We also consider the more generic problem of obtaining one persistent graph from another. From an autonomous agent point of view, this corresponds to an on-line reorganization of the agent formation. The subsequent analysis leads us then to the definition of different "distances" between persistent graphs (the distance between two graphs being the smallest number of operations needed to obtain one from the other). Note that although the notion of persistence has been also defined in three or higher dimensions [6, 12, 13], the present analysis only concerns two-dimensional per- sistence, i.e., the persistence of graphs representing the information structure of a formation evolving in a two-dimensional space. Extension to the three dimen- sional case may be difficult; even for undirected graphs, Henneberg sequence theory is effectively incomplete. In Section 2, we review the main properties of minimally rigid and minimally persistent graphs. We present the two basic undirected operations - vertex ad- dition and edge splitting - involved in the Henneberg sequences, together with their natural extension to directed graphs. We show that although these di- rected operations preserve minimal persistence, they are not sufficient to build all minimally persistent graphs. This analysis is done by reasoning on reverse construction of persistent graphs using reverse operations. In Section 3, we show how the goal of building all minimally persistent graph can be reached by introducing a third local directed operation - edge reversal. We see that, unlike when building minimally rigid undirected graphs with Henneberg sequences, the required number of operations is not uniquely determined by the size of the graph. We show in Section 4 that this drawback can be avoided by using only directed operations equivalent to the vertex addition and the edge split- tings from an undirected point of view. However, we prove that a set of such operations allowing one to build all minimally persistent graphs always contain at least one non-confined operation, i.e. an operation reversing the directions of (possibly several) edges that are not affected by the corresponding operation for undirected graphs. We provide then such a set of four operations, and analyze the relations between this set and the set of three operations treated in Section 3. Finally, this paper ends with the concluding remarks of Section 5. 4 2 Directed and undirected Henneberg sequences In this section, we recall some results about (minimal) rigidity and (minimal) persistence. We also describe the Henneberg sequences for undirected graphs and show why their obvious adjustment to the directed case is not sufficient to build all minimally persistent graphs. 2.1 Minimally rigid graphs Note that in this section, all graphs are considered as undirected, but in the rest of this paper, they are always assumed to be directed. However, although all the definitions and results of this section are given for undirected graphs, they can also be applied to directed graphs. If G is a directed graph, we call the underlying undirected graph of G the undirected graph obtained by ignoring the directions of the edges of G. The rigidity of a graph has the following intuitive meaning: Suppose that each vertex represents an agent in a formation, and each edge represents an inter-agent distance constraint enforced by an external observer. The graph is rigid if for almost every such structure, the only possible continuous moves are those which preserve every inter-agent distance. Note that this notion also represents the rigidity of a framework where the vertices correspond to joints and the edges to bars. For a more formal definition, the reader is referred to [5, 10]. In ℜ2, there exists a combinatorial criterion to check if a given graph is rigid (Laman's theorem [8,11]). A minimally rigid graph is a rigid graph such that no edge can be removed without losing rigidity. From Laman's Theorem, it is possible to deduce the following criterion: Proposition 1. A graph G = (V, E) (V > 1) is minimally rigid if and only if E = 2 V − 3 and for all E′′ ⊆ E, E′′ 6= ∅, there holds E′′ ≤ 2 V (E′′) − 3. We say that a pair of unconnected vertices defines an implicit edge in a graph G = (V, E) if their connection would create a subgraph G′ = (V ′, E′) with E′ > 2 V ′ − 3. Intuitively, this means that the addition of such an edge would not improve the rigidity of the graph, i.e., the constraint that this edge would enforce is a linear combination of already present constraints. One can easily prove that two unconnected vertices define an implicit edge in a graph if and only if there is a minimally rigid subgraph containing both of them. By extension, we sometimes call an edge of a graph an explicit edge. In a (mini- mally) rigid graph, every pair of vertices is connected by either an explicit or implicit edge. But, if one removes an (explicit) edge in a minimally rigid graph, the corresponding pair of vertices never defines an implicit edge in the graph obtained. Let j, k be two distinct vertices of a minimally rigid graph G = (V, E). A vertex addition operation consists in adding a vertex i, and connecting it to j and k, as shown in Figure 4(a). One can see using Proposition 1 that this 5 k j i l k j i j k l k j (a) (b) Figure 4: Representation of (a) undirected vertex addition operation and (b) edge splitting operation. operation preserves minimal rigidity. Moreover, if a vertex i has degree 2 in a minimally rigid graph, one can always perform the inverse vertex addition op- eration by removing it (and its incident edges) and obtain a smaller minimally rigid graph. Let j, k, l be three vertices of a minimally rigid graph such that there is an edge between j and k. An edge splitting operation consists in removing this edge, adding a vertex i and connecting it to j, k and l, as shown in Figure 4(b). This operation provably preserves minimal rigidity [10]. The reverse operation is less straightforward than the reverse vertex addition operation. Given a vertex i connected to j, k and l, the minimal rigidity of the graph is preserved if one removes i and adds one edge among (j, k), (k, l) and (l, j). However, one cannot always freely choose any one of these edges to add. One has indeed to make sure that the added edge does not already belong to the graph, and also that its addition does not create a subgraph G′ = (V ′, E′) with E′ > 2 V ′ − 3, i.e., that the pair of vertices does not define an implicit edge in the graph obtained after deletion of i. Figure 5 shows an example of such an unfortunate added edge selection. Suppose indeed that the vertex 5 is removed from the minimally rigid graph 5(a). The pair (1, 4) does provably not define an implicit edge, and its addition leads thus to a minimally rigid graph, which is represented in Figure 5(b). However, if (1, 6) is added instead of (1, 4), the graph obtained contains a subgraph G′ = (V ′, E′) with V ′ = {1, 2, 3, 6} such that 6 = E′ > 2 V ′−3 = 5, as shown in Figure 5(c). The pair (1, 6) defines thus an implicit edge. It is pos- sible to prove that at least one among the three possible pairs of vertices does not define an actual nor an implicit edge [8, 10]. One can thus always perform a reverse edge splitting on any vertex with a degree 3. A Henneberg sequence is a sequence of graphs G2, G3, . . . , GV with G2 be- 6 1 1 1 2 3 4 5 2 3 4 2 3 4 6 (a) 6 6 (b) (c) Figure 5: Example of unfortunate added edge selection in reverse edge splitting. After the removal of the vertex 5 from the minimally rigid graph (a), minimal rigidity can be preserved by the addition of the edge (1, 4) but not of (1, 6), as shown respectively on (b) and (c). The pair (1, 6) defines an implicit edge in the minimally rigid subgraph induced by 1, 2, 3 and 6. ing the complete graph on two vertices K2 and each graph Gi (i ≥ 3) can be obtained from Gi−1 by either a vertex addition operation or an edge splitting operation. Since these operations preserve minimal rigidity and since K2 is minimally rigid, every graph in such a sequence is minimally rigid. A simple degree counting argument shows that every minimally rigid graph GV = (V, E) with more than 2 vertices contains at least one vertex with degree 2 or 3. One can thus always perform either a reverse vertex addition or a reverse edge splitting operation and obtain a smaller minimally rigid graph GV −1. Doing this recursively, one eventually obtains a minimally rigid graph on two vertices, which can only be K2. It is straightforward to see that the sequence K2 = G2, G3, . . . GV is then a Henneberg sequence. We have thus proved the following result [10]: Theorem 1. Every minimally rigid graph on more than one vertex can be obtained as the result of a Henneberg sequence. The result of Theorem 1 provides a way to exhaustively enumerate all min- imally rigid graphs. One can thus use it to obtain an upper bound on the number of minimally rigid graph having a certain number of vertices. However, this only provides an upper bound for the Henneberg sequence allowing one to build a certain minimally rigid graph is usually not unique. The graph in Figure 1(b) can for example be obtained from K2 by either two vertex additions or one vertex addition followed by one edge splitting. 2.2 Minimally persistent graphs Consider a group of autonomous agents represented by vertices of a graph. To each of these agents, one assigns a (possibly empty) set of unilateral distance constraints represented by directed edges: the notation (i, j) for a directed edge connotes that the agent i has to maintain its distance to j constant during any 7 continuous move. The persistence of the directed graph means that provided that each agent is trying to satisfy its constraints, the distance between any pair of connected or non-connected agents is maintained constant during any continuous move, and as a consequence the shape of the formation is preserved. A formal definition of persistence is given in [5]. In a two-dimensional space, an agent having only one distance constraint to satisfy can move on a circle centered on its neighbor, and has thus one degree of freedom. Similarly, an agent having no distance constraint to satisfy can move freely in the plane and has thus two degrees of freedom. We call the number of degrees of freedom of a vertex i the (generic) dimension of the set in which the corresponding agent can chose its position (all the other agents being fixed). It represents thus in some sense the decision power of this agent. The number of degrees of freedom of a vertex i is given by max (0, 2 − d+(i)) (where d+(i) and d−(i) represent respectively the in- and out-degree of the vertex i). A graph is minimally persistent if it is persistent and if no edge can be re- moved without losing persistence. The following result provides a swift criterion to decide minimal persistence : Proposition 2. [5] A graph is minimally persistent if and only if it is minimally rigid and no vertex has an out-degree larger than 2. As a consequence of Proposition 2, the number of degrees of freedom of a vertex i in a minimally persistent graph is 2−d+(i). By Proposition 1, it follows after summation on all the vertices that the total number of degrees of freedom present in such a graph is always 3. This result is consistent with the intuition, there are indeed three degree of freedom to chose the position and orientation of a rigid body in a 2-dimensional space. Let j, k be two distinct vertices of a minimally persistent graph G = (V, E). A directed vertex addition [3, 4] consists in adding a vertex i and two directed edges (i, j) and (i, k) as shown in Figure 6(a). Since it is a vertex addition oper- ation, it preserves minimal rigidity. Besides, the added vertex has an out-degree 2 and the out-degree of the already existing vertices are unchanged. By Propo- sition 2, the directed vertex addition thus preserves the minimal persistence. Moreover, if a vertex has an out-degree 2 and an in-degree 0 in a minimally persistent graph, one can always perform a reverse (directed) vertex addition by removing it, and obtain a smaller minimally persistent graph. Let (j, k) be a directed edge in a minimally persistent graph and l a distinct vertex. A directed edge splitting [3, 4] consists in adding a vertex i, an edge (i, l), and replacing the edge (j, k) by (j, i) and (i, k), as shown in Figure 6(b). Again, this operation preserves minimal rigidity since it is an edge splitting operation from an undirected point of view, and since the added vertex has an out-degree 2 and the out-degree of the already existing vertices are unchanged, it also preserves minimal persistence. But, unlike in the case of directed vertex 8 k j i l k j i j k l k j (a) (b) Figure 6: Representation of the directed vertex addition (a) and edge splitting (b). addition, the reverse operation cannot always be performed. Suppose indeed that we have a vertex i with out-degree 2 and in-degree 1, and call its neighbors j, k and l. The reverse operation consists in removing i and its incident edges, and adding either (j, k) or (j, l) (note that k and l are interchangeable). Adding any other edge such as (k, l) or (l, k) would indeed prevent the operation from being out-degree preserving, and one could then not guarantee the minimal persistence of the graph obtained (by Theorem 2). But, it can happen that both pairs (j, l) and (j, k) are already connected by explicit or implicit edges. In such a case, minimal rigidity is only preserved by addition of an edge between k and l, which as explained above may not preserve persistence. We now show that the vertex addition and edge splitting operations do not allow one to grow all minimally persistent graphs from an initial seed. Consider the graph in Figure 7 (for n ≥ 1). One can verify by Theorem 1 that it is mini- mally rigid. Moreover, no vertex has an out-degree larger than 2; by Proposition 2 it is thus minimally persistent. Observe that no vertex has an in-degree 0; it is thus impossible to perform a reverse vertex addition operation. Moreover, only the vertex 2n satisfies the required conditions about the in- and out-degree in order to offer the possibility of removal by a reverse edge splitting operation, and one can verify that this operation cannot be performed due to the presence in the graph of the edges (2n + 1, 2n − 1) and (2n − 2, 2n − 1). Since this is true for any n, we have an infinite class of graphs on which none of the two above defined reverse operations can be performed (Note that there provably exists other such infinite classes in which the graphs have one vertex with two degrees of freedom and one vertex with one degree of freedom instead of three vertices with one degree of freedom). As a consequence, it is not possible to build every minimally persistent graph by performing a sequence of directed vertex addition or edge splitting operations on some seed graph taken in a finite set of graphs. However, we have the following less powerful result (as argued in [5]). 9 2 4 2n−2 2n 1 3 5 2n−1 2n+1 Figure 7: Class of graphs on which no reverse vertex addition or edge splitting can be performed. Proposition 3. It is possible to assign directions to the edges of any minimally rigid graph such that the obtained directed graph is minimally persistent and can be obtained by performing a sequence of vertex additions and edge splittings on an initial graph of two vertices connected by one directed edge (called a "leader- follower seed"). Proof. Let G be a minimally rigid (undirected) graph. By Theorem 1, it can be obtained by performing a sequence of undirected vertex additions and edge splittings on K2. By performing the same sequence of the directed version of these operations on an initial leader-follower seed, one obtains a directed graph having G as underlying undirected graph. Moreover, since this initial seed is trivially minimally persistent (by Proposition 2), and since the directed versions of both vertex addition and edge splitting preserve minimal persistence, the obtained graph is minimally persistent. In the following sections, we examine different possibilities of additional op- erations that allow the construction of all minimally persistent graphs. In order to avoid confusion, we shall sometimes refer to the directed version of vertex addition and edge splitting as standard vertex addition and standard edge split- ting. We denote by S the set consisting of these two operations and S−1 the one consisting of their inverses (the same convention is used in the sequel for all the operations set). Note that it is always possible to perform an operation of S on a minimally persistent graph, but we have seen that this is not true for operations of S−1. 3 A purely directed operation We introduce here a third persistence-preserving operation: the edge reversal. Unlike those of S, does not affect the underlying undirected graph. We then define to macro-operations which help us to prove that the edge reversal is suf- ficient to obtain any minimally persistent graph from any other one having the same underlying undirected graph, and show how this implies that this opera- tion combined with those of S is sufficient to obtain any minimally persistent graph from a unique initial seed. 10 3.1 Edge reversal Let (i, j) be an edge such that j as at least one degree of freedom, i.e., d+(j) = 0 or d+(j) = 1. The edge reversal operation consists in replacing the edge (i, j) by (j, i). As a consequence, one degree of freedom is transferred from j to i. This operation is its auto-inverse and preserves minimal persistence since it does not affect the underlying undirected graph and the only increased out-degree d+(j) remains no greater than 2. From an autonomous agent point of view j transfers its decision power (or a part of it) to i. 3.2 Path reversal Given a directed path P between a vertex i and a vertex j such that j has a positive number of degrees of freedom, a path reversal consists in reversing the directions of all the edges of P . As a result, j loses a degree of freedom, i acquires one, and there is a directed path from j to i. Moreover, the number of degrees of freedom of all the other vertices remain unchanged. Note that i and j can be the same vertex, in which case the path either has a trivial length 0 or is a cycle. In both of these situations, the number of degrees of freedom is preserved for every vertex. The path reversal can easily be implemented with a sequence of edge re- versals: Since j has a degree of freedom, one can reverse the last edge of the path, say (k, j), such that j loses one degree of freedom while k acquires one. One can then iterate this operation along the path until i, as shown in Figure 8. At the end, i has have an additional degree of freedom, j has lost one, and all the edges of the path have have been reversed. Note that the sequence of edge reversals can usually not be performed in another order, for the condition about the availability of a degree of freedom would not be satisfied. The final result would be the same, but all the intermediate graphs would not necessarily be minimally persistent. We now show that this operation allows one to reposition the three degrees of freedom of a minimally persistent graph onto any chosen vertices (with at most two degrees of freedom on a single vertex). For this purpose, we need the following result (which is a particular case of a result available in [6, 13]): Proposition 4. Let G be a minimally persistent graph, i and j two vertices of G with d+(i) ≥ 1 and d+(j) ≤ 1. Then, there is a directed path from i to j. Proof. Suppose (to obtain a contradiction) that we have a minimally persistent (and thus minimally rigid) graph G = (V, E), a vertex i with positive out-degree and a vertex j with a positive number of degrees of freedom such that there is no directed path connecting i to j. Let V ′′ ⊂ V be the set of vertices that can be reached from i, and E′′ the set of edges that leave vertices of V ′′. Obviously, every edge of E′′ is incident to two vertices of V ′′, and because d+(i) > 0, we have V ′′ ≥ 2 and E′′ ≥ 1. Moreover, the sum on the vertices of V ′′ of 11 i i i i * k k * k k j * j j j * Figure 8: Implementation of the path reversal by a sequence of edge reversals. The symbol "" represents one degree of freedom. the numbers of degrees of freedom (which we denote F (V ′′)) is smaller than 3. There are indeed only three degrees of freedom in a minimally persistent graph as explained in Section 2.2, and the vertex j which has at least one of them does by hypothesis not belong to V ′. Since every vertex has an out-degree smaller no greater than two in a minimally persistent graph (by Proposition 2), we have thus a subgraph G′′ = (V ′′, E′′) such that E′′ = X k∈V ′′ d+(k, G′′) = 2 V ′′ − F (V ′′) > 2 V ′′ − 3, which by Proposition 1 is impossible for a subgraph of a minimally rigid graph. Let us now suppose that one wants to transfer a degree of freedom from some vertex j to some vertex i which has at most one degree of freedom (transferring a degree of freedom to a vertex that has already two degrees of freedom would indeed be impossible as there is no edge inwardly incident). By Proposition 4, there exists a directed path from i to j. The transfer can then be done by reversing this path, which leaves the positions of all the other degrees of freedom unchanged. By doing this at most three times, one can thus reposition the three degrees of freedom onto any chosen vertices. As a consequence, we have the following result. Proposition 5. Let GA and GB be two minimally persistent graphs having the same underlying undirected graph. By applying a sequence of at most three path reversals on GA, it is possible to obtain a minimally persistent graph G′ A in which every vertex has the same number of degrees of freedom as in GB 12 l l l l i i P m * P' m i * P' m i P m * Figure 9: Implementation of cycle reversal. The "" represents one degree of freedom. 3.3 Cycle reversal A cycle reversal consists in reversing all the edges of a directed cycle. This operation does not affect the number of degrees of freedom of any vertex nor the underlying undirected graph, and preserves therefore minimal persistence. A cycle reversal on a minimally persistent graph can be implemented by a sequence of edge reversals. Let us indeed first suppose that there is a vertex i in the cycle that has at least one degree of freedom. In that case, the cycle reversal is just a particular case of the path reversal, with i = j. We now assume that no vertex in the cycle has a degree of freedom. Let l be a vertex in the cycle, and m a vertex that does not belong to the cycle but has a degree of freedom. By Proposition 4, it follows that there exists a directed path from l to m. Let i be the last vertex in this path belonging to the cycle. There is trivially a path P from i to m such that every other vertex of this path does not belong to the cycle. The implementation of a cycle reversal by three path reversals is then represented in Figure 9. One begins by reversing the path P into P ′ such that i acquires a degree of freedom. As explained above, the cycle can then be reversed since it is a particular case of path reversal, and finally, one reverses the path P ′ back to P such that the degree of freedom acquired by i is re-transmitted to m. Note that an alternative equivalent approach is to reverse the path from l to m containing i and one part of the cycle, and then to reverse the newly created path from m to l containing i and the other part of the cycle. Remark 1. Both cycle reversal and path reversal are their auto-inverse, as is the case for edge reversal. Moreover, the fact that they can be implemented using only edge reversals is another way to show that they preserve minimal persistence. We now prove that from any minimally persistent graph, one can obtain any other minimally persistent graph having the same underlying undirected graph and allocation of degrees of freedom by a sequence of cycle reversals. For this purpose, we need the following result. Lemma 1. Let GA = (V, EA) and GB = (V, EB) be two graphs having the same underlying undirected graph and such that every vertex has the same out-degree 13 in both graphs. If an edge of GA has the opposite direction to that in GB, it belongs to (at least) one cycle of such edges in GA. Proof. Suppose that (i0, i1) ∈ EA and (i1, i0) ∈ EB (i.e., this edge has opposite directions in GA and GB); then there exists at least one vertex i2 6= i0 such that (i1, i2) ∈ EA and (i2, i1) ∈ EB. For if the contrary holds, we would have d+(i1, GA) = d+(i1, GB) − 1, which contradicts our hypothesis. Repeating this argument recursively, we obtain an (infinite) sequence of vertices i0, i1, i2, . . . such that for each j ≥ 0, (ij, ij+1) ∈ EA and (ij+1, ij) ∈ EB. Since there are only a finite number of vertices in V , at least one of them will appear twice in this sequence. By taking the subsequence of vertices (and induced edges) appearing in the infinite sequence between any two of its occurrences we obtain then a cycle having the required properties. Proposition 6. Let GA = (V, EA) and GB = (V, EB) be two minimally per- sistent graphs having the same underlying undirected graph and such that every vertex has the same number of degrees of freedom in both of them. Then it is possible to obtain GB from GA by a sequence of at most EA /3 = EB /3 cycle reversals. Proof. Suppose that GA 6= GB, and let Eo denote the set of edges of GA that do not have the same direction as in GB. Since both graphs have the same underlying undirected graph and since all the vertices have the same out-degrees in both of them it follows from Lemma 1 that there exists a cycle of edges of Eo. Eo is thus strictly decreased by reversing this cycle. Doing this recursively leads then to Eo = 0, i.e, to a situation where GA = GB. Moreover, since every cycle has at least three edges. The number of cycle reversals is at most E /3 3.4 Obtaining all minimally persistent graphs using three primitive operations Using the results of the two previous subsections, we can now show the following Proposition. Proposition 7. By applying a sequence of edge reversals to a given minimally persistent graph, it is possible to obtain any other minimally persistent graph having the same underlying undirected graph. Moreover, all the intermediate graphs are then minimally persistent. Proof. Since both path reversal and cycle reversal can be implemented by a sequence of edge reversals (which preserves minimal persistence), this result is a direct consequence of Propositions 5 and 6. From an autonomous agent formation perspective, suppose that a reorgani- zation of the distance constraints distribution needs to be performed, and that this reorganization preserves the structure of constraints from an undirected point of view, i.e., the reorganization is solely one involving changes of some 14 directions. Proposition 7 implies that this can be done by a sequence of lo- cal degree of freedom transfers, in such a way that during all the intermediate stages, the formation shape is guaranteed to be maintained as a result of per- sistence being preserved. Let T be the set of operations containing vertex addition, edge splitting, and edge reversal. A leader-follower seed is a minimally persistent graph on two vertices. It contains only one edge, leaving a vertex called "the follower", and arriving at the other one, called "the leader". The next theorem states that one can obtain any minimally persistent graph from an initial leader-follower seed using only operations of T . Theorem 2. Every minimally persistent graph can be obtained by applying a sequence of operations of T to an initial leader-follower seed. Moreover, all the intermediate graphs are minimally persistent. Proof. Consider a minimally persistent graph G. This graph is also minimally rigid. By Proposition 3, there exists thus a (possibly different) minimally persis- tent graph having the same underlying undirected graph that can be obtained by performing a sequence of operations of S ⊂ T on an initial leader follower seed. By Proposition 7, G can then be obtained by applying a sequence of edge reversals on this last graph. Moreover, since all the operations of T preserve minimal persistence, all the intermediate graphs are minimally persistent. As an illustration of Theorem 2, consider the graph G represented in the right hand side of Figure 10(c), which is an instantiation of the graph of Figure 7 with n = 2. As explained in Section 2.2, it cannot be obtained by applying a vertex addition or an edge splitting on a smaller minimally persistent graph. However, by Theorem 2, it can be obtained by applying a sequence of operations of T on an initial leader-follower seed. Let us take 1 and 2 as respectively leader and follower of this initial seed. One can begin by adding 3, 4 and 5 using three vertex additions as shown in Figure 10(a). The graph obtained has the same underlying undirected graph as G, but the degrees of freedom are not allocated to the same vertices. By reversing the path P ({5, 4, 2, 1}) (using a sequence of edge reversals), one can then transfer one degree of freedom from 1 to 5 as shown in Figure 10(b) such that in the obtained graph, all the vertices have the same number of degrees of freedom as in G. As stated in Proposition 6, any edge of this graph that does not have the same direction as in G belongs to a cycle of such edges. The only such cycle here is C. By reversing it (using a sequence of edge reversals), one finally obtains the graph G, as shown in Figure 10(c). Note that consistently with Theorem 2, all the intermediate graphs are minimally persistent. Theorem 2 also proves that it is always possible to obtain a leader-follower pair from any minimally persistent graph by applying an appropriate sequence of operations of T −1. This can be also stated as follows: Theorem 3. Let G be a minimally persistent graph. By applying a (possibly empty) sequence of edge reversals on G, it is always possible to obtain a mini- 15 2 4 1 ** 3 (a) 2 * P 4 1 ** 3 5 1 * 2 * 4 (b) C 3 1 * 5 * 1 * (c) 2 * 2 * 5 3 3 4 4 5 * G 5 * Figure 10: Example of obtaining of a minimally persistent graph by applying a sequence of operations of T on a leader-follower seed. The graph G is obtained from the leader-follower seed by (a) three vertex additions, (b) the reversal of the path P and (c) of the cycle C. mally persistent graph on which at least one operation of S−1 (i.e, reverse edge splitting or reverse vertex addition) can be performed. Starting from a minimally persistent graph, one can thus first use operations of T −1 to obtain a leader-follower pair, and then use operations of T to obtain any other minimally persistent graph. This method is generally not optimal in terms of the number of operations. However, the argument proves the following corollary. Corollary 1. Every minimally persistent graph can be transformed into any other minimally persistent graph using only operations of T ∪ T −1. This result allows us to define a distance on the minimally persistent graphs (on more than one vertex) by saying that the distance between two of them is the minimal number of operations of T ∪ T −1 needed to transform one into the other. Propositions 3, 5 and 6 imply that the distance between two graphs is quadratically bounded by their size, the quadratic character coming from the cycle reversing operations (the others requires only a linear number of opera- tions). However, a better bound is likely to exist. Remark 2. Observe that the three operations of T are relatively basic ones and are performed locally. They could thus easily be implemented in a local way on an autonomous agent formation. It might be however possible to improve this 16 l k j l k j i l k j i Figure 11: Implementation of the edge splitting by a vertex addition and an edge reorientation. The vertex i is first added with two out-going edges by vertex addition, and the edge (j, k) is then reoriented and becomes (j, i). basic character using for example an operations such as an edge reorientation, i.e., an operation consisting in changing the arrival vertex of an edge. As shown in Figure 11, a vertex addition operation and an edge reorientation operation can indeed implement an edge splitting operation which could thus be discarded. However, this would require an efficient and simple criterion to determine when such an edge reorientation operation could be performed. 4 An alternative set of four primitive operations As explained in Section 3, every minimally persistent graph can be obtained by applying a sequence of operations belonging to T on an initial leader-follower seed, in such a way that all the intermediate graphs are minimally persistent. However, unlike in the case of an undirected Henneberg sequence (see Section 2), the number of vertices in the final graph does not determine uniquely the required number of intermediate graphs, but only an upper bound on it (see Section 3). In this section, we focus on sets of operations equivalent to those of S from an undirected point of view and that allow one to build all minimally persistent graphs (the number of intermediate graphs being thus uniquely de- termined by the the number of vertices of the final graph since each operation adds one vertex). It is proved that those sets always contain at least one oper- ation involving the reversal of edges that are not affected by the corresponding operation for undirected graphs. We then provide such a set A of four types of operations and show how it allows one to build any minimally persistent graph G = (V, E) by applying V − 2 operations to an initial leader-follower seed. Finally we study the relations between the two sets A and T . 4.1 Necessary involvement of external edges. In the sequel, we adopt the terms generalized vertex addition and generalized edge splitting for any operation which is equivalent to a vertex addition or an edge splitting from an undirected point of view. Such an operation is said to be confined if it only affects edges that are involved in the corresponding undi- rected operation. For example, all the operation of S are confined, while the 17 edge reversal operation defined in Section 3.1 is not. Suppose that one wants to remove a vertex (without losing persistence) from the provably minimally persistent graph represented in Figure 12 using a generalized reverse edge splitting or reverse vertex addition. The only ones that can be removed are those with a label "+", and due to their total degree, this could only be done by a generalized reverse edge splitting operation. Suppose now that one wants to use a confined version of this operation. One would then remove one of the vertices with a label "+" and connect two of its neighbors by a directed edge. Observe that among the three pairs of neighbors of any vertex with a label "+", two are already connected, and the last pair contains two vertices with an out-degree 2. Adding an edge between a pair of neighbors of the removed vertex without reversing the direction of any other edge would thus imply the presence of either a vertex with out-degree 3 (which by Theorem 2 is impossible in a minimally persistent graph) or of a cycle of length 2 (which by Proposition 1 cannot appear in a minimally rigid graph). This removal should therefore be performed by a non-confined reverse generalized edge splitting. The following result is thus proved. Proposition 8. If a set exists of generalized vertex additions and edge splittings allowing one to build all minimally persistent graphs from an initial leader- follower seed, such a set must always contain a non-confined edge splitting. The existence of confined operations that would not be equivalent to vertex addition or edge splitting, but that would however preserve minimal persistence and allow one to build all minimally persistent graphs with V vertices in V −2 operations (starting with a leader-follower seed) remains an open question. Note that such operations would have to be proved to preserve minimal rigidity. 4.2 Description of a set A of four primitive operations We define here a new set A of four operations. The first two are the standard vertex additions and edge splitting as described in Section 2.2, which implies that S ⊂ A. The two others are atypical versions of these. Let j, k be two vertices of a minimally persistent graph such that j has at least one degree of freedom. The atypical vertex addition operation consists in adding the vertex i, the edges (j, i) and (i, k), as shown in Figure 13(a). As a result, j loses a degree of freedom, and i appears with one. The reverse atypical vertex addition operation consists in removing a vertex with in- and out-degree 1. Proposition 9. Atypical vertex addition and reverse atypical vertex addition preserve minimal persistence. Proof. Since these operations are respectively a generalized vertex addition and a reverse generalized vertex addition, they preserve minimal rigidity as explained 18 + + * + + * + * + Figure 12: A minimally persistent graph no vertex of which can be removed (without losing persistence) by a reverse (generalized) vertex addition or a con- fined (generalized) reverse edge splitting. The symbol "" represents one degree of freedom. Vertices that are candidate to be removed by a reverse generalized edge splitting are labelled "+". in Section 2.1. Moreover, the reverse atypical vertex addition does not increase the out-degree of any vertex, while the atypical vertex addition only increases by one an out-degree that is smaller than 2. In both situations the graph obtained after performing the operation does not contain any vertex with an out-degree larger than 2 and is thus minimally persistent (by Proposition 2). Let j, k and l be three vertices of a minimally persistent graph such that there is a (simple) directed path from j to k and (k, l) ∈ E. The atypical edge splitting operation consists in removing (k, l), adding a vertex i and the edge (j, i), (i, k) and (i, l), and reversing the direction of all the edges belonging to the path from j to k, as represented in Figure 13(b). Proposition 10. Atypical edge splitting preserves minimal persistence. Proof. This operation is a generalized edge splitting and thus preserves minimal rigidity. Since it does not affect the out-degree of any already existing vertex and adds a vertex with out-degree 2, it also preserves minimal persistence (by Proposition 2). Consider a vertex i with out-degree 2 and in-degree 1 in a minimally persis- tent graph, and call its neighbors j, k and l as in Figure 13(b). Suppose that in the graph obtained after deletion of i, there is a path from k (or equivalently l) to j and the pair (k, l) is not connected by an implicit nor an explicit edge. The reverse atypical edge splitting consists then in removing i, reversing all the edges of the path from k to j to obtain a path from j to k, and adding the edge (k, l). Proposition 11. Reverse atypical edge splitting preserves minimal persistence. 19 j k j k i* l k j i l k j (a) (b) Figure 13: Representation of the atypical (a) vertex addition operation and (b) edge splitting operation, both belonging to A. The symbol "" represents one degree of freedom. Proof. From an undirected point of view, this operation consists in removing one vertex incident to three edges, and then connecting a pair of unconnected vertices that does not define an implicit edge in the intermediate graph. It thus preserves minimal rigidity. Moreover, it does not affect the out-degree of any remaining vertex. It follows from Proposition 2 that reverse atypical edge splitting preserve minimal persistence. The conditions in which the reverse atypical edge splitting can be performed are not always easy to check. However, the following result holds: Lemma 2. In a minimally persistent graph, a vertex with in-degree 1 and out- degree 2 can always be removed by either a reverse standard edge splitting or a reverse atypical edge splitting. Proof. Consider a minimally persistent graph G = (V, E) and a vertex i ∈ V with d+(i) = 2, d−(i) = 1. We call its neighbors j, k and l such that (j, i), (i, k), (i, l) ∈ E, as in Figure 14(a) Let us assume that i cannot be removed by a reverse standard edge splitting, i.e., that j is connected to both k and l by an explicit or implicit edge in G \ {i}. As already mentioned in Section 2.1, k and l are in such a case never connected by an implicit nor an explicit edge in G \ {i}, and the graph obtained by connecting them after removing i from G is therefore minimally rigid. It remains to prove the existence of a directed path from k or l to j (note that k and l are interchangeable) in order that an edge splitting operation can be applied. For this purpose, we are going to construct a minimally persistent graph G′ close to G and in which j has a degree of freedom. As explained below, Proposition 4 guarantees then the existence of a directed path from either k or l to j. It will be proved that this implies the existence of such a path in G \ {i}. 20 k l k l qQ k l P p i j P' p j * P'Q P' p Q j * (a) (b) (c) Figure 14: Representation of different paths and graph involved in Lemma 2. (b) shows the minimally persistent graph G′ obtained from the graph G represented in (a) by removing i, adding (k, l) and reversing P . (c) shows then the path Q (in G′) from k to j in a case where it has a non empty intersection with P ′. Consider a vertex p having at least one degree of freedom in G. Since d+(i) = 2 in G, Proposition 4 guarantees the existence (still in G) of a (cycle- free) directed path from i to p. Without loss of generality, let us assume that the second vertex of this path is k (it has indeed to be a neighbor of i, and k and l are interchangeable). There exists thus a directed path P from k to p to which i does not belong. We build G′ by reversing the path P (which becomes P ′), removing i and adding the edge (k, l), as shown in Figure 14(a) and (b). As already mentioned, any graph obtained by removing i and connecting k to l is minimally rigid. Moreover, after the reversal of P and addition of (k, l), p loses one degree of freedom, i.e., its out-degree in G′ is increased by one with respect to its out-degree in G (which is smaller than two). On the other hand, j acquires a degree of freedom. No vertex has thus an out-degree larger than 2 in G′, which is therefore minimally persistent. In G′, k has by construction a positive out-degree, and j has at least one degree of freedom. By Proposition 4, there exists thus a (cycle-free) directed path Q in G′ from k to j. In order to prove the existence of such a path in G \ {i}, we consider three cases. Observe that the only edges of G′ that do not exist in G \ {i} are (k, l) and those of P ′ (which exist but with the opposite direction). • Q and P ′ have no common edge, (k, l) 6∈ Q: In that case, the path Q also exists in G \ {i}. • Q and P ′ have no common edge, (k, l) ∈ Q: Since the (simple) path Q does not contain any cycle, (k, l) must be the first edge of Q. By removing this edge, one obtains a directed path from l to j having no intersection with P ′ and that does not contain (k, l). This path exists thus also in G \ {i}. • Q and P ′ have some common edge(s): Let q be the last vertex of Q that also belongs to P ′. The edges and vertices of Q which are after q constitute a directed path from q to j. By definition, it does not intersect P ′, and 21 it does not contain (k, l) for this would mean that Q contains a cycle. It exists therefore also in G. Moreover, since q belongs to the path P ′ in G′, it belongs to the path P in G \ {i}, and there is thus a directed path from k to q in G \ {i}. By taking the union of the path from k to q and the one from q to j, one obtains then a directed path from k to j in G \ {i}, as shown in Figure 14(c). In any of these three cases, there is thus a directed path from k or l to j in G \ {i}. As explained above, this implies that one can perform the reverse atypical edge splitting on i if one cannot perform the reverse standard one. 4.3 Obtaining all minimally persistent graphs using A Let G = (V, E) be a minimally persistent and therefore minimally rigid graph with more than two vertices. By Proposition 1, there holds 4 V − 6 = E = Pi∈V d−(i) + d+(i). Moreover, it can be shown (using Proposition 1) that such a graph never contains any vertex with a total degree smaller than 2. A counting argument shows then that it always contain at least one vertex with either (d−, d+) = (0, 2), (1, 1) or (1, 2). In the first two cases, one can perform on this vertex a reverse standard or atypical vertex addition, while in the last case, it follows from Lemma 2 that either a reverse standard edge splitting or a reverse atypical one can be performed. It is thus always possible to obtain a minimally persistent graph G′ = (V ′, E′) with V ′ = V − 1 by performing an operation of A−1 on G. Doing this recursively, one can obtains after V − 2 operation a minimally persistent graph on two vertices, i.e., a leader-follower seed. The reverse sequence of operations allows then one to obtain G from this seed. Since the operations of A and A−1 preserve minimal persistence, all the intermediate graph of such a sequence are persistent. We have thus proved the following theorem: Theorem 4. Every minimally persistent graph G = (V, E) (V > 1) can be obtained by performing V −2 operations of A on an initial leader-follower seed. Moreover, all the intermediate graphs are minimally persistent. Using the same argument as for Corollary 1, we obtain the following result. Corollary 2. Every minimally persistent graph can be transformed into any other minimally persistent graph using only operations of A ∪ A−1. As in Section 3.4, one can use the set A ∪ A−1 to define a distance on the minimally persistent graphs (on more than one vertex). But, as a consequence of Theorem 4, the distance between the graph G = (V, E) and G′ = (V ′, E′) is never greater than V + V ′ − 4. Remark 3. The non-confined character of the atypical edge splitting makes it more complicated to implement in an autonomous agent formation. It can in- deed involve the direction reversal of a potentially large number of edges that are 22 not involved in the corresponding operation for undirected graph. Proposition 8 states that there always is such a non-confined operation in a set of generalized vertex addition and edge splitting operations which allows one to build all min- imally persistent graphs. However, the example in Figure 12 only requires one edge to be reversed, and no example was found yet where it was necessary to reverse more than one edge. There might thus exist a set of operations having the same properties as A (with respect to the building of all minimally persistent graphs) but in which the non-confined operation only involves the reversal of a number of edges bounded independently of V . Remark 4. It is possible to show that among the four operations of A (resp. A−1), none can be removed without being replaced by some alternative new operation if the operation set is to produce all minimally persistent graphs (resp. contain for each minimally persistent graph an operation that can be performed on it). For each operation of A−1, one can indeed find a graph where none of the three other operations can be performed. However, the set of generalized vertex additions and edge splittings that we present here is just one among the several sets that we have found allowing one to build all minimally persistent graphs. It offers the advantage that its non-confined operation has a more local character than those contained in the other sets (which are not described here). 4.4 Relations between A and T We examine here the relation between the two sets A and T . Let X and Y be two sets of operations. We say that X ≤ Y if all the operations of X can be implemented by a sequence of operations of Y. If X ≤ Y and X ≥ Y, we say that X = Y. If X ≤ Y and X 6= Y, we say that X < Y. One can see that X −1 ≤ Y −1 if and only if X ≤ Y. Lemma 3. An atypical vertex addition can be implemented using one standard vertex addition and one edge reversal. Proof. Let k and j be two distinct vertices in a minimally persistent graph such that k has at least one degree of freedom. Instead of adding a vertex i and the edge (i, j) and (k, i) (atypical vertex addition), one can equivalently add the vertex i and the edges (i, j) and (i, k) (standard vertex addition), and then reverse the edge (i, k) (edge reversal). Note that this edge reversal can be performed because k has a degree of freedom. Lemma 4. An atypical edge splitting can be implemented using one standard edge splitting and one or more edge reversal(s). Proof. Let j, k and l be three vertices of a minimally persistent graph G satis- fying the conditions required to perform an atypical edge splitting (see Section 4.2), and let G′ be the graph obtained by performing an atypical edge splitting on (k, l) in G as represented in Figure 13(b). By performing a standard edge splitting on (k, l) in G such that the added vertex is also connected to j, one 23 obtains a minimally persistent graph having the same underlying undirected graph as G′. It is then a consequence of Proposition 7 that this last graph can be obtained by a sequence of edge reversals. Proposition 12. A < T , and equivalently A−1 < T −1 Proof. Observe first that all the operations of A increase the number of vertices in the graph, while edge reversal does not. Thus edge reversal cannot be imple- mented by a sequences of operations of A, and A 6≥ T . Since the operations of S (standard vertex additions and edge splitting) be- long to both A and T , and since by Lemmas 3 and 4, the operations of A \ S (atypical vertex addition and atypical edge splitting) can be implemented using operations of T , we have A ≤ T , which together with A 6≥ T implies that A < T . Since the set T of operations is more powerful than A, Theorem 4 is a stronger result than Theorem 2. However, if we look at the sets containing both normal and inverse operations, the results are different. Suppose indeed that a graph G′ is obtained by performing an operation of T ∪ T −1 on a minimally persistent graph G. Since both graphs are minimally persistent, Corollary 2 implies that G′ can also be obtained by applying a sequence of operations of A ∪ A−1 on G. Any operation of T ∪ T −1 can thus be implemented by a sequence of operations of A ∪ A−1. Conversely, any operation of A ∪ A−1 can be implemented by a sequence of operations of T ∪ T −1. We have thus shown the following result: Proposition 13. A ∪ A−1 = T ∪ T −1 Both sets A and T allow one to enumerate exhaustively all minimally persis- tent graphs. However, as explained in Section 2.1, since the sequence that can build a certain minimally persistent graph is not unique, this enumeration can allow one to compute an upper bound on the number of minimally persistent graphs having a certain number of vertices, but not their exact number. 5 Conclusions and future work In this paper, we have extended the Henneberg sequence concept to directed graphs. From an autonomous agent point of view, this provides a systematic approach to sequentially obtain or reorganize a minimally persistent agent for- mation. We also exposed some natural restrictions to these extensions, the main one being the impossibility of building all minimally persistent graphs using only confined generalized vertex additions or edge splittings. We proposed two sets of operations, each of which allows one to obtain any minimally persistent graph from a leader-follower seed. The first one (T in Sec- tion 3) contains the two standard vertex additions and edge splittings already introduced in [3] and a purely directed operation (i.e. a neutral operation from 24 an undirected point of view). The second set (A in Section 4) contains, in ad- dition to the two standard operations, two atypical versions of them, among which one is not a confined operation. It involves indeed the reversal of a path of undetermined length in the graph. However, it is still an open question to know if similar results could be obtained using operations involving a number of reversals fixed or bounded independently of the size of the graph. Note that for the second set, the number of operations required to build a minimally per- sistent graphs is uniquely determined by the size of the graph. From an autonomous agent point of view, it would be useful to study how these various operations could be performed in a decentralized way in order to efficiently construct or reorganize a formation. For this purpose, the operations of T could be preferred for their simplicity and because the successive modifica- tions are only local modifications. This study could also imply the development of an optimal algorithm (using one of the two sets of operations) to reorganize a persistent formation. An improvement could come from the use of an even more simple operation such as edge reorientation, which would consist in changing the arrival point of an edge. However, the conditions under which minimal rigidity is preserved by this operation are not known yet. As a final remark, note that we have only focused on the transformations of minimally persistent graph into other minimally persistent graphs. Several practical issues concerning autonomous agent formations arise relating to the merging of such graphs, or their repair after the loss of some vertices or edges. It would thus be worthwhile to study these particular problems as well. References [1] J. Baillieul and A. Suri. Information patterns and hedging brockett's theo- rem in controlling vehicle formations. In Proc. of the 42nd IEEE Conf. on Decision and Control, volume 1, pages 556 -- 563, Hawaii, December 2003. [2] A. Das, J. Spletzer, V. Kumar, and C. Taylor. Ad hoc networks for lo- calization and control. In Proc. of the 41st IEEE Conf. on Decision and Control, volume 3, pages 2978 -- 2983, Las Vegas, NV, December 2002. [3] T. Eren, B.D.O. Anderson, A.S. Morse, and P.N. Belhumeur. Informa- tion structures to secure control of rigid formations with leader-follower structure. In Proc. of the American Control Conference, pages 2966 -- 2971, Portland, Oregon, June 2005. [4] J.M. Hendrickx, B.D.O. Anderson, and V.D. Blondel. Rigidity and persis- tence of directed graphs. In Proceedings of the 44th IEEE Conference on Decision and Control, Seville, Spain, December 2005. [5] J.M. Hendrickx, B.D.O. Anderson, J.-C. Delvenne, and V.D. Blondel. Di- rected graphs for the analysis of rigidity and persistence in autonomous 25 agents systems. To appear in International Journal of Robust and Nonlin- ear Control's special issue on Communicating-Agent Systems. [6] J.M. Hendrickx, B. Fidan, C. Yu, B.D.O. Anderson, and V.D. Blondel. In Rigidity and persistence of three and higher dimensional formations. Proceedings of the First International Workshop on Multi-Agent Robotic Systems (MARS 2005), pages 39 -- 46, Barcelona, Spain, September 2005. [7] L. Henneberg. Die graphische Statik der starren Systeme. Leipzig, 1911. [8] G. Laman. On graphs and rigidity of plane skeletal structures. J. Engrg. Math., 4:331 -- 340, 1970. [9] R. Olfati-Saber and R.M Murray. Graph rigidity and distributed formation stabilization of multi-vehicle systems. In Proceedings of the 41st Confer- ence on Decision and Control, volume 3, pages 2965 -- 2971, Las Vegas, NV, December 2002. [10] T. Tay and W. Whiteley. Generating isostatic frameworks. Structural Topology, (11):21 -- 69, 1985. [11] W. Whiteley. Some matroids from discrete applied geometry. In Matroid theory (Seattle, WA, 1995), volume 197 of Contemp. Math., pages 171 -- 311. Amer. Math. Soc., Providence, RI, 1996. [12] C. Yu, J.M. Hendrickx, B. Fidan, and B.D.O. Anderson. Structural per- sistence of three dimensional autonomous formations. In Proceedings of the First International Workshop on Multi-Agent Robotic Systems (MARS 2005), pages 47 -- 55, Barcelona, Spain, September 2005. [13] C. Yu, J.M. Hendrickx, B. Fidan, B.D.O. Anderson, and V.D. Blondel. Three and higher dimensional autonomous formations: Rigidity, persis- tence and structural persistence. To appear in Automatica. 26
1611.06858	1	1611	2016-11-21T16:00:59	What Do We Elect Committees For? A Voting Committee Model for Multi-Winner Rules	[ "cs.MA", "cs.GT" ]	We present a new model that describes the process of electing a group of representatives (e.g., a parliament) for a group of voters. In this model, called the voting committee model, the elected group of representatives runs a number of ballots to make final decisions regarding various issues. The satisfaction of voters comes from the final decisions made by the elected committee. Our results suggest that depending on a decision system used by the committee to make these final decisions, different multi-winner election rules are most suitable for electing the committee. Furthermore, we show that if we allow not only a committee, but also an election rule used to make final decisions, to depend on the voters' preferences, we can obtain an even better representation of the voters.	cs.MA	cs	What Do We Elect Committees For? A Voting Committee Model for Multi-Winner Rules∗ Piotr Skowron University of Oxford Oxford, UK [email protected] Abstract We present a new model that describes the process of electing a group of representatives (e.g., a parliament) for a group of voters. In this model, called the voting committee model, the elected group of representatives runs a number of ballots to make final decisions regarding various issues. The satisfaction of voters comes from the final decisions made by the elected committee. Our results suggest that depending on a decision system used by the committee to make these final decisions, different multi-winner election rules are most suitable for electing the committee. Furthermore, we show that if we allow not only a committee, but also an election rule used to make final decisions, to depend on the voters' preferences, we can obtain an even better representation of the voters. 1 Introduction There are various scenarios where a group of representatives is selected to make decisions on behalf of a larger population of voters. The examples of such situations include parliamentary elections, elections for supervisory or faculty board, elections for the trade union, etc. In such scenarios, the elected group of representatives, also referred to as a committee, runs a sequence of ballots making decisions regarding various issues. For instance, these can be decisions made by the elected parliament regarding financial economics, national health-care system, retirement age, or changes in the specific law acts. In such case, it is natural to judge the quality of the elected committee based on the quality of its decisions. In this paper we explore this idea and introduce new approach that allows for a normative comparison of various multiwinner election rules. We introduce and study a new formal model, hereinafter referred to as the voting committee model. In this model we assume that an elected committee runs a sequence of independent bal- lots, in which it makes collective decisions regarding a given set of issues. We assume that the ultimate satisfaction of the voters depends solely on the final decisions made by the committee. ∗The preliminary version of this paper was presented at the 24th International Joint Conference on Artificial Intelli- gence (IJCAI-2015) and on the 13th Meeting of the Society for Social Choice and Welfare in 2016 (SSCW-2016). 1 Consequently, each voter ranks candidates for the committee based on how likely it is that they vote according to his or her preferences. The voting committee model allows to numerically assess qual- ities of committees, depending on what election rule (refered to as a decision rule) these committees use to make the final decisions. Intuitively, the numerical quality of a committee S estimates how likely it is that S makes decisions consistent with voters' preferences. There are many works that study scenarios in which a group of representatives is elected to make certain decisions. To the best of our knowledge, however, starting with the famous Condorcet's Jury Theorem [11] most of these works [47, 18, 19, 32] model scenarios where there exists some ground truth and when the decisions made by the elected committee are either objectively correct or wrong. Thus, these models refer to the process of selecting a group of experts. Our approach is different -- there is no ground truth and the voters can differ with their opinions regarding various issues. Informally speaking, a committee is good if it well represent voters' subjective opinions rather than if it can solve certain specific problems having objectively good or bad solutions. Our Contribution In this work we introduce a new formal probabilistic model that relates voters' satisfactions from a committee to their satisfactions from the committee's decisions. In our model we define the notion of optimality of a multi-winner election rule, given that a certain decision rule is used by the elected committee to make the final decisions. We consider two approaches in our probabilistic voting committee model. In the first approach we consider voters which can be correlated with respect to their preferences regarding various issues. We study two specific cases, namely: (i) when voters can be represented as uniformly distributed points on the interval (in such case points can represent, for instance, positions of the corresponding voters in the left-right political spectrum), and (ii) when preferences of voters are taken from publicly available datasets containing distributions of votes in numerous real-life voting scenarios [27]. In this approach we use the voting committee model, and through computer simulations we obtain several interesting conclusions. In particular, we observe that representation-focused multi- winner election rules, such as the Chamberlin -- Courant rule [9], make better decisions with respect to voters' preferences in comparison to the other committee selection rules that we study. Somehow surprisingly, according to our simulations the Chamberlin -- Courant rule is even superior to Pro- portional Approval Voting, a multiwinner extension of the d'Hondt proportional method of appor- tionment. Further, we observe that in our experiments proportional committees make significantly better decisions than a single representative would make. Last, but not least, we argue that decisions made by the elected committee through majority rule are better than those made by the committee through the random dictatorship rule. We also consider the specific case when each voter has a single, primarily important for him or her, issue. In this case the probabilistic model collapses to the deterministic one. Here, we ob- tain theoretical justification for the following claims. We argue that the top-K rules, i.e., scoring rules [46] that select K candidates with the highest total score assigned by voters, are suitable for electing committees that use the random dictatorship rule to make decisions. It might seem that the significance of this observation is compromised by the low applicability of randomized election rules. Such randomized decision-making processes, however, model situations where decisions are 2 made by individual members of the committee but we are uncertain which issues will be considered by which individuals. We also observe that the median OWA rule [41] is suitable for electing com- mittees that make final decisions by majority voting. Informally speaking, in the median OWA rule a voter is satisfied with a committee S if he or she is satisfied with at least half of the members of S. We recall the formal definition of the class of the OWA rules in Section 4. Further, our analysis sug- gests that representation focused multi-winner election rules are particularly well-suited for electing committees that need to make unanimous decisions. Unanimity is often required in situations where making a wrong decision implies severe consequences, e.g., in case of juries voting on convictions. Indeed, in such cases we are particularly willing to ensure that minorities are represented in the committees to avoid biased decisions. In the second approach in our probabilistic model we consider independent voters. In such case we prove the existence of the optimal voting rule and show how to construct one. Discussion of this approach leads to the new very interesting concept -- to the notion of a full multiwinner rule. Informally speaking, a full multiwinner rule is the concept that allows voters to elect not only committees, but also the rules used by these committees to make decisions. We discuss the existence of optimal full multiwinner rules in Section 6.3. Related Work Our research is most closely related to the work of Koriyama et al. [24], who alike assume that the elected committee makes a number of decisions in a sequence of ballots. In such model, they compute the frequency, with which the will of each individual is implemented. The difference is, however, that Koriyama et al. consider the party list setting and the problem of apportionment, i.e., of allocating seats to the parties according to the numbers of votes the parties receive from voters. The authors show that certain assumptions regarding how the society evaluates the frequencies with which the opinions of multiple individuals is implemented, justify degressive proportionality, an interesting principle of proportional apportionment. Similarly to our work, Casella [7, 8] considers the model where a committee meets regularly over time to make a sequence of binary decisions. She defines Storable Votes, a multiple-issue vot- ing system aimed at promoting rights of minorities. The strategic behavior of committee members voting on various issues is also captured in an innovative model by Colonel Blotto games [26]. Our work extends the literature on properties of multi-winner election rules [39, 2, 14, 1, 21, 13]. This literature includes, e.g., the works of Barber´a and Coelho [2], where the authors define prop- erties that "good" multi-winner rules should satisfy. Elkind et al. [14] argue that the desirability of many natural properties of multi-winner rules must be evaluated in the context of their specific ap- plications. Our paper extends this discussion by showing an intuitive model and concrete examples concerning this model, for which different multi-winner rules are particularly well applicable. Similarly to Fishburn [20], we explore the idea of comparing multi-winner election systems. Further, our perspective is conceptually close to the ones given by Christian et al. [10], who study computational problems related to lobbying in direct democracy, where the decisions are made directly by the voters who express their preferences in open referenda (there are no representatives). Many multi-winner election systems are defined as functions selecting such committees that optimize certain metrics, usually related to voters satisfaction [4, 9, 31, 5, 23, 13]. These different 3 optimization metrics capture certain desired properties of election systems. Our paper complements these works by presenting specific metrics that are motivated by the analysis of decision-making processes of committees. Finally, we note that there exists a broad literature on voting on multi-attribute domains, where agents need to make collective decisions regarding a number of issues [6, 25, 44, 45]. We differ from these works by considering an indirect process of decision-making, through an elected committee. 2 Illustrative Example Consider a multiwinner election with six voters, eight candidates, and where the goal is to select a committee of three representatives. Assume that for each voter there exists a single important issue that he or she cares about; an issue can be, for instance, the decision of decreasing the taxes, of decreasing the retirement age, or the decision of making a specific change in the national health care system. The elected committee will vote on each issue and make decision of either accepting or rejecting it; assume that the elected committee will make such decisions by majority voting. Further, assume that the voters know the views of the candidates on the issues and that a voter i approves of a candidate a if and only if i and a have the same view on the i's important issue. Consider example preferences of voters over candidates illustrated in Table 1a. Observe that committee S = {c1, c2, c3} would make decisions consistent with preferences of voter i1. Indeed, for the i1's important issue, the majority of S, namely candidates c1 and c2, have preferences con- sistent with i1. Similarly, for the issues important for voters i2, i3, i4 and i6, committee S would make decisions consistent with the preferences of these voters. Informally speaking, five out of six voters would be satisfied with the decisions made by committee S. Any other committee would make decisions satisfying less voters. Consequently, in this scenario committee S should win the election. Table 1: Example of preferences of the voters over candidates. Values '0' and '1' in the fields denote that corresponding voters approve of or disapprove of the corresponding candidates, respectively. For instance, voter i1 in Table (a) approves of candidates c1, c2, c5, and c8. The same voter in Table (b) approves of candidates c1 and c8. c1 1 1 1 0 1 0 c2 1 1 0 1 0 1 c3 0 1 1 1 0 1 i1 i2 i3 i4 i5 i6 (a) c4 0 1 1 0 1 0 c5 1 1 0 0 1 0 c6 0 1 0 0 1 1 c7 0 0 1 1 0 0 c8 1 0 0 0 0 1 c1 1 0 1 0 1 0 c2 0 1 0 0 0 1 c3 0 0 0 1 0 0 i1 i2 i3 i4 i5 i6 (b) c4 0 1 0 0 0 0 c5 0 0 0 0 1 0 c6 0 0 0 0 0 0 c7 0 0 0 1 0 0 c8 1 0 0 1 0 1 The above arguments can be used to design a voting rule that would be optimal for dichotomous voters' preferences. In such a rule, a voter i approves of a committee S if majority of the members 4 of S are approved by v; the committee approved by most voters is announced as the winner. We will formalize this argument later in Theorem 1. Now, let us change our initial assumptions and consider the case when the elected committee uses random dictatorship rule to make decisions regarding issues. Let us recall that according to the random dictatorship rule, for each issue a single committee member is selected uniformly at random, and the selected member makes a decision on the issue. Observe that in such case, voter i1 is less satisfied with committee S = {c1, c2, c3} than before. Indeed, on i1's important issue, committee S will make decision consistent with i1's preferences only with probability equal to 2/3. The expected number of voters satisfied with the decisions made by committee S is equal to four, and for any other committee the expected number of satisfied voters is even lower. We conclude that in this case it is socially more beneficial if the elected committee uses majority voting rather than random dictatorship rule. The natural question arises whether this is always the case. To answer this question consider preferences of voters from Table 1b. For these preferences, the decisions made by the optimal committee via majority rule will satisfy only two out of six voters. On the other hand, committee S′ = {c1, c2, c8} using random dictatorship would satisfy, in expectation, 22/3 voters. The above observations suggest that it is justified to consider committee selection rules where not only the winning committee, but also the rule used by the elected committee to make decisions, depends on the voters' preferences. In Section 6.3 we formalize this concept, by introducing full multiwinner rules. A full multiwinner rule is a function that given voters' preferences returns a pair: a given-size committee and a randomized decision rule over the set of two alternatives (accept and reject). In the further part of this paper we will explain how to generalize the above reasoning beyond dichotomous preferences of the voters. Briefly speaking, we will assume that the voters do not know how the candidates are going to vote over the issues. Instead, we will assume that each voter has some intuition about how well he or she can be represented by a given candidate; formally, we will assume that each voter v is able to assess the probability with which a given candidate will vote according to v's preferences. This captures, for instance, a common scenario where the issues are not known in advance, thus the voters must use their beliefs to decide on how well given candidates will represent them. 3 The Voting Committee Model In this section we describe a voting committee model that formalizes intuitions given is Section 2, and that allows for a normative comparison of various multi-winner election rules. For each set X, by 1X we denote the indicator function of X, i.e., 1X(x) = 1 if x ∈ X, and 1X(x) = 0 if x /∈ X. For simplicity, we write 1X instead of 1X(x) whenever x is clear from the context. Let N = {1, 2, . . . , n} be a set of voters, and let C = {c1, c2, . . . , cm} be a set of candidates. We assume there is a set D = {D1, D2, . . . , Dr} of r issues; each issue Dj is a binary set consisting of two alternatives Dj = {A, R}. Intuitively, the issue can be either accepted or rejected; A and R corresponds to accepting and rejecting it, respectively. Voters have strict preferences over the alter- natives within each issue; by dj i we denote the preferred alternative from Dj from the perspective of voter i. The issues might differ in their importance to different voters. We consider two types of 5 attitudes: a voter i can consider an issue Dj either as important or as insignificant. We denote the set of all the issues important for voter i as Dim(i) ⊆ D. In the first stage in our model, a committee S is selected through a multi-winner election rule Rmult; the selected committee consists of K members. In the second stage, the committee S runs r independent ballots, for each ballot using the same (randomized) decision rule Rdec. In the i- th ballot, 1 ≤ i ≤ r, the committee makes a collective decision regarding issue Di -- for Di the committee makes either decision A or decision R. The final outcome of this two-stage process is described by a vector of r decisions. Intuitively, the first stage of elections might correspond to e.g., parliamentary elections, elections for supervisory or faculty board, etc. An election in the second stage might be viewed as e.g., a parliamentary ballot on an issue regarding, e.g., financial and monetary economics, education politics, changes in national health-care system, etc. A multiwinner election rule Rmult takes as input a matrix of scores that voters assign to the candidates and returns a subset of K candidates. The scores might be provided directly by the voters or extracted from their ordinal preferences through a positional scoring function (we will discuss this issue in more detail in the subsequent sections). A (randomized) decision rule Rdec is a function Rdec : N → R, that for each natural number a, corresponding to the number of votes casted on A by the committee members, returns the probability that value A is selected. We require that Rdec is symmetric with respect to decisions A and R, i.e., that for each a, 0 ≤ a ≤ K, it holds that Rdec(a) = 1 − Rdec(K − a). The ultimate satisfaction1 of voter i depends solely on the final outcome of the r ballots. Intu- itively, voter i considers committee S as good, if for i's important issues, i.e., issues from Dim(i), S is likely to make decisions consistent with i's preferences. The ultimate satisfaction of a voter i from committee S is measured by value PS,Rdec(i), the probability that S makes a decision consis- tent with i's preferences assuming that S uses rule Rdec to make final decisions. Throughout the paper we will consider several specific variants of the voting committee model which will differ in a way the value PS,Rdec(i) is defined. Let us now define the central notion of this paper that we will use in further analysis. Definition 1. Let Rdec be a decision rule used by the committee to make final decisions, and let K denote the size of the committee to be elected. A committee S is optimal if: S ∈ argmaxS ′⊆C:S ′=KXi PS ′,Rdec(i). (1) In the above definition we take the utilitarian approach -- the optimal rule aims at maximizing the sum of the ultimate satisfaction of the voters. Analogously, we can define committees optimal in the egalitarian sense, by replacing "sum" with "min" in Definition 1, optimizing the ultimate satisfac- tion of the least satisfied voter. For the sake of concreteness, in this paper we focus on the utilitarian case only. Definition 1 implicitly introduces the way of comparing different committees, explored PS ′,Rdec(i). in this paper. A committee S is preferred over a committee S′ ifPi 1We write "ultimate satisfaction" instead of "satisfaction" to distinguish these values, representing utilities that voters get from decisions of a committee from the scores that voters assign to individual candidates, and which quantify the level of appreciation of voters to individuals. PS,Rdec(i) >Pi 6 4 Overview of Election Rules In this section we recall definitions of several known decision and multi-winner election rules that we use in our further analysis. 4.1 Decision Rules We recall that a (randomized) decision rule Rdec is a function Rdec : N → R, that for each natural number a, corresponding to the number of votes casted on A (accept), returns the probability that the value A will be selected. We recall that we require Rdec to be symmetric, that is for each a, 0 ≤ a ≤ K, it must hold that Rdec(a) = 1 − Rdec(K − a). Below we recall the definition of popular decision rules that we will use in our analysis. The uniformly random dictatorship rule selects an alternative d with probability proportional to the number of committee members who vote for d. The majority rule deterministically selects A if at least half of the committee members vote for A (in order to avoid issues related to tie- breaking, we will always use the majority rule for the odd number of votes). Additionally, we consider one rule which is not symmetric with respect to decisions A and R -- the unanimity rule returns deterministically A if and only if all committee members vote for A. 4.2 Multiwinner Election Rules For the description of multiwinner rules we assume that voters express their preferences over can- didates by providing scores: for a voter i and candidate c, by ui,c we denote the score that i assigns to c. Intuitively, ui,c quantifies the level of appreciation of voter i for candidate c. In the description of our results in the subsequent sections we will discuss where these scores come from, and their specific structure. There are two particularly interesting types of scores: Approval scores. We say that scores are approvals if for each voter i and each candidate c, we have ui,c ∈ {0, 1}. We call the setting with approval scores the approval model. In the approval model, we say that a voter i approves of a candidate c if ui,c = 1. Otherwise we say that i disapproves of c. Borda scores. In the model with Borda scores, for each voter i and each candidate c, we have ui,c ∈ {0, . . . , m − 1}, and ui,c 6= ui,c′ for each c 6= c′. In other words, each voter assigns his or her most preferred candidate score of m − 1, to his or her second most preferred candidate score of m − 2, etc. One way in which the Borda scores can be elicitated, is by asking voters for their rankings of candidates, and by applying Borda positional scoring functions to such ordinal preferences (cf. the work of Young [46]). A score profile is a matrix of the scores of all voters over all candidates. We denote the set of all score profiles as U . For each j ∈ N, by Pj(S) we denote the set of all subsets of S of size j. We refer to the elements of PK (C) as to committees. A multi-winner election rule is a function Rmult : U → PK(C) that for a given score profile of voters returns a committee of size K. A significant part of results provided in this paper concerns OWA rules (OWA stands for an ordered weighted average operator). OWA rules in the context of approval model were first 7 mentioned in the 19th century in the early works of Danish astronomer and mathematician Thor- vald N. Thiele [43] and than generalized to arbitrary score profiles by Skowron et al. [41]. Below, we recall the definition of this remarkably general class of rules and describe several concrete examples of OWA rules. For each voter i, each committee S, and each number j, 1 ≤ j ≤ K, let ui,S(j) de- note the score of the j-th most preferred candidate from S, according to i. In other words: {ui,S(1), . . . , ui,S(K)} = {ui,c : c ∈ S}, and ui,S(1) ≥ · · · ≥ ui,S(K). For instance, ui,S(1) is the score that i assigns to his or her most preferred candidate in S, ui,S(2) is the score that i assigns to his or her second most preferred candidate in S, etc. For each voter i, each committee S, and each K-element vector α = hα1, . . . , αK i, we define the α-satisfaction of i from S as the ordered weighted average (OWA) of the scores of the members of S: α(i, S) = αjui,S(j). K Xj=1 The α-rule selects a committee S that maximizes total α-satisfaction of the votersPi α(i, S). Intuitively, the α vector provides a way of aggregating scores of the individual members of the committee to obtain the score of the committee as the whole. Indeed, many known multiwinner rules are in fact OWA rules: Top-K rule and K-Borda rule. If αTopK = h1, 1, . . . , 1i we get a top-K rule that selects K can- didates with the highest total scores. Such rules are referred to as the weakly separable rules [14]. For Borda scores, the top-K rule collapses to the well known K-Borda rule [13]. Chamberlin -- Courant rule. Chamberlin -- Courant rule [9] is an OWA rule defined by the weight vector αCC = h1, 0, . . . , 0i. Informally speaking, according to Chamberlin and Courant a voter cares only about his or her most preferred candidate in the committee; such a most preferred candidate is a representative of the voter in the elected committee. Initially, Cham- berlin and Courant defined their rule for Borda scores. Using Chamberlin -- Courant rule in the approval model was first suggested by Thiele [43]. Elkind et al. [14] considered the Chamberlin -- Courant rule for different types of score profiles, and refer to these types of rules as representation focussed rules [14]. Proportional Approval Voting (PAV). PAV is defined as an OWA rule with the weight vector αPAV = h1, 1/2, . . . , 1/Ki. This rule, developed by Thiele, has been used for a short pe- riod in Sweden during early 1900's. This specific sequence of harmonic weights ensures a certain level of proportionality of the rule [23, 1]; as a matter of fact, PAV is often considered as an extension of the d'Hondt method of apportionment [35] to the case when voters can vote for individual candidates rather than for political parties. k-median rule. The k-median rule is the OWA rule defined by the vector α which has 1 on the k-th positions and 0 on the others. In other words, according to the k-median rule, the satisfaction of a voter from a committee S is his or her satisfaction from the k-th most preferred member of S. 8 The class of OWA-based rules captures many other interesting election systems such as top-K- counting rules [17], separable rules [14] etc. For more discussion on OWA based rules and their applications beyond voting systems we refer the reader to the work of Skowron et al. [41]. Sequential OWA Rules. Unfortunately, for many OWA-based rules, finding the winners is a com- putationally hard problem2. Sequential OWA rules form an appealing, computationally easy, alter- native for OWA rules. Let α denote the vector of K weights. The sequential α-rule proceeds as follows. It starts with an empty solution S = ∅, and in each of the K consecutive steps it adds to S the candidate that increases the α-satisfaction of the voters most. Usually, sequential OWA rules provide a good way of approximating their optimal counterparts [41], sometimes exhibiting even more interesting properties than the original OWA rules [14]. 5 Correlated Voters in the Voting Committee Model In this section we consider voters which are not independent with respect to their preferences over the set of issues. We start our analysis by considering voters and candidates which can be rep- resented as points on the line of preferences. The concept of representing voters and candidates as points in the Euclidean space dates back to 1966 [12, 33], and since than, due to its many natural interpretations, it received a considerable amount of attention in the social choice litera- ture [15, 16, 29, 30, 38]. Let us consider the following illustrative example. Consider the population where one third of voters are left-wing and these voters are represented on the line by point 0, one third are right-wing and represented by point 1, and one third are centrists and represented by point 1/2. There are three issues L, C, and R: the left-wing voters would like L accepted, R rejected, and they do not care about C; the right-wing voters would like R accepted, L rejected, and they consider issue C insignificant; the centrists voters only care about C being accepted. A left-wing candidate will vote for accepting L, C and R with probabilities 1, 1/2, and 0, respectively. A right-wing candidate will vote for accepting these issues with probabilities 0, 1/2, and 1, respectively. A centrist candidate will vote for accepting them with probabilities 1/2, 1, and 1/2, respectively. In particular, left-wing voters are perfectly represented by left-wing candidates and perfectly misrepresented by right-wing candidates. Consider the two committees: S which consists of a left-wing, a centrist, and a right- wing candidate, and Q that consists of three centrist candidates. Proportional committee S Centrist committee Q The probability that committee S makes decision consistent with preferences of left-wing voters is equal to 1/2. The same probability for right-wing voters is also equal to 1/2, and for the centrists voters to 3/4 (the probability that at least one of the two committee members, the left-wing or the 2Computational hardness of the Chamberlin -- Courant rule was first proved by Procaccia et al. [34]. Betzler et al. [3] showed that this problem is also hard from the perspective of parameterized complexity theory. For arguments referring to other weight vectors we refer the reader to to the work of Skowron et al. [41] 9 right-wing, will vote for accepting issue C). The ultimate satisfaction of the voters from committee S is thus equal to n/3 · (1/2 + 1/2 + 3/4). The centrist committee Q, however, will make decision consistent with left-wing, centrist and right-wing voters with probabilities equal to 1/2, 1, and 1/2, respectively. Thus, committee Q results in a higher ultimate satisfaction of the voters. The surpris- ing conclusion of the above example is that the centrist committee, in some situations, can represent the voters better than the proportional one. In particular, this example shows that it is not clear how the best decision-making committee should be aligned on the preference line. Below, we will explore this idea further, we will argue that the described phenomenon is specific to our example, and that in the considered cases the proportional committee is usually, in some sense, superior. 5.1 Uniform Distribution on the Euclidean Line Let us first consider the case where voters and candidates are uniformly distributed on the [0, 1] interval. For this distribution with n = 500 voters and m = 500 candidates, we run the following computer simulations: 1. For each voter v we assumed that there exists one issue which can be represented on the line by the same point as the corresponding voter. 2. For each issue i we uniformly at random selected a value p ∈ [3/2, 5/2]. Intuitively, values of p closer to 3/2 make the issue more likely to be preferred by a majority of voters, thus such an issue can be considered as objectively desired. Values of p closer to 5/2 suggest that the issue is likely to be preferred only by a minority of voters. Specifically, for each voter and for each candidate we randomly selected their preference over the issue. For an issue i and a voter (or a candidate) x we took 1 − p · i − x as the probability of x accepting i, where i − x is the distance between points corresponding to i and x on the line. In particular, each voter always accepts his or her corresponding issue; further, for p = 2 and a centrist issue i, in expectation half of the voters (and half of the candidates) accept i. 3. For each issue we computed decisions made by different committees with respect to such issues, and we assigned to each voter v the satisfaction equal to the fraction of issues for each a given committee made decisions consistent with v's preferences. 4. We repeated such experiment 500 times and for each voter we computed the average ultimate satisfaction. In these simulations we considered five different committees. Three of the considered committees consisted of 51 candidates and were selected by the top-K rule, and by the sequential variants of PAV, and Chamberlin -- Courant rules. To compute committees according to these rules, we assumed that the scores that voters assign to candidates are proportional to one minus the distance between the respective points. For a better intuition, below we depict example committees returned by the three considered multiwinner election rules (for the sake of readability of the diagram, below we depict the case of 10 rather than 51 committee members). Further, we considered one committee that consisted of a single centrist candidate, and one com- mittee that consisted of all voters, i.e., the committee which represents the direct democracy -- the 10 Chamberlin -- Courant PAV Top-K case where all decisions are made in referenda. Finally, we considered two different decision rules that the elected committees used to make final decisions with respect to issues (cf. point 3 above), namely the majority rule and the random dictatorship rule. 0.8 0.6 0.4 0.2 n o i t c a f s i t a s e t a m i t l u 0 0 0.8 0.6 0.4 0.2 n o i t c a f s i t a s e t a m i t l u 0 0 Referendum CC PAV Top-k Single Winner 0.2 0.4 0.6 0.8 1 voters (a) majority Single Winner Top-k PAV CC Referendum 0.2 0.4 0.6 0.8 1 voters (b) random dictatorship Figure 1: The results of the simulations for voters and candidates uniformly distributed on the Euclidean interval [0, 1]. All issues are important. Points on the x-axis correspond to voters. The results of our simulations are depicted in Figure 1. Our conclusions are the following: 1. For the case of majority rule, the committee elected by PAV strictly dominates the single- winner committee. This justifies the decision of selecting a collective body rather than en- trusting the whole power to a single representative. Further, this observation suggest a conjec- ture that an analogous variant of Condorcet's jury theorem can be formulated for the voting committee model. We recall that the main difference is that in our model there is no absolute criterion suggesting that some voter is right or wrong with respect to a given issue. Even for a high value of parameter p (cf. point 2 in the description of simulations), where the issue is relatively unpopular, a voter who is close to such issue will still prefer it to be accepted. 2. The areas below the plots for all four multiwinner committees are almost the same. Yet, the committee selected by the Chamberlin -- Courant rule ensures more fairness to the voters; the two remaining multiwinner rules tend to favor centrist voters. This observation suggests the conjecture that the distribution of the members of the elected committee in a spatial model should resemble the distribution of the voters. 3. Decisions made by the majority rule are in expectation better than decisions made by the random dictatorship rule. Next, we wanted to check how the intensities of voters' preferences are reflected in the voting committee model. Intuitively, a voter has stronger feelings about issues which are either very close 11 to him or her, or which are far away on the line of preferences. A voter cares more about decisions made with respect to such issues in comparison to issues for which he or she is relatively indifferent. We modeled intensities of voters' preferences by introducing insignificant issues to our simulations. For an issue i and voter v we assumed that i is insignificant for v if p·i− x ∈ [2/5, 3/5] -- intuitively, in such case i is insignificant because the probabilities of v willing to accept and reject the issue are relatively similar. The results of simulations with this modification showed almost the same shapes as in Figure 1. These results confirm our previous conclusions. Even more, in this case the area below the plot for the Chamberlin -- Courant committee was by 4% larger than for the top- K committee. This suggests that the Chamberlin -- Courant committee is not only more fair to the voters, but it is also superior to the centrist committee with respect to the total satisfaction of voters. 5.2 Simulations With Datasets Describing Real-Life Preferences As the next step we run simulations for numerous datasets describing people's preferences. To this end we used PrefLib [27], a reference library of preference data. PrefLib contains over 300 datasets describing different scenarios where voters have (weak) preferences over candidates. We filtered out datasets with less than 15 voters and with less than 20 candidates. For each of the remaining datasets we run the following procedure: 1. For each voter v we introduced one issue iv. 2. For an issue iv and each voter v′ we computed the Kendal-Tau distance [22] between rankings v′ and v. We scaled these distances so that the average distance between iv and all voters was equal to 1/2. For each candidate c we defined the distance between iv and c as the position of c in v's preference ranking. We also scaled these distances so that the average distance between iv and all candidates was equal to 1/2. 3. We selected a value p, uniformly at random from the interval [0, 3]. For each issue i and for each voter (respectively, each candidate) x we randomly selected their preference over the issue. We took 1 − p · i − x as the probability of x accepting i, where i − x denotes the scaled distance between the voter (the candidate) and the issue, defined in the previous point. 4. Similarly as in Section 5.1, we computed decisions made by different committees with respect to different issues, and for each voter v we computed the average fraction of issues for each a given committee made decisions consistent with v's preferences. 5. We repeated such experiment 10 times and for each voter we computed the average ultimate satisfaction. To compute the five winning committees, we took scores that voters assign to candidates, which were extracted from their ordinal preferences by using the Borda positional scoring function. The results of these experiments are depicted in Figure 2. Since in this case the voters are not represented by points on the line, in Figure 2 we sorted the voters according to the their ultimate satisfactions. For instance a point on the blue line with x- and y- coordinates equal to 0.8 and 12 0.8 0.6 0.4 0.2 n o i t c a f s i t a s e t a m i t l u 0 0 0.8 0.6 0.4 0.2 n o i t c a f s i t a s e t a m i t l u 0 0 Referendum CC PAV Top-k Single Winner 0.2 0.4 0.6 0.8 1 voters (a) majority Referendum CC Top-k PAV Single Winner 0.2 0.4 0.6 0.8 1 voters (b) random dictatorship Figure 2: The results of the simulations for PrefLib [27]. 0.7 respectively, denotes that 80% of voters had ultimate satisfaction below 0.7, and 20% of voters above 0.7. The results of these simulations confirm our conclusions from Section 5.1. One significant difference is that now, we can objectively claim the Chamberlin -- Courant rule to be superior, not only with respect to the fairness, but also with respect to the total ultimate satisfaction of the voters. 5.3 Single Issue & Deterministic Case: Theoretical Results In this subsection we consider scenarios where each voter has a single important issue that is known in advance, and where voters know preferences of the candidates over the issues. Thus, when electing a committee, it is most natural to consider approval scores of the voters over candidates -- a voter i approves of a candidate c only if the preference of c over the i's important issue is consistent with the preference of the voter. We recall that if a voter i approves of a candidate c, we say that it assigns to c score equal to 1, and denote it by ui,c = 1. Otherwise, we say that he or she assigns to c score of 0, and denote it by ui,c = 0. These scores form an input for the multiwinner election rule Rmult, and thus they can be viewed as the only information that rule Rmult can use to select a committee of K representatives. We recall that for each committee S, and for each voter i, by PS,Rdec(i) we denote the prob- ability that S makes a decision consistent with i's preferences, and that we consider PS,Rdec(i) the ultimate satisfaction of voter i from committee S. Since we consider the model with approval scores, it holds that: PS,Rdec(i) = Rdec(cid:16)Pc∈S ui,c(cid:17). In our subsequent discussion we will show that several known multi-winner election rules can be viewed as optimal in the deterministic variant of our voting committee model. Each such a rule is optimal for different decision system Rdec, used by the selected committee to make final decisions. Thus, our results give intuition regarding the applicability of different multi-winner election rules, depending on for what kind of decision making the committee is elected for. 13 Definition 2. Let Rdec be a decision rule used in the second stage of the election model. A multi- winner election rule Rmult is optimal for Rdec if for each preferences of voters it elects an optimal committee. Below, we explain which multiwinner rules are optimal for majority, random dictatorship, and unanimity decision systems. Theorem 1. Assume K is odd. In the deterministic model: (i) the K+1 optimal for the majority rule, and (ii) the top-K rule is optimal for the random dictatorship rule. 2 -median election system is Proof. We start by considering the case when Rdec is the majority rule. We calculate PS,Rdec(i), the ultimate satisfaction of a voter i from a committee S, assuming S uses majority rule to make final decisions. A committee member c ∈ C votes according to i's preferences if and only if c is approved by i. Thus, a committee S makes decisions consistent with i's preferences, if it contains at least K+1 2 members approved by i. The satisfaction PS,Rdec(i) of i from S is equal to 1 if S contains at least K+1 2 members approved by i, otherwise it is equal to 0. The same formula defines satisfaction of i from S in the K+1 2 -median election system. Fi- 2 -median election system, the committee that maximizes voters' total nally, we note that in the K+1 satisfaction is selected. Let us now move to the case when Rdec is the random dictatorship rule. For a committee S, let apprvi(S) denote the number of candidates from S that are approved of by i. For each committee S, and each member c ∈ S the probability that during the uniformly random dictatorship ballot regarding the issue, the final decision will be made by the committee member c is equal to 1/K. The probability that S will make decision according to i's preferences is, thus, equal to: PS,Rdec(i) =Xc∈S 1 K · 1i approves of c = apprvi(S) K . Consequently, a committee S is optimal if it maximizes the value of the formula Pi apprvi(S). Exactly such committee is elected by the top-K rule. Theorem 1 is quite powerful in a sense that it claims the optimality of the K+1 2 -median and top-K multi-winner rules, for the corresponding decision rules, irrespectively of the preferences of the voters over the issues. Unfortunately, this is not always the case, which is illustrated in the following example. Example 1. Consider an election with two voters, three candidates, and where the goal is to select a committee with two candidates (K = 2). Assume the elected committee will use the unanimity rule to make decisions on issues. Further, assume the deterministic model with two issues, D1 and D2; issue D1 is important only for voter i1 and issue D2 only for voter i2. Consider preferences of the voters and of the candidates over issues given in the left table below. i1 D1 R D2 i2 c2 c1 c3 R R A R A A R i1 D1 A D2 14 i2 c2 c1 c3 A A R A R R A Observe that voter i1 approves of candidates c1 and c2 and that voter i2 approves of candidate c3. In this case the optimal committee would be {c1, c3} or {c2, c3}. Both committees will reject both issues, and so they will satisfy both voters. Now, consider the preferences from the right table. Similarly as before, in this case voter i1 approves of candidates c1 and c2 and voter i2 approves of candidate c3. However, in this case the optimal committee is {c1, c2}. This committee will accept issue D1, hence satisfy one voter. Any other committee will reject both issues satisfying no voter. Example 1 suggests that for the unanimity rule there exists no optimal multiwinner election method. This is an artifact of the fact that unanimity is not symmetric with respect to decisions A and R. For the unanimity rule we can obtain a much weaker claim, yet our claim will also have an interesting interpretation. Before we proceed further, we introduce two new definitions that describe two extreme classes of voters' preferences over the issues. We say that voters are rejection-oriented if for each voter i, di = R, meaning that each voter gets satisfaction only from rejecting issues. Analogously, we say that the voters are acceptance-oriented if for each voter i, di = A. Proposition 1. For the unanimity system in the deterministic model: (i) the Chamberlin -- Courant system with approval votes is optimal for rejection-oriented voters, and (ii) the K-median system with approval votes is optimal for acceptance-oriented voters. Proof. We provide the proof for rejection-oriented voters. The proof for acceptance-oriented voters is analogous. Our reasoning is very similar to the proof of Theorem 1. The ultimate satisfaction of a voter i from a committee S, PS,Rdec(i), is equal to 1 if S contains at least one candidate approved by i, and is equal to 0 otherwise. This is equivalent to the definition of voters' satisfaction in the Chamberlin -- Courant election rule, which completes the proof. Proposition 1 suggest that Chamberlin -- Courant rule and K-median rule are suitable for electing committees that have veto rights. One should choose one of them depending on the voters' satis- faction model. For instance, if passing a wrong decision has much more severe consequences than rejecting a good one (which is captured by modeling the voters as rejection-oriented), a committee should be selected with Chamberlin -- Courant rule, and it should use unanimity rule to make final decisions. This is very often the case when there exists a well established status quo which should be changed only in indisputable cases. On the other hand, if passing a wrong decision has relatively low cost compared to rejecting a good one, the committee should be selected with K-median rule. 6 Independent Voters: Theoretical Analysis In this section we consider voters whose preferences over issues can be viewed as independent random variables. We will show that if the voters are independent, there exist optimal rules, also in the nondeterministic case. In what it follows, we relax two assumptions from Section 5.3. First, we assume that each voter can consider multiple issues as important. Second, we assume that the voters do not know how the candidates are going to vote. Nevertheless, they are provided with some form of intuition which allows them to recognize that some candidates would better represent them in the elected committee than others. To capture this intuition, for each voter i and each candidate c we define the probability of representation of i by c, denoted by pi,c. This is the probability that c will 15 vote according to i's preferences on an i's important issue. We write qi,c = (1 − pi,c). Clearly, value pi,c measures how well i feels represented by c, thus we say that voter i assigns score ui,c = pi,c to candidate c. These scores are given as an input to the multiwinner election rule Rmult. Further, for each voter i, each important for i issue Dj ∈ Dim(i), each committee S, and each committee vote v ∈ DK j , let PS(v) denote the probability that members of S cast vote v: PS(v) = Yc∈S(cid:16)1 v[j]=dj i pi,c + 1 v[j]6=dj i qi,c(cid:17) . Recall that Rdec : N → R is a randomized decision rule used by the committee to make decisions over issues. By PRdec(iv) we denote the probability that the rule Rdec, given vote v, makes decision di j, i.e., decision consistent with i's preferences: PRdec(iv) = 1 = 1 dj dj 1d=A(cid:17) + 1 i =ARdec(cid:16)Xd∈v i =ARdec(cid:16)Xd∈v 1d=A(cid:17) + 1 i(cid:17). d=dj 1 = Rdec(cid:16)Xd∈v dj dj 1d=A(cid:17)(cid:17) i =R(cid:16)1 − Rdec(cid:16)Xd∈v i =R(cid:16)Rdec(cid:16)Xd∈v 1d=R(cid:17)(cid:17) We recall that for each committee S, each voter i, and each important issue Dj ∈ Dim(i), by PS,Rdec(i) we denote the probability that S makes decision consistent with i's preferences. Thus: PS,Rdec(i) = Xv∈DK j PS(v)PRdec(iv). PS,Rdec(i) can be also viewed as the expected fraction of issues important for i, for which the committee C would make decisions consistent with i's preferences. Similarly as in the deterministic model, we will call PS,Rdec(i) the ultimate satisfaction of a voter i from the committee S. Example 2. Consider the illustrative example from Section 5 where one third of voters are left-wing, one third are right-wing, and one third are centrists. Recall that left-wing, right-wing, and centrists candidates can be represented by points 0, 1/2, and 1 on the line, respectively. For this example the probability of representation pi,c is equal to pi,c = 1 − d(i, c), where d(i, c) is the distance between the points corresponding to voter i and candidate c. The approval model is a special case of the probabilistic model where there are only two "al- lowed" values of scores. Thus, in the approval model we assume that there exists two values p, q ∈ [0, 1] with q = p−1, such that for each voter i and each candidate c we have pi,c, qi,c ∈ {p, q}. This model captures scenarios when each voter can view each candidate as belonging to one of the two extreme categories: each candidate can be either "good" or "bad" from i's point of view. This model is appealing because in many real-life scenarios the voters express their preferences by as- signing candidates to one of the two groups: either by approving or by disapproving them. 16 In contrast to the previous section, here we assume that the scores that the voters assign to candidates are scaled: we assume that a voter i assigns score 1 to candidate c if pi,c = p and score 0 if pi,c = q. These scores will be given as an input to the multiwinner election rule Rmult. The fact that we use rescaled scores will not affect our further results, but will enable us to use a more readable notation. 6.1 Optimality of Known Decision Rules under Approval Voting We start our analysis with the approval model, i.e., from the case when the voters have two types of scores only. We start by observing that the characterization of the optimal rule for the random dictatorship method from Theorem 1 can be generalized to the approval model. Proposition 2. The top-K rule is optimal for the random dictatorship rule in the approval model. Proof. Similarly as in the proof of Theorem 1, for a committee S we define apprvi(S) as the number of candidates from S that are approved by i. Let p and q denote the probabilities of repre- sentation of i by candidates approved of and disapproved of by i, respectively. Naturally, p > q. For each committee S, and each member c ∈ S the probability that during the uniformly random dictatorship ballot regarding an issue, the final decision will be made by the committee member c, is equal to 1/K. The probability that S will make decision according to i's preferences is equal to: PS,Rdec(i) =Xc∈S 1 K · (p1i approves of c + q1i disapproves of c) = apprvi(S) · = apprvi(S) · + (K − apprvi(S)) · q K p K p − q K + q. Consequently, a committee S is optimal if it maximizesPi apprvi(S). Exactly such committee is elected by the top-K rule. Below, we show a more general result that characterizes a class of optimal election systems. Theorem 2. For each decision rule Rdec, there exists a K-element vector α, such that α-rule is optimal for Rdec in the approval model. Proof. Let Sℓ,i denote a committee that has exactly ℓ members approved by i, and let P(i, s, ℓ) be the probability that exactly s members of Sℓ,i will vote accordingly to i's preferences. We have: P(i, s, ℓ) = K Xx=1 1x≤ℓ1x≤s1s−x≤K−ℓ ·(cid:18)ℓ x(cid:19)pxqℓ−x(cid:18)K − ℓ s − x(cid:19)qs−xpK−ℓ−s+x. We can see that P(i, s, ℓ) does not depend on i. Further, let PRdec(is) be the probability that the rule Rdec makes decision consistent with i's preferences on issue Dj, assuming s members of Sℓ,i vote accordingly to i's preferences. PRdec(is) = Rdec(s)1 i =A + (1 − Rdec(K − s))1 dj i =R. dj 17 i = A or dj Since either dj i = R and since Rdec(s) = (1 − Rdec(K − s)), we get that PRdec(is) = Rdec(s), and thus PRdec(is) does not depend on i. Consequently, we can calculate the ultimate satisfaction PSℓ,i,Rdec(i) of a voter i from a committee Sℓ,i, so that this value does not depend on i: PSℓ,i,Rdec(i) = K Xs=1 P(i, s, ℓ)PRdec(is). Because PSℓ,i,Rdec(i) does not depend on i, we will denote it as PSℓ,Rdec. Naturally, since p ≥ q, we have PSℓ+1,Rdec ≥ PSℓ,Rdec, for each ℓ. Now, we can see that the following vector: α =DPS1,Rdec, (PS2, Rdec − PS1,Rdec), (PS3,Rdec − PS2,Rdec), . . . , (PSK ,Rdec − PSK−1,Rdec)i satisfies the requirement from the thesis. Indeed, in the α-rule the satisfaction of a voter from a committee with ℓ approved members is the sum of first ℓ coefficients of α, which is PCℓ,Rdec. This completes the proof. The above theorem can be viewed as an evidence of the expressiveness and power of OWA election rules. Unfortunately, OWA election rules are not sufficiently expressive to describe the non-approval model, which we address in the next subsection. 6.2 Optimality of Known Decision Rules under Probabilistic Model Let us move to the probabilistic model in its full generality. To get characterization similar to the one given in Theorem 2, but for arbitrary scores, we would need to consider a more general class of election rules, that is rules in which the satisfaction of a single voter i from a committee C is expressed as a linear combination of products of i' scores as- signed to individuals, i.e., as a linear combination of the values from the set(cid:8)Qc∈C ′ ui,c : C ′ ⊆ C(cid:9) (in contrast to a linear combination of scores only). Such rules that consider inseparable committees were considered e.g., by Ratliff [37, 36]. We formalize this observation in the following theorem. Theorem 3. Let Rdec : N → R be the rule that the elected committee S uses to make final decisions. Let us consider the function v : PK(C) × Πn(C) → R defined in the following way: S v(S, huiii∈N ) =Xi∈N Xj=1 Rdec(j) XSap⊆Pj (S) Yc∈Sap ui,c Yc /∈Sap(cid:0)1 − ui,c(cid:1)!, (2) The rule that for each score profile huiii∈N selects the committee C that maximizes v(C, huiii∈N ) is optimal for Rdec. Proof. The thesis follows from the fact that v(C, huiii∈N ) is the sum of ultimate expected satis- factions over all voters. Indeed, for each possible value of j representing the number of committee 18 members voting according to i's preferences, the expression: XSap⊆Pj (S) Yc∈Sap ui,c Yc /∈Sap(cid:0)1 − ui,c(cid:1)! gives the probability that exactly j committee members will vote according to i's preferences. In the above formula Sap represents the set of j committee members who vote according to i's preferences. Nevertheless, for the case of the uniformly random dictatorship rule used to make final deci- sions, we can get a result similar to the one given in Proposition 2 even for the case of arbitrary scores. Proposition 3. Top-K rule is optimal for the random dictatorship rule in the probabilistic model. Proof. The reasoning is similar to the proof of Proposition 2. The probability that a committee S will vote according to i's preferences is equal to: PS,Rdec(i) = pi,c. 1 K Xc∈S The rule that maximizes the total score of selected candidates, also maximizes their total expected ultimate satisfaction. This completes the proof. Theorem 3 gives us a constructive way of defining an optimal rule in the most general case. Unfortunately, such rule is often hard to compute. We briefly discuss computational issues in Ap- pendix A. 6.3 Optimality of Full Multiwinner Rules In the previous subsection we studied the optimality of multi-winner election rules, given infor- mation on what decision system the committee will use to make final decisions. In this section we show that in our probabilistic model we can compare qualities of the pairs of multi-winner and decision systems. Thus, our results not only suggest which multi-winner election rule is suitable for electing a committee, but also indicates which decision rule should be used by the committee to make final decisions. We start with introducing the definition of the full multiwinner rule. This definition introduces a novel concept to the literature on social choice: it allows the voters to elect not only committees, but also decision rules that these committees use to make final decisions. Definition 3. Let SW be the set of all symmetric randomized decision rules. A full multiwinner rule is a function F : Π → P(A) × SW that for each score profile π ∈ Π returns a pair F(π) = (S, Rdec), where S ∈ P(C) is the elected committee, and Rdec ∈ SW is the decision rule that S will use when making final decisions. We use the notation F(π)[1] = S and F(π)[2] = Rdec. 19 Below we define a partial order on full multiwinner rules. This definition provides a way to compare full multiwinner rules, in particular it allows us to extend the concept of optimality to full multiwinner rules. Definition 4. A full multiwinner rule F weakly dominates a full multiwinner rule G, if for each score profile π ∈ Π(N ), F returns a committee and a decision rule that gives the total expected ultimate satisfaction of the voters at least as high as the one given by a committee and a decision rule returned by G: Xi PF (π)[1],F (π)[2](i) ≥Xi PG(π)[1],G(π)[2](i). (3) F strongly dominates G for P if it weakly dominates G and if there exists profile π ∈ Π for which Inequality (3) is strict. A full multiwinner rule F is optimal if it weakly dominates every other full multiwinner rule. In the previous sections we assumed that the decision rule used by the committee is fixed, and that we use voters' preferences only to determine the winning committee of K candidates. This traditional definition can be mapped to the new model. Definition 5. For a multi-winner rule Rmult and a decision rule Rdec, by Rmult followed by Rdec we call the full multiwinner rule that selects committee using Rmult and, independently of voters' preferences, always uses Rdec to make final decisions. Now, we show how to apply Definition 4 in perhaps the simplest variant, that is in the deter- 2 -median rule ministic model. First, we define a new election rule COMB as a combination of K+1 followed by majority rule with the K-approval rule followed by the random dictatorship rule. Definition 6. The rule COMB is defined as follows. Let S and S′ be committees elected by K- approval rule and by K+1 2 -median rule, respectively. Let apprv be the total approval score of S and let owa be the total OWA score of S′. If apprv K > owa, then COMB returns the pair (C, random dictatorship). Otherwise, COMB returns the pair (C ′, majority). It is remarkable that with this simple idea we obtained the new rule, COMB, that strongly dom- inates both rules that it is derived from. Proposition 4. In the deterministic model, COMB strongly dominates K+1 majority, and K-approval followed by random dictatorship. 2 -median followed by Proof. Let apprv and owa be defined as in Definition 6. Repeating the analysis from the proofs of Theorems 1 and Proposition 2, we get that the total ultimate satisfaction of voters under K+1 2 - median followed by majority is equal to owa, and that the total ultimate satisfaction of voters under K-approval followed by random dictatorship is equal to apprv K . The total ultimate satisfaction of voters and under COMB is equal to max( apprv K , owa). It is easy to see that there exist profiles where apprv K is strictly greater than owa and vice versa. This completes the proof. 20 The natural question is whether we can find an optimal rule in our general probabilistic model. Interestingly, for the deterministic model the answer is positive. The question whether this result can be extended to the general probabilistic model is still open. Theorem 4. In the deterministic model there exists an optimal full multiwinner rule. Proof. The proof is constructive. We recall that a single-winner rule Rdec can be described by K values Rdec(1), . . . , Rdec(K). We use notation from the proof of Theorem 2; let Sℓ,i denote a committee that has exactly ℓ members approved by i, and let P(i, s, ℓ) denote the probability that exactly s members of Sℓ,i will vote accordingly to i's preferences. In the deterministic model P(i, s, ℓ) = 1s=ℓ. Consequently, for a committee Sℓ,i with ℓ ap- proved members, we have PSℓ,i,Rdec = Rdec(ℓ). Let 1ℓ,i denote a function such that 1ℓ,i(S) = 1 if S contains ℓ elements approved by i, and 1ℓ,i(S) = 0 otherwise. For a given committee S we can find an optimal decision rule Rdec by solving the following linear program: maximize Xi∈N PS,Rdec(i) =Xi∈N Rdec(ℓ)1i approves ℓ elements of S subject to: (a) : Rdec(ℓ + 1) ≥ Rdec(ℓ), (b) : 0 ≤ Rdec(ℓ) ≤ 1, (c) : Rdec(ℓ) = 1 − Rdec(K − ℓ), 1 ≤ ℓ ≤ K − 1 1 ≤ ℓ ≤ K 1 ≤ ℓ ≤ K (c) : K Xℓ=1 Rdec(ℓ) = 1. The optimal full multiwinner rule tries all committees and selects such that gives the best so- lution to the integer program. From the solution of the integer program we can extract values Rdec(1), . . . , Rdec(K) that describe the optimal decision rule that should be used to make final decisions. 7 Discusion & Conclusion We defined a new model, called the voting committee model, which explores scenarios where a group of representatives is elected to make decisions on behalf of the voters. This model links utilities of the voters from the elected committee to their utilities from the committee's decisions regarding various matters. Intuitively, a satisfaction of voter i from a committee S is proportional to the probability that for the issues important for i, S's decisions are consistent with i's preferences. Our results give positive support for employing representation-focused multiwinner election rules -- indeed under several assumptions, the decisions made by such committees reflect the pref- erences of the population best. Most importantly, however, our model introduces a new framework for normative comparison of multiwinner voting rules. Indeed, the analysis presented in this paper can be repeated for more specific scenarios, where the relation between voters' preferences over candidates and respective probabilities of representation exhibits some specific structure. 21 We introduced the notion of a full multiwinner rule, in which voters elect a decisive committee along with the decision rule to be used. Under some assumptions, we showed that there exists an optimal full multiwinner rule and we described the way how it can be constructed. We believe that this is an interesting concept that deserves further attention. Among many natural open questions, there is one we consider particularly appealing: is it possible to provide theoretical analysis suggesting which committee is the best for the case of spatial model, such as the one considered in Section 5? Acknowledgements. The author thanks Jean Franc¸ois Laslier, Marcin Dziubi´nski and Piotr Fal- iszewski for their comments and for the fruitful discussion. The author was supported by the European Research Council grant ERC-StG 639945. References [1] H. Aziz, M. Brill, V. Conitzer, E. Elkind, R. Freeman, , and T. Walsh. Justified representa- tion in approval-based committee voting. In Proceedings of the 28th Conference on Artificial Intelligence (AAAI-2015), 2015. [2] S. Barber´a and D. Coelho. How to choose a non-controversial list with k names. Social Choice and Welfare, 31(1):79 -- 96, 2008. [3] N. Betzler, A. Slinko, and J. Uhlmann. On the computation of fully proportional representa- tion. Journal of Artificial Intelligence Research, 47:475 -- 519, 2013. [4] H. Bock, W. H. E. Day, and F. R. McMorris. Consensus rules for committee elections. Math- ematical Social Sciences, 35(3):219 -- 232, 1998. [5] S. J. Brams, D. M. Kilgour, and M. R. Sanver. A minimax procedure for electing committees. Public Choice, 132(3 -- 4):401 -- 420, 2007. [6] S. J. Brams, D. M. Kilgour, and W. S. Zwicker. The paradox of multiple elections. Social Choice and Welfare, 15(2):211 -- 236, 1998. [7] A. Casella. Storable votes. Games and Economic Behavior, 51(2):391 -- 419, 2005. [8] A. Casella. Storable Votes: Protecting the Minority Voice. Oxford University Press, Oxford, UK, 2011. [9] B. Chamberlin and P. Courant. Representative deliberations and representative decisions: Pro- portional representation and the Borda rule. American Political Science Review, 77(3):718 -- 733, 1983. [10] R. Christian, M. Fellows, F. Rosamond, and A. Slinko. On complexity of lobbying in multiple referenda. Review of Economic Design, 11(3):217 -- 224, 2007. [11] Condorcet. Essai sur l'application de l'analyse `a la probabilit´e des d´ecisions rendues `a la pluralit´e des voix. Imprimerie Royale, Paris, 1785. 22 [12] O. A. Davis and M. J. Hinich. A mathematical model of preference formation in a democratic society. In J. Bernd, editor, Mathematical Applications in Political Science II, pages 175 -- 208. Southern Methodist University Press, 1966. [13] B. Debord. An axiomatic characterization of Borda's k-choice function. Social Choice and Welfare, 9(4):337 -- 343, 1992. [14] E. Elkind, P. Faliszewski, P. Skowron, and A. Slinko. Properties of multiwinner voting rules. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS-2014), May 2014. Also presented in FIT-2013. [15] J. M. Enelow and M. J. Hinich. The spatial theory of voting: An introduction. CUP Archive, 1984. [16] J. M. Enelow and M. J. Hinich. Advances in the spatial theory of voting. Cambridge University Press, 1990. [17] P. Faliszewski, P. Skowron, A. Slinko, and N. Talmon. Multiwinner analogues of the plurality rule: Axiomatic and algorithmic views. In Proceedings of the 29th Conference on Artificial Intelligence (AAAI-2016), 2016. [18] T. Feddersen and W. Pesendorfer. Voting behavior and information aggregation in elections with private information. 65:1029 -- 1058, 1997. [19] T. Feddersen and W. Pesendorfer. Convicting the Innocent: The Inferiority of Unanimous Jury Verdicts under Strategic Voting. The American Political Science Review, 92(1):23 -- 35, 1998. [20] P. C. Fishburn. An analysis of simple voting systems for electing committees. SIAM Journal on Applied Mathematics, 41(3):499 -- 502, 1981. [21] W. Gehrlein. The Condorcet criterion and committee selection. Mathematical Social Sciences, 10(3):199 -- 209, 1985. [22] M. G. Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81 -- 93, 1938. [23] D. Kilgour. Approval balloting for multi-winner elections. In J. Laslier and R. Sanver, editors, Handbook on Approval Voting, pages 105 -- 124. Springer, 2010. [24] Y. Koriyama, J. F. Laslier, A. Mac´e, and R. Treibich. Optimal Apportionment. Journal of Political Economy, 121(3):584 -- 608, 2013. [25] D. Lacy and E. M. S. Niou. A problem with referendums. 12(1):5 -- 31, 2000. [26] J. Laslier and N. Picard. Distributive politics and electoral competition. Journal of Economic Theory, 103(1):106 -- 130, 2002. [27] N. Mattei and T. Walsh. Preflib: A library for preferences http: //www.preflib.org. In Proceed- ings of the 3rd International Conference on Algorithmic Decision Theory (ADT-2013), pages 259 -- 270, 2013. 23 [28] N. Mattei and T. Walsh. PrefLib: A library of preference data. In Proceedings of the 3rd International Conference on Algorithmic Decision Theory (ADT-2013), pages 259 -- 270, 2013. [29] R. D. McKelvey, P. C. Ordeshook, et al. A decade of experimental research on spatial models of elections and committees. Advances in the spatial theory of voting, pages 99 -- 144, 1990. [30] S. Merrill and B. Grofman. A unified theory of voting: Directional and proximity spatial models. Cambridge University Press, 1999. [31] B. Monroe. Fully proportional representation. American Political Science Review, 89(4):925 -- 940, 1995. [32] R. B. Myerson. Extended Poisson games and the Condorcet jury theorem. Games and Eco- nomic Behavior, 25:111 -- 131, 1997. [33] C. R. Plott. A notion of equilibrium and its possibility under majority rule. The American Economic Review, 57(4):787 -- 806, 1967. [34] A. Procaccia, J. Rosenschein, and A. Zohar. On the complexity of achieving proportional representation. Social Choice and Welfare, 30(3):353 -- 362, April 2008. [35] F. Pukelsheim. Proportional Representation: Apportionment Methods and Their Applications. Springer, 2014. [36] T. C. Ratliff. Some startling inconsistencies when electing committees. Social Choice and Welfare, 21(3):433 -- 454, 2003. [37] T. C. Ratliff. Selecting committees. Public Choice, 126(3-4):343 -- 355, 2006. [38] N. Schofield. The spatial model of politics. Routledge, 2007. [39] B. Simeone and F. Pukelsheim. Mathematics and Democracy: Recent Advances in Voting Systems and Collective Choice. Studies in Choice and Welfare. Springer Berlin Heidelberg, 2006. [40] P. Skowron and P. Faliszewski. Fully proportional representation with approval ballots: Ap- proximating the maxcover problem with bounded frequencies in FPT time. In Proceedings of the 28th Conference on Artificial Intelligence (AAAI-2015), pages 2124 -- 2130, 2015. [41] P. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From proportional In Proceedings of the 28th Conference on multirepresentation to group recommendation. Artificial Intelligence (AAAI-2015), 2015. [42] P. Skowron, P. Faliszewski, and A. Slinko. Achieving fully proportional representation: Ap- proximability result. Artificial Intelligence, 222:67 -- 103, 2015. [43] T. N. Thiele. Om flerfoldsvalg. Oversigt over det Kongelige Danske Videnskabernes Selskabs Forhandlinger, pages 415 -- 441, 1895. 24 [44] L. Xia, V. Conitzer, and J. Lang. Voting on multiattribute domains with cyclic preferential In Proceedings of the 23rd National Conference on Artificial Intelligence - dependencies. Volume 1, Proceedings of the 23rd National Conference on Artificial Intelligence (AAAI- 2008), pages 202 -- 207, 2008. [45] L. Xia, V. Conitzer, and J. Lang. Aggregating preferences in multi-issue domains by using maximum likelihood estimators. In Proceedings of the 9th International Conference on Au- tonomous Agents and Multiagent Systems (AAMAS-2010), pages 399 -- 406, 2010. [46] H. P. Young. Social choice scoring functions. SIAM Journal on Applied Mathematics, 28(4):824 -- 838, 1975. [47] H. P. Young. Optimal voting rules. The Journal of Economic Perspectives, 9(1):51 -- 64, 1995. A Computational Aspects of Comparing Committees Unfortunately, finding optimal committees is often computationally hard. For example, the problem of finding winners under Chamberlin -- Courant rule is known to be NP-hard [34] and hard from the point of view of parameterized complexity theory [3]. On the positive side, this problem can be effectively solved in practice by using high-quality polynomial-time [42] and fixed-parameter- tractable-time [40] approximation algorithms. Regrettably, not all variants of the problem of finding optimal committees can be well-approximated. For instance, Skowron el al. [41] showed that there exists no polynomial-time constant-approximation algorithm for the problem of finding winners under the K+1 2 -median rule. Even though the problem of finding optimal committees is computationally intractable, we can use our findings to compare committees, even in the case of arbitrary utilities. Indeed, Proposition 5 below shows that deciding which one of the given two committees is better according to our model is solvable in polynomial time. This observation can be used to make decisions about which from the several shortlisted committees should be selected in a particular instance of elections, or to derive more general conclusions about applicability of different multi-winner rules by comparing their qualities on real data describing voters' preferences [28]. Proposition 5. For a decision rule Rdec : N → R, a committee S, and a utility profile huiii∈N , the value of the function v defined in Equation (2) can be computed in polynomial time. Proof. For each voter i ∈ N and each committee S, the expected ultimate satisfaction of i from S can be computed by dynamic programming. Let us sort the members of S in some arbitrary order, so that S = {c1, c2, . . . , cK }. For each j, 0 ≤ j ≤ K and for each ℓ, 0 ≤ ℓ ≤ K, we define the value A[j][ℓ] as the probability that exactly ℓ members from the set {c1, . . . , cj} will vote according to i's preferences. Naturally, A[0][0] = 1, and for ℓ > 0, A[0][ℓ] = 0. To compute the remaining values of the table A we can use the following relation: A[j][ℓ] = ui,cj · A[j − 1][ℓ − 1] + (1 − ui,cj ) · A[j − 1][ℓ]. 25 The value of the expected ultimate satisfaction of i can be obtained by computing a linear com- ℓ=1 Rdec(ℓ) · A[K][ℓ]. This completes the bination of the values A[K][ℓ], i.e., by computing PA proof. 26
1311.6229	1	1311	2013-11-25T08:35:17	Intelligent Agent for Prediction in E- Negotiation: An Approach	[ "cs.MA" ]	With the proliferation of web technologies it becomes more and more important to make the traditional negotiation pricing mechanism automated and intelligent. The behaviour of software agents which negotiate on behalf of humans is determined by their tactics in the form of decision functions. Prediction of partners behaviour in negotiation has been an active research direction in recent years as it will improve the utility gain for the adaptive negotiation agent and also achieve the agreement much quicker or look after much higher benefits. In this paper we review the various negotiation methods and the existing architecture. Although negotiation is practically very complex activity to automate without human intervention we have proposed architecture for predicting the opponents behaviour which will take into consideration various factors which affect the process of negotiation. The basic concept is that the information about negotiators, their individual actions and dynamics can be used by software agents equipped with adaptive capabilities to learn from past negotiations and assist in selecting appropriate negotiation tactics.	cs.MA	cs	Intelligent Agent for Prediction in E- Negotiation: An Approach Mohammad Irfan Bala, Sheetal Vij Department of Computer Engineering Maharashtra Institute of Technoology Pune, India {mirfan508,sheetal.sh}@gmail.com Debajyoti Mukhopadhyay Department of Information Technology Maharashtra Institute of Technology Pune, India [email protected] Abstract— With the proliferation of web technologies it becomes more and more important to make the traditional negotiation pricing mechanism automated and intelligent. The behavior of software agents which negotiate on behalf of humans is determined by their tactics in the form of decision functions. Prediction of partner’s behavior in negotiation has been an active research direction in recent years as it will improve the utility gain for the adaptive negotiation agent and also achieve the agreement much quicker or look after much higher benefits. In this paper we review the various negotiation methods and the existing architecture. Although negotiation is practically very complex activity to automate without human intervention we have proposed architecture for predicting the opponents behavior which will take into consideration various factors which affect the process of negotiation. The basic concept is that the information about negotiators, their individual actions and dynamics can be used by software agents equipped with adaptive capabilities to learn from past negotiations and assist in selecting appropriate negotiation tactics. Keywords— Electronic negotiation, decision functions, agent negotiation, neural networks I. INTRODUCTION Negotiation is a form of interaction in which a group of agents, with conflicting interests and a desire to cooperate try to come to a mutually acceptable agreement on the division of scarce resources. These resources do not only refer to money but also include other parameters like product quality features, guaranty features, way of payment, etc. Electronic negotiations have gained heightened importance due to the advance of the web and e‐commerce. The tremendous successes of online auctions show that the dynamic trade based on e-negotiation will gradually become the core of e-commerce. Whether it is a case of B to B purchase or a case of online shopping [11], it is important to make the traditional negotiation pricing mechanism automated and intelligent. The automation saves human negotiation time and computational agents are sometimes better at finding deals in combinatorally and strategically complex settings. Traditionally e-negotiation processes have been carried out by humans registering at certain web pages, placing bids, making offers and receiving counter offers from other participants. One major disadvantage with this way of e- negotiation is that the knowledge and experience is kept within the human minds [11]. So humans were replaced by negotiation agents in the process of negotiation. However various problems are faced by the negotiation agents such as limited and uncertain knowledge and conflicting preferences. Also agents may have inconsistent deadline and partial overlaps of zones of acceptance [13]. Moreover, multilateral negotiations are more complicated and time consuming than bilateral negotiations. These factors make it difficult to reach consensus. So decision making mechanism is required to overcome this problem. In Figure 1 we show various modes of interactions in a typical market negotiation framework. Fig. 1. Market negotiation framework The need is that the agents should be equipped with a decision making mechanism which allows them to adapt to the behavior of the negotiation partner [3]. Intelligent systems for negotiation aim at increasing the negotiators abilities to understand the opponent‟s needs and limitations. This ability helps to predict the opponent‟s moves which can be a valuable tool in negotiation tasks. Various negotiation strategies have been proposed which are capable of predicting the opponent‟s behavior. The research presented here focuses on the online prediction of the other agent‟s tactic in order to reach better deals in negotiation. While the extensive coverage of all the prediction methods employed in negotiation is beyond the scope of the current work, it is useful to mention several key studies. In this paper we are also proposing a new architecture for prediction of opponent‟s behavior. II. RELATED WORK Predicting the agent‟s behavior and using those prediction results to maximize agent‟s own benefits is one of the crucial issues in the negotiation process. It is necessary for an agent to produce offers based on his own criteria because an agent has limited computational power and incomplete knowledge about opponents. Various approaches [1,2,10,15,16,18] have been proposed for predicting the opponent‟s negotiation behavior. We reviewed some of the approaches to come up with certain conclusions regarding the efficiency of each approach and their short comings. Initially game theory was used in the negotiation process. It treats negotiation as a game and the negotiation agents are treated as players of the game. Zeng and Sycara [9] used game‐theoretic approach with Bayesian belief revision to model a negotiation counterpart. However game theory has two main drawbacks [1] which make it unsuitable for use in the negotiation process. First is that it assumes the agent has infinite computational power and secondly it assumes all the agents have common knowledge. These limitations of the game theory were overcome by the decision functions. The decision functions produce offers based on the amount of time remaining, resource remaining or the opponent‟s behavior. Faratin [18] proposed a bilateral negotiation model in which the two parties negotiate on an issue like price, delivery time, quality etc. The two parties adopt opposite roles (buyer and seller) and use one of the three families of negotiation tactics namely: Time dependant tactics, Resource dependant tactics and behavior dependant tactics. The offers exchanged between the agents are represented as . This is the offer generated by agent „a‟ for agent „b‟ at time „t‟. All the offers are restricted in between mina and maxa which specifies the range of all possible offers of „a‟. Each agent has a scoring function Va which assigns a score to each offer produced. An agent may respond to the offer by any of the three ways: withdraw, accept or offer is the counter offer generated by agent „a‟ in response to the offer is the deadline for of agent „b‟. agent „a‟ by which the negotiation should be complete. Offers generated use one of the three families of tactics [18]. In time dependant tactics time is the predominant factor and each offer generated depends on the amount of time remaining. In resource dependant family of tactics offers depend on how a resource is being consumed. Offers become more and more cooperative as the quantity of the resource diminishes. In behavior dependant family of tactics agent imitates the behavior of the opponent. These tactics differ depending on the behavior of the opponent they imitate and to what degree. Chongming Hou [1] proposed to use non linear regression approach for the prediction of the opponent‟s tactics. It could predict the approximate value of opponent‟s deadline and reservation values. The performance of the agent improved by using this approach as it reduced the number of negotiation breakdowns and caused early termination of unprofitable negotiations. But this approach is restricted for bilateral negotiations only and can be used only when the agent is sure that the opponent is using one of the above mentioned families of tactics for negotiation. E-negotiation can be classified into three types: one-to-one negotiation, one-to-many negotiation and many-to-many negotiation. Hsin Rau, Chao-Wen Chen, Wei-Jung Shiang and Chiuhsiang Joe Lin [6] focused on one-to-many negotiation architecture and integrated two commonly used coordination strategies i.e. Desperate Strategy and Patient Strategy, to develop a new coordination strategy. Fig. 2. One-to-many negotiation In one-to-many negotiation (Figure 2), buyer is represented by a combination of one coordinating agent and multiple sub- buyer agents while each supplier is represented by supplier agent. They may use one of the coordination strategies: Desperate strategy and patient strategy. In desperate strategy agents want to complete negotiation process as early as possible. The negotiation process is terminated as soon as any of the sub-buyer agents is successful in negotiation. In case several proposals are found at the same time, the proposal with highest utility gain is accepted. However in patient strategy all the sub-buyer agents are allowed to complete their negotiation process. The sub-buyer agents finishing early are made to wait for other sub-buyer agents till all complete their negotiation. After all the sub-buyer agents finish negotiation, coordination agent selects the best proposal to make the contract. Hsin Rau, Chao-Wen Chen, Wei-Jung Shiang and Chiuhsiang Joe Lin proposed a new strategy called “adapted coordination strategy” [6] which integrates the advantages of above two strategies. Many other prediction approaches have been proposed which are based on machine learning mechanism. Most of the work devoted to the learning approach is focused on learning from previous offers i.e. offline learning. It includes: Bayesian learning, Q-learning, case-based reasoning and evolutionary computation. They require training data and such agents need to be trained in advance. However this approach may not always work well for the agents whose behavior has been excluded from the training data. Also such data may not be always available. This issue was overcome by Fenghui Ren and Minjie Zhang [5] who proposed three regression functions namely linear, power and quadratic to predict agent‟s behavior. These regression functions use only data about historical offers in the current negotiation thread instead of the using training data which may not always be available. These three regression function are given below and cover most of the negotiation behaviors of the general agents. Linear Regression Function u = b * t + a Power Regression Function U = a * tb Quadratic Regression Function U = a * t2 + b * t + c where u denotes the estimated value for an agent‟s utility, denotes the negotiation time and a, b and c are the parameters which need to be calculated. Parameters a, b and c are independent of t. Brzostowski and Kowalczyk [10] presented a way to estimate partners‟ behaviors by employing a classification method. They used a decision making mechanism which allows agents to mix time-dependant tactics with behavior dependant tactics using weights which can result in quite complex negotiation behavior. However this approach only works for the time dependent agent and the behavior-dependent agent, which limits its application domains. Gal and Pfeffer presented a machine learning approach based on a statistical method [14,17]. The limitation of this approach is the difficulty of training the system perfectly. Therefore, for some unknown kind of agents whose behaviors are excluded in the training data, the prediction result may not reach the acceptable accuracy requirements. I. Roussaki, I. Papaioannou, M. Anagnostou [13] proposed an approach based on learning technique which has been employed by Client Agents and uses a feed-forward back- propagation neural network with a single output linear neuron and three hidden layer‟s neurons. These neural networks require minimal computational and storage resources making it ideal for mobile agents. Also the system does not require information about the previous records. Only information about the current negotiation process is taken into consideration. The agents use a fair relative tit-for-tat negotiation strategy and the results obtained were evaluated via numerous experiments under various conditions. The experiments indicated an average increase of 34% in reaching agreements [13]. This approach has excellent performance when the acceptable interval of the negotiation issue overlaps irrespective of the concession rate. On the other hand if the acceptable intervals‟ overlap is limited and the deadline is quite high, this approach is likely to fail. Also this work was restricted to single issue and bilateral negotiations only. III. PROPOSED ARCHITECTURE We are proposing the architecture of behavior prediction module in the form of web services as depicted in Figure 3. It has already been established in [4] that providing negotiation as a service (NaaS) is a completely innovative application model of software which provides services through internet. Its benefits are , we can obtain stable visiting quantity, user need not concern about maintenance and upgrade of system as it is done on the server independently, saving human and material resources, automated negotiation system can make use of the existing basic facilities provided by e-commerce platform i.e. security, authentication, transaction management etc. ,saving costs of development. Working: The seller will advertise itself through a well known service registration center which will make it visible to all the interested buyers. All the available services at any point of time are stored in the service registration center. A buyer looking for some product will query the service registration center to discover the product of his interest. Once the preferences are matched, buyer and seller will directly communicate with each other and start negotiation. Each buyer and seller has its own module for behavior prediction. Complexity of the behavior prediction module may vary depending on the number of issues taken into consideration. Also the negotiations may be bilateral or multi lateral which will make the prediction process more complicated. Here we have taken seven issues into consideration during prediction: Original price, age, culture, time, quantity, quality and feedback although the number of issues may increase or decrease with corresponding increase or decrease in the complexity of the behavior prediction module. Fig. 3. Proposed Architecture for behavior prediction The behavior prediction logic will use artificial neural networks which have been proved to be universal approximators when provided with sufficient hidden layer neurons and assuming that the activation function is bounded and non-constant. Neural networks also possess the abilities of being self adaptive and self learning. During the first few iterations of negotiation, behavior prediction module will not be used and all the offers and counter offers will be stored in the database and an attempt will be made to try to find the decision function used by the opponent. Later the data stored in the database is used to train the negotiating agent which can predict the opponent‟s offers. Fig. 4. Structure of the artificial neural network The structure of the artificial neural network used in the proposed system is given. All the issues included in the behavior prediction logic are given as input to the system and each issue is assigned some weight depending on its importance. The input for each neuron is not the input variable or the signal coming from previous neuron but it is the signal multiplied by the weight assigned to that issue. Issues with continuous values should be provided discrete values first to make the process of behavior prediction easy. Example: Instead of using the continuous values for „age‟, it should be grouped as youth, middle-aged and old for age group of [10- 25], [25-50], [50+] correspondingly.re you begin to format your paper, first write and save the content as a separate text file. Keep your text and graphic files separate until after the text has been formatted and styled. Do not use hard tabs, and limit use of hard returns to only one return at the end of a paragraph. Do not add any kind of pagination anywhere in the paper. Do not number text heads-the template will do that for you. is for bilateral negotiations. The architecture shown However it can be extended to support multi lateral negotiations where each pair of agents has similar architecture in between them. A. Sample Negotiation Scenario Consider a sample negotiation scenario where the two companies are trying to secure a contrac t between them. Suppose company A wants to sell aircrafts and which company B is considering to purchase. The company A will register itself with the well known registration center and advertise itself making it visible to all buyers. Company B will use the registration center to find the company A which can satisfy its requirements. Once the companies meet they will start communication and will first decide the issues of conflict. Here we consider only three issues: Price, quantity and warranty. Both the companies will be using some strategy for negotiation which is private and is not disclosed during the negotiation process. Each company will assign some weight to these issues such that total of all the weights is 100. Negotiation issue Price Quantity Warranty Company A weights 50 20 30 Company B weights 60 15 25 Table 1. Relative ratings of the two companies Each issue has one or more options, for example, price has three options: $1 million, $1.1 million and $1.2 million. Each option should also be rated like the negotiation issues were rated. Total rating of the options should preferable add up to 100 although not necessary. Similarly quantity may have two options: 3 and 5, and warranty may have 4 options: no warranty, 6 months, 1year and 2 years. Each option of each issue should be rated before starting the negotiation process. Also the rating of the least desired option of each issue should have zero rating making it a threshold value beyond which negotiation will terminate. The agent will generate all possible permutations of the offers and calculate the total profit of each offer in advance. None of the given two companies knows about the preference structure of the other company and at no point are these preferences revealed. Once the assignment of ratings is complete, any of the two companies may start the negotiation process. Whenever a counter offer is received, utility function is used to calculate the concession offered over the previous offer. The utility function is given as: Where n is the number of issues, „rating of offered option of ith issue‟ is the rating of the i th issue in the proposed offer with maximum rating of each issue given on the denominator. The negotiation process continues with several offers and counter offers. Each offer should provide increased profit than the previous offer. Increase in profit after each negotiation round indicates that the negotiation process is converging and the possibility of negotiation process resulting in agreement increases. In case multiple offers are received with decreasing utility function, the agent may decide to end the negotiation process early to reduce the cost of unsuccessful negotiations. Negotiation can also terminate if the opponent crosses the threshold values for any of the issues. IV. CONCLUSION AND FUTURE WORK This work reviews the various methods used for predicting the opponent‟s behavior and then proposes architecture for behavior prediction using artificial neural networks. It proposes the use of database for storing the results and suggests various issues that can be taken into consideration while predicting the opponent‟s behavior. The proposed intelligent agent based architecture is for bilateral negotiations and may be extended in future to multi lateral negotiations. The given architecture is for general use and may not produce optimal results in all situations. So a situation specific architecture is required in every case of negotiation, where the negotiation issues are selected accordingly. In future we would be making the system to simulate above architecture with the application of agent‟s behavior prediction in web based negotiation. We plan to test it vigorously and do the necessary comparative study and analysis with above mentioned related systems. We can also extend our research in behavior prediction of the agents in the direction of multilateral negotiations after successful completion of bilateral system. [13] Roussaki, I.; Papaioannou, I.; Anagnostou, M.,”Employing neural networks to assist negotiating intelligent agents”, 2nd IET International Conference on Intelligent Environments, Volume:1 pp101 -110,2006 [14] Sanghyun Park; Sung-Bong Yang,” An automated system based on Incremental learning with applicability toward multilateral negotiations “SICE-ICASE International Joint Conference, pp: 6001-6006, 2006 [15] Ning Liu; DongXia Zheng; YaoHua Xiong,” Multi-agent negotiation model based on RBF neural network learning mechanism ”, International Symposium on Intelligent Information Technology Application Workshops, pp 133-136, 2008 [16] Hamid Jazayeriy, Masrah Azmi-Murad, Md. Nasir Sulaiman and Nur Izura Udzir,” A review on soft computing techniques in automated negotiation”, Academic journals for Scientific research and essays, Volme 6(24), pp 5100-5106, 2011 [17] Bangqing Li; Yulan Ma,” An auction-based negotiation model in intelligent multi-agent system ”, International Conference on Neural Networks and Brain, volume 1, pp 178-182, 2005 [18] Faratin P,”Automated service negotiation between autonomous compositional agents”, PhD thesis, Queen Mary & Westfield college, University of London,UK, 2000 Eason, B. Noble, and I.N. Sneddon, “On certain integrals of Lipschitz-Hankel type involving products of Bessel functions,” Phil. Trans. Roy. Soc. London, vol. A247, pp. 529 - 551, April 1955. (references) REFERENCES [1] C. Hou, “Predicting agents‟ tactics in automated negotiation,” Proc. IEEE/WIC/ACM Int‟l Conf. Intelligent Agent Technology (IAT ‟04), pp. 127-133, 2004. [2] Beheshti, R.; Mozayani, N. “Predicting opponents offers in multi-agent negotiations using ARTMAP neural network” , Second International Conference on Future Information Technology and Management Engineering, 2009. FITME '09. Pp 600-603, 2009 [3] Réal Carbonneau , Gregory E. Kersten & Rustam Vahidov,” Predicting opponent‟s moves in electronic negotiations using neural networks”, Expert Systems with Applications: An International Journal , Volume 34 Issue 2 ,2008 [4] Debajyoti Mukhopadhyay, Sheetal Vij, Suyog Tasare, "NAAS: Negotiation automation architecture with buyers behavior pattern prediction component", Advances in Intelligent Systems and Computing Volume 176, pp 425-434,2012 [5] Fenghui Ren and Minjie Zhang,"Prediction of partners behaviors in agent negotiation under open and dynamic environments", proceedings of International Conferences on Web Intelligence and Intelligent Agent Technology, pp 379-382, 2007 [6] Hsin Rau, Chao-Wen Chen, Wei-Jung Shiang, Chiuhsiang Joe Lin, “Develop an adapted coordination strategy for negotiation in a buyer - driven E-marketplace”, Proceedings of the Seventh International Conference on Machine Learning and Cybernetics, pp 3224 -3229, 2008 [7] Farhana H. Zulkernine and Patrick Martin,"An adaptive and intelligent SLA negotiation system for web services",IEEE Transactions on service computing, VOL. 4,pp 31-43, 2011 [8] Galit Haim, Sarit Kraus and Yael Blumberg,"Learning human negotiation behavior across cultures",Second International Working Conference on Human Factors and Computational Models in Negotiation, 2010 [9] Zeng, D., & Sycara, K. “Bayesian learning in negotiation”. International Journal of Human‐Computer Studies 48, 125‐141.1998 [10] J. Brzostowski and R. Kowalczyk, “Predicting partner‟s behaviour in agent negotiation,” Proc. Int‟l Joint Conf. Autonomous Agents and Multiagent Systems , pp. 355-361, 2006. [11] Cao Mukun,” Multi-agent automated negotiation as a service”, 7th International Conference on Service Systems and Service Management (ICSSSM), pp 1-6,2010 [12] Raz Lin and Sarit Kraus,"Magazine communications of ACM",Volume 53 Issue 1, January 2010 the
1803.03654	1	1803	2018-03-09T19:00:15	Directing Chemotaxis-Based Spatial Self-Organization via Biased, Random Initial Conditions	[ "cs.MA" ]	Inspired by the chemotaxis interaction of living cells, we have developed an agent-based approach for self-organizing shape formation. Since all our simulations begin with a different uniform random configuration and our agents move stochastically, it has been observed that the self-organization process may form two or more stable final configurations. These differing configurations may be characterized via statistical moments of the agents' locations. In order to direct the agents to robustly form one specific configuration, we generate biased initial conditions whose statistical moments are related to moments of the desired configuration. With this approach, we are able to successfully direct the aggregating swarms to produced a desired macroscopic shape, starting from randomized initial conditions with controlled statistical properties.	cs.MA	cs	March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Directing Chemotaxis-Based Spatial Self-Organization via Biased, Random Initial Conditions 1 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen∗ {spg63,lb353,awm32,david}@cs.drexel.edu +1 215-895-2669 Department of Computer Science Drexel University 3141 Chestnut Street Philadelphia, PA 19104 USA This research was funded by National Science Foundation grants CCF-0636323 and IIS-0845415. 8 1 0 2 r a M 9 ] A M . s c [ 1 v 4 5 6 3 0 . 3 0 8 1 : v i X r a * Corresponding author. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems The International Journal of Parallel, Emergent and Distributed Systems Vol. 00, No. 00, Month 2011, 2–31 RESEARCH ARTICLE Directing Chemotaxis-Based Spatial Self-Organization via Biased, Random Initial Conditions Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen∗ Department of Computer Science, Drexel University Philadelphia, PA 19104 USA (Received 00 Month 200x; in final form 00 Month 200x) Inspired by the chemotaxis interaction of living cells, we have developed an agent-based ap- proach for self-organizing shape formation. Since all our simulations begin with a different uniform random configuration and our agents move stochastically, it has been observed that the self-organization process may form two or more stable final configurations. These differing configurations may be characterized via statistical moments of the agents' locations. In order to direct the agents to robustly form one specific configuration, we generate biased initial conditions whose statistical moments are related to moments of the desired configuration. With this approach, we are able to successfully direct the aggregating swarms to produced a desired macroscopic shape, starting from randomized initial conditions with controlled sta- tistical properties. Keywords: spatial self-organization, directed emergence, genetic programming, agent-based system, statistical moments 1. Introduction Motivated by the ability of cells to form into specific shapes and structures, in pre- vious work we developed chemotaxis-inspired software agents for self-organizing shape formation [1, 2]. The actions of the agents, which we call Morphogenetic Primitives (MPs), are based on the behaviors exhibited by living cells. Cells emit chemicals into the environment. Neighboring cells detect the overall chemical con- centration at their surfaces and respond to the chemical stimulus by moving along the chemical field's gradients [3]. Similarly, in our system the agents emit a virtual chemical, with its concentration defined by an explicit mathematical expression. A set of agents start with an initial random configuration and stochastically follow the gradient of the cumulative concentration field. These chemotaxis-based local interactions direct the agents to self-organize into user-specified shapes (Figure 1). Since the behaviors of MPs are based on local information and interactions, they could provide a distributed, scalable approach for controlling the movements of a robotic swarm. ∗Corresponding author. Email: [email protected], +1 (215) 895-1626 March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 3 In some cases, we have observed though that the agents do not spatially self- organize into a unique shape, but instead form two or more stable final configu- rations. If MPs are used to control the motions of individual robots, it would be extremely useful to direct the outcome of these bifurcating spatial self-organization processes towards a consistent outcome. This would allow us to guarantee that all of our MP aggregations produce a single, desired shape. This is a property that is essential for robust and predictable swarm control algorithms, one that would make the algorithm reliable for engineering applications. Towards this end, we have analyzed whole swarm populations at a global level in search of macroscopic, distin- guishing attributes. This analysis identified features, based on statistical moments of the agents' positions, that have significantly different values for different out- comes of a swarm aggregation. In earlier work we discovered that these statistical moments can be used to accurately predict the outcome of the self-organization process at an early stage of the shape aggregation [4]. Given these differentiating moments, the work described here investigated techniques for directing the out- come of our self-organizing system via biased, random initial conditions in order to consistently produce a desired final configuration. Through our study of the dynamics of a swarm's statistical moments during the aggregation process we noted the connection between initial conditions and the final shape configuration of the swarm. We discovered that biased, random initial conditions that meet specified constraints, i.e. have well-defined statistical properties, robustly yield simulations with a unique final outcome. For those agent interactions that ultimately produce bifurcating/multiple shapes, we have identi- fied for each shape, the most distinguishing macroscopic statistical moment of the evolving swarm. It is possible to generate random distributions of MPs that have specific statistical moments. Our work empirically shows that for bifurcating self- organizing, non-linear, dynamical systems (e.g. a swarm of Morphogenetic Prim- itives) one final outcome can be consistently generated by enforcing a constraint on the value of a single moment when generating the swarm's initial conditions. Given this feature of our system, we are able to control the final outcome of the simulation by simply thresholding the value of a statistical moment for a partic- ular starting distribution, i.e., we constrain the random initial conditions to have specific statistical properties. 2. Related Work Research on distributed agent-based systems that can form spatial patterns and shapes, as well as swarm behaviors, has been conducted for several decades. Reynolds [5] proposed the seminal model for simulating flocking and schooling behaviors based on the local interactions of "boids". Fleischer and Barr [6, 7] ex- plored a cell-based developmental model for self-organizing geometric structures. Theraulaz and Bonabeau [8, 9] presented a modeling approach based on the swarm- ing behavior of social insects. They combined swarm techniques with 3D cellular automata to create autonomous agents that indirectly interact in order to create complex 3D structures. Viscek et al. [10] investigated a particle-based model re- lated to the Reynolds model and found that macroscopic phase changes occurred March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 4 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen in the particle system when introducing noise in the local interactions. Jadbabaie et al. [11] further explored the Vicsek model and provided a theoretical explana- tion for the model's observed behavior, as well as convergence results for classes of switching signals and arbitrary initial heading vectors. The initial work in this area of research created distributed, locally-interacting agent-based systems, then observed and characterized their behaviors. Later work explored techniques for directing these distributed, self-organizing systems. Eggen- berger Hotz [12, 13] proposed the use of genetic regulatory networks coupled with developmental processes for use in artificial evolution and was able to evolve simple shapes. Bonabeau et al. [14] applied genetic algorithms to the stigmergic swarm- based 3D construction method of Theraulaz and Bonabeau in order to evolve interactions that produce user-acceptable structures. Nagpal et al. [15, 16] pre- sented techniques to achieve programmable self-assembly. Cells are identically- programmed units which are randomly distributed and communicate with each other within a local area. In this approach, global-to-local compilation is used to generate the program executed by each cell, which has specialized initial parame- ters. Stoy and Nagpal [17] presented an approach to self-reconfiguration based on directed growth, where the desired configuration (which is stored in each module) is grown from an initial seed module. Spare modules move along recruitment gradi- ents emanating from attached modules to create the final shape. Gradients derived from global potential fields have also been investigated for directing robot swarms. Both Rimon and Koditschek [18] and Hsieh and Kumar [19] demonstrated that robot paths and controls can be computed from these fields, which lead the robots to form a pre-defined shape. Shen et al. [20] proposed a Digital Hormone Model for directing robot swarms to perform such tasks as surrounding a target, covering an area and bypassing barri- ers. The model relies on local communications between identical agents, but it also has each agent move towards a single global target. Swarm chemistry, proposed by Sayama [21, 22] and based on Reynolds' model, is an approach for designing spatio- temporal patterns for kinetically interacting, heterogeneous agents. An interactive evolutionary method, similar to Sims' [23], has been used to define system param- eters that lead to agent segregation and structure formation. Mamei et al. [24] proposed a distributed algorithm for robots that are attracted to and aggregate around targets sensed over short distances. By electing leader(s) as barycenter(s), propagating gradients of varying structure and using these gradients as instruction conditionals, a swarm of simulated robots are able to self-organize into a number of simple shapes such as a circle, ring, and lobes. Von Mammen and Christian [25] de- scribed swarm grammars, an agent-based extension of Lindenmayer systems, that are capable of adapting to their environment and evolve agent parameters in order to create structures that incorporate aspects of developmental design and morpho- genesis. The field of Guided Self-Organization (GSO) [26] has developed techniques for steering self-organizing systems towards desired outcomes, while still attempt- ing to not constrain the system's configuration space during its evolution. Doursat [27, 28] proposed a model for artificial development which combines proliferation, differentiation, self-assembly, pattern formation and genetic regula- tion. Via genetic-like regulation at the agent level, the agents can self-organize into a number of patterned shapes and structures. Werfel et al. [29] proposed a de- March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 5 centralized multi-agent system approach, inspired by mound-building termites, for building user-defined structures. A user specifies a desired structure, and the sys- tem automatically generates low-level rules for independent climbing robots that guarantee production of the structure. A single "seed" brick is used as a land- mark to identify where the structure is going to be built, and defines the origin of a shared coordinate system for the robots. Gerling and Von Mammen [30] pro- vided a context for this type of work in a summary of self-organized approaches to construction. Our previous and latest work on agent-based shape formation stands apart from related work in that it utilizes a chemotaxis-based interaction paradigm inspired by the behaviors of living cells which leads to the formation of tissues. Given the goal of recreating the properties of cells, MPs were designed with principles that should make them scalable and robust [2]. These design principles define MPs as identical, distributed agents that are not directed by a 'master designer', exchange information locally, carry no representation of the shape to be formed, and have no information about their global location. The macroscopic shape of the swarm emerges from the aggregation of local interactions and behaviors. Our approach is therefore novel compared to previous work in that it contains all of the following features. 1) All morphogenetic primitives are randomly placed in the environment, are identical, and perform the same simple actions, unlike Nagpal et al. [15, 16]), Mamei et al. [24], and Doursat [27, 28]. They require no differentiated behaviors or customized initialized states. 2) No initialization of spatial information is needed in the computational environment, unlike Stoy and Nagpal [17], Shen et al. [20], and Werfel et al. [29]. 3) Individual MPs do not know their location in any external/global coordinate system, unlike Stoy and Nagpal [17], and [29]. 4) MPs do not contain or utilize a representation of the predefined global shape that is being composed, unlike Stoy and Nagpal [17], Rimon and Koditschek [18], and Hsieh and Kumar [19]. 5) We utilize genetic programming to discover the MP concentration field functions that lead to the formation of a user-specified shape. Chemotaxis then provides a straightforward mechanism for determining the motion of MPs, in contrast to the difficult-to-program approaches of Shen et al. [20] and Sayama [21, 22]. Note that our new work described here enhances our previously developed system [2] to make MPs more robust, consistent and reliable. 3. Background Material 3.1 Agent-based Shape Formation Like Pfeifer et al. [31] we turn to biology and self-organization for insights into the design of autonomous robots, robotic swarms in our case. Our previous work in self-organizing shape formation [2, 32] is inspired by developmental biology [33] and morphogenesis [34], and builds upon a chemotaxis-based cell aggregation simulation system [35]. Morphogenesis is the process that forms the shape or structure of an organism through cell shape change, movement, attachment, growth and death. We have explored chemotaxis as a paradigm for agent system control because the motions induced by chemotaxis (one of the mechanisms of morphogenesis) may produce patterns, structures or sorting of cells [36]. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 6 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen 3.1.1 Morphogenetic Primitives Morphogenetic Primitives are initially placed inside a 2D environment with a random uniform distribution. Each MP is represented by a small disc and emits a 'chemical' into the environment within a fixed distance relative to its own local coordinate system. Every MP emits the identical local chemical field. An MP de- tects the cumulative chemical field at eight receptors on its surface, and calculates the field gradient from this input. MPs move in the direction of the field gradient with a speed proportional to the magnitude of the gradient. By employing these relatively simple chemotaxis-inspired behaviors MPs are able to self-organize into user-specified macroscopic shapes. This process is schematically presented in the bottom 2/3 of Figure 2. In the middle left of the figure, a close-up of an MP is provided showing its eight chemical sensors and the range of the finite chemical field that it emits. Numerous MPs are randomly placed in the computational arena, and are provided as initial conditions to a chemotaxis-based cell aggregation simulator. The simulator then computes an aggregation simulation based on the one chemical field that is associated with all MPs. The self-organization of the agent swarm is shown in the middle right of the figure. The bottom flowchart of the figure outlines the steps taken by the cell simulator for each cell. The bottom left image shows a representative MP chemical field, with the chemical concentration visualized with gray-scale colors. Isolines are added to highlight the structure of the chemical field. The image to its right shows a cumulative chemical field given the contributions of all of the MPs in the arena. The top part of the figure will be explained later in the paper. 3.1.2 Genetic Programming for Discovering Local Interactions fundamental While MPs' interactions are based on a chemotaxis-inspired paradigm, we do not limit their behaviors/properties to be physically realistic or completely consistent with biology. Instead, developmental biology provides a mo- tivating starting point for MPs. As a way to customize chemotaxis-inspired agents for shape formation, we alter the chemical concentration fields around individual cells. Instead of the chemical concentration dropping off only as an inverse func- tion of distance d from the cell's surface (e.g. 1/d), in our system we define the concentration field with an explicit function of d and θ, the angular location in the cell's local coordinate system. Currently, there is no prescriptive way to specify a particular local field function that will direct MPs to form a specific macroscopic shape, we therefore employ genetic programming [37] to produce the mathematical expression that explicitly specifies the field function. In order to meet the substantial computational re- quirement imposed by our evolutionary computing approach, we have implemented a master-slave form of the distributed genetic programming process [1]. The fit- ness measure associated with each individual field function is based on the shape that emerges from the chemical-field-driven aggregation simulation, and determines which functions will be passed along to later generations. The genetic process stops once an individual (i.e., a mathematical expression) in the population produces the desired shape via a chemotaxis simulation, or after a certain number of generations have been produced and evaluated. Figure 3 illustrates this approach. See [1], [2] and [38] for more details on MPs and the software system that implements them. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 7 With this algorithm, we have successfully evolved local MP chemical field func- tions for a number of simple shapes [2]. These results support the proposition that biological phenomena offer paradigms for designing cellular primitives for self-organizing shape formation. While the resulting explicit chemical fields are not biologically/chemically plausible, they do provide an approach for controlling robot swarms that communicate wirelessly over short distances and share minimal infor- mation with each other. Thus the agents in the swarm do not require significant compute power to self-organize. Additionally, evolutionary computing techniques, specifically genetic programming, have been crucial for discovering the detailed local interactions that lead to the emergence of the swarm's macroscopic structure. However, given the MPs' initial random configurations and the stochastic nature of the self-organization process, the outcomes of the simulations with a specific field function are not always the same. We have found that the shape formation simulations, which include random displacements of the MPs and noise in their movements, can generate bifurcating results. For some field functions, if we run numerous simulations each starting with a different random uniform distribution of MPs, two sets of final configurations will be formed. In most cases an equal num- ber of each configuration are produced, but in a few cases the ratio of the numbers is not one. Since it would be useful to control the outcomes of the self-organization process, we have developed methods for directing the final configuration of a bi- furcating simulation by starting the simulation with biased initial conditions [39]. 3.2 Outcome Prediction The first step towards developing methods that direct the outcome of a swarm simulation involved identifying spatial features that are correlated with and can differentiate the final, different swarm configurations. Our initial effort towards achieving this goal investigated methods for predicting the final configuration of a bifurcating simulation at an early stage of the aggregation process. Our rea- soning was that if certain spatial features can be used to predict the outcome of an aggregation, then they represent unique attributes of the swarm that could be manipulated to direct the swarm. In order to predict the final outcome of a self- organizing shape formation simulation, we first extracted features that capture the spatial distribution of the MPs. Moments provide a quantitative way to describe a distribution. Since MPs are defined as small discs, we use the center of each disc to represent each MP's location. We therefore can simplify the collection of MP locations as a set of 2D points, and apply moment analysis to this set over the duration of the MP simulation. We calculated the mean (first moment), variance (second central moment), skew- ness (third central moment) and kurtosis (fourth central moment) from the x and y coordinates of the MP centers. We analyzed the locations Xi of all points (MPs) as a whole, rather than tracking the location and movement of each individual point. The population size of the agents is denoted as n, (n = 500), and the formulas of March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 8 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen the four moments M1 to M4 are given in Equations 1 to 4, i=1 n(cid:88) n(cid:88) n(cid:88) n(cid:88) i=1 i=1 i=1 M1 = M2 = 1 n 1 n M3 = [ M4 = [ 1 n 1 n Xi, (Xi − M1)2, (Xi − M1)3]/(M2)3/2, (Xi − M1)4]/(M2)2. (1) (2) (3) (4) These statistical moments provide quantitative information about the shape of histograms/distributions. When computed for the x and y coordinates of the MPs, these moments capture the asymmetry and shape of the spatial distribution of the whole population. We have not found it necessary to compute cross moments, with the first four moments providing sufficient information for our analysis. Since the x and y coordinates of the points change over time, so do the four moments of the distribution of the x and y values. The change of the moments as a function of simulation time also provides insight into the dynamic nature of a particular MP simulation. At each simulation time t, the four moments Mi(t) (i = 1 to 4) of the overall distribution are calculated. We then approximate the time derivative of the mo- ments as the slope of a linear interpolating function of consecutive moment values. By calculating the moments and their time derivatives for both the x and y coor- dinates of the point set, at a given time t, we obtain a 16-dimensional vector to represent the distribution, Mx1(t), My1(t), Mx2(t), My2(t), Mx3(t), My3(t), Mx4(t), My4(t), kx1(t), ky1(t), kx2(t), ky2(t), kx3(t), ky3(t), kx4(t), ky4(t). Given the sensitivity of non-linear dynamical systems to initial conditions [40], it makes it extremely difficult, if not impossible, to predict the outcome of our complex, self-organizing system from its initial, random spatial configuration. We therefore attempted to predict the final spatial configuration at an early stage of the aggregation, usually before it is visually evident what shape will emerge from the process. We considered prediction of the bifurcating outcomes as a classification problem and utilized support vector machines (SVMs) [41] to solve it. We have found that applying SVMs to the distribution feature vector at a simulation time that is a small percentage of the total time needed for the final aggregated shape to form produced acceptable results. Given 200 MP simulations for a variety of bifurcating self-organizing shapes we found that we could predict the outcome of the aggregation at a time point 5% to 10% into the simulation with an accuracy of 81% to 91%. We view these results as satisfactory because they demonstrate that a strong correlation between a swarm's moments and its final formed shape does March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 9 exist. More details about this study may be found in [39] and [4]. 4. Directing Spatial Self-Organization Since the outcome of an MP simulation can be predicted at an early stage of aggregation using the moments of the agents' positions, we then explored methods for controlling the swarm via manipulating the moments of the swarm's initial configuration. The general strategy is to create random initial configurations for the MP simulations, but with constrained, biased moments. We have observed that this strategy can consistently direct the swarm to aggregate into specific final configurations. The first step of the strategy analyzes the bifurcating simulations to determine which of the moments diverge the most for the two different outcomes. This is the moment that will be biased in the swarm's initial, random configuration. 4.1 Moment Analysis One of our aggregations, which produces what we call the quarter-moon shape, provides an example of the moment analysis. Of the 200 simulations starting with a uniform random, unbiased initial condition, 100 produce left-pointing structures and 100 produce right-pointing structures. Figure 5 shows a typical swarm aggre- gation for this shape. The four moments of both x and y coordinates are calculated over all 35,000 simulation steps. Additionally the mean and standard deviation of each statistical moment are calculated for the two categories, i.e. left-pointing and right-pointing, over the simulation time steps. Plotting the mean and the mean ± standard deviation of the moments over time immediately highlights the moments which are the most differentiating and may be used to identify specific shapes. For the quarter-moon example the third x moment (skewness) is the one with the greatest separation of values for the two possible outcomes, as seen in Figure 6. The solid and dashed lines are the mean of the skewness of the x coordinate for the two outcomes. The dotted and dot-dashed lines are mean ± standard deviation. The dashed curve is produced from structures that are right-pointing and follow the path in the top of Figure 5. The solid curve is produced from the left-pointing structures, with a typical aggregation presented at the bottom of Figure 5. By analyzing the time series in Figure 6, we see that the skewness of the x co- ordinates of the two classes starts at about the same value, approximately −0.05 to 0.05, at time step 0. The values should be near zero, since all simulations begin with uniform random configurations. The skewness of the two classes first separates by increasing or decreasing, followed by a zero crossing and then a reversed trend appears until they reach their final states at step 35,000. Observing the values for the solid and dashed curves over all simulation time steps, we can identify three regions in the plot: a region occupied by solid/dotted curves only, a region occu- pied by dashed/dot-dashed curves only and an overlapping region. To be specific, considering the values of x skewness in Figure 6 (by projecting the curves into the y axis), the range of [−0.195,−0.150] is covered by solid/dotted curves only; [0.150, 0.190] is covered by dashed/dot-dashed curves only and [−0.150, 0.150] is covered by both outcomes. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 10 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen When determining the appropriate threshold value for a constrained moment, we start with the mean value of the non-overlapping moment range for a particular shape, and then adjust if needed. For the region covered only by the quarter-moon's dashed/dot-dashed curves (for right-pointing shapes) the mean value of skewness is 0.170. While using this value for the moment constraint produced reasonable results (94% of the biased initial conditions produced the desired shape), we found, via multiple experimental runs, that the threshold value on the skewness had to be increased to produce a consistent result. In general this was the process employed for determining the moment constraint thresholds needed to generate the desired outcomes. 4.2 Generating Constrained Biased Distributions Once the most significant distinguishing moment (the one with the greatest value difference in the final configuration) for a shape is identified, the information is utilized to direct the shape aggregation by imposing constraints on this moment in the initial conditions. We assume that the MPs' x and y coordinates are indepen- dent, and therefore create two probability density functions, each representing the x and y coordinate. One probability density function is created for the constrained coordinate and the other coordinate is considered to be uniformly random. Sam- ples are drawn independently from the two distributions to produce a single (x, y) location. This approach generates random distributions, i.e. 2D random initial configura- tions for the shape simulation, that meet a constraint on a particular moment in the x or y coordinate. We have found that constraining one of the eight moments (mean, variance, skewness and kurtosis for x and y) is sufficient for producing satisfactory results. Constraining multiple moments does not significantly improve the outcomes, and would further complicate the process of generating initial con- ditions. Once one significant moment of a distribution and its threshold value have been identified, the remaining three moments for that spatial coordinate (x or y) may be set to values observed in uniform random distributions. These values are mean = 500 (the center of our computational arena), variance = 15, 000, skewness = 0 (to make the distribution symmetric), and kurtosis = 2. Once the four moments for one of the coordinates have been specified, a probability density function (PDF) with those moments is defined. The values for the other coordinate are generated from a uniform random distribution. For the constrained dimension we create a Gram-Charlier expansion of the nor- mal distribution (chosen for its convergence properties) with specified moments [42]. This Gaussian-expanded probability density function is then discretized with 100,000 samples with values falling within the range of 0 to 1000 (the range of the computational arena). Slice sampling [43], a Markov chain sampling method chosen for its efficiency, is then utilized to draw 500 samples from this discretized distribution. Theoretically, the specified PDF can be sampled to produce a distri- bution that has the same moments as the PDF. Our experience has shown that the sample size needs to be quite large (on the order of 1 million) for this to be true. For our sample size of 500 (the number of MPs in an aggregation simulation) and heavily biased PDFs, the resulting moments of the finite sample set do not March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 11 necessarily match the ones desired for a particular shape. As the distribution generation process cannot guarantee that a sampled distribu- tion will meet the required moment constraints, the four moments of the generated distributions for the constrained coordinate are computed to determine if the con- straint is actually met. If the value of the significant moment is not above or below (depending on the constraint to be enforced) the specified threshold, the sampled distribution is rejected. If the constraint is met, the distribution is accepted as the initial conditions for a simulation computation. Since our agents are not points, but in fact are discs with a fixed radius, we maintain a distance of 2R (where R is the radius of a disc) between sample points, to ensure that the MPs do not overlap. Therefore, if an (x, y) pair is generated that is less than 2R distance to a previously generated location, it is rejected and another (x, y) pair is calculated. This process continues until a sufficient number of MP locations are generated for the initial conditions. We have found that it is more computationally efficient to generate numerous smaller sample sets and then merge them into a single point set, rather than at- tempt to compute a single, large point sample, when composing biased initial con- ditions for our aggregation simulations. We apply this approach by drawing 50 subsets, Tk=1,...,50 of 10 samples, rather than 1 set of 500 samples. Each sample drawn is checked for overlap with existing samples, discarded if overlap exists, and otherwise added to the current Tk. Once each Tk has 10 samples, it is checked for conformance to the moment restrictions prior to insertion into the final set, S, of 500 samples. If a given Tk fails, a new Tk is drawn. An acceptable Tk is merged with S, which is then checked for conformance to the moment restrictions. If the updated S does not pass, it is reverted to its prior state, S − Tk, and a new Tk is drawn. This process continues until S = 500. The algorithm for generating biased initial condition, once the initial sampling does not meet the moment requirements, is diagrammed in the flowchart at the top of Figure 2. Via this approach, we are able to generate initial conditions for our computational experiments in a few sec- onds, as opposed to several hours, when attempting to generate all 500 points of S at once for certain "extreme" biased conditions, e.g. low kurtosis. 5. Results We have applied our method for directing spatial self-organizations, which gener- ates biased initial configurations, to a number of bifurcating shape aggregations. We refer to the resulting shapes as the quarter-moon, ellipse, discs, and two parallel line segments. These shapes (and their associated chemical fields) were utilized in an earlier study [4], and had shown not to produce a single final, aggregated result. In our initial MP work it was not uncommon for the output of the evolution pro- cess to generate chemical fields that led to the formation of different shapes from a single field. These four were chosen because they generated two different shapes in equal proportions (except for the parallel lines shape) from uniformly random initial conditions. The chemical field functions that direct MPs to form into these shapes are detailed in Table 1. Given biased initial conditions the aggregation simu- lations produce one final outcome in almost all of cases. Moreover, by thresholding March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 12 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen Shape Field Function quarter-moon 1.0/(ln(ln(ln(exp(sin(ln(cos(θ) ∗ (((d + 0.761214) ∗ ln(d)) + ln(d − θ)))) ∗ (θ − eθ))) − ((θ − d) ∗ (d/θ))) − ((θ − eθ) ∗ ln(ln(exp(sin(ln(cos(θ) ∗ (((d + 0.761214) ∗ ln(d)) + ln(d − θ)))) ∗ (((θ − eθ) − ((θ − eθ) ∗ (d/θ))) + 0.432846))) − ((θ − d) ∗ (d/θ)))))) ellipse 1.0/ cos(θ) + ln(d) discs 1.0/(cos(sin(θ) − (cos(θ) − (ln(−0.367378)/(θ + d)))) + ln(d)) lines 1.0/(ln(ln(ln((cos(cos(d)) + (d ∗ (3d + ln(θ) + ln(ln(θ)))))+ (((ln(θ)/(0.285192)) ∗ ((1.0/θ) + 0.423969))/d)) + d) + d)) Table 1. The field functions for the MPs utilized in this study. Note that exp(x) signifies ex. the moment constraints on the biased initial conditions it is possible to control which shape is produced by a simulation. Figure 4 (top row) shows biased initial conditions for a number of shape aggregations created with this method. The bot- tom row illustrates the final outcome of each MP simulation that is produced from the associated biased starting configuration. In order to identify the significant, distinguishing moments for each shape, ag- gregation simulations (usually several hundred) with unbiased initial conditions are first performed. The shapes of the final outcomes are visually inspected and placed into categories. For each final shape, the mean and standard deviation of the four statistical moments of the evolving system are computed over the entire shape aggregation process. For simulations starting with uniform random, unbiased initial conditions, the final outcomes of the quarter-moon, ellipse and discs shapes evenly split into two categories, with roughly 50% of the final outcomes belonging to each class. The outcomes of the parallel line shapes are unbalanced, with 84.1% belonging to the majority class (two lines) and 15.9% belonging to the minority class (one line). These results, as well as the details that follow in the remainder of the section, are summarized in Table 2. A typical unbiased aggregation for the quarter-moon shape is shown in Figure 5. The simulation reaches a stable state by 35,000 simulation steps. We identified skewness in the x coordinate to be the significant macroscopic feature for this shape. See Figure 6 for the evolution of this feature over the course of the aggregation given uniform random initial conditions. Two sets of biased initial conditions (each with 100 examples) were generated with constrained skewness values in the x coordinate, March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 13 Quarter-moon Shapes Unbiased Percentage left-pointing right-pointing 50% 50% Thresholded Moment Skewnessx ≥ 0.315 Skewnessx ≤ −0.315 Ellipse Shapes Unbiased Percentage single ellipse non-single ellipse two ellipses 50% 50% − Thresholded Moment Kurtosisy ≥ 2.18 Kurtosisy ≤ 1.85 Multiple Discs Shapes three discs no three discs Shapes three discs four discs five or more discs Unbiased Percentage 0% 100% Unbiased Percentage 0% 50% 50% Thresholded Moment V ariancex ≤ 10, 270 V ariancex ≥ 15, 500 Thresholded Moment 1.99 ≤ Kurtosisx ≤ 2.09 Parallel Line Segments Shapes two lines one line Unbiased Percentage 84.1% 15.9% Thresholded Moment Kurtosisx ≤ 1.90 Kurtosisx ≥ 2.29 Biased Percentage 100% 100% Biased Percentage 100% 0% 100% Biased Percentage 100% 100% Biased Percentage 4% 75% 21% Biased Percentage 100% 100% Table 2. Table summarizing results generated with unbiased and biased initial conditions. with the thresholds set as greater than 0.315 (2 standard deviations from the mean) and less than −0.315. Since the distributions are generated stochastically, they do not have the exact targeted skewness value. So our acceptance test is based on a threshold. We performed simulations with the quarter-moon interaction function for these 200 biased initial conditions. Of the 100 initial conditions with a thresholded third x moment below −0.315, 100% of the final outcomes are right- pointing structures. Of the 100 initial conditions with a thresholded third x moment above 0.315, 100% are left-pointing structures. Figure 7 presents the evolution of this feature over the course of the aggregation given the biased initial conditions. A typical unbiased aggregation for the ellipse shape is shown in Figure 8, with half of the unbiased initial conditions producing a single "perfect" ellipse, with the other half producing either two ellipses or a deformed "blob". The unbiased simulation is computed for 10,000 steps. If the simulations are run for 50,000 steps they all will March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 14 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen produce a single ellipse. Kurtosis in the y direction was found to be the significant macroscopic feature for this shape. See Figure 9 for the evolution of this feature over the course of the aggregation given uniform random initial conditions. 100 simulations were performed with initial conditions that had their y coordinates' kurtosis thresholded to be above 2.18. Given these biased initial conditions all simulation (100%) produced a perfect single ellipse by step 7,500. 100 additional simulation were performed with initial conditions that had their y kurtosis set below 1.85. All 100 simulations produced two ellipses by step 7,500. Figure 10 presents the evolution of this feature over the course of the aggregation given the biased initial conditions. These results show that not only can biased initial conditions direct the outcomes of the simulations, but they can also significantly speed up the formation of the desired result, with the single ellipse being guaranteed to form by 50,000 steps in the unbiased case and by 7,500 steps given biased initial conditions. The discs dataset, when run with 200 unbiased initial conditions, produces 100 four discs structures and 100 structures of five or more discs, with a typical shape aggregation shown in Figure 11. The simulation reaches a stable state by 15,000 steps. We identified variance in the x direction to be the significant macroscopic feature for this shape. See Figure 12. In our experiments as the variance of the initial conditions was lowered, we found that we could generate a new shape (one that did not appear with unbiased initial conditions), that contained only three discs. Thresholding the x variance of the initial conditions to be less than 10,270 would always (100%) produce a 3-disc result. A typical 3-disc shape is presented in Figure 4(c). We were unable to consistently generate a 4-disc result for most simulations by thresholding the variance. Thresholding the x kurtosis to be greater than 1.90 and less than 2.09 did lead to an increased number of 4-disc results (75%), which we deemed as less than consistent or robust. Figure 13 presents the evolution of x variance over the course of the 3-disc aggregation given the biased initial conditions. Note that no 3-disc results were produced when keeping the x variance above 15,500 in the biased initial conditions. The parallel line dataset, when run with unbiased initial conditions, contains 526 instances of two vertical parallel line segments (84.1%) and 100 instances of one vertical line (15.9%), as seen in Figure 14. The simulation reaches a stable state by 50,000 steps. We identified kurtosis in the x coordinate to be the significant macroscopic feature for the shape. See Figure 15. 100 simulations were performed with initial conditions that had their x coordinates' kurtosis thresholded to be below 1.90. Given these biased initial conditions all simulation (100%) produced the two line structure. 100 simulations were then performed with initial conditions that had their x coordinates' kurtosis thresholded to be above 2.29. These biased conditions produced results that consistently (100%) created the minority class structure of a single line by 10,000 steps. Figure 16 presents the evolution of x kurtosis over the course of the aggregation given the biased initial conditions. 6. Discussion Our experiments show that our agents (MPs) can be reliably directed to form into large-scale, macroscopic structures using local-only behaviors, based on chem- March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems Parallel, Emergent and Distributed Systems 15 ical diffusion fields and biased initial conditions; thus producing a stigmergic phe- nomenon [44]. This type of global outcome is of particular interest to those looking for a robust, adaptable, and independent self-organizing system. A review con- ducted by the European Space Agency has shown that for their harsh working environment (space, Mars, etc...) the robustness of a local-only system is a key consideration and would allow for relatively simple (and easy to transport) satel- lites/equipment to form into a larger and more complex system that would be impossible to transport as a monolithic structure [45]. Systems using global infor- mation, with centralized communication between primitives, or a command-and- control structure, can form more complex shapes more quickly than local-only approaches. However these systems can fail if this communication is interrupted or the command-and-control structure breaks down [46]. The reasons that certain biased initial conditions may be used to direct the outcome of an aggregation process are frequently visually obvious. For example, the biased initial conditions seen in Figure 4(a) are clearly skewed to the right side of the arena. So it is clear that the majority of the agents are already amassed around the center of the object to be formed. This can also be seen in Figure 4(c), where lowering the variance of the X component makes the initial distribution of the MPs cluster around the central axis which the three discs will form along. The visual evidence that biasing initial conditions effectively pre-starts the aggregation process towards a desired shape is not as evident in the examples that constrain kurtosis ((b) and (d)). A higher Y kurtosis value in example (b) means that there should be a higher concentration of agents along the Y = 500 axis. In (d), a lower X kurtosis value means that there should be more agents distributed away from the center of the arena. But in both of these cases, while statistically this is true, it is not visually evident. It should be noted that not all shapes are stable throughout the simulation. The vertical lines being a good example; up to approximately 5,000 steps the 1-line and 2-lines shapes look very similar and have similar statistical moments, however between 10,000 and 45,000 steps 15.9% of the simulations will converge into a single line. This shows that some shapes may appear to be stable at one point during a simulation, when actually they have not yet stabilized. Or in other words, our agents may produce more than one type of distinct shape during their evolution. Thus, our experiments show that we are able to consistently produce certain self- organized swarms at a specific point in time during their aggregation. For these types of self-organizing systems, it is clearly desired to have a com- pletely local solution. Given that the methods described here require the compu- tation of global system-level quantities (moments of the entire distribution) and the manipulation of the locations of the agents in order to meet some global con- straint, we have not achieved this goal. In the future, we intend to explore if our genetic programming approach to local chemical field evolution, that leads to the formation of macroscopic shapes, can also be used to find chemical fields (i.e. local interactions) that direct a swarm that is uniformly randomly distributed into one that has specific statistical properties. This would lead to a two-step approach that is truly based on local-only interactions. In this case, the agents, which have been uniformly randomly placed in an environment, would follow chemical fields that move them into a biased initial condition. Then they would switch to a field that March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 16 Sean Grimes, Linge Bai, Andrew W.E. McDonald, and David E. Breen robustly directs them to form in a specific macroscopic shape. Additional future work with this system will involve robustness testing of the parameters of the initial conditions. Previous work in this area has defined a lower- bound for the number of primitives required to successfully create a single ellipse (400), but has not defined an upper-bound [47]. An upper and lower bound on the number of primitives (which of course is related to MP density) in the simula- tion is important as more complex shapes are created. Additionally future work will involve further analysis of the aggregation processes that cannot be completely con- trolled. These investigations should reveal new features that differentiate swarms that can be controlled via the reported method and those that cannot. We imagine that manipulation of other features will further enhance our ability to direct our self-organizing system. Ultimately we would like to implement our spatial self-organization approach in a swarm robotics system. We believe that this would be feasible with robots that communicate locally via Bluetooth. Given that all robots emit the same field, a single robot can compute the cumulative chemical field and its gradient at its location simply by knowing the distances and angular relationships of neighboring robots from itself. If each robot has a unique broadcasted ID, we imagine that distance could be derived from signal strength and angular information could be computed from inputs from multiple antennae. We have had discussions with a local roboticist about the possibility of using her swarm robotics platform to investigate our methods for distributed control. 7. Conclusions We have previously developed an agent-based self-organizing shape formation sys- tem. The agents perform identical behaviors based on sensing local information emitted into the environment by the agents. Genetic programming may be used to discover local interaction rules that lead the agents to self-organize into a number of user-specified shapes. However, since the agents are initially uniformly randomly placed in the environment and they stochastically follow prescribed rules, the aggre- gation simulations do not always produce the same final results. In order to develop methods that could be used to direct the agents to robustly form one specific con- figuration, we explored the relationships between an agent swarm's moments and its final configuration. After having shown that these moments could be used to predict the outcome of an MP aggregation in previous work, we demonstrate in this work that biasing the swarm's initial conditions based on these moments can be used to consistently direct the swarm to produce a desired macroscopic shape. By analyzing the statistical moments of the agents' positions over the entire shape aggregation process, we have identified significant, distinguishing moment features, and utilize them as constraints on simulation initial conditions for a number of bifurcating shapes. Biased initial conditions may be generated that meet these moment constraints, which then affect the resulting shape outcomes. In almost all of our examples we can completely control the result of the self-organization process. In some other cases we can significantly increase the likelihood of producing a desired configuration. In a more general sense, our work also indicates that complex, March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 17 non-linear dynamical self-organizing systems may be controlled by manipulating their initial conditions. Disclosure/Conflict-of-Interest Statement The authors declare that the research was conducted in the absence of any com- mercial or financial relationships that could be construed as a potential conflict of interest. Acknowledgments The authors would like to thank Robert Gilmore, Christian Kuehn and Santiago Ontan´on for many helpful discussions and suggestions. This research was funded by National Science Foundation grants CCF-0636323 and IIS-0845415. References [1] L. Bai, M. Eyiyurekli, and D. Breen, Automated Shape Composition Based on Cell Biology and Distributed Genetic Programming, in Proc. Genetic and Evolutionary Computation Conference, 2008, pp. 1179–1186. [2] L. Bai and D. Breen, Chemotaxis-inspired cellular primitives for self- in Morphogenetic Engineering: Toward Pro- organizing shape formation, grammable Complex Systems, chap. 9, Springer, 2012, pp. 209–237. [3] E. Eisenbach, et al., Chemotaxis, Imperial College Press, London, 2004. [4] L. Bai, R. Gilmore, and D. Breen, Predicting Spatial Self-Organization with Statistical Moments, in Proc. Spatial Computing Workshop of the AAMAS Conference, 2014, p. Article 2. [5] C. Reynolds, Flocks, herds and schools: A distributed behavioral model, in Proc. SIGGRAPH '87, 1987, pp. 25–34. [6] K. Fleischer and A. Barr, A Simulation Testbed for the Study of Multicellular Development: The Multiple Mechanisms of Morphogenesis, in Proc. Artificial Life III, 1994, pp. 389–408. [7] K. Fleischer, Investigations with a Multicellular Developmental Model, in Proc. Artificial Life V, 1996, pp. 389–408. [8] G. Theraulaz and E. Bonabeau, Coordination in distributed building, Nature 269 (1995), pp. 686–688. [9] G. Theraulaz and E. Bonabeau, Modeling the collective building of complex ar- chitectures in social insects with lattice swarms, Journal of Theoretical Biology 177 (1995), pp. 381–400. [10] T. Vicsek, A.C. E. Ben-Jacobabd , I. Cohen, and O. Shochet, Novel type of phase transition in a system of self-driven particles, Physical Review Letters 75 (1995), pp. 1226–1229. [11] A. Jadbabaie, J. Lin, and A. Morse, Coordination of groups of mobile au- tonomous agents using nearest neighbor rules, IEEE Trans. on Automatic Control 48 (2003), pp. 988–1001. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 18 REFERENCES [12] P. Eggenberger, Evolving Morphologies of Simulated 3D Organisms Based on Differential Gene Expression, in Proc. 4th European Conference on Artificial Life, 1997, pp. 205–213. [13] P. Hotz, Combining developmental processes and their physics in an artificial evolutionary system to evolve shapes, in On Growth, Form and Computers, K. S. and P. Bentley, eds., Academic Press, 2003, pp. 302–318. [14] E. Bonabeau, S. Guerin, D. Snyers, P. Kuntz, and G. Theraulaz, Three- dimensional architectures grown by simple 'stigmergic' agents, BioSystems 56 (2000), pp. 13–32. [15] R. Nagpal, Programmable self-assembly using biologically-inspired multiagent control, in Proc. 1st International Joint Conference on Autonomous Agents and Multiagent Systems: part 1, 2002, pp. 418–425. [16] R. Nagpal, A. Kondacs, and C. Chang, Programming Methodology for Biologically-Inspired Self-Assembling Systems, in Proc. AAAI Spring Sympo- sium on Computational Synthesis, 2003, pp. 173–180. [17] K. Stoy and R. Nagpal, Self-reconfiguration using directed growth, in Proc. Internatioanl Symposium on Distributed Autonomous Robotic Systems, 2004, pp. 3–12. [18] E. Rimon and D. Koditschek, Exact robot navigation using artificial potential functions, IEEE Trans. on Robotics and Automation 8 (1992), pp. 501–518. [19] M. Hsieh and V. Kumar, Pattern generation with multiple robots, in Proc. IEEE International Conference on Robotics and Automation, Orlando, Florida, 2006, pp. 2442–2447. [20] W.M. Shen, P. Will, A. Galstyan, and C.M. Chuong, Hormone-inspired self- organization and distributed control of robotic swarms, Autonomous Robots 17 (2004), pp. 93–105. [21] H. Sayama, Swarm chemistry, Artificial Life 15 (2009), pp. 105–114. [22] H. Sayama, Robust morphogenesis of robotic swarms, IEEE Computational Intelligence 5 (2010), pp. 43–49. [23] K. Sims, Interactive Evolution of Dynamical Systems, in Proc. 1st European Conference on Artificial Life, 1992, pp. 171–178. [24] M. Mamei, M. Vasirani, and F. Zambonelli, Experiments of morphogenesis in swarms of simple mobile robots, Applied Artificial Intelligence 18 (2004), pp. 903–919. [25] S. Von Mammen and C. Jacob, The evolution of swarm grammars-growing trees, crafting art, and bottom-up design, IEEE Computational Intelligence Magazine 4 (2009), pp. 10–19. [26] M. Prokopenko (ed.), Guided Self-Organization: Inception, Springer, Berlin, 2014. [27] R. Doursat, The growing canvas of biological development: Multiscale pattern generation on an expanding lattice of gene regulatory nets, in Unifying Themes in Complex Systems, Springer Verlag, 2008, pp. 205–210. [28] R. Doursat, Organically grown architectures: Creating decentralized, au- tonomous systems by embryomorphic engineering, in Organic Computing, R. Wurtz, ed., chap. 8, Springer Verlag, 2008, pp. 167–200. [29] J. Werfel, K. Petersen, and R. Nagpal, Designing collective behavior in a termite-inspired robot construction team, Science 343 (2014), pp. 754–758. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 19 [30] V. Gerling and S. Von Mammen, Robotics for Self-Organised Construction, in Foundations and Applications of Self* Systems, IEEE International Workshop on, 2016, pp. 162–167. [31] R. Pfeifer, M. Lungarella, and F. Iida, Self-organization, embodiment, and biologically inspired robotics, Science 318 (2007), pp. 1088–1093. [32] L. Bai, M. Eyiyurekli, P. Lelkes, and D. Breen, Self-organized sorting of het- erotypic agents via a chemotaxis paradigm, Science of Computer Programming 78 (2013), pp. 594–611. [33] S. Gilbert, Developmental Biology, 10th ed., Sinauer Associates, Inc., Sunder- land, MA, 2013. [34] J. Davies, Mechanisms of Morphogenesis: The Creation of Biological Form, Elsevier, Amsterdam, 2005. [35] M. Eyiyurekli, P. Manley, P. Lelkes, and D. Breen, A computational model of chemotaxis-based cell aggregation, BioSystems 93 (2008), pp. 226–239. [36] T. Sekimura, S. Noji, N. Ueno, and P. Maini (eds.), Morphogenesis and Pattern Formation in Biological Systems, Springer, Tokyo, 2003. [37] J. Koza, Genetic Programming: On the Programming of Computers by Means of Natural Selection, MIT Press, Cambridge, MA, 1992. [38] L. Bai, Self-Organizing Primitives for Automated 2D Shape Composition, , Master's thesis, Drexel University, Philadelphia, PA (2008). [39] L. Bai, Chemotaxis-based Spatial Self-Organization Algorithms, Ph.D. thesis, Drexel University, Philadelphia, PA, 2014. [40] S. Wiggins (ed.), Introduction to Applied Nonlinear Dynamical Systems and Chaos, 2nd Edition, Springer, New York, 2003. [41] N. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector Ma- chines and Other Kernel-Based Learning Methods, Cambridge University Press, 2000. [42] S. Seabold and J. Perktold, Statsmodels: econometric and statistical modeling with python, in Proc. 9th Python in Science Conference, 2010, pp. 57–61. [43] R.M. Neal, Slice sampling, Annals of Statistics (2003), pp. 705–741. [44] G. Theraulaz and E. Bonabeau, A brief history of stigmergy, Artificial Life 5 (1999), pp. 97–116. [45] Swarm intelligence (2017), Retrieved October 23, 2017, http://www.esa.int/gsp/ACT/ai/projects/swarm.html. [46] L. Pettazzi, D. Izzo, and S. Theil, Swarm navigation and reconfiguration using electrostatic forces, in Proc. 7th International Conference On Dynamics and Control of Systems and Structures in Space, 2006, pp. 257–268. [47] L. Bai, M. Eyiyurekli, and D. Breen, An Emergent System for Self-Aligning and Self-Organizing Shape Primitives, in Proc. IEEE International Conference on Self-Adaptive and Self-Organizing Systems Conference, 2008, pp. 445–454. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 20 Figures REFERENCES Figure 1. Morphogenetic Primitives self-organizing into a star shape. Initially published in [2]. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 21 Figure 2. Schematic diagram of the directed self-organization process based on specifying biased initial conditions for Morphogenetic Primitives. r2r3r4r6r5r8r7RMaxrcr1Emit ChemicalSense Cumulative Chemical FieldCalculate Gradient and Step Move (Take Step)Collision?Take Random StepYesNo CellSimulatorStep 0Step 1,000Step 5,000Step 10,000Select ShapeGenerate biasedPDFGenerate 10pointsPointsokay?Add toexistingAllpointsokay?Removelatest 10500points?DoneYesNoYesYesNoNoMarch 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 22 REFERENCES Figure 3. The genetic programming process that produces the local chemical field functions of the shape primitives. Initially published in [1]. March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 23 (a) (b) (c) (d) Figure 4. (top row) Biased initial conditions ((a) x skewness = -0.315, (b) y kurtosis = 2.150, (c) x variance = 9,596, (d) x kurtosis = 1.88) that robustly evolve (bottom row) into (a) a right-pointing quarter-moon, (b) a single ellipse, (c) three discs and (d) two line segments. BaiandBreenDirectingChemotaxis-BasedSelf-Organization(a)(b)(c)(d)Figure11.(toprow)Biasedinitialconditions((a)xskewness=-0.268,(b)ykurtosis=2.150,(c)xvariance=9,596,(d)xkurtosis=1.88)thatrobustlyevolve(bottomrow)into(a)aright-facingquarter-moon,(b)asingleellipse,(c)threediscsand(d)twolinesegments.Thisisaprovisionalfile,notthefinaltypesetarticle20March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 24 REFERENCES Figure 5. Shape aggregation of the quarter-moon MPs starting from random initial conditions. Step 0Step 5,000Step 2,500Step 10,000Step 35,00050%50%March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 25 Figure 6. Skewness of the x coordinate of the unbiased quarter-moon shape aggregations over time. Figure 7. Skewness of the x coordinate of the biased quarter-moon shape aggregations over time. 0500010000150002000025000300003500040000Simulation Time0.200.150.100.050.000.050.100.150.20Moment Rangesolid = left, dashed = right skewness0500010000150002000025000300003500040000Simulation Time0.40.20.00.20.4Moment Rangesolid = left, dashed = right skewnessMarch 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 26 REFERENCES Figure 8. Shape aggregation of the ellipse MPs starting from random initial conditions. Step 0Step 1,250Step 250Step 2,500Step 10,00050%50%March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 27 Figure 9. Kurtosis of the y coordinate of the unbiased ellipse shape aggregations over time. Figure 10. Kurtosis of the y coordinate of the biased ellipse shape aggregations over time. 0200040006000800010000Simulation Time1.41.61.82.02.2Moment Rangesolid = one ellipse, dashed = two ellipses kurtosis010002000300040005000600070008000Simulation Time1.52.02.53.0Moment Rangesolid = one ellipse, dashed = two ellipses kurtosisMarch 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 28 REFERENCES Figure 11. Shape aggregation of the discs MPs starting with random initial conditions. Step 0Step 1,500Step 500Step 6,000Step 15,00050%50%March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 29 Figure 12. Variance of the x coordinate of the unbiased discs shape aggregations over time. Figure 13. Variance of the x coordinate of the biased discs shape aggregations over time. 02000400060008000100001200014000Simulation Time90001000011000120001300014000150001600017000Moment Rangesolid = four discs, dashed = not four variance02000400060008000100001200014000Simulation Time200040006000800010000120001400016000Moment Rangesolid = three, dashed = not three varianceMarch 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems 30 REFERENCES Figure 14. Shape aggregation of the line segment MPs starting from random initial conditions. Step 0Step 25,000Step 1,000Step 40,000Step 50,00084.1%15.9%March 13, 2018 IJPEDS17arXiv 0:48 The International Journal of Parallel, Emergent and Distributed Systems REFERENCES 31 Figure 15. Kurtosis of the x coordinate of the unbiased line segment shape aggregations over time. Figure 16. Kurtosis of the x coordinate of the biased line segment shape aggregations over time. 01000020000300004000050000Simulation Time1.001.251.501.752.002.252.502.75Moment Rangesolid = one, dashed = two kurtosis0200040006000800010000Simulation Time20246810Moment Rangesolid = one, dashed = two kurtosis
1907.11852	1	1907	2019-07-27T04:50:45	G-flocking: Flocking Model Optimization based on Genetic Framework	[ "cs.MA", "cs.RO" ]	Flocking model has been widely used to control robotic swarm. However, with the increasing scalability, there exist complex conflicts for robotic swarm in autonomous navigation, brought by internal pattern maintenance, external environment changes, and target area orientation, which results in poor stability and adaptability. Hence, optimizing the flocking model for robotic swarm in autonomous navigation is an important and meaningful research domain.	cs.MA	cs	XXXX, VOL. XX, NO. XX, XXXX 2019 1 G-flocking: Flocking Model Optimization based on Genetic Framework Li Ma, Weidong Bao, Xiaomin Zhu, Member, IEEE, Meng Wu, Yuan Wang, Yunxiang Ling, and Wen Zhou an obstacle populated planar workspace under a single leader- multiple followers architecture. Abstract -- Flocking model has been widely used to con- trol robotic swarm. However, with the increasing scala- bility, there exist complex conflicts for robotic swarm in autonomous navigation, brought by internal pattern main- tenance, external environment changes, and target area orientation, which results in poor stability and adaptability. Hence, optimizing the flocking model for robotic swarm in autonomous navigation is an important and meaningful research domain. Index Terms -- Robotic swarms, flocking model, multia- gent systems (MASs). I. INTRODUCTION R OBOTIC swarms pose an attractive and scalable solution to accomplish complicated missions such as search-and- rescue [1], mapping [2], target tracking [3], and full coverage attacking, which can prevent human beings from boring, harsh and dangerous environment. One main advantage of a robotic swarm solution is the simple local interactions between individuals within a complex system that could generate some new properties and phenomena observed at the system level, such as a set of collective behaviors [4]. Much like their biological counterparts such as fish schools [5], bird flocks [6], ant colonies [7], and cell populations [8], the resulting collective patterns are robust and flexible to agents joining in and dropping out, especially when accidents like obstacles, dangers, and new missions emerge. While robotic swarm en- joys numerous advantages, the large-scale autonomous robotic swarm incurs a high robot-failure probability due to real- life conditions when delays, uncetainties, and kinematic con- straints are present. This phenomenon is even more noticeable in military projects like Gremlins [9] and LOCUST [10], since they are built on small, low-cost and semi-autonomous UAVs whose failure probability is expected to be much higher. Among researches of flocking models with obstacle avoid- ance, Wang et al. [11] proposed an improved fast flocking algorithm with obstacle avoidance for multi-agent dynamic systems based on Olfati-Sabers algorithm. Li et al. [12] studied the flocking problem of multi-agent systems with obstacle avoidance, in the situation when only a fraction of the agents have information on the obstacles. Vrohidis et al. [13] considered a networked multi-robot system operating in This work was supported in part by the National Natural Science Foundation of China under Grants 61872378, 91648204, 71702186, in part by Postgraduate Research Innovation Project in Hunan Province under grant CX2018B021, in part by the Scientific Research Project of National University of Defense Technology under Grants ZK17-03-48 Fig. 1. Genetic flocking optimizing framework To the best of our knowledge, few previous literatures studied the model that satisfies both stability and adaptivity of the autonomous robotic swarm. Thus, we design a novel genetic flocking optimizing framework that can achieve both stability and adaptivity of the robotic swarms. As shown in Fig. 1, a robot is seperated into three layers, including sensing layer, decision layer, and action layer, which supports the basic navigation function. Besides, we generalize velocity updating formula by rules described with weight parameters, which evolves through interaction with the environment. The environment is divided into two layers: the evaluation layer and the evolutionary layer, where the former provides fitness function for the latter. II. GENERALIZED FLOCKING MODEL As shown in Fig. 2, a robot agent i has four detection areas: repulsion area, alignment area, attraction area, and obstacle avoidance area. The velocity updating fomula is described as follows: the weight parameters a, b, c, d, e ∈ (0, 1), which is used to flexibly handle the generalization formula. In equation (1), we define Sensing layerDecision layerAction layerTarget orientationObstacle avoidanceRepulsionAlignmentAttractionGeneralized velocity update formulaEvaluation layerEvolution layerAverage timeDeath rateAnisotropy Aggregation UniformityInitial populationFitness testSelect best parentsChildrenBREEDMUTATEUltrasonicLaser radarCameraCommunicationOne GenerationRules Rules Robot(Agent)Environment2 XXXX, VOL. XX, NO. XX, XXXX 2019 whole navigation process, while N is the agent number of the agents that arriving at target area. We define the stability of the robotic swarm as the variance of the γt sequence, which describes whether the flock structure of this swarm is stable. (cid:80)T t=0 (γt − ¯γ)2 (cid:113)(cid:0)px (cid:1)2 +(cid:0)py T s2 γ = (cid:80) t j − rx N j γt = , j − ry t (cid:1)2 . (3) (4) Fig. 2. avoidance area (Zobs). Pattern-formation areas (Zrep, Zali, Zatt) and obstacle- (cid:88) ∆vi = a (R0 − rij) (cid:88) vjvj + c j∈Zali (R3 − rik) pi − pj rij (cid:88) j∈Zatt pi − pk rik j∈Zrep 1 Nali (cid:88) j∈Zobs + b + d (R2 − rij) ptar − pi + e ritar pj − pi rij + v(t). Tunning the model above means that we propose four rules referring to classic reynolds' boids model and we optimize the parameters there. Note that the parameter space is 20- dimensional; tuning, global optimization methods, or parameter sweeping would be generally much time-consuming. therefore, manual III. ORDER PARAMETERS AND FITNESS FUNCTION In order to select the set of parameters that perform the best in the simulation process, we propose a fitness function com- posed by several order parameters, which helps us to abstract the mathematical model of single objective optimization. The fitness function is described as flollows: − T start ) × Ndeath Ntotal j − ry t )2 t )2 + (py j j (cid:80) (cid:113) j (T arrive (cid:80) (cid:80) N j − rx (cid:80) (px N T t (γt − ¯γ) (cid:80) (cid:80) j − δt)2 j t t T j (θt T F = × × × × α, j where T start is the time when the navigation is triggered, j and T arrive is the time when robotic agent j reaches the target area. Ndeath represents the number of the dead agent, and Ntotal represents the number of the total agents in the robotic swarm. rx t are the abscissa and ordinate of the position of the swarm's centroid at time t. T is the total time of the t and ry We define anisotropic index to describe the variation of population velocity direction. Specifically, it needs to calculate the average angle of each individual velocity direction and flock velocity direction at a certain time, and then calculate the average value of the whole process, which is the index of anisotropic index. The variance of the average angle of the whole process represents the variation range of anisotropic index, and the formula of anisotropy is as follows: (cid:80) θj δt = j N . (5) (1) With this method, we created a single-objective optimization scenario, which can be solved using genetic algorithm. IV. THE G-FLOCKING ALGORITHM This research adopts Parameter Tuning of Flocking Model based on classical genetic algorithm (GA) framework. The main algorithm is described as follows: Algorithm 1 The G-flocking algorithm Require: Rexp: a set of traditional expert's rules; P (0): randomly generate an initial population of rules; M: max- imum number of iterations; Np: number of the population; LR: length of the rules; Ns: number of seed used to produce the next generation; r: member mutation rate; Evaluate fitness of P (t) for i = 1 → M do Ensure: Ropt: an set of optimal rules; 1: t = 0 2: while t <= T do 3: 4: 5: 6: 7: 8: 9: 10: end for for i = 1 → M/2 do end for for i = 1 → M do Select operation to P (t) Crossover operation (2) to P (t) = Mutation operation to P (t) = M utation(P (t), r) Crossover(Ns, Np) end for for i = 1 → M do 11: 12: 13: 14: 15: 16: 17: 18: 19: end while end for t = t + 1 end for for i = 1 → M do P (t + 1) = P (t) R0zrepR2R3zalizattR1zobsLI MA et al.: PREPARATION OF PAPERS FOR XXXX (2019) 3 In the algorithm, the random rules Rexp are represented (cid:9) , i = 0, ei 0 0, ci 0, di (cid:9), and Ri the optimal rules for the 0, R2 0, R3 0, R4 0 1, 2, 3, 4. as: (cid:8)R1 rules: Ropt = (cid:8)R1 (cid:8)ai (cid:9), and Ri 0 = (cid:8)ai (cid:9) , i = 1, 2, 3, 4. opt, bi Once optimized, we can get opt, R2 opt opt, di opt, ei opt, ci 0, bi flocking which composing the BRIAN model. The outputs of G-flocking are also a set of optimized opt = opt, R4 opt, R3 opt V. EXPERIMENT ANALYSIS To reveal the performance improvements of BRIAN, we compare it with basic rule-based model (BREAM). BREAM derives from the classical Reynolds' flocking model that has been widely used. To apply Reynolds' flocking model to more complex environment, the comprehensive obstacle avoidance strategies are integrated into BREAM. (a) BREAM (20 robots) (b) BRIAN (20 robots) (c) BREAM (60 robots) (d) BRIAN (60 robots) (e) BREAM (100 robots) (f) BRIAN (100 robots) Fig. 3. BREAM and BRIAN traces of the robotic swarm in autonomous navigation in test scenery ( We tested 3 groups of experiments using 20, 60 and 100 robotic agents for simulation.) In order to clearly observe the impacts of different parame- ters of the formula for velcocity updating, we compare the performance of basic rule-based model (BREAM) and our optimized flocking model for robotic swarm in navigation (BRIAN) in scenary with three basic environmental elements including tunnel obstacle, non-convex obstacle and convex obstacle. Fig. 3 shows directly that BRIAN performs better than BREAM in uniformity and stability. Fig. 3(a) and Fig. 3(b) represent the performance of these two models with 20 robotic agents, Fig. 3(c) and Fig. 3(d) with 60 robots, Fig. 3(e) and Fig. 3(f) with 100 robots. Specific performance indicators are shown in Table I. We record the values of each evaluation index of the two models in three situations of the number and scale of robots. Generally, all the indicators of BRIAN model perform better (the smaller, the better). Specifically, aggregation of BRIAN is 56% lower than that of BREAM while the reduction of other indicators (anisotropy, average- time, uniformity, deathrate, and fitness function) are 88.61%, 32.55%, 89.69%, 100%, and 99.92%, respectively. Fig. 4. changes through time. The uniformity of the robotic swarm with each experiment Fig. 4 shows the change of uniformity in the whole time step. The total time step of each group of experiments is not the same, but it can be seen from the figure that the data of each group of BRIAN are stable between 0 and 1, which means that the stability and tightness of the cluster are very good during the whole cruise. When BREAM passes through obstacles, it can be seen that there will be large fluctuations near step 31 and step 71. Such fluctuations represent the situation of low cluster tightness and stability when cluster passes through narrow and non-convex obstacles, and the formation is not well maintained. At the same time, it can be seen that BRIAN has completed the whole task in about 84 seconds, while BREAM has completed the whole task. VI. CONCLUSIONS AND FUTURE WORK We presented in this paper an optimized flocking model for robotic swarm in autonomous navigation. This model is obtained through G-flocking algorithm proposed by us, which is extended from the classical genetic algorithm and rule-based flocking model in most relative researches. This is the first of its kind reported in the literatures; it comprehensively ad- dresses the reliability, adaptivity and scalability of the robotic swarm during completing the navigation tasks. The following issue will be addressed in our future work: First, we will extend our experiment to the real-world sys- tems such as unmanned aerial systems and unmanned ground systems. Second, we will take more uncertainties of scenaries into the model to verify the correctness of our model, such as adding the moving obstacle, the irregular barriers, and even fluid barriers. 051015202505101520250510152025051015202505101520250510152025051015202505101520250510152025051015202505101520250510152025010203040506070809010011002468BREAM20BRIAN20BREAM60BRIAN60BREAM100BRIAN1004 XXXX, VOL. XX, NO. XX, XXXX 2019 COMPARISONS BETWEEN BREAM & BRIAN WITH 20, 60 AND 100 ROBOTS TABLE I Aggregation Evaluation Algorithm BREAM BRIAN BREAM BRIAN BREAM BRIAN BREAM BRIAN BREAM BRIAN BREAM BRIAN 0.0079 Num-20 0.0371 Num-60 Num-100 0.0317 130.0500 128.7667 115.8000 38.6869 42.8201 50.1455 84.5000 83.9167 84.2500 6.1110 7.6107 92.8195 0.0000 0.0000 0.0000 0.8481 0.8639 1.1681 0.4666 0.4326 0.4557 5.3094 4.6934 4.9927 0.2864 0.3195 0.2100 0.0076 0.0435 0.0330 0.3500 0.3333 2.7367 Fitness Function Averagetime Uniformity Anisotropy Deathrate REFERENCES [1] Murphy, R. R., Tadokoro, S., Nardi, D., Jacoff, A., Fiorini, P., & Choset, H., et al. (2008). Search and rescue robotics. Springer Handbook of Robotics, 1151 -- 1173. [2] Dirafzoon, A. , & Lobaton, E. . (2014). Topological Mapping of Un- known Environments using an Unlocalized Robotic Swarm. IEEE/RSJ International Conference on Intelligent Robots & Systems. IEEE. [3] Parker, L. E., Rus, D., & Sukhatme, G. S. (2008). Multiple Mobile Robot Systems. springer Handbook of Robotics, 921 -- 941. [4] Brown, D. S. , Kerman, S. C. , & Goodrich, M. A. . (2014). [acm press the 2014 acm/ieee international conference - bielefeld, germany (2014.03.03-2014.03.06)] In Proceedings of the 2014 acm/ieee interna- tional conference on human-robot interaction - hri 14 - human-swarm interactions based on managing attractors. 90-97. [5] Krause, J., Hoare, D., Krause, S., Hemelrijk, C. K., & Rubenstein, D. I. (2015). Leadership in fish shoals. Fish & Fisheries, 1(1), 82-89. [6] Nagy, Mt, kos, Zsuzsa, Biro, D. , & Vicsek, Tams. (2010). Hierarchical group dynamics in pigeon flocks. Nature, 464(7290), 890-893. [7] Feinerman, O. , Pinkoviezky, I. , Gelblum, A. , Fonio, E. , & Gov, N. S. . (2018). The physics of cooperative transport in groups of ants. Nature Physics. [8] Cheung, K. J., Gabrielson, E., & Werb, Z., et al.(2013). Collective invasion in breast cancer requires a conserved basal epithelial program, Cell, 155(7). [9] Talal Husseini.(2018). Gremlins are enters Phase technology.com/features/gremlins-darpa-uav-programme/. its UAV programme. III of coming: DARPA https://www.army- [10] (2018). Raytheon gets $29m for work on US Navy LOCUST UAV https://navaltoday.com/2018/06/28/raytheon-wins-contract- prototype. for-locus-inp/. [11] Wang, J., & Xin, M. (2013). Flocking of multi-agent system using a unified optimal control approach. Journal of Dynamic Systems Mea- surement & Control, 135(6), 061005. [12] Li, J. , Zhang, W. , Su, H. , & Yang, Y. . (2015). Flocking of partially- informed multi-agent systems avoiding obstacles with arbitrary shape. Autonomous Agents and Multi-Agent Systems, 29(5), 943-972. [13] Vrohidis, C. , Vlantis, P. , Bechlioulis, C. P. , & Kyriakopoulos, K. J. . (2018). Reconfigurable multi-robot coordination with guaranteed con- vergence in obstacle cluttered environments under local communication. Autonomous Robots, 42(4), 853-873. Li Ma is currently working toward his Ph.D. de- gree in the College of Systems Engineering, Na- tional University of Defense Technology. Contact him at [email protected]. Xiaomin Zhu is currently an Associate Profes- sor in the College of Systems Engineering at Na- tional University of Defense Technology, Chang- sha, China. Contact him at [email protected]. Meng Wu is currently pursuing the M.S. degree in the College of Systems Engineering, National University of Defense Technology, China. Con- tact her at [email protected]. College Yuan Wang is currently a Ph.D Candidate of Engineering, National University of Defense Technology, Changsha, at [email protected]. Systems Contact China. him of Yunxiang Ling is currently a professor in Offi- cers college of PAP, Chengdu, China. Contact him at [email protected]. Weidong Bao is currently a professor in the Col- lege of Systems Engineering at National Univer- sity of Defense Technology, Changsha, China. Contact him at [email protected]. Wen Zhou is currently an Assistant Professor in the College of System Engineering at National University of Defense Technology, Changsha, China. Contact him at [email protected].
cs/0512003	1	0512	2005-12-01T04:09:22	Societal Implicit Memory and his Speed on Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms	[ "cs.MA", "cs.AI" ]	In order to overcome difficult dynamic optimization and environment extrema tracking problems, We propose a Self-Regulated Swarm (SRS) algorithm which hybridizes the advantageous characteristics of Swarm Intelligence as the emergence of a societal environmental memory or cognitive map via collective pheromone laying in the landscape (properly balancing the exploration/exploitation nature of our dynamic search strategy), with a simple Evolutionary mechanism that trough a direct reproduction procedure linked to local environmental features is able to self-regulate the above exploratory swarm population, speeding it up globally. In order to test his adaptive response and robustness, we have recurred to different dynamic multimodal complex functions as well as to Dynamic Optimization Control problems, measuring reaction speeds and performance. Final comparisons were made with standard Genetic Algorithms (GAs), Bacterial Foraging strategies (BFOA), as well as with recent Co-Evolutionary approaches. SRS's were able to demonstrate quick adaptive responses, while outperforming the results obtained by the other approaches. Additionally, some successful behaviors were found. One of the most interesting illustrate that the present SRS collective swarm of bio-inspired ant-like agents is able to track about 65% of moving peaks traveling up to ten times faster than the velocity of a single individual composing that precise swarm tracking system.	cs.MA	cs	Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 1 Societal Implicit Memory and his Speed on Tracking Extrema in Dynamic Environments using Self-Regulatory Swarms Vitorino Ramos1, Carlos Fernandes2,3, Agostinho C. Rosa2 1 CVRM-IST, Technical Univ. of Lisbon (IST), Av. Rovisco Pais, 1, 1049-001, Lisbon, PORTUGAL [email protected] 2 LaSEEB-ISR-IST, Technical Univ. of Lisbon (IST), Av. Rovisco Pais, 1, TN 6.21, 1049-001, Lisbon, PORTUGAL {cfernandes,acrosa}@laseeb.org 3 EST-IPS, Setúbal Polytechnic Institute (IPS), R. Vale de Chaves - Estefanilha, 2810, Setúbal, PORTUGAL Abstract— In order to overcome difficult dynamic optimization and environment extrema tracking problems, we propose a Self-Regulated Swarm (SRS) algorithm which hybridizes the advantageous characteristics of Swarm Intelligence as the emergence of a societal environmental memory or cognitive map via collective pheromone laying in the landscape (properly balancing the exploration/exploitation nature of the search strategy), with a simple Evolutionary mechanism that trough a direct reproduction procedure linked to local environmental features is able to self-regulate the above exploratory swarm population, speeding it up globally. In order to test his adaptive response and robustness, we have recurred to different dynamic multimodal complex functions as well as to Dynamic Optimization Control (DOC) problems. Measures were made for different dynamic settings and parameters such as, environmental upgrade frequencies, landscape changing speed severity, type of dynamic (linear or circular), and to dramatic changes on the algorithmic search purpose over each test environment (e.g. shifting the extrema). Finally, comparisons were made with traditional Genetic Algorithms (GA) as well as with more recently proposed Co-Evolutionary approaches. SRS, were able to demonstrate quick adaptive responses, while outperforming the results obtained by the other approaches. Additionally, some successful behaviors were found: SRS was able not only to achieve quick adaptive responses, as to maintaining a number of different solutions, while adapting to new unforeseen extrema; the possibility to spontaneously create and maintain different subpopulations on different peaks, emerging different exploratory corridors with intelligent path planning capabilities; the ability to request for new agents over dramatic changing periods, and economizing those foraging resources over periods of stabilization. Finally, results prove that the present SRS collective swarm of bio-inspired agents is able to track about 65% of moving peaks traveling up to ten times faster than the velocity of a single ant composing that precise swarm tracking system. This emerged behavior is probably one of the most interesting ones achieved by the present work. ORIGINAL PAPER WITH FULL PICTURES AVAILABLE IN: http://alfa.ist.utl.pt/~cvrm/staff/vramos/Vramos-NIAS06.pdf I. INTRODUCTION M (EC) and swarm OST research in evolutionary intelligence (SI) computation focuses on optimization of static, non-changing problems. Many real-world optimization problems, however, are dynamic, and optimization is needed that are capable of continuously adapting the solution to a changing environment. In fact, many real-world problems are actually dynamic: new jobs have to be added to the schedule, machines may break down or wear down slowly, raw material is of changing quality, etc [8]. If the optimization problem is dynamic, the goal is no longer to find the extrema, but to track their progression through the space as closely as possible. One method for achieving this is to evolve a function off-line that models the dynamics of the environment directly [58] while another is to use the dynamics of the evolutionary -or Swarm Intelligent- process itself on-line to track the progression of the extrema. As described by Angeline [1] evolving a function that describes extrema dynamics is preferable when the dynamics can be modeled accurately off-line. When a single function cannot describe the dynamic accurately enough, an on-line approach using the implicit dynamics of the self- adaptive method is preferable. However, while evolutionary computation methods that evolve mutation variances perform well in static environments, it is not clear that these methods are beneficial when the gradient at each point is constantly in flux as in most dynamic environments [1], apart from significant, well-know and promising attempts. For these reasons, it seems appropriate to keep the suggestion of on-line it with different simultaneously attempting methods, Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 2 computational paradigms such as those derived from Swarm Intelligence which allows for distributed real-time self- organization of solutions while maintaining strong adaptive capabilities. Following this new research path, we propose a Self-Regulated Swarm (SRS) algorithm which hybridizes the advantageous characteristics of Swarm Intelligence as the emergence of a societal environmental memory or cognitive map [50,49] via collective pheromone laying in the landscape (balancing the exploration/exploitation nature of the search strategy), with a simple evolutionary mechanism that trough a direct reproduction procedure linked to local environment features is able to self-regulate the above exploratory swarm population, speeding it up globally [17]. Our present proposal is fully discussed in section II. A. Approaches Dynamic Optimization (DO) problems, on a more abstract level can be characterized by several situations; this might mean that the optimization function, the problem instance, or some restrictions may change, and thus the optimum to that problem might change as well. If any of these events are to be taken into account in the optimization process, we call the problem dynamic or changing terms are used (both synonymously and often in the literature, also the term non- stationary is used). Since many approaches are possible, Branke and Schmek [8] suggested in 2002, a classification of DO problems, surveying it, while classifying a number of the most widespread techniques that have been published in the literature so far to make evolutionary algorithms (EAs) suitable for changing optimization problems. As pointed by them, one standard approach to deal with these dynamics is to regard each change as the arrival of a new optimization problem that has to be solved from scratch [48]. However, this simple approach is often impractical: solving a problem from scratch without reusing information from the past might be too time consuming, a change might not be identifiable directly, or the solution to the new problem should not differ too much from the solution of the old problem. Thus, as in the on-line tracking process suggested by Angeline [1], Branke [8,9,10] recommended that would be nice to have an optimization algorithm that is capable of continuously adapting the solution to a changing environment, reusing the information gained in the past. Since in nature adaptation is a continuous and continuing process, and EAs have much in common with natural evolution, they seem to be a suitable candidate. However, these type of evolutionary approaches typically converge to an optimum and thereby lose their diversity necessary for efficiently exploring the search space and consequently also their ability to adapt to a change in the environment when such a change occurs [8,10]. The problem here can be stated as seeking an appropriate and difficult balance between two contradictory characters of the search procedure; those between the exploring (ideal for gathering new solutions) and exploiting (making the best use of past solutions) nature of the algorithm. Over the past few years, a number of authors have addressed loss of the problem of convergence and subsequent adaptability in many different ways. According to Branke and Schmeck [8], most of these approaches could be grouped into one of the following three categories established by them: I - React on Changes: The EA is run in standard fashion, but as soon as a change in the environment has been detected, explicit actions are taken to increase diversity and thus facilitating the shift to the new optimum. II - Maintaining Diversity throughout the run: Convergence is avoided all the time and it is hoped that a spread-out population can adapt to changes more easily. III – Memory-based Approaches: The EA is supplied with a memory to recall useful information from past generations, which seems especially useful when the optimum repeatedly returns to previous locations. Techniques such as Hypermutation (Cobb, [13]) or Variable Local Search (VLS) (Vavak, [66,67]) pursue category I, keeping the whole population after a change but increasing population diversity by drastically increasing the mutation rate for some number of generations. VLS on the other hand, increases mutation gradually after a change has been detected [66], where the mutation range itself can be adapted [67]. Another variant with significant improvements in convergence speed and solution quality have been found if the altered (old) individuals are reused (see Bierwirth [5,6], Lin [39], Mattfeld [41], or Reeves et al. [55]). Artificial Neoteny techniques proposed for EAs by Ramos [52] can also be considered among this approach. Neoteny, also know as Paedomorphosis, can be defined in biological terms as the retention by an organism of juvenile or even larval traits into later life. In some species, all morphological development is retarded; the organism is juvenilized but sexually mature. Such shifts of reproductive capability would appear to have adaptive significance to organisms that exhibit it, and the technique is used within EAs by simple re-injecting old artificial DNA which was randomly captured on the first generations. Finally, the A approach ends up with a number of authors suggesting to leave the response to a change in the environment up to the strategy itself, by relying on his self- adaptiveness. This is the case in works by Angeline [1], Bäck [4], Grefenstette [20] and Stephens [60]. Maintaining diversity is another possible approach (II). Grefenstette [21] introduced the method Random Immigrants where in every generation, the population is partly replaced by randomly generated individuals, while Andersen [2] uses genotypic and phenotypic sharing to increase the Genetic track optima Algorithm ability to in slowly changing environments. A very interesting sharing scheme was introduced by Liles and DeJong [38], based on Tag bits. There, tag bits are appended to each genotype, and only individuals with equal tag bits are allowed to mate, which somehow can be regarded as introducing subpopulations of varying size [59]. The neighborhood used for sharing is then the number of individuals with the same tag bit, i.e. individuals with rare tag bits are favored. The approach is able to maintain different subpopulations on different peaks, Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 3 however for a simple environment with two peaks of changing weights. Other techniques include a crowding-like replacement scheme entitled “Worst among Most Similar” by Cedeno [11], by taking in account each individual’s age, modifying the fitness function fmod = g (fold, age) as proposed by Ghosh [18] or finally by using the Thermodynamical Genetic Algorithm (TDGA) proposed by Mori [45] addressing the diversity in the population explicitly controlling a measure entitled “free energy” F. For a minimization problem this term is calculated as F = <E>-TH, where <E> stands for the average population fitness and H is a measure for the diversity in the population. The temperature T is a parameter of the algorithm and reflects the emphasis of diversity (adjusting T is discussed in [46], a problem similar in many respects in what we found to control scheduling temperatures in Simulated Annealing, e.g. [52]). Finally, a very interesting and recent approach was implemented by Huang and Rocha [30,56] outperforming traditional Genetic Algorithms via obtaining greater phenotypic plasticity. They proposed a co- evolutionary agent based model of genotype editing (ABMGE), constructed using several genetic editing characteristics that are gleaned from the RNA editing system as observed in several organisms. The incorporation of editing mechanisms provides a means for artificial agents with genetic descriptions to gain greater phenotypic plasticity. By allowing the family of editors and the genotypes of agents to co-evolve using the re-generation of editors as a control switch for environmental changes, the artificial agents in ABMGE can discover proper editors to facilitate the tracking of the extrema in dynamic environments. Other approaches on this line are fully discussed in [8,9,10] by Branke and Schmeck. Another kind of approach is to supply the algorithm with some sort of memory (approach III) storing good partial solutions in order to reuse them later. This can be advantageous in cases where the environment is changing periodically, and repeated situations occur. However they also could be counterproductive if the environment changes dramatically with open-ended novelty. Memory may be provided in two general ways: implicitly by using redundant representations, or explicitly by introducing an extra memory and formulating strategies to deposit and retrieve solutions later. Generally, the most prominent approach to implicit memory and redundant representation is multiploidy [19] (e.g., Ryan [57], Lewis [37] and Dasgupta [16]). On the other hand, while redundant representations allow the EA to implicitly store some useful information during the run, it is not clear that the algorithm actually uses this memory in an efficient way. As an alternative, the following approaches use an explicit memory in which specific information is stored and reintroduced into the population at later generations. The memory can be used to transfer individuals from one EA run to seed the initial population on another EA after a single change in the environment, as suggested by Louis and Xu [40], or trough the use of a knowledge base to memorize successful individuals in a permanent memory, assuming that the system can measure the environmental conditions (Ramsey and Grefenstette [54]). Other examples include a procedure by Trojanowski [64] where each individual is extended with additional memory for a number of its ancestors, or via a variation of evolutionary elitism within Thermodynamical Genetic Algorithms [44,45,46] – i.e., in every generation the best individual is stored in the memory, and another individual is deleted from the memory depending on its age and contribution to this memory population’s diversity (measured as variance over bit positions). Finally, Branke [9] compared a number of replacement strategies for inserting new individuals into a memory stressing the importance of diversity for memory-based approaches. As an example, Branke found out that a simple replacement of the most similar individual performed almost equivalently to a strategy that replaces the worse of the two individuals in the memory closest to each other. Although the memory/search population approach was quite successful, it soon became obvious that a strategy based on memorization would be too restricted to adapt successfully to a wide range of dynamic environments. As an alternative, Branke and Schmek [8] developed an approach with multiple populations acting as Self-Organizing Scouts (SOS), watching over the changing landscape. The basic idea of SOS is that once a peak has been found (i.e. the population converged to a high-performance region), the population should split, while the “child population” should “watch” over that peak, the remainder should spread out searching for new peaks. Since our current approach is also based in Self-Organization [50,53], trough the use of stigmergic mechanisms as well as Swarm Intelligence [7], there are in fact some similarities as well differences between the Self-Regulated Swarms SRS approach and SOS that will be interesting to discuss later (see section V-B). Finally some recent proposals have been made using a Swarm Intelligent (SI) [7, 33] approach to attempt to solve these dynamic problems. Generally, Swarm Intelligence can be regarded as the property of a system whereby the collective behaviors of (unsophisticated) entities interacting locally with their environment cause coherent functional global patterns to emerge. SI provides a basis with which it is possible to explore collective (or distributed) problem solving without centralized control or the provision of a global model (Stan Franklin, Coordination without Communication, talk at Memphis Univ., USA, 1996) [49]. These entities can be either regarded as bio-inspired ant-like agents in which self- organization occurs trough trail formation via pheromone deposition and evaporation [7,50,17,49,23,53,51] giving rise to the well know Ant Colony Systems (ACS) and Ant Colony Optimization (ACO) algorithms by Dorigo et al. [7], or as physical particles embodied with direction, velocity and intrinsic memory for best global and local position [33,2,3,15,32] know as Particle Swarm Optimization (PSO) algorithms developed by Kennedy, Eberhart et al. [33]. Apart from the SI research branch taken (ACS or PSO), only very recently there are being made efforts for dynamic optimization problems using Swarm Intelligence. Via ACS, Guntsch and Middendorf [23], applied population based ACO algorithms for tracking extrema in dynamic environments, while others like Ramos, Fernandes and Rosa [50,17,49] developed distributed pheromone laying over the dynamic environment itself, in order to track different peaks. They show that the self-organized algorithm is able to cope and Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 4 quickly adapt to unforeseen situations even when over the same cooperative foraging period, the community is requested to deal with two different and contradictory purposes (different extrema), comparing their results with those from Passino et al. [47] which developed a Bacterial Foraging Optimization Algorithm (BFOA) for distributed optimization and control. Finally also via ACS, Guntsch, Middendorf and Schmeck [22] developed an ACO introducing local variance where needed, and a heuristic repair of solutions. They applied successfully the novel algorithm for the combinatorial dynamic TSP (Travelling Salesman Problem). Meanwhile via the PSO [33] approach, Carlisle and Dozier [2,3], adapted it for dynamic environments. The process consists of causing each particle in the swarm to reset its record of its best position as the environment changes, to avoid making direction and velocity decisions on the basis of outdated information. Thus, the authors explicitly discard memory from the standard PSO, and call it Adaptive PSO. On the other hand, Parrott and Li [15] developed a PSO model for tracking a three-peak multimodal environment. To achieve this, a form of speciation allowing development of parallel subpopulations is used. The model employs a mechanism to encourage simultaneous tracking of multiple peaks by preventing overcrowding at peaks. Others like Janson and Middendorf [32], followed an hierarchical strategy adjusting the PSO accordingly and developing the first hierarchical PSO for dynamic problems. Other authors like Altshuler and Wagner [69], presented a multi-agent system for a dynamic cleaning problem on the area of Swarm Cooperative Robotics. In order to find the minimal cleaning time possible for a given “contaminated” shape, they have designed the system in such a way that all the cleaning agents are controlled by a central unit (referred to as the queen). Upon initialization, the queen is given the complete information regarding the contaminated shape to be cleaned. While the agents are traveling along the grid, the queen is immediately aware of any new information discovered by the agents. The queens orders as to the next desired movements of the agents are also immediately transferred to the agents, which carry them our automatically. B. Dynamism Characterization and Benchmarks Branke and Schmeck [8,10] suggested a number of criteria along which dynamic environments could be categorized as well as tested. Since many other authors on dynamic optimization and control are following them, we will adopt them as well, having in mind different comparison purposes. Four main aspects should be considered. The first two are usually adopted: I – Frequency of change: This criterion establishes how often the environment changes (starting from very rare changes up to continuous change). A parameter uf (update frequency) is introduced. If for instance uf = C, then at every C iterations (or generations, in case of EA’s) the landscape will suffer changes. II – Severity of change: This criterion establishes how strongly the system is changing. From slight changes to completely new situations. The severity parameter S will be defined later. III – Predictability of change: The present aspect gives a notion if there is a pattern or trend in the changes. That is, depending on the problem at hands, it is somehow possible to predict direction, time or, the severity of the next change given the changes encountered so far? IV – Cycle length: This criterion measures how often the optimum returns to previous locations or at least gets close to them. Since these to measure some of it seems difficult characteristics, comparisons between different optimization problems may still be out of reach, but at least the above characteristics can be varied on a single problem such that their qualitative influence on a specific approach can be examined [8]. In our different experiments (section III) we will use criterions of type IV (section III.A), type I (section III.B) and of type II (section III.C), or with combinations with the formers. Moreover, in order to compare the relative abilities of different algorithmic approaches for on-line tracking of dynamic extrema, Angeline [1] suggested a number of dynamical environments created from a single multi-dimensional function, referred to as the base function below. Dynamic environments were then obtained by translating the base function along a number of distinct temporal trajectories (fig. 1). This setup allows complete control of the dynamics and generates simple yet non-trivial dynamic functions with know properties. Tests under this specific benchmark were performed in section III.A. Angeline [1] has made use of a simple convex parabolic function in three dimensions, while we tested our algorithm over two complex multimodal functions, well know in evolutionary computation performance evaluation: the Ackley (fig. 3) and Schaffer F7 functions. The Ackley function was used as our base function for section III.A tests. b) c) Fig. 1. Example dynamics used for the trajectory of the base function over time in this study along with their projection onto the (x,y)-plane (grey); pictures and proposal from Angeline [1]. a) Linear dynamic given by Eq. 1; b) Circular dynamic given by Eq. 2; In c) an example of random dynamic using Gaussian noise given by Eq.3. As described by Angeline, the temporal dynamics applied to the base function involve three distinct parameters: dynamic type, step size and update frequency. Dynamic type identifies one of three distinct parameterized methods for translating the base function. Severity is a parameter used as input to each of the dynamics to determine the amount the base function is displaced from its current position. Finally, update frequency determines the number of iterations (or generations) between Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 5 each movement of the base function. Angeline [1] tested three different dynamic types in his experiments: linear, circular and random. Given a severity of S, the linear dynamic updates the current offset for dimension k, denoted as Δ´ k, as follows (Eq.1): +Δ=Δ´ k k S which simply adds a constant displacement to the offset in each dimension equivalent to the severity parameter’s setting. On the other hand, the circular dynamic is computed as follows (Eq.2): ´ +Δ=Δ k k sin. 3,1 = S k ⎛ ⎜ ⎝ ⎛ cos ⎜ ⎝ t 2 π C t 2 π C ⎞ , ⎟ ⎠ ⎞ , ⎟ ⎠ ´ +Δ=Δ k k S . k = 2 (1) (2) (3) where t is the number of applications of the dynamic thus far. Equation 2 describes a trajectory that translates the base function in a circular path through all three dimensions. The equation is designed to cycle the values of the offset every C applications with the severity parameter determining the radius of the circular path. Finally, for the random dynamic, random noise is added to the offset as follows (Eq. 3): ( )1,0 ´ +Δ=Δ k k NS . where N(0,1) is a Gaussian random variable with mean 0 and variance 1. Here, the severity parameter determines the variance of the noise added to the base function offset. Figure 1 (from [1]) shows an example of a random trajectory generated by this dynamic. In our tests (section III) we have made use of linear and circular dynamics (Eq.1,2) using the Ackley multimodal complex function (fig.3) as our base function. C. Our Proposal Many structures built by social insects are the outcome of a process of self-organization [50,49,51], in which the repeated actions of the insects in the colony interact over time with the changing physical environment to produce a characteristic end state [25]. A major mediating factor is stigmergy [63], the elicitation of specific environment-changing behaviors by the sensory effects of local environment changes produced by previous and past behavior of the whole community. Stigmergy is a class of mechanisms that mediate animal- animal interactions through artifacts or via indirect communication, providing a kind of environmental synergy, information gathered from work in progress, a distributed incremental learning and memory among the society. In fact, the work surface is not only where the constituent units meet each other and interact, as it is precisely where a dynamical cognitive map could be formed, allowing for the embodiment of emergent adaptive memory, cooperative learning and perception [50,49]. Constituent units not only learn from the environment as they can change it over time. Its introduction in 1959 by Pierre-Paul Grassé1 made it possible to explain what had been until then considered paradoxical observations: In an insect society individuals work as if they were alone while their collective activities appear to be coordinated. The stimulation of the workers by the very performances they have achieved is a significant one inducing accurate and adaptable response (check applications in [53,51]). Keeping in mind these characteristics we will present a stigmergic self- regulated model to tackle the collective adaptation of a social swarm for dynamic tracking, based on real ant colony behaviors. II. SELF-REGULATED SWARMS; THE SRS ALGORITHM As mentioned above, the distribution of the pheromone represents the memory of the recent history of the swarm (his social cognitive map), and in a sense it contains information which the individual ants are unable to hold or transmit [50]. There is no direct communication between the organisms but a type of indirect communication through the pheromonal field. In fact, ants are not allowed to have any local memory and the individual’s spatial knowledge is restricted to local information about the whole colony pheromone density. In order to design this behaviour, one simple model was adopted [12], and extended due to specific constraints of the present proposal, in order to deal with 3D dynamic environments. As described by Chialvo and Millonas, the state of an individual ant can be expressed by its position r, and orientation θ. Since the response at a given time is assumed to be independent of the previous history of the individual, it is sufficient to specify a transition probability from one place and orientation (r,θ) to the next (r,θ) an instant later. In a previous works by Millonas [42,43], transition rules were derived and generalized from noisy response functions, which in turn were found to reproduce a number of experimental results with real ants. The response function can effectively be translated into a two-parameter transition rule between the cells by use of a pheromone weighting function (Eq.4): W ( ) σ = ⎛ ⎜⎜ ⎝ 1 + 1 β ⎞ σ ⎟⎟ γσ + ⎠ (4) This equation measures the relative probabilities of moving to a cite r (in our context, to a cell in the grid habitat) with pheromone density σ(r). The parameter β is associated with the osmotropotaxic sensitivity, recognised by Wilson [68] as one of two fundamental different types of ant’s sense-data 1 Grassé, P.P.: La reconstruction du nid et les coordinations inter-individuelles chez Bellicositermes natalensis et Cubitermes sp. La théorie de la stigmergie : Essai d’interpretation des termites constructeurs. Insect Sociaux (1959), 6, 41- 83. The phrasing of his introduction to the term stigmergy is worth noting (translated to English in [OK30]): The coordination of tasks and the regulation of constructions do not depend directly on the workers, but on the constructions themselves. The worker does not direct his work, but is guided by it. It is to this special form of stimulation that we give the name Stigmergy (stigma - wound from a pointed object, and ergon - work, product of labor = stimulating product of labor). Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 6 TABLE I HIGH-LEVEL DESCRIPTION OF THE SELF-REGULATED SWARM (SRS) ALGORITHM PROPOSED /* Initialization / For all ants do Place ant at randomly selected site r e[ant]=1.0 End For / Main loop / For t = 1 to tmax do For all ants do / According to Eqs. 4 and 5 (section II) / Compute W(σ) and Pik Move to a selected neighboring site not occupied by other ant / According to Eq. 6 (section II) / Increase pheromone Pr at site r: Pr= Pr+[η+p(Δ[r]/Δmax)] / Reproduction procedure (section II - A) / Compute n, the number of occupied surrounding cells If ant meets ant (i.e., n ≥ 1) then Determine P(n) Compute reproduction probability P = P*(n) [Δ(r)/Δmax] If real random [0, 1] < P then Create one ant with e[ant]=1.0 and place it randomly on one of the free cells surrounding the main parent at site r End If End If End For Evaporate pheromone by K, at all grid sites For all ants do Decrease ant energy: e[ant]=e[ant]- Δe If e[ant] ≤ 0.0 then Kill that ant End If End For Print location of all agents Print pheromone distribution at all sites End For /* Values of parameters used in experiments / k = 1.3, η = 0.07, β=3.5, γ=0.2, Δe=0.1, p = 1.9, tmax = 100 or 400 time steps. / Constant values / P(0) = P(8) =0, P(4) = 1, P(5) = P(3) =0.75, P(6) = P(2) =0.5, P(7) = P(1) = 0.25 / Useful references / Check [50], [17], [49], [53] and [12]. processing. Osmotropotaxis, to a kind of related is instantaneous pheromonal gradient following, while the other, klinotaxis, to a sequential method (though only the former will be considered in the present work as in [12]). Also it can be seen as a physiological inverse-noise parameter or gain. In practical terms, this parameter controls the degree of randomness with which each ant follows the gradient of pheromone. On the other hand, 1/γ is the sensory capacity, which describes the fact that each ant’s ability to sense pheromone decreases somewhat at high concentrations. In addition to the former equation, there is a weighting factor w(Δθ), where Δθ is the change in direction at each time step, i.e. measures the magnitude of the difference in orientation. As an additional condition, each individual leaves a constant amount η of pheromone at the cell in which it is located at every time step t. This pheromone decays at each time step at a rate k. Then, the normalised transition probabilities on the lattice to go from P ik = ( ) ( ) W w Δ σ i i ) ( ( ∑ W w / σ Δ j kj ) j (5) (6) cell k to cell i are given by Pik (Eq. 5, [12]), where the notation j/k indicates the sum over all the surrounding cells j which are in the local neighbourhood of k. Δi measures the magnitude of the difference in orientation for the previous direction at time t-1. Since we use a neighbourhood composed of the cell and its eight neighbours, Δi can take the discrete values 0 through 4, and it is sufficient to assign a value wi for each of these changes of direction. Chialvo et al, used the weights of w0 =1 (same direction), w1 =1/2, w2 =1/4, w3 =1/12 and w4 =1/20 (U-turn). In addition coherent results were found for η=0.07 (pheromone deposition rate), k=0.015 (pheromone evaporation rate), β=3.5 (osmotropotaxic sensitivity) and γ=0.2 (inverse of sensory capacity), where the emergence of well defined networks of trails were possible. Except when indicated, these values will remain in the following test framework. As an additional condition, each individual leaves a constant amount T η += p [ ] i Δ maxΔ η of pheromone at the cell in which it is located at every time step t. Simultaneously, the pheromone evaporates at rate k, i.e., the pheromonal field will contain information about past movements of the organisms, but not arbitrarily in the past, since the field forgets its distant history due to evaporation in a time τ ≅ 1/k. As in past works, toroidal boundary conditions are imposed on the lattice to remove, as far as possible any boundary effects (e.g. one ant going out of the grid at the south-west corner, will probably come in at the north-east corner). In order to achieve emergent and autocatalytic mass behaviours around specific extrema locations (e.g., peaks or valleys) on the habitat, instead of a constant pheromone deposition rate η used in [12], a term not constant is included. This upgrade can significantly change the expected ant colony cognitive map (pheromonal field). The strategy follows an idea implemented earlier by Ramos [53], while extending the Chialvo model into digital image habitats, aiming to achieve a collective perception of those images by the end product of swarm interactions. The main differences to the Chialvo work is that ants, now move on a 3D discrete grid, representing the functions which we aim to study (figs. 3,4, section III) instead of a 2D habitat, and the pheromone update takes in account not only the local pheromone distribution as well as some features of the cells around one ant. In here, this additional term should naturally be related with specific characteristics of cells around one ant, like their altitude (z value or function value at coordinates x,y), having in mind our present aim. So, our pheromone deposition rate T, for a specific ant, at one specific cell i (at time t), should change to a dynamic value (p is a constant = 1.93) expressed by equation 6. In this equation, Δmax = zmax – zmin , being zmax the maximum altitude found by Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 7 the colony so far on the function habitat, and zmin the lowest altitude. The other term Δ[i] is equivalent to (if our aim is to minimize any given landscape): Δ[i] = zi – zmax , being zi the current altitude of one ant at cell i. If on the contrary, our aim is to maximize any given dynamic landscape, then we should instead use Δ[i] = zi – zmin . Finally, notice that if our landscape is completely flat, results expected by this extended model will be equal to those found by Chialvo and Millonas in [7], since Δ[i]/Δmax equals to zero. In this case, this is equivalent to say that only the swarm pheromonal field is affecting each ant choices, and not the environment - i.e. the expected network of trails depends largely on the initial random position of the colony, and in trail clusters formed in the initial configurations of pheromone. On the other hand, if this environmental term is added a stable and emergent configuration will appear which is largely independent on the initial conditions of the colony and becomes more and more dependent on the nature of the current studied dynamic landscape itself. As specified earlier, the environment plays an active role, in conjunction with continuous positive and negative feedbacks provided by the colony and their pheromone, in order to achieve a stable emergent pattern, societal memory and distributed learning by the community [50,53]. A. Reproduction procedure In addition to the above advantageous characteristics of Swarm Intelligence as the emergence of a societal environmental memory or cognitive map [50,49] via collective pheromone laying in the dynamic landscape, we hybridized it with a simple evolutionary mechanism that trough a direct reproduction procedure linked to some local environmental features is able to self-regulate the above exploratory swarm population, speeding it up globally. The full SRS strategy adapted for dynamic extrema tracking then finally consists of using Swarms with Varying Population Size (SVPS) proposed and analyzed earlier by Fernandes, Ramos and Rosa [17]. This characteristic is achieved by allowing ants to reproduce and die through their evolution in the landscapes. To be effective, the process of variation must incorporate some kind of environmental pressure towards successful behavior, that is, ants that reach peaks/valleys must have some kind of reward, by staying alive for more generations – generating more offspring - or by simply having a higher probability of generating offspring at each time step. In addition, the population density in the area surrounding the parents must be taken into account during a reproduction event. When one ant is created (during initialization or by the reproduction process) a fixed energy value is assigned to it (e[ant] = 1). Every time step, the ant’s energy is decreased by a constant amount of Δe (usually 0.1). The ant’s probability of survival after a time step is proportional to its energy during that same iteration, which means that after ten generations (with Δe = 0.1), this and other ants will inevitably die (e[ant] = 0). Within these settings one ant that is, for instance, 7 iterations old, has a probability of 0.3 to survive through the current time step. Meanwhile, for the reproduction process, we assume the following heuristic: an ant triggers a (main parent) reproduction procedure if it finds at least another ant (Moore occupying one of its 8 surrounding cells neighborhood is adopted). The probability P (Eq.7) of generating offspring – one child for each reproduction event – is computed in two steps (see table I). * P = ⎡ ( ) nP . ⎢ ⎣ Δ Δ ( ) r max ⎤ ⎥ ⎦ (7) First, the surrounding area is inspected in order to see if it is too crowded. Being n the number of occupied cells around this ant, the probability to reproduce P is set to the values shown in table II. Notice that: 1) an ant completely surrounded by other ants, or isolated (n=8, n=0) do not reproduce; 2) the maximum probability is achieved when the area that ant is half occupied (n=4). After the probability P TABLE II REPRODUCTION PROBABILITY P ACCORDING TO N MOORE NEIGHBORS n Moore neighbors Reproduction probability P(n) n = 0 or n = 8 0.00 n = 4 1.00 n = 5 or n = 3 0.75 n = 6 or n = 2 0.50 n = 7 or n = 1 0.25 is set to one of the previous values, the final probability is computed according to Eq.7, with the help of Δ(r) and Δmax (similarly as in Eq. 6). This operation guarantees that any ant reaching the higher/lower peaks/valleys has more chance to produce offspring (notice that one ant in the higher/lower site has P=1 if n=4 and will reproduce for certain). If this ant passes the reproduction test (table I), then a new agent is created occupying one of his vacant cells around the main parent. None infant ants are allowed to be allocated in places where other ants are. Finally notice that when n=0 or n=8 no reproduction takes place and that higher/lower (maximization/ minimization) ants have more chance to reproduce. B. Shifting the Extrema - Past work results One of the features of Swarms with Varying Population Size SVPS discussed in [17] was the ability to adapt to sudden changes in the roughness of the landscape. These changes were simulated by abruptly replacing one test function by another after the swarm reached the desired regions of the landscape. Another way of simulating changes in the environment consists on changing the task from minimization to maximization (or vice-versa). The swarm performance was convincing and reinforced the idea that the system is highly adaptable and flexible. Further tests using SVPS concluded that varying population size increases the capability of the swarm to react to changing landscapes [17]. Figure 2 shows SVPS trying to find the lower values of Passino F1 [47] until t=250, and then searching for the higher values. Comparisons with fixed sized swarms and Bacterial Foraging Optimization Algorithms (BFOA, [47]) were made in [49,50]. a) 3D view b) 2D view Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 8 t = 280 t = 260 t = 250 t = 10 t = 500 t = 320 t = 300 Fig. 2. SVPS evolving on a complex multimodal function seen in a-b) [17,47,49]: the self-organized swarm emerges a characteristic flocking migration behavior between one deep valley (south region) and one peak (north region), surpassing in intermediate steps (Mickey Mouse shape at t = 300) some local optima. Over each foraging step, the population self- regulates. From t=0 to t=250 the swarm is induced to search the lowest valleys of the landscape. After t=250 the task changes (target peak moves to the north of the territory) and the swarm must find the higher values of the function. Check for detailed results and extended analysis in [17]. III. DYNAMIC ENVIRONMENT TESTBED AND RESULTS In order to compare the relative behavior of our present Self-Regulated Swarms (SRS) approach for on-line tracking of dynamic extrema, we followed diverse kind of test beds, reflecting different dynamism characterization and benchmarks described earlier in section I.B, namely according to: dynamic type (section III.A), severity and speed tests (section III.B), as well as the application of the current self- organized algorithm to dynamic optimal control problems (section IV). A. Dynamic Type Tests Over this specific test bed, dynamic environments were obtained by translating the Ackley complex multimodal base function (fig.3) along a number of distinct linear and circular temporal trajectories (fig.1). This setup allows complete control of the dynamics and generates simple yet non-trivial dynamic functions with know properties. The Ackley function [14,28] (Eq.8) is considered as a minimization problem. Originally this problem was defined for two dimensions, but the problem has been generalized to N dimensions [62]. Formally, this problem can be described as finding a string xi={x1, x2, …,xN}, under the domain xi → (-32.768, 32.768), that minimizes equation 8. In order to define an instance of this function we need to provide the dimension of the problem (in here, n=2). The optimum solution of the problem is the vector u = (0,...,0) with F(u)=0. We will apply it to several swarm dynamic tests extending past analysis [50,17]. ⎛ ⎞ .12.0 n ⎛ ⎞ → ) ( ∑ ⎜ ⎟ 2 xF −=⎟ −⎟ ⎜ ⎜ n ⎝ ⎠ ⎝ ⎠ i 1 = 1 ( ( .2 π n (8) exp.20 exp cos 20 a i + e x i x i − a i ⎛ ⎜ ⎝ n ∑ i 1 = − ⎞ ) ) +⎟ ⎠ For our specific tests we have made use of the domain x,y → [-2.0,2.0],[-2.0,2.0], containing 16 valleys. In order to change it over time, we have defined a target path line over our domain, from northwest (-2,2) to southeast (2,-2). Figure 3, shows three typical minimal target points along this line, A, B and C: A(-1.5;1.0), B(0.0;0.0) and C(1.0;-1.5). To introduce the dynamics into the problem, the base function minimal target point starts to move from B, moving continuously to southeast (passing C). Arriving at the southeast corner, the target point then moves to the northwest corner (since our habitat is toroidal), then passing A, and B again, over and over − again has times passes, for several specific test speeds. Note that in equation 8, ai defines the point coordinates where the Ackley function has his minimal target point, thus introducing a more severe change than translation itself, since around that minimal point all the function changes differently depending on where this minimal point precisely is. This collection of dynamics when used with the complex multimodal Ackley base zmin=0 at (x,y)=-1.5,1 (normal view) zmin=0 at (x,y)=0,0 (normal view) zmin=0 at (x,y)= 1,-1.5 (normal view) zmin=0 at (x,y)=-1.5,1 (from above) zmin=0 at (x,y)=0,0 (from above) zmin=0 at (x,y)=1,-1.5 (from above) Fig. 3. The Ackley complex multimodal function seen from different perspectives and with respective global minimum zmin=0 at A(-1.5;1.0), B(0.0;0.0) and C(1.0;-1.5), over the domain x,y → [-2.0,2.0],[-2.0,2.0] containing 16 valleys. function has a number of nice features for testing the performance of self-adaptive algorithmic optimization. The linear dynamic will always move away from the position that the population is converging towards with the speed parameter determining how quickly it moves away. The SRS algorithm was then tested for different test speeds: v=0 (static environment), v=0.5, v=1, v=1.5, v=2, v=3, v=5 and v=10. In order to have an idea of the severity of this parameter, for instance at v=2, the target valley is traveling to southeast 2 habitat cells at each t time step starting in the B(0.0;0.0) mid- point, while each SRS ant can only move one cell for each time step t. SRS parameter values were the usual ones as indicated in table 1, section II. Figure 4 shows SRS agents space-time distribution (a-c) and their pheromone distribution (b-d) for v=0 (a-b) (the target valley is constantly at B=(0;0)) and for v=0.5 (c-d) (the valley is traveling to southeast one cell at each two time steps the B mid-point), for several starting at time steps t=1,5,10,15,20, 30,50,70 and 100. In (b-d) black pixels show where each ant agent is placed in the environment at a precise time step, starting initially by a random distribution at t=0. On the other hand, pheromone distribution represents the societal memory and the recent history of the swarm (considered as his emergent social cognitive map [12,49,50]), and in a sense it contains distributed information embedded within the dynamic environment itself. This pheromone distribution is represented in grey levels, i.e., points in space with higher levels of pheromone are represented by darker pixels. The mapping from the amount of pheromone at each cell and its correspondent gray level is linear and coded in 8 bit images (i.e., 256 gray levels). Figures 5, 6 and 7 shows SRS results for speeds v=1, v=1.5, v=2, v=3, v=5 and v=10. These tests and respective figures show that for low values of dynamic environment speed (v=0, v=0.5, v=1 and v=2), the swarm can consistently track the target extrema while he moves through the entire space. This is perfectly visible for speeds v=0 and v=0.5 (fig.4) where the target remains moving to the southeast Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 9 corner over his 100 time steps run. As the target moves the swarm as a whole keeps following and converging to it, simultaneously self-regulating his population, i.e., downsizing it. At the same time, the whole swarm emerges a strong pheromone concentration at the right location of the target, increasing this concentration as times passes. For v=1 and v=1.5 however, due to speed, the target passes once at the southeast corner and reappears at t=50, 40 respectively for v=1, 1.5, at the northwest corner. In here, the swarm as a whole evolves a different global behavior. For v=1 and t=50 (fig.5), the swarm quickly adopts a short path (moving northeast) instead of moving ahead to northwest (where the target is), taking profit of the toroidal habitat. The same is true for v=1.5 and for t=40, whereas in here the swarm splits into two different clouds or clusters, one with the above short path strategy, while the other moving directly to it, following backwards the linear path (NW → SE) of the target dynamic. b) v=0 c) v=0.5 d) v=0.5 a) v=0 Fig. 4. SRS agents space-time distribution (a-c) and pheromone distribution (b-d) for v=0 (a-b) (the valley is constantly at B=[0.0;0.0]) and for v=0.5 (c- d) (the valley is traveling to southeast one cell for each two t time step starting at the B mid-point), for several time steps t = 1,5,10,15,20,30,50,70 and 100. When after some time steps the target is captured again the 2 clouds of agents congregate in one big cluster, and the chase continues with good tracking results. Somehow this emergent short path strategy is followed by the swarm, splitting in different clouds when needed for test speeds equal to v=2 or above that value (figs. 6,7). For dramatic values of speed however (v=5), the swarm is constantly and properly following the target but due to the value of severe speed, he is loosing the run. In order to overcame this lack of control the swarm itself self-regulates its population, and for dramatic speed tests of v=10 (fig.7c) or above, he explodes its population, trying to collect much spatial information as possible. Figure 8 can help us understand more profoundly the swarm behavior facing these different speeds. Fig. 8 (1st row) shows the population size as time passes. d) v=1.5 a) v=1 b) v=1 Fig. 5. SRS agents space-time distribution (a-c) and pheromone distribution (b-d) for v=1 (a-b) and for v=1.5 (c-d) for several time steps. In (a-b) t = 1, 5, 10, 15, 20, 30, 50, 70 and 100. In (c-d) t = 1, 5, 10, 15, 20, 30, 40, 50, 60, 70, 80, 90 and 100. Every test started with a number of ant-like agents equal to ⅓ of the habitat size (100 x 100 cells) as in [12,53]. We see that generally for speed tests up to v=5, the population more or less c) v=1.5 a) v=2 d) v=3 c) v=3 b) v=2 Fig. 6. SRS agents space-time distribution (a-c) and pheromone distribution (b-d) for v=2 (a-b) and for v=3 (c-d), for several time steps t = 1,5,10,15, 20,30,40,50,60,70,80,90 and 100. Note for e.g. v=2, that since the habitat is toroidal, after leaving the southeast corner at t=20 (columns a-b, 5th row) and at t=70 (columns a-b, 10th row) the target valley reappears at the northwest corner at t=30 (a-b, 6th row) and at t=80 (a-b, 8th row). a) v=5 d) v=10 c) v=10 b) v=5 Fig. 7. SRS agents space-time distribution (a-c) and pheromone distribution (b-d) for v=5 (a-b) and for v=10 (c-d), for several time steps t = 1,5,10,15, 20,30,40,50,60,70,80,90 and 100. For dramatic speed tests of v=10 (fig.7c) or above, the swarm looses control and explodes its population, trying to collect much spatial information as possible. Fig. 8. Population size versus time (1st row) and mean altitude (averaged fitness) versus time (2nd row), for several runs under different environmental speeds: v=0, 0.5, 1, 1.5, 2, 3, 5 and v=10. Check the space-time distribution of these ant agents and their pheromone allocation (social cognitive map or societal memory [12,49,50]) in figs. 4-5-6-7. Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 10 Fig. 9. Mean Altitude (averaged fitness) versus time for several runs. For comparison purposes, we plotted bold curves for static Ackley functions with respective global minimum at A(-1.5;1.0), B(0.0;0.0) and C(1.0;-1.5). The other curves are for different landscape change updated frequencies: uf=50, 25, 10 and 5. self-regulates under a common value, indicating a possible good tracking behavior accompanied by agent’s specialization (exploitation nature). On the other hand, figure 8 (2nd row) shows what happens to mean agent’s altitude over the Ackley landscape (F value on Eq. 8) as time go by. For speed tests up to v=2, the swarm can without difficulty track the extrema. For speed tests above that value, the behavior is more oscillatory, since at the toroidal borders the function assumes - due to simple and normal constraints in our representation - abrupt values from one border cell to their neighbors on “the other side”. While the extrema is reasonably well tracked, ant- like agents very near the minimal extrema (near borders) could easily have high altitude “on the other side”. Thus, even if the extrema is well tracked, mean altitude values achieve a visible increment. In addition, the swarm reaction speeds are undeniably very fast and the mean altitude quickly decreases often to values equal of those obtained with test speeds less or equal than v=2. This is particularly evident if we conduct a different and more dramatic test. In figure 9 we have updated the changes in the environment for different updated frequencies uf. For uf=5 (worst case) the base function minimal target point changes abruptly from B to C, from C to A, and from A to B, etc, at every 5 time steps. We see that even for this dramatic uf values, the swarm as a whole can still obtain performances, which in some cases outperform those obtained for static environments (bold curves). Reactions speeds are also quite high, generally in the range order of four to nine time steps. The SRS performance can also be measured and analyzed taking in account the best minimum value found among all agents at each time step (zero is equivalent to full performance or the perfect capture of the right extrema). In figure 10 we have plotted these values for different test speeds. As expected, as speed increases the swarm passes more time away from the right extrema or close to it. The swarm however is able to capture the right extrema in perfection, for the most part of the run-time (the oscillatory behavior is again mainly due to our toroidal borders). From these results we can obtain the SRS success rate (fig. 11), defined as the number of times the colony as a whole was able not only to track, but to capture the right target over a run of 100 time steps, for each test speed. Remarkably the swarm is generally able to capture the right perfect extrema around 65% of the time, up to environmental speeds of 10 times the speed of a single ant- like agent or even for speeds greater than those values. Finally, very similar results were found for circular dynamics as introduced on section I.B. Random dynamics were not tested. B. Speed Tests and Severity As discussed earlier in section I.B, Angeline [1] as well as Bäck [61,62], Branke, Schmeck [8] and others, studied not only how adaptive algorithms behave in drastic changing environments (as in the preceding subsection), but also on how the degree of these changes affect different approaches. v = 0.0 v = 0.5 v = 1.0 v = 1.5 v = 2.0 v = 3.0 v = 5.0 v = 10.0 Fig. 10. Best minimum value found by SRS among all agents at each time step (zero is equivalent to full performance or the perfect capture of the right extrema). Plots are for different test speeds. Fig.11. Averaged success rate (10 runs for each test speed). The success rate indicates the number of times the colony as a whole was able to track the right target over a run of 100 time steps, at each speed. δ=0 (e.g. s=0.1,T=0 or s=1,T=0) δ=0.1 (e.g. s=0.1,T=1) δ=0.3 (e.g. s=0.1,T=3) δ=0.5 (e.g. s=0.1,T=5) δ=0.6 (e.g. s=0.1,T=6) δ=0.7 (e.g. s=0.1,T=7) Fig. 12. Sketches from the Schaffer F7 complex multimodal function seen for different values of T related to the severity tests (here, s=0.1). The question is important, since can help us understand on how strongly the system changes are and how severe this change is going to constrain the search nature of an algorithm. For these types of tests as well as for comparison purposes we have decided to implement a test function used recently by Huang and Rocha [56,30,29] with a Co-Evolutionary agent- based model of Genotype Editing (ABMGE). This testbed is a dynamic version of the modified Schaffer’s F7 function studied in [29]. Several sketches of this multimodal function for different parametric values (different time) are illustrated in fig. 12. Being s instances in representing a parameter to set the severity and Xi = xi + δ(t), with -1 ≤ xi ≤ 1 for i = 1,2, a possible test problem [30] can be described by Equation 9: ⎛ → Xf ⎜ ⎝ ⎞ =⎟ ⎠ 5.2 − ( XX 2 + 1 ) 25.02 2 2 sin ⎡ ⎢⎣ ⎛ ⎜ ⎝ ( .50 XX 2 + 1 ) 1.02 2 ⎞ +⎟ ⎠ ⎤ 1 ⎥⎦ (9) In here, X and Y axis represent the index of the sample points in parameters x1 and x2 that are used to compute f(x), which is then represented on z axis, being our aim to maximize it. The example uses linear dynamics with severity s: Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 11 ( ) 0 δ ( ) T δ = = ,0 ( T δ ) 1 +− s (10) Fig. 13. Best-so-far performance of a standard GA and the co-evolutionary ABMGE proposed by Rocha and Huang [56] on the linear dynamic Schaffer F7 function with s=0.1 and uf=50, achieved by Huang and Rocha in [30]. Fig. 14. Best-so-far performance (best maximum found among all agents at each iteration) of the self-regulated swarm SRS approach against different severity values s=0.1, s=0.2, s=0.3, s=0.5, s=1.0, and s=1.5 (uf=50). Fig. 15. SRS swarm population size at each iteration, for tests in figure 14. As explained in [30], note that T is used as index for the environmental state; whenever the environment changes (e.g., every 50 generations in [30]), T is increased by 1. With these settings, f(x) has an optimal value of 2.5 among all ranges of s, here tested (this happens for s=0.1). This equation (10) will be used for the dynamic test function studied in the next section (IV), as well. In here we tested the SRS algorithm against severity values of s=0.1, s=0.2, s=0.3, s=0.5, s=1.0, and s=1.5 (for a constant updated frequency, uf = 50), as well as for different updated frequency’s uf = 50, 25, 10 and 5 (worst dramatic case), for a constant severity of s=1. SRS parameter values were the usual ones as indicated in table 1, section II. Figure 13 displays the results achieved by Huang [30]; averaged best-so-far performance for the Schaffer’s F7 (Eq. 9) dynamic problem, with traditional genetic algorithms as well as with the co-evolutionary RNA editing model (ABMGE) [56], for a single test with s=0.1 and uf=50. As outlined by the authors, these results are encouraging since the co- evolutionary ABMGE model consistently outperforms the traditional GA in tracking the extrema. Meanwhile, figure 14 displays the results achieved by SRS against not only s=0.1 as well as for other values of severity s (all with uf=50). We see that the self-regulated swarm is not only able to track perfectly the correct extrema, nearly at all iterations for s=0.1, as achieves similar results for increasing values of severity. Please note that when s increases, equation 10 returns a different domain for our test dynamic function (different sample points for different T), thus the right respective extrema have optimal values less and obviously different than those present for s=0.1 (z=2.5). Meanwhile, figure 15 displays the SRS population size over the same tests. Since generally, the optimal peak form is the same (due to translation), for different severities (which did not happen in section III.A tests), we now see SRS self-regulating its population size similarly for different test runs (SRS tend to converge his population to values equal of those optimal and near optimal areas; circular concentric regions in fig. 12). As usual, SRS tends also to use more exploratory resources as severity increases. Finally, we tested SRS against different updated frequencies. Respective results of performance and population size can be seen in figures 16 and 17. We see that even for the worst case scenario (uf=5 and s=1), SRS swarms are able to track the correct extrema in the majority of the test run-time. Naturally, there are oscillatory behaviors when the frequency updates, but again the swarm reaction speed is quite fast: in the order of ten to fifteen time-steps. As usual, at these kind of critical periods (environmental phase transitions) the swarm explodes his population (fig. 17), in order to overcome the abrupt changes in the landscape, using intelligently more resources to explore the novel sudden habitat. When again the extrema is correctly tracked after ten to fifteen time-steps, SRS self- regulates the number of their agents, downsizing it. Indeed, after the initial shock the resource economy is now so dominant and wise, that for certain cases the population decreases about nine times their earlier maximum size, settling down only at the correct peak. Fig. 16. Best-so-far performance (best maximum found among all agents at each iteration) of the self-regulated swarm SRS approach against different updated frequencies uf=50, 25, 10 and 5 (s=1). Fig. 17. SRS swarm population size at each iteration, for tests in figure 16. IV. DYNAMIC OPTIMAL CONTROL AND PROBLEM SOLVING BY EMERGENCE Optimal Control Theory [34,70] aims to determine the control signals that will cause a process to satisfy the physical constraints and at the same time minimize (or maximize) some performance criterion. Practically, this kind of nonlinear and multiple local optimal problems (generally) appear in all engineering and science fields, and have been well studied from both theoretical and computational perspectives [30]. Once more, for comparison purposes we adopt an earlier artificial optimal control problem designed in [29], further studied and reported in [30]. The constraints of the artificial optimal control problem are: ( ( ) tzd 2 dt 2 sin = ( ) tz sin + ( ) ut . 2 + 1 ( tz ,2 ′ = + ( ) tdz ( ) ) tz . dt ( ) ut cos . sin 2 + 2 [ ]1,0 ) t ,2 = ∈ 0 o cos ( ) t . sin ( ) uut . . 1 , 2 ( ( ) ) ( ) 3 tz tz . = (11) δ=4 (e.g. s=1,T=4) → region E in δ=0 (e.g. s=1,T=0 or s=0.1,T=0) → region A in fig.12. fig.12. δ=6 (e.g. s=1,T=6) → region G in δ=2 (e.g. s=1,T=2) → region C in fig.12. fig.12. δ=7 (e.g. s=1,T=7) → region H in δ=3 (e.g. s=1,T=3) → region D in fig.12. fig.12. Fig. 18. Sketch of the dynamic optimal control problem. Please note that the vertical z(tf)2 scale has changed for sketches T=6,7 since the highest peak for T=6 has now values slightly above 300, and slightly above 650 for T=7. An animated GIF can be found in [26]. where Ui = ui + δ(t), with -5 ≤ ui ≤ 5 for i = 1,2. As in [30], two examples are studied in this subsection that use linear dynamics with severity s=0.1 and s=1, according to equation 10 in the previous subsection, respectively. The goal is to Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 12 maximize z(tf)2 by searching for two control variables, u1 and u2 (-5 ≤ u1, u2 ≤ 5). Sketches of this dynamic functions for δ=0, 2, 3, 4, 6 and 7 are respectively illustrated in figure 18. X and Y represent the index of sample points in parameters u1 and u2, that are used to compute z(tf)2 , which is then plotted on the Z-axis. In order to compute z(tf)2 , the 2nd order differential equation (DE) in (11) must be converted into an equivalent system of two 1st order ordinary differential equations. The standard optimal control problem should now be represented by max J(u1, u2) = z(tf)2 , with tf = 1, z(t0)=2, y(t0)=2, dz(t)/dt = y(t), and dy(t)/dt given by Eq. 11. In order to numerically solve it for the different range values of δ (see Eq. 10), we have made use of a MATLAB version, working in reverse communication, of the Ordinary Differential Equation (ODE) solver DOPRI853 developed and coded in FORTRAN by Hairer et al. [24]. As in [30] we have completed tests for 400 iterations, with s=0.1 and s=1. Every 50 generations (uf=50), T (Eq. 10) is increased by 1. c) s=1.0 d) s=1.0 a) s=0.1 b) s=0.1 Fig. 19. SRS agents space-time distribution (a-c) and pheromone distribution (b-d), with uf=50, for s = 0.1 (a-b) and for s = 1 (c-d). Maximal regions travel in waves from southeast to northwest. Figures are for several time steps t = 1, 25, 75, 100, 125, 150, 200, 250, 300, 350,375,400. An animated GIF can be found in [27]. The SRS self-regulated swarm space-time agent distribution and respective pheromone distribution are displayed in figure 19. We have used SRS values of Δe=0.01 (remaining parameter values were the usual ones as indicated in table 1, section II). Figures 20 and 22, respectively show traditional GA and ABMGE results for s=0.1 and 1, achieved by Huang and Rocha [30,56]. They show that RNA editors [56], provide adaptive advantage for the ABMGE approach in tracking the extrema, while comparing it with traditional GAs. Fig. 20. Best-so-far performance of a standard GA and the co-evolutionary ABMGE proposed by Rocha and Huang [56] on the dynamic optimal control function (11) with s=0.1 and uf=50, achieved by Huang and Rocha in [30]. Fig. 21. Best-so-far performance (best maximum found among all agents for each iteration) of the self-regulated swarm SRS approach on the dynamic optimal control function (11) with s=0.1 and uf=50. On the other hand, figures 21 and 23, respectively show SRS results for s=0.1 and 1, achieved by our self-regulated swarm, clearly outperforming not only traditional GAs as the ABMGE model when facing dynamic environments. An exception however, should be made for t → [160,200] over the s=0.1 severity test, were SRS population exploded, due to sudden upcoming novel peak values over the domain. Figure 24 shows SRS population size as time passes, for the s=0.1 and s=1 tests. Generally, the standard GA and ABMGE does not have time, within 50 generations to exploit better peaks. As we can see Fig. 22. Best-so-far performance of a standard GA and the co-evolutionary ABMGE proposed by Rocha and Huang [56] on the dynamic optimal control function (11) with s=1 and uf=50, achieved by Huang and Rocha in [30]. Fig. 23. Best-so-far performance (best maximum found among all agents for each iteration) of the self-regulated swarm SRS approach on the dynamic optimal control function (11) with s=1 and uf=50. Regions A,B, …, H correspond to regions in figure 18. Fig. 24. Population size as time passes, of the self-regulated swarm SRS approach on the dynamic optimal control function (11) with s=0.1 and s=1 (uf=50). from these pictures see the SRS reaction speed is generally higher, maintaining good results for different time-steps. V. DISCUSSION A. Emergence of a Societal Memory and Intelligent Path Planning via Stigmergy As used in our Self-Regulated Swarms algorithm (SRS), stigmergy as a coordination mechanism is characterised by a lack of planning using implicit undirected communication between entities, a fact that makes it extremely flexible and robust for the exploration of large medium. The main idea relies on a form of asynchronous interaction and information interchange between entities mediated by an “active” environment. What characterizes stigmergy from other means of communication is, (1) the physical nature of the information released by the communicating insects, which corresponds to a modification of physical environmental states visited by the insects, and (2) the local nature of the released information, which can only be accessed by insects that visit the state in which it was released (or some neighbourhood of that state). Agents engage on what is known as a perception-action loop [36]. Actions in the environment can influence agent’s sensors, creating a loop. Relevant information flows from the environment via sensors to actuators, thus connecting them. The perception-action loop is important for understanding of adapted agents and to capture certain phenomena, like imprinting information onto the environment, offloading and later reacquisition of information. If we view sensors as capturing information and agent actuators as being capable of imprinting information on the dynamic environment, we can Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 13 then treat agents as creating, maintaining and using various information flows, both internal and external. The view may be quite useful since there are strong indications that biological agents are partly driven by the necessity of acquiring, exchanging and also concealing information [36]. One of the possibilities is uncovering hidden information, which was demonstrated by Kirsh and Maglio [35] using the perception-action loop. In this work [35], advanced Tetris game players are shown to quickly rotate a falling block while it still is not visible completely. This active modification of the environment allows the players to discover the shape of the block before it actually becomes completely visible on the screen. Other possibility is offloading of information into the environment. This can be illustrated by the fact that we often write notes or reminders, which we then later look at to reacquire some information. A good account of such information flow between several people is given in the analysis of how members of an airliner crew indirectly communicate using cockpit controls as a medium [31]. For example, long before landing one of the pilots calculates proper flap settings for various speeds and then based on the results sets special markers on the airspeed indicator. Later, during the landing phase, the markers allow the crew to quickly and reliably find out what flap settings to use for the momentary speed. In a wider context, indirect communication via the environment is of high importance in distributed Fig. 25. The curve shows the averaged agent’s y spatial coordinate as time passes, for the test realized in figure 2 (section II B) [17]. After t=250, the swarm keeps moving north, emerging an intelligent path planning and showing a strong behavioral phase transition. Grey curves show upper and lower bounds for the entire population, at each time step. is possible, since environmental systems. Flexibility perturbations are directly connected to ant’s behaviour: when the environment changes because of an external perturbation, the insects respond appropriately to that perturbation, as if it were a modification of the environment caused by the colonies activities. In other words, the colony can collectively respond to the perturbation with individuals exhibiting the same behaviour. When it comes to artificial agents, this type of flexibility is priceless: it means that agents can respond to a perturbation without being reprogrammed to deal with that particular instability. Over our dynamic tracking context, this means the possibility of achieving highly adaptive responses. Ants change the perceived environment of other ants (their cognitive map, according to Chialvo and Millonas [12,42,43]), and in every example, the environment serves as medium of communication. The problem in itself (dynamic environments) is implicitly used as a form of communication between agents trying to solve the problem. The environment is thus used as a global meta external memory, which is changing over time. This implicit use of the environment as memory allows ants to produce certain behaviour as a consequence of the effects produced in the local environment by previous behaviour. This meta global memory can be useful in cases where, for instance, a extrema reappears at a previous location or near from it, or when there is a necessity to maintain diversity. In the former case this environmental distributed implicit memory could recall that location and instantaneously move the population to the new optimum. As was pointed by Branke and Schmeck [8], it might also guide system’s evolution to promising areas after a re-initialization (if adopted). But while the memory might allow exploitation of knowledge gained in the past, it might as well mislead evolution and prevent it from exploring new regions and discovering of new extrema [8]. Nevertheless, we believe that the positive feedback (allocation of pheromone) / negative feedback (evaporation) based stigmergic-reproductive highly adaptive approach implement on our proposal helps reducing that risk. For instance, an incoming new and bigger peak to the environment, even if initially yet not recognized, may contribute to higher altitude on agents near by that region peak, thus triggering a major evaporation rate on other regions, and pressuring those agents to leave that old region or to die from it. On the other hand, SRS does not use explicit representation of goals, and the dynamics of group behaviour are emergent and self-organizing, allowing for the appearance of strong phase transitions in behaviour (fig. 25). Distributed cognition and intelligent path planning is one of these interesting behaviours. A good illustration of this behaviour (among others available from different tests in sections III and IV) can be accessed from a previous test [17] (section II-B). In here, swarms with varying population size were submitted to a the function. Simulating changes test particular in environment consisted on changing the task from minimization (t<250) to maximization (t>250). We see that swarm performance was convincing and reinforced the idea that the system is highly adaptable and flexible. Moreover we see the appearance of an emergent intelligent path planning. After minimization is completed (t<250), for t>250 swarms move from local optima to local optima peaks, avoiding completely valleys on their path. First the swarm divides into two exploratory clusters of agents, one taking a northeast direction and then dying out due to inappropriate local conditions, while the other moves first to northwest, taking advantage of some local peaks and then moving to northeast, from an intermediate phase to the last final optimum peak. B. Self-Regulated Swarms (SRS) and SOS As we have seen on the introductory sessions, many authors have tried to implement algorithmic solutions via the introduction of memory. Memory may be provided in two general ways: implicitly by using redundant or emergent representations (as in our SRS case), or explicitly by introducing an extra memory and formulating strategies to deposit and retrieve solutions later. In [9], Branke have compared a number of ways to organize an explicit memory (over a search population approach), and examined the usefulness of memory in different environments. As a result, he concluded that explicit memory is only useful when the environment repeatedly returns to a small set of previously Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 14 experienced solutions. As they advance in [8], although the memory / search population approach was quite successful, it soon became obvious that a strategy based on (explicit) memorization would be too restricted to adapt successfully to a wide range of experiments. As an alternative, Branke and Schmeck [8] developed more recently an approach with multiple populations acting as Self- Organizing Scouts (SOS), watching over the changing landscape. The novel idea is not only interesting as there are (SRS) to our Self-Regulated-Swarm similarities some interesting to point out, and from which future Self- Organizing approaches could ideally be designed. The basic idea of SOS is that once a peak has been found (i.e. the population converged to a high-performance region), the population should split: a small fraction, called the “child population” should “watch” over that peak, while the remainder of the population (“base population”) should spread out and continue searching for new peaks. When a watched peak moves, the child population may follow it through space, and even request reinforcement. Since the population size is limited, unpromising peaks may be abandoned. To answer questions like, how to determine peaks that justify a population split-off, when should sub-populations ask for request, how many individuals should stay at each peak, or when should a peak be abandoned, Branke and Schmeck borrowed some ideas from the Forking Genetic Algorithm (fGA) as proposed by Tsutui et al. [65]. From these global SOS features, we can already detect some prominent similarities with SRS, at the level of our main objectives (reinforcement, division of labor, the number of individuals at each prominent peak, removal of individual with low fitness, etc). However, there are many differences on how these goals are achieved. Even if both algorithms rely on principles of Self-Organization, Self-Regulated Swarms (SRS) achieve these goals mainly by emergence, and several decisions, like reinforcement of a particular region, the number of individuals at each peak, division of labor (population split –off) and individual removal from the global system, are completely done via implicit mechanisms, through pheromone local allocation and global evaporation (distributed cognitive map [12,50,49]), as well as on the reproductive capabilities of each foraging agent. VI. CONCLUSIONS In order to overcome difficult dynamic environment extrema tracking, we have proposed a Self-Regulated Swarm (SRS) advantageous the algorithm which hybridizes characteristics of Swarm Intelligence as the emergence of a societal environmental memory or cognitive map via collective pheromone laying in the landscape (balancing the exploration/exploitation nature of the search strategy), with a simple evolutionary mechanism that trough a direct reproduction procedure linked to local environment features is able to self-regulate the above exploratory swarm population, speeding it up globally. The different experiments carry out on sections III and IV, demonstrate that SRS is able for quick adaptive response, outperforming results not only with standard Genetic Algorithms, as well as comparing it with very recently successful approaches based on Co-Evolution. From the different SRS features, we highlight some successful behaviors found: (1) SRS is able to maintain a number of different solutions while adapting to new peaks appearing in the landscape. (2) The experiments show that the SRS approach is able to spontaneously create and maintain (for certain instable changing conditions), different subpopulations on different peaks, emerging different exploratory corridors with intelligent path planning capabilities. In addition, these population split-offs only occurs when needed. For instance, in between static conditions temporal windows (when the upgrade frequency determines that the changes will only occur after some T time steps), the population quickly converges to a single exploratory cluster. When however, the changes arrive, and for certain environmental conditions, the swarm may split again in two or more subpopulations. (3) For dramatic conditions, the swarm is able to request more exploratory agents, reproducing more and scaling up its population. After this is not necessary and redundant (a highly-performing peak is found) however, the swarm can self-regulate down their number, scaling it down, thus economizing its search resources. (4) The approach allows self-organized injection of larger diversity when needed, as we can perceive from grey curves in fig. 25. (5) For certain environmental conditions, the swarm not only is able to smoothly track the dynamic extrema, as for many time-steps over some different temporal windows, is able to capture it in perfection. (6) Last but not least, we prove that our SRS collective swarm of ant-like agents is able to track about 65% of moving peaks traveling up to ten times faster than the velocity of a single ant composing that precise swarm tracking system. This emerged behavior is probably one of the most interesting ones achieved by the present work. We expect the framework proposed in this work to advance the current state of research on Swarm Intelligence and Evolutionary Computation in dynamic Environments. ACKNOWLEDGMENT The authors would like to thank to Professor Fernando Durão (CVRM-IST, Technical Univ. of Lisbon, Portugal) for providing the MATLAB version, working in reverse communication, of the Ordinary Differential Equation (ODE) solver DOPRI853 developed and coded in FORTRAN by E. Hairer, S.P.Norset and G.Wanner [24]. The second author wishes to thank FCT, Ministério da Ciência e Tecnologia - Portugal, for his research fellowship SFRH/BD/18868/2004. Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 15 REFERENCES [1] Angeline, P.J., “Tracking Extrema in Dynamic Environments”, in Angeline, Reynolds, McDonnell and Eberhart (Eds.), Proc. of the 6th Int. Conf. on Evolutionary Programming, pp. 335-345, LNCS, Vol. 1213 , Springer, 1997. [2] Anthony Carlisle, Gerry Dozier, “Adapting Particle Swarm Optimization to Dynamic Environments”, in ICAI´00, International Conference on Artificial Intelligence, Las Vegas, Nevada, USA, 2000. [3] Anthony Carlisle, Gerry Dozier, “Tracking Changing Extrema with Adaptive Particle Swarm Optimizer”, in ISSCI, World Automation Congress, Orlando, Florida, USA, June, 2002. [4] Bäck, T., “On the Behavior of Evolutionary Algorithms in Dynamic Environments”, in IEEE Int. Conf. on Evolutionary Conference, pp. 446- 451, 1998. [5] Bierwirth C., Kopfer, H., “Dynamic task scheduling with Genetic Algorithms in Manufacturing Systems”, Technical Report, Departmen of Economics, University of Bremen, 1994. [6] Bierwirth C., Mattfeld, D.C., “Production scheduling and rescheduling with Genetic Algorithms”, Evolutionary Computation Journal, 7(1), pp. 1-18, 1999. [7] Bonabeau, E., Dorigo, M., Theraulaz, G., Swarm Intelligence: From Natural to Artificial Systems, Santa Fe Institute series in the Sciences of Complexity, Oxford Univ. Press, New York, Oxford, 1999. [8] Branke J., Schmeck, H., “Designing Evolutionary Algorithms for Dynamic Optimization problems”, In S. Tsutsui and A. Ghosh (Eds.), Theory and Application of Evolutionary Computation: Recent Trends, pp. 239-262. Springer, 2002. [9] Branke, J, “Memory enhanced Evolutionary Algorithms for Changing Optimization Problems”, in Congress of Evolutionary Computation, CEC’99, vol. 3, pp. 1875-1882, IEEE Press, 1999. [10] Branke, J, Evolutionary Optimization in Dynamic Environments, Kluwer, 2002. [11] Cedeno, W., Vemuri, V.R., “On the use of Niching for Dynamic Landscapes”, in Int. Conf. on Evolutionary Conference, pp. 361-366, IEEE, 1997. [12] Chialvo, D.R., Millonas, M.M., “How Swarms build Cognitive Maps”, In Steels, L. (Ed.): The Biology and Technology of Intelligent Autonomous Agents, 144, NATO ASI Series, 439-450, 1995. [13] Cobb, H.G., “An investigation into the use of hypermutation as an adaptive operator in Genetic Algorithms having continuous, time- dependent nonstationary environments”, Technical Report AIC-90-001, Naval Research Laboratory, Washington DC, 1990. [14] D. H. Ackley. “A Connectionist Machine for Genetic Hillclimbing”. Boston: Kluwer Academic Publishers, 1987. [15] Daniel Parrott, Xiaodong Li, “A Particle Swarm Model for Tracking Multiple Peaks in a Dynamic Environment using Speciation”, in CEC´04, Proc. of the 2004 Congress on Evolutionary Computation, p.98 - 103, IEEE Piscataway, 2005. [16] Dasgupta D., McGregor D.R., “Nonstationary function optimization using the structured Genetic Algorithm”, in Manner R., Manderick B. (Eds.), Parallel Problem Solving from Nature, pp. 145-154, Elsevier Science, 1992. [17] Fernandes, C., Ramos, V., Rosa, A.C., “Varying the Population Size of Artificial Foraging Swarms on Time Varying Landscapes”, in W. Duch, J. Kacprzyk, E. Oja, S. Zadrozny (Eds.), Artificial Neural Networks: Biological Inspirations, LNCS series, Vol. 3696, Part I, pp. 311-316, Springer-Verlag, Sept. 2005. [18] Ghosh A., Tstutsui S., Tanaka H., “Function Optimization in Nonstationary Environment using steady state Genetic Algorithms with aging of Individuals”, in IEEE Int. Conf. on Evolutionary Computation, pp. 666-671, 1998. [19] Goldberg, D.E., Smith R.E., “Nonstationary function optimization using Genetic Algorithms with dominance and diploidy”, in Grefenstette J.J. (Eds.), 2nd Int. Conf. on Genetic Algorithms, pp. 59-68, Lawrence Erlbaum Associates, 1987. [20] Grefenstette J.J., “Evolvability in Dynamic Fitness Landscapes”: A in Congress on Evolutionary Genetic Algorithm approach”, Computation, CEC 99, vol. 3, pp. 2031-2038. IEEE, 1999. [21] Grefenstette, J.J., “Genetic Algorithms for Changing Environments”, in Maener R. and Manderick B. (Eds.), Parallel Problem Solving from Nature 2, pp. 137-144, North-Holland, 1992. [22] Guntsch, M., Middendorf M., Schmeck H.: “An Ant Colony Optimization Approach to Dynamic TSP”, in L. Spector et al. (eds.) Genetic and Evolutionary Computation Conference, San Francisco, CA: Morgan Kaufmann pages 860-867, 2001. [23] Guntsch, M., Middendorf, M., “Applying Population Based ACO to Dynamic Optimization Problems”, in Ant Algorithms, Proc. of Third International Workshop ANTS 2002, Brussels, Belgium, Springer Verlag, LNCS 2463, pp. 111-122, 2002. [24] Hairer E. Norset S.P., Wanner G., “Solving Ordinary Differential Equations: I. Nonstiff Problems”, 2nd Edition, Springer-Verlag, Springer Series in Computational Mathematics vol. 8, 1993. [25] Holland, O., Melhuish, C.: Stigmergy, “Self-Organization and Sorting in Collective Robotics”, Artificial Life, Vol. 5, n. 2, MIT Press, 173, 1999. [26] http://alfa.ist.utl.pt/~cvrm/staff/vramos/DOCs=1land.gif. [27] http://alfa.ist.utl.pt/~cvrm/staff/vramos/DOCs=1P.gif. [28] http://tracer.lcc.uma.es/problems/ackley/ackley.html. [29] Huang, C.-F., A study of Mate Selection in Genetic Algorithms, Doctoral dissertation. Ann Arbor, MI: University of Michigan, Electrical Eng. And Computer Science, 2002. [30] Huang, C.-F., Rocha, Luis M., “Tracking Extrema in Dynamic Environments using a Co-Evolutionary Agent-based Model of Genotype Edition”, in Hans-Georg Beyer et al. (Eds.), GECCO’05 - Genetic and Evolutionary Computation Conference, pp. 545-552. ACM, 2005. [31] Hutchins, E., “How a cockpit remembers its speeds. Cognitive Science”, 19(3):265.288, 1995. [32] Janson, S., Middendorf, M., “A Hierarchical Particle Swarm Optimizer for Dynamic Optimization Problems”, in 1st European Workshop on Evolutionary Algorithms in Stochastic and Dynamic Environments, Coimbra, Portugal, 2004. [33] Kennedy, J. Eberhart, Russel C. and Shi, Y., Swarm Intelligence, Academic Press, Morgan Kaufmann Publ., San Diego, London, 2001. [34] Kirk, Donald E., Optimal Control Theory: An Introduction, Prentice Hall Publ., USA, 1970. [35] Kirsh D., Maglio P., “On distinguishing Epistemic from Pragmatic Action”. Cognitive Science, 18(4):513.549, 1994. [36] Klyubin, A.S, Polani, D., Nehaniv, C.L., “Tracking Information Flow through the Environment: Simple Cases of Stigmergy”, in Pollack J., Bedau M., Husbands P. and Ikegami T. (Eds.), ALIFE IX, Proc. of the 9th Int. Conf. on the Simulation and Synthesis of Living Systems, pp. 563-568, MIT Press, 2004. [37] Lewis J., Hart E., Ritchie G., “A Comparison of dominance mechanisms and simple mutation on non-stationary problems”, in Eiben, Bäck et al (eds.), Parallel Problem Solving from Nature, LNCS, Vol. 1498, pp. 139-148, Springer, 1998. [38] Liles W., DeJong K., “The usefulness of tag bits in Changing Environments”, in Congress on Evolutionary Computation, CEC 99, vol. 3, pp. 2054-2060. IEEE, 1999. [39] Lin, S.-C., Goodman E.D., Punch W.F., “A Genetic Algorithm approach to Dynamic job shop scheduling problems”, in Bäck, T. (Eds.), 7th Int. Conf. on Evolutionary Computation, Morgan Kaufmann, pp. 481-488, 1997. [40] Louis S.J., Xu Z., “Genetic Algorithms for open shop scheduling and rescheduling”, in Cohen M.E., Hudson D.L. (eds.), ISCA 11th Int. Conf. on Computers and their Applications, pp. 99-102, 1996. [41] Mattfeld D.C, Bierwirth C., “Minimizing job tardiness: Priority rules vs. adaptive scheduling”, In Parmee, I.C. (Ed.), Proc. of ACDM, pp. 59-67, Springer, 1998. [42] Millonas, M.M., “A Connectionist-type model of Self-Organized Foraging and Emergent Behavior in Ant Swarms”, J. Theor. Biol., nº 159, 529, 1992. [43] Millonas, M.M., “Swarms, Phase Transitions and Collective Intelligence”, In Langton, C.G. (Ed.): Artificial Life III, Santa Fe Institute, Studies in the Sciences of Complexity, Vol. XVII, Addison- Wesley, Reading, Massachusetts, 417-445, 1994. [44] Mori, N., Imanishi S., Kita H., “Adaptation to Changing Environments by means of the Memory based thermodynamical Genetic Algorithm”, in Bäck, T. (Ed.), 7th Int. Conf. on Evolutionary Computation, Morgan Kaufmann, pp. 299-306, 1997. [45] Mori, N., Kita H., Nishikawa Y., “Adaptation to a changing environment by means of the thermodynamical Genetic Algorithm”, in Voigt H.-M. (Eds.), Parallel Problem Solving from Nature, Vol. 1141, LNCS, pp. 513-522, Springer, 1996. Ramos, V., Fernandes, C., Rosa, A.C., Tracking Extrema over Dynamic Environments using Self-Regulatory Swarms. 16 [69] Yaniv Altshuler, Israel A. Wagner, Alfred M. Bruckstein, “On Swarm Optimality In Dynamic And Symmetric Environments”, in ICINCO´05, 2nd International Conference on Informatics in Control, Automation and Robotics, 2005. [70] Zhou K., Doyle J.C., Glover K., “Robust and Optimal Control”, in Automatica Journal, Elsevier Science, Vol. 33, Number 11, pp. 2095- 2095(1), Nov. 1997. ORIGINAL PAPER WITH FULL PICTURES AVAILABLE IN: http://alfa.ist.utl.pt/~cvrm/staff/vramos/Vramos-NIAS06.pdf [46] Mori, N., Kita H., Nishikawa Y., “Adaptation to a changing environment by means of the feedback thermodynamical Genetic Algorithm”, in Eiben, Bäck et al (eds.), Parallel Problem Solving from Nature, LNCS, Vol. 1498, pp. 149-158, Springer, 1998. [47] Passino, K.M., “Biomimicry of Bacterial Foraging for Distributed Optimization and Control”, IEEE Control Systems Magazine, pp. 52-67, June 2002. [48] Raman, N, Talbot, F.B, “The Job Shop tardiness Problem: a Decomposition approach”, European Journal of Operation Research, vol. 69, pp. 187-199, 1993. [49] Ramos V., Fernandes C., Rosa A.C., “On Ants, Bacteria and Dynamic Environments”, in NCA-05, Natural Computing and Applications Workshop, IEEE Computer Press, Timisoara, Romania, Sep. 2005. [50] Ramos V., Fernandes C., Rosa A.C., Social Cognitive Maps, Swarm Collective Perception and Distributed Search on Dynamic Landscapes, to appear in Brains, Minds & Media – Journal of New Media in Neural and Cognitive Science, NRW, Germany, 2006. [51] Ramos V., Muge F., Pina P., “Self-Organized Data and Image Retrieval as a Consequence of Inter-Dynamic Synergistic Relationships in Artificial Ant Colonies”, in Javier Ruiz-del-Solar, Ajith Abraham and Mario Köppen (Eds.), Soft Computing Systems - Design, Management and Applications, Frontiers in Artificial Intelligence and Applications series, vol. 87, pp. 500-509, IOS Press, Netherlands, 2002. [52] Ramos, V., “The Biological Concept of Neoteny in Evolutionary Colour Image Segmentation - Simple Experiments in Simple Non-Memetic Genetic Algorithms”, in Applications of Evolutionary Computation, (Eds.), EuroGP´01 – European Conf. on Genetic Progr., LNCS, Vol. 2037, pp. 364-378, Springer-Verlag, Berlin-Heidelberg, 2001. [53] Ramos, V., Almeida, F., “Artificial Ant Colonies in Digital Image Habitats: A Mass Behavior Effect Study on Pattern Recognition”, In Dorigo, M., Middendorf, M., Stuzle, T. (Eds.): From Ant Colonies to Artificial Ants - 2nd Int. Wkshp on Ant Algorithms, 113-116, 2000. [54] Ramsey, C.L., Grefenstette, “Case-based Initialization of Genetic Algorithms”, in Forrest S. (Ed.), 5th Int. Conf. on Genetic Algorithms, pp. 84-91, Morgan Kaufmann, 1993. [55] Reeves C., Karatza H., “Dynamic sequencing of a multi-processor system: A Genetic Algorithm approach”, in Albrecht R.F:, Reeves C., Steele N.C. (Eds.), Artificial Neural Networks and Genetic Algorithms, pp. 491-495, Springer, 1993. [56] Rocha, Luis M., Huang, C.-F., “The Role of RNA Editing in Dynamic Environments”, in ALIFE9, 9th Int. Conf. on the Simulation and Synthesis of Living Systems, Boston, Massachusetts, Sep. 12-15th 2004. [57] Ryan C., “Diploidy without Dominance”, in Alander J.T. (Ed.), 3rd Nordic Workshop on Genetic Algorithms, pp. 63-70, 1997. [58] Schwefel, H.-P., Evolution and Optimum Seeking, John Wiley & Sons, New York, 1995. [59] Spears, W., “Simple subpopulation schemes”, in Evolutionary Programming Conference, pp. 296-307, World Scientific, 1994. [60] Stephens C.R., Olmedo I.G., Vargas J.M., Waelbroeck H., “Self- Adaptation in Evolving Systems”, Artificial Life, 4(2), 1998. [61] T. Bäck., “On the Behavior of Evolutionary Algorithms in Dynamic Environments”, in IEEE Int. Conf. on Evolutionary Computation, pp. 446-451, 1998. [62] T. Bäck., Evolutionary Algorithms in Theory and Practice, Oxford University Press, 1996. [63] Theraulaz, G., Bonabeau, E., “A Brief History of Stigmergy”, Artificial Life, Vol. 5, n. 2, MIT Press, 97-116, 1999. [64] Trojanowski K., Michalewicz Z., Xiao, J., “Adding Memory to the Evolutionary Planner/Navigator”, in IEEE Int. Conf. on Evolutionary Computation, pp. 483-487, 1997. [65] Tstutsui S., Fujimoto Y., Ghosh A., “Forking Genetic Algorithms: GAs with search space division schemes”, Evolutionary Computation, 5(1), pp. 61-80, 1997. [66] Vavak, F., Jukes K., Fogarty, T.C., “Adaptive Combustion Balancing in Multiple Burner Boiler using a Genetic Algorithm with variable range of local search”, in Bäck, T. (Eds.), 7th Int. Conf. on Evolutionary Computation, Morgan Kaufmann, pp. 719-726 1997. [67] Vavak, F., Jukes K., Fogarty, T.C., “Learning the local search range for Genetic Optimization in Nonstationary Environments”, in IEEE Int. Conf. on Evolutionary Conference, ICEC’ 97, pp. 355-360, IEEE, 1997. [68] Wilson, E.O., The Insect Societies, Cambridge, MA., Belknap Press, 1971.
1204.6089	1	1204	2012-04-27T00:34:00	Multi-model-based Access Control in Construction Projects	[ "cs.MA", "cs.CE", "cs.CY", "cs.SE" ]	During the execution of large scale construction projects performed by Virtual Organizations (VO), relatively complex technical models have to be exchanged between the VO members. For linking the trade and transfer of these models, a so-called multi-model container format was developed. Considering the different skills and tasks of the involved partners, it is not necessary for them to know all the models in every technical detailing. Furthermore, the model size can lead to a delay in communication. In this paper an approach is presented for defining model cut-outs according to the current project context. Dynamic dependencies to the project context as well as static dependencies on the organizational structure are mapped in a context-sensitive rule. As a result, an approach for dynamic filtering of multi-models is obtained which ensures, together with a filtering service, that the involved VO members get a simplified view of complex multi-models as well as sufficient permissions depending on their tasks.	cs.MA	cs	Multi-model-based Access Control in Construction Projects Frank Hilbert TU Dresden Raimar J. Scherer TU Dresden Larissa Araujo USP Institute of Construction Informatics Dresden University of Technology Dresden, Germany Institute of Construction Informatics Dresden University of Technology Dresden, Germany Engineering School of Sao Carlos University of Sao Paulo Sao Paulo, Brazil [email protected] [email protected] [email protected] During the execution of large scale construction projects performed by Virtual Organizations (VO), relatively complex technical models have to be exchanged between the VO members. For linking the trade and transfer of these models, a so-called multi-model container format was developed. Considering the different skills and tasks of the involved partners, it is not necessary for them to know all the models in every technical detailing. Furthermore, the model size can lead to a delay in communication. In this paper an approach is presented for defining model cut-outs according to the current project context. Dynamic dependencies to the project context as well as static dependencies on the organizational structure are mapped in a context-sensitive rule. As a result, an approach for dynamic filtering of multi-models is obtained which ensures, together with a filtering service, that the involved VO members get a simplified view of complex multi-models as well as sufficient permissions depending on their tasks. 1 Introduction For the processing of large and complex construction projects, single independent organizations work together as Virtual Organizations (VO) [1], combining their core competencies [2, 3]. To manage the complexity in construction in which the costs incurred within a project vary, both physical and functional information split into various domain-specific application models (e.g. building model, cost model, time model, etc.) on the basis of which the work is running, typically discipline-specific [4]. The use of Building Information Modelling (BIM) is becoming increasingly important for collaborative processing [5]. The elements of the various specialized models are implicitly linked together (e.g. the component inside wall xy from the model building with the process concrete first floor from the process model). To transfer application models together with their links, in the German Project MEFISTO 1 and in the international Project HESMOS 2 , a multi-model container format (MMC) was developed which sum- marizes application models with their dependencies [6]. When using this MMC approach, however, the following aspects must be considered: • (A1) The MMC, based on the number and granularity of the application models, becomes relatively large. Because of distributed processing in Virtual Organizations, these multi-models are shared very often and the size of the model can lead to the obstruction of communication (e.g. the size of the IFC building model of a simple high-rise building is about 40 MB). • (A2) For the processing of individual project tasks only a relatively small part of the subject model is usually interesting. According to their current project responsibilities the agents are only in- terested, in most cases, in the specific aspects (e.g. building services, building spatial structures, structural analysis, etc.). 1http://www.mefisto-bau.de 2http://hesmos.eu Jeremy Bryans and John Fitzgerald (Eds.): Formal Aspects of Virtual Organisations 2011 (FAVO 2011) EPTCS 83, 2012, pp. 1 -- 9, doi:10.4204/EPTCS.83.1 c(cid:13) F. Hilbert, R. J. Scherer, L. Araujo This work is licensed under the Creative Commons Attribution License. 2 Multi-model-basedAccessControlinConstructionProjects • (A3) It is not necessary, and for reasons of data protection, may also not be desirable that all partners involved must know all the technical details of the project model. From these considerations, it is useful to facilitate the working with multi-models, the transmitting of model cut-outs or the generating of model views. For this purpose, a central filter and mapping service are currently being developed at the TU Dresden, which generate partial application models while maintaining the link structure between the application models [7]. This filtering service is used within this approach to enforce the definition of model views and covers the aspect of access and privacy when accessing multi models within virtual organizations. In this manner, an initiated approach, as to the integration of these filter services in a context-sensitive fine-grained access control, is described. The next section briefly describes the structure of the multi-model used. On this basis, in Chapter 3, different multi-model filter designs are considered; first the uses of multi-model templates for the generation of model views, and second, the filtering of elementary models for creating cut-outs. Then, in chapter 4, the idea of an extended context-sensitive role model for virtual organizations is pursued, and in chapter 5, a proposed methodology for using this model for the filtered access to multi-models follows. The last chapter outlines practical scenarios and gives an outlook into further development. 2 Basic Concepts 2.1 The multi-model container format In order to distribute project information together from the various semantically and structurally inho- mogeneous domain expert models and, on the other side, externalize the implicit relationship between the model elements, a generic link model was developed, which will be shown together with the asso- ciated application models within a multi-model. Then the link model in the form of an XML document is stored, the format of the application models is not specified. The multi-model was changed in a so- called multi-model container (MMC) which is within the multi-model container and separated in each model, described by metadata (see Figure 1). These multi-model containers are read into the specialized applications of the VO partners, modified and shared with other partners. More information about the structure and the use of MMC is given by Fuchs in [6]. MMC MulƟ- model container Building Model BIM OperaƟon Model COM Costs Model SPM Project Schedule TSM LM Container metadata Figure 1: Example Structure of the Multi-Model-Container F.Hilbert,R.J.Scherer,L.Araujo 3 2.2 Filter Concepts for Multi-Models In the filtering of multi-models it is necessary to distinguish between the pure selection of application models as they are defined for model-views (domain-views) with any setting of the desired granularity (see Figure 2) and of the filter model to generate model cut-outs (see Figure 3). MMC MulƟ-model container MMC MulƟ-model container Building Model BIM Building Model BIM OperaƟon Model COM OperaƟon Model COM Costs Model SPM Costs Model SPM Project Schedule TSM Project Schedule TSM LM LM Container metadata Container metadata Figure 2: Example of a Multi-Model-View Figure 3: Example of a Multi-Model-Cutout For handling different tasks, it is often not necessary for all application models to be available. The use of model views serves not only for the simplified representation of complex issues but also to increase the security against unauthorized data access and is achieved by the simplest form of filtering through the use of multi-model templates. Figure 4: filtering with multi-model templates In the MEFISTO Project seven exemplary different multi-model templates have been defined which are based on the phases of the project and selected in the first step of each different subset of application models (Building, Operation, Performance, Costs and Construction Site Model). In a second step, the associated link model can be reduced to the remaining links of the tray model elements. Then, the tem- plates demand not only the existence of application models in the multi-model, but claim additionally the compliance model-dependent levels of detail and processing status. It is conceivable that in a production model, the application building model (BM) and operation model (OM) are structured much more com- prehensively than within the model range. In a third step, these additional conditions are checked using the metadata within the container. In Table 1, examples of granularities are outlined (e.g. in the appli- cation model the process model-column templates of the frame schedule (0.1), coarse schedule (0.2)) to detail schedule (1)). Other various project plans are discussed in more detail in [4]. 4 Multi-model-basedAccessControlinConstructionProjects Project Phase Offer phase Offer phase Contract phase Contract phase Execution phase Task Offer request Offer delivery Negotiation Commissioning Scheduling MM-Template Tender Model Offer Model Negotiation Model Contract Model Tender Model BM OM PM CM CSM 0,5 0,5 0,5 1 1 0,1 0,2 0,2 1 1 0 0 0 0 0,7 0,1 1 1 1 1 0 0,5 0,7 1 1 Table 1: LOD of Application Models in Multi-model templates For the selection of individual object fields for further processing or to facilitate semantic multi- model filter, the use of specialized model cut-outs is desirable. Then these excerpts can turn into various, in part dynamic, criteria to be defined, for example as geometric part-models, time periods or (external) object properties of the model elements. The generation of task-specific ad-hoc multi-model views, which are combined to link models of BIM data with information from other application models is relatively complex and is a topic of current research. Katranuschkov describes in [7] an approach for the generation of BIM-based multi-model views, and a reference implementation, based on the use of open IFC toolset from the University of Weimar and HOCHTIEF that, with the use of the generic Model Subset Definition Schema (GMSD), focuses on these goals. Another challenge is the generation of partial geometric models, here under certain circumstances broken up into objects, because it must be generated at different levels of the objects detail. This aspect is the subject of current research within the project HESMOS and is not described in this work. 3 Context-dependent Access Control Camarinha-Matos differentiated in [8], exogenous and endogenous VO roles, with the former relation- ships reflecting outwards, such as interaction with the environment, customers, competitors and potential partners, and on the other side endogenous roles which are represented by the relationships within the VO. These endogenous roles handle the VO participants during the exercising of their cooperation and are defined in the VO-initiation, as well as other descriptive attributes, in the Organization Model. The classic role-based access control (RBAC) [9] defines the relationship in a simple way: WHO (project role) may WHAT (authorizations)? Hence, VO participants may bring different talents or skills into a VO, so that the VO actors are assigned with so-called potential roles. Both in [10] and in [11], it was found that the role assignment of the VO members is not rigidly fixed in practice, but contextual and dynamic role assignments are subject to restrictions and must therefore be context-dependently modelled and dynamically evaluated. Only with this combination, the observance of the principle of Separation of Duty [9] and the Principle of Least Privilege [12] can be ensured. We use an expanded role model for which the access control including the access context offers and determines illustrated in simplified form: WHO (current role) may WHAT (authorizations) with WHOM (object) in which SITUATION (context)? F.Hilbert,R.J.Scherer,L.Araujo 5 An example scenario: A player is entrusted with the tasks within the VO plan creation and plan evaluation. To ensure the Separation of Duty, this actor should not be allowed to control himself, so both roles can only be selected when accessing different objects. 3.1 The use of context information To add context-sensitive abilities to security mechanisms, context attributes must be described in a for- mal context model. The context attributes are not individually described within this context model, but they are assigned real world objects. This so-called deputy approach describes a basic concept for the representation of contextual information [13] and link distributed context attributes of the subject and the object model together with the environment context attributes. Using a context model now we can com- bine previously individually held context information so that situations can be precisely described. The relevant relationships for access decisions can be complex and involve various types and combinations of context information. For the modelling of the knowledge about their relationships, the impact of this information has to be formally described. For the definition of the access control policies, we use the Web Rule Language (SWRL) (Horrocks et al., 2003) to achieve flexibility and interoperability as well as easy administration. Some examples of SWRL rules: • Role assignment: (Actor(?a), Role(?e), Resource(?r), hasTarget(?a,?r), hasRole(?a,?e), hasOwner(?r,?a)) ->PermittedRole(?e) • Ownership: (Action(?a), Resource(?r), hasTarget(?a,?r), hasActor(?a,?x), hasOwner(?r,?x)) ->PermittedAction(?a) • Sstate dependence: (Action(?a), Resource(?r), hasTarget(?a,?r), hasObjectRestriction(?a,?x), has- PassRestriction(?r,?x)) ->PermittedAction(?a) 3.2 Structure of the VO model We have modelled the RBAC-Model in our VO-Ontology model with the classes Actor, Role, Permission and Restriction. Objects of the Class Restriction define the context sensitive Assignment-Restriction in the form of SWRL-rules. The Relation sr is modelled by the bijective mapping of the ObjectPropertys hasActorRoleAssignment and hasRoleActorAssignment, which connect the classes Actor and Role. Both, the ObjectPropertys hasPermissionRoleAssignment and hasRolePermissionAssignment con- nect the classes Role and Permission and reflect the Relation sr. The SubPropertys hasRoleRestriction and hasPermissionRestriction link the Class Restriction to the ObjectPropertys hasActorRoleAssignment and hasPermissionRoleAssignment and realize the functions fc(sr) and gc(pr) (in Figure 3 dotted). The Restriction-Objects contain the restrictions to be evaluated, where the Open World Assumption is valid, i.e. the modelled Relations sr and pr are true, if there are no Restrictions or the Restrictions can be evaluated to be true. 3.3 Dynamic Authorization Decisions Scholz describes in [15] mutual trust as the basis for VOs and the transparent handling of explicit rules and standards as the basis for this confidence. For the rules of a context-based access control for multi-models, we use the Fine Grain Access Control (FGAC) rules described by Franzoni in [16] 6 Multi-model-basedAccessControlinConstructionProjects Figure 5: VO Model with a context-sensitive expansion, which additionally accesses the context in which access decisions are integrated. The actual access role when accessing an object instance from a dynamic access control according to the access context, consists of various determined static (potential roles with permissions) and dynamic conditions (object status and subject status, subject-object relationships and project status). This consists mainly of an empty MMC with pre-populated metadata descriptions. Using these templates it is easy to determine whether a padded MMC corresponds to a template. In MEFISTO a VO-based ser- vice platform is used which registers all created MMC templates and the metadata of all transmitted application models. MMT Filter MMC MMC* Figure 6: Access Control with Multi-Model-Filterl We propose a context-sensitive access control in 5 steps: 1. The role of a requesting actor is determined by evaluating its potential VO roles and the access context. F.Hilbert,R.J.Scherer,L.Araujo 7 2. The User ID is checked in the VO registry and connected access rules are evaluated. 3. With the Task description in the core ontology, the necessary permissions are evaluated 4. The task-dependent Template is searched on the VO-platform. If there is no template, then it will return an unfiltered Multi-Model, according to the preferences, or the access will not be permitted. If there is a template, it is created on the basis of a container template a Multi-Model-Container through the base service. Then, the registered trade models use the service from the platform and search in the model registry for an application model in the desired resolution. 5. If an application model does not exist in the required granularity, the subject-specific model type of filter methods and services are requested and an application model with higher granularity and the desired level of detail are given as an argument. The application model produced is inserted into the Multi-Model-Container. 6. This Multi-Model-Container is transferred to the actor. 4 Related work The common standardised access model role-based access control model (RBAC) [9] is based on the con- cept of the user role as an intermediary between subjects and permissions and is purely subject-based. The properties of the objects have no effect on the access decision. Therefore, role assignments may be wrong after a state change of the entities (objects as well as subjects). Various proposals have extended the RBAC model with attribute-based access control capabilities [17] which also consider environment attributes (e.g. system status, time, date), subject attributes (e.g. age, location, proofs of identity) as well as object attributes (e.g. size, value, location, state of objects). Priebe et al. [18] proposes an attribute-based access control approach, but roles are subject- dependently modelled. Kumar et al. use in [19] context filters for role assignment, based on hash tables describing user context and object context- attributes. Franzoni et al. presents in [16] a fine-grained access control model, with focus on Semantic Databases, using ontologies to describe the subject-object relationships. Last but not least, Goslar de- scribes a context model with dynamic real-world object linking in an economic focus [13]. Common to all these approaches is an attribute based assignment of subjects to roles while the roles-permissions assignment remains static. 5 Outlook It has outlined an approach which covers the aspects mentioned in chapter 1: (A1) Due to one filter of the Multi-Model-Container, the communication between the partners was simplified and accelerated. Example, it is now possible to use Multi-Model-Container also on mobile devices. (A2) Only those application models necessary for the current processing situation were transferred to the Multi-Model- Container. (A3) The transferred application models have the exact resolution (size) needed for further processing. Both the major components of the filtering services as well as most of the required multi-model templates are still under development, making the approach only roughly sketched. Currently in the focus of the research at the TU Dresden are the methods for the generation of geometric partial models and scheduling time spans. 8 Multi-model-basedAccessControlinConstructionProjects Another development is focused on the filtering of MMC for transitive object properties of model elements. An example scenario for this comes from subsequent processing. Here, the complete reduction of MMC to a supplementary multi-model is desired which contains only subsequent positions in the application model specifications, corresponding to a reduced geometry and link model. Such application model sections, due to the dynamic component, cannot be generated with template models. References [1] MOESLEIN K., (2001): Die virtuelle Organisation: Von der Idee zur Wettbewerbsstrategie, In: Rohde, Markus , Rittenbruch, Markus , AufdemWegzurvirtuellenOrganisation, Physica-Verlag Heidelberg. [2] FISCHER K., (2007): Intelligente Agenten fur das Management virtueller Unternehmen, In: Information Management 1/96, Computerwoche Verlag GmbH, Muenchen. [3] KELLER M., (2007)): IT supported co-operation in construction projects, PhD thesis BerichtedesInstituts fuerBauinformatik, Heft6,978-3-86780-004-4 [4] KOECHENDOERFER B., LIEBCHEN J.H., VIERING M.G., (2006): Projektorganisation - Bau-Projekt- Management, Teubner,978-3-8351-9075-7 [5] EASTMAN C. M., (1999): Building Product Models: Computer Environments, Supporting Design and Construction, CrcPrInc.,0849302595 [6] FUCHS S., KATRANUSCHKOV P. and SCHERER R. J., (2010): A framework for multi-model collabora- tion and visualization, In Proc. ECPPM 2010, Cork, Rep. of Ireland, TaylorandFrancis,UK [7] KATRANUSCHKOV P., WEISE M., WINDISCH R., FUCHS S. und SCHERER R. J. , (2010): Bim-based generation of multi-model views, In Proc. CIB W78 2010. [8] CAMARINHA-MATOS L.M., AFSARMANESH H., (2008): On reference models for collaborative net- worked organizations, In InternationalJournalProduction,Vol.46,No9 [9] FERRAIOLO D. F., J. A. C., KUHN D. R., (1995): Role-based access control (rbac): Features and motiva- tions, In Proceedingsof11thAnnualComputerSecurityApplication. [10] SOHR K., DROUINEAUD M., AHN G.-J., (2005): Formal Specification of Role-based Security Poli- In Proceedings of the 2005 ACM symposium on Applied, cies for Clinical Information Systems, doi:10.1145/1066677.1066756. [11] HILBERT F., KATRANUSCHKOV P. and SCHERER R. J., (2010): Formal Fine-Grained Information Ac- In Proc. eChallenges e-2010 Conference, 27-29 October, Warsaw, Poland, cess in Virtual Organisations, IIMCInternationalInformationManagementCorp.Ltd.,Dublin,Ireland. [12] SALTZER J. H., SCHROEDER M.D., (1975): Protection of Information in Computer Systems, In Pro- ceedingsoftheIeee,doi:10.1109/PROC.1975.9939. [13] GOSLAR, K., (2006): Ein Integrations- und Darstellungsmodell fur verteilte und heterogene kontextbezo- gene Informationen, PhDTheses. [14] HORROCKS I., PATEL-SCHNEIDER P.F., BOLEY H., TABET S., GROSOF B. and DEAN M.., (2004): SWRL: A Semantic Web Rule Language Combining OWL and RuleML, In W3CMemberSubmission,. [15] SCHOLZ C., (1994): Die virtuelle Organisation als Strukturkonzept der Zukunft?, in DiskussionsbeitragNr. 30desLehrstuhlsfrBetriebswirtschaftslehre,UniversitaetdesSaarlandesSaarbrcken. [16] FRANZONI S., MAZZOLENI P., VALTOLINA S., BERTINO E., (2007): Towards a Fine-Grained Access Control Model and Mechanisms for Semantic Databases, In WebServices,2007.ICWS2007.IEEEInter- nationalConference,pp.993-1000,doi:10.1109/ICWS.2007.176. F.Hilbert,R.J.Scherer,L.Araujo 9 [17] YUAN E. and TONG J., (2005): Attributed based access control (ABAC) for web services, In IEEE International Conference on Web Services, Vols 1 and 2, Proceedings, 2005, pp. 561-569, doi:10.1109/ICWS.2005.25. [18] PRIEBE T., DOBMEIER W. and KAMPRATH N., (2006): Supporting attribute-based access control In Proc. 1st Int. Conf. on Availability, Reliability and Security, 2006, pp. 465472, with ontologies, doi:10.1109/ARES.2006.127. [19] KUMAR. A, KARNIK N. and CHAFLE G., (2002): Context Sensitivity in Role Based Access Control, In ACMSIGOPSOperatingSystemReview36(3),vol.36,2001,pp.53 -- 66,doi:10.1145/567331.567336.
1802.10446	2	1802	2018-11-12T18:13:01	Identifying Sources and Sinks in the Presence of Multiple Agents with Gaussian Process Vector Calculus	[ "cs.MA", "cs.AI" ]	In systems of multiple agents, identifying the cause of observed agent dynamics is challenging. Often, these agents operate in diverse, non-stationary environments, where models rely on hand-crafted environment-specific features to infer influential regions in the system's surroundings. To overcome the limitations of these inflexible models, we present GP-LAPLACE, a technique for locating sources and sinks from trajectories in time-varying fields. Using Gaussian processes, we jointly infer a spatio-temporal vector field, as well as canonical vector calculus operations on that field. Notably, we do this from only agent trajectories without requiring knowledge of the environment, and also obtain a metric for denoting the significance of inferred causal features in the environment by exploiting our probabilistic method. To evaluate our approach, we apply it to both synthetic and real-world GPS data, demonstrating the applicability of our technique in the presence of multiple agents, as well as its superiority over existing methods.	cs.MA	cs	Identifying Sources and Sinks in the Presence of Multiple Agents with Gaussian Process Vector Calculus Adam D. Cobb Richard Everett Department of Engineering Science Department of Engineering Science University of Oxford Oxford, United Kingdom [email protected] Andrew Markham Department of Computer Science University of Oxford Oxford, United Kingdom [email protected] University of Oxford Oxford, United Kingdom [email protected] Stephen J. Roberts University of Oxford Oxford, United Kingdom [email protected] Department of Engineering Science ABSTRACT In systems of multiple agents, identifying the cause of observed agent dynamics is challenging. Often, these agents operate in di- verse, non-stationary environments, where models rely on hand- crafted environment-specific features to infer influential regions in the system's surroundings. To overcome the limitations of these inflexible models, we present GP-LAPLACE, a technique for locating sources and sinks from trajectories in time-varying fields. Using Gaussian processes, we jointly infer a spatio-temporal vector field, as well as canonical vector calculus operations on that field. Notably, we do this from only agent trajectories without requiring knowl- edge of the environment, and also obtain a metric for denoting the significance of inferred causal features in the environment by exploiting our probabilistic method. To evaluate our approach, we apply it to both synthetic and real-world GPS data, demonstrating the applicability of our technique in the presence of multiple agents, as well as its superiority over existing methods. CCS CONCEPTS • Computing methodologies → Gaussian processes; Unsuper- vised learning; Multi-agent systems; • Mathematics of computing → Bayesian nonparametric models; • Applied computing; KEYWORDS Gaussian processes; potential fields; multi-agent systems; GPS data; animal tracking 1 INTRODUCTION Inferring the possible cause of an agent's behaviour from observing their dynamics is an important area of research across multiple Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. KDD '18, August 19 -- 23, 2018, London, United Kingdom © 2018 Association for Computing Machinery. ACM ISBN 978-1-4503-5552-0/18/08...$15.00 https://doi.org/10.1145/3219819.3220065 domains [2]. For example, in ecology, the increasing availability of improved sensor technology and GPS data enables us to learn the motivations behind animal movement [9, 31], assisting with animal conservation and environmental efforts. Other domains, such as robotics, use apprenticeship learning to construct reward functions to allow them to mimic their observations [1]. Typical solutions for inferring the cause of an agent's behaviour tend to exploit Inverse Reinforcement Learning (IRL) [25] and in- clude learning preference value functions [7] as well as utility, value, or policy surfaces from observed actions in a space. Unfortunately, these solutions often rely on the problem being easily framed as a Markov decision process [21] which is not always appropriate. In- stead, one may be interested in identifying an interpretable potential function, defined in continuous space, that can explain trajectories made by agents in a multi-agent system (MAS). This approach can be seen in previous work where agents are modelled as particles with their dynamics determined by a potential field [4, 5, 20]. In this paper, we consider observed trajectories influenced by a time- and space-varying potential function, and infer the spatio- temporal potential function from observed movement alone. To accomplish this, we present GP-LAPLACE: Gaussian Processes for Locating Attractive PLACEs1. By constructing Gaussian processes (GPs) [23], our method jointly infers a spatio-temporal vector field as well as canonical vector calculus operations on that field, al- lowing us to estimate a time-varying map of sources and sinks of potential influence on agents. There are three notable advantages to our approach. First, it is able to reason about the interaction between agents and the envi- ronment from only agent trajectories, without requiring knowledge of the environment. Second, it allows the potential field to be non- stationary, which more accurately reflects the real-world. Finally, by exploiting a probabilistic method, we obtain a metric for denoting the significance of inferred causal features in the environment. To demonstrate the generality of our method as a tool for ex- plaining agent and animal behaviour, we apply it to two distinct data sets. As an illustrative example, we evaluate our approach on a synthetic data set which we can compare to the true potential function (which isn't accessible in the real-world). Next, motivated by the existence of long-term GPS data, we apply our method to 1Code and data: https://github.com/AdamCobb/GP-LAPLACE We define Equation (4) as the predictive mean of the nth derivative with respect to any test points x∗ and Equation (5) as its correspond- ing variance. The choice of covariance selected throughout this paper is the squared exponential kernel, kSE(x∗, xi) = l2 exp (cid:18) (x∗ − xi)⊤Λ−1(x∗ − xi) −1 2 , (6) (cid:19) with x∗ and xi corresponding to a test point and training point respectively. The hyperparameter Λ, is a diagonal matrix of input scale lengths, where each element determines the relevance of its corresponding dimension. The output scale, denoted by l2, controls the magnitude of the output [24]. The choice of kernel is motivated by the desire to obtain smooth measures over arbitrary derivative functions and the ease by which the kernel can be differentiated and used in combination with Equations (4) and (5). The following formulae define the squared exponential kernel for the first and second order derivatives [17]: ∂kSE(x∗, xi) ∂x∗ = −Λ−1(x∗ − xi)kSE(x∗, xi) = Λ−1(cid:16)(x∗ − xi)(x∗ − xi)⊤Λ−1 − I (cid:17) ∂2kSE(x∗, xi) ∂x2∗ kSE(x∗, xi). (8) We note at this point that estimating derivatives using a joint GP model over the function and derivatives [6, 10] offers a benign noise escalation in comparison to numerical differentiation. (7) 2.1 Vector calculus with GPs We define a scalar potential function ϕ(x, t) of space x and time t. Furthermore, we define the time-dependent gradient of ϕ(x, t) according to (9) t for x ∈ R2. We can model this time-varying vector value function by a multi-input, multi- output GP with a three-dimensional input tuple consisting of X = (x, t). This GP is constructed by introducing a separable kernel [3], with respect to x, where Vt =(cid:2)Vx Vy such that(cid:20) Vx (cid:3)⊤ (cid:20) kx(X,X′) ∇x ϕ(x, t) = Vt (cid:18)(cid:20) µx (cid:21)(cid:19) (cid:21) (cid:21) ∼ GP Vy , µy 0 0 ky(X,X′) contains an independently learned kernel for each output dimen- sion. Applying Equation (4) to this GP model, by jointly inferring the derivatives in the x and y directions, gives the time dependent posterior for each random variable in the following tuple: (cid:18) Vx , Vy , ∂Vx ∂x , ∂Vy ∂y , ∂Vx ∂y , ∂Vy ∂x (cid:19) . t We combine these predictive derivatives using Equations (10) and (11) to infer probability distributions over the divergence and curl of the vector field2 V: ∇ · V = ∂Vx ∂x + ∂Vy ∂y , (10) 2Vectors i, j, k denote unit vectors of a 3-D Cartesian coordinate system. Figure 1: GP-LAPLACE tracking a time-varying attractor from observing four agent trajectories. The black and white marker indicates the true location of an attractor that is in- creasing in strength. a real-world GPS data set of pelagic seabirds [19], discovering a number of attractors which influence their behaviour. The rest of our paper is organised as follows: Section 2 intro- duces the GP and its derivative manipulations for vector calculus, followed by Section 3 where we present our model. In Section 4 we evaluate our method on a synthetic data set where we know the true potential function. We then demonstrate how our approach can interpret real-world data by applying it to the Scolopi's shear- water GPS data set in Section 5. Finally, we highlight the novelty of our work compared to the existing literature in Section 6 before concluding with future work in Section 7. 2 PRELIMINARIES As a requirement for our model, we introduce the Gaussian process, defined by its mean function µ(x) and covariance function K(x, x′) [23]. The mean and covariance functions are parameterised by their hyperparameters and encode prior information into the model, such as smoothness, periodicity and any known underlying behaviour. In our work, µ(x) is set to zero as our data is preprocessed to have a zero mean. We define a function, distributed via a GP, as follows: (1) Manipulating Gaussian identities using Bayes' rule gives formulae for the GP's posterior mean, f(x) ∼ GP(cid:0)µ(x), K(x, x′)(cid:1) . E[f (x∗)] = k⊤ xx∗(Kxx + σ 2I)−1y, (2) and posterior covariance, V[f (x∗)] = kx∗x∗ − k⊤ (3) where x∗ is a test point under question and σ 2 is the noise variance hyperparameter. xx∗(Kxx + σ 2I)−1kxx∗ , Any affine transformation of Gaussian distributed variables re- main jointly Gaussian distributed. As differentiation is an affine operation, applying this property to any collection of random vari- ables distributed by a GP, gives jointly Gaussian distributed deriva- tives, f′(x). For a test point x∗ and corresponding output f (x∗), the derivatives associated with the GP in Equation (1) are distributed with posterior mean, (cid:21) (cid:20) ∂n f (x∗) (cid:21) ∂xn∗ = ∂2nkx∗x∗ ∂xn∗ ∂xn∗ = ∂nk⊤ xx∗ ∂xn∗ − ∂nk⊤ xx∗ ∂xn∗ E (cid:20) ∂n f (x∗) ∂xn∗ V and posterior covariance, (Kxx + σ 2I)−1y, (Kxx + σ 2I)−1 ∂nkxx∗ ∂xn∗ . (4) (5) 2 -19.56-16.51-13.46-10.40-7.35-4.30-1.251.80(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∇ × V = i ∂ ∂x Vx j ∂ ∂y Vy k ∂ ∂z 0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = k (cid:18) ∂Vy ∂x (cid:19) . − ∂Vx ∂y (11) Additionally, by simple application of the appropriate operators, we may readily define the time-varying spatial Laplacian: ∇ · (∇ϕ) ≡ ∇2ϕ. (12) This (time-varying) Laplacian is of key importance as it defines sources and sinks in the spatial domain. It can be thought of as indicative of the flow in a vector field. Positive values denote attrac- tive regions, where the vector field indicates flow towards these regions, whereas negative values denote repulsive regions due to the vector field pointing in the opposite direction. 3 MODEL Our model builds upon the theory introduced in Section 2. The objective is to design a model that can indicate influential features in an agent's environment from observing their trajectories. A trajectory ζa for agent a is defined as a collection of times- tamped locations x ∈ R2 or tuples (x, t). The elements of the vector x are referred to as x and y for the purposes of this model and we continue to use the tuple X = (x, t) to refer to the domain of space and time. We also make the assumption that each agent acts ac- cording to a utility or potential function ϕ(x, t), which is dependent on space and time. Whilst interacting with the environment, each agent aims to maximise this utility function at all times. (cid:21)(cid:19) (cid:18) (cid:20) kx(t, t′) 3.1 Fitting to the agent trajectories The first component of GP-LAPLACE uses Equations (2) and (3) to fit a GP to each trajectory ζa. Using a single GP with a sepa- rable kernel, Equation (13) defines our GP prior over the x and y components of the path (cid:174)f: 0 p((cid:174)f) = GP . 0 0, ky(t, t′) (13) For each of the x and y components of (cid:174)f, a set of hyperparameters are learnt for the separable kernel. The input space of this GP model is time t and the output space consists of the x and y components of (cid:174)f. If an agent trajectory, ζa, consists of N data points, we can then further apply Equation (4) to infer higher order derivatives for each of the n data points, where we denote { (cid:219)fx , (cid:219)fy} and { (cid:220)fx , (cid:220)fy} as the first and second-order time derivatives in the x and y directions. Second-order derivatives are inferred at this stage, as we make the assumption that an agent acts in accordance with a second- order dynamical system, i.e. the agent obeys Newtonian dynamics. This assumption means that an agent's underlying utility induces a "force" of influence on the agent, thus generating an acceleration (we here take the 'influence-mass' of the agent as unity). More formally, this induces an acceleration equal to the derivative of the agent utility or potential: (cid:220)fx = (cid:220)fy = (14) , . ∂ϕ ∂x ∂ϕ ∂y Although we choose to infer second-order derivatives, the model is not limited to the assumption in Equation (14). The flexibility of our model means that we can also infer first-order terms, (cid:219)fx and 3 (cid:219)fy, along with other higher-order terms. Therefore, throughout the rest of this section, references to (cid:220)fx and (cid:220)fy can be considered easily interchangeable with these other derivative terms. When dealing with a multi-agent system of M homogeneous agents, a trajectory model can be calculated for each agent to form the set of joint distributions, (cid:110) p((cid:174)f, (cid:219)fx , (cid:219)fy , (cid:220)fx , (cid:220)fy ζa)(cid:111)M . a=0 From the posterior GP model we are able to jointly predict the ve- locity and acceleration at any point on an agent's trajectory for M agents. At this stage, we now have a collection of posterior deriva- tives and their corresponding location in X. If each of the agent trajectories has length N , then the size of X is M×N . The next layer of our model combines the outputs from the set of posterior distri- butions to construct a probability distribution over the extended agent environment. 3.2 Inferring the vector field and Laplacian In order to infer the gradient of the potential function, ∇ϕ(x, t), the set of inferred second derivatives for all M agents is propagated through a second GP, which also has a separable kernel model, as below: p (cid:32)(cid:34) (cid:16)(cid:174)V(x, t) { (cid:220)fx , (cid:220)fy , ζa}M (cid:35) (cid:34) The vector (cid:174)V(x, t) = (cid:2)Vx Vy post x post y = GP µ µ a=0 k , (cid:17) (cid:3)⊤ consists of two random variables (X,X′) 0 0 (X,X′) (cid:35)(cid:33) post y post x (15) k . that model the acceleration in the two axes and the superscript label 'post' refers to the calculated posterior mean and covariance. Equation (15) combines the M multiple agent paths into one model and enables predictions to be made at different points in space that are not constrained to a single agent trajectory as in Equation (13). The input-output pairs for this GP model are the x, y and t values in each ζa that correspond to the (cid:220)fx and (cid:220)fy values. cluded as The Newtonian assumption made in Section 3.1 is formally in- ∇ϕ(x, t) ∝ (cid:174)V(x, t). The distribution over the partial derivatives, (cid:20) ∂Vx ∂x (cid:21) , ∂Vy ∂y , ∂Vx ∂y , ∂Vy ∂x , can then be calculated from Equation (15) by applying Equations (4) and (5). We thus calculate a distribution over the divergence of (cid:174)V. It follows that this divergence is proportional to the Laplacian under the same assumption, ∇2ϕ(x, t) ∝ ∇ · (cid:174)V(x, t). (16) In particular, our interest lies in the estimation of the Laplacian of the utility function, as it indicates sources and sinks of the poten- tial function in the environment. In this context, we regard sinks as agent attractors and sources as agent repellers. We have therefore in- troduced a novel framework, which enables us to infer sources and sinks, in an unsupervised manner, to offer an explanation behind multiple observed agent trajectories. 3.3 Metric for locating significant attractors and repellers: Kullback -- Leibler divergence We now require a metric that is able to take advantage of having ac- cess to both the posterior mean and variance over sources and sinks in the environment. Therefore, our metric of change from prior field to posterior field is measured via the Kullback -- Leibler (KL) divergence [13]. The motivation for selecting the KL divergence comes from its ability measure a distance between two distributions. This provides a natural indication of the informativeness of spatial locations, at given times, and in the context of our application offers a measure of trajectory-influencing locations. Given the model at time t, each point in space has an associ- ated potential field distribution, defined via the GP posterior as a univariate normal distribution. The KL divergence can be readily calculated as the difference between two univariate normal distri- butions, namely the prior and posterior [8], as below: DKL(pprior pposterior) = 1 2 (µpo − µpr)2 (cid:18) σpo (cid:32) σ 2 − 1 + ln (cid:19)(cid:33) + , σpr pr σ 2 po σ 2 po (17) where (cid:17) , (cid:18) ∂Vx (cid:19) ∼ N(cid:16) pprior = N(µpr, σ 2 pr), pposterior = N(µpo, σ 2 po). following prior Laplacian at location Xi in space-time: We refer back to Equation (10) and (12) in order to calculate the ∂2kx(Xi ,X′ i) ∂2ky(Xi ,X′ i) + + + 0, ∂x ∂y2 ∂x 2 ∂y2 h2 y λ2 y ∂Vy ∂y where, ∂2kx(Xi,X′ i) . The hyperparameters hx and λx are the output and input scale lengths of the x-part of the separable kernel in the GP model, with hy and λy corresponding to the y-part. + ∂2ky(Xi,X′ i) ∂x2 = h2 x λ2 x As our interest lies in determining attractors and repellers in the field, a further addition to the KL divergence in Equation (17) is to multiply it by the sign of the posterior mean of the Laplacian. This multiplication carries over the prediction of negative sinks and positive sources, whilst measuring the divergence from the zero prior mean. We refer to this extension as the signed KL divergence: SDKL(pprior pposterior) = sign(µpo) DKL(pprior pposterior). (18) Large values in the SDKL indicate significant influential features in the agent environment, whereas small values around zero indicate that the model is likely to have reverted to its prior, the sensible prior being that there is no significant feature present. 3.4 Computational complexity In order to use GP-LAPLACE on large data sets, we must overcome the computational complexity associated with inverting the N × N covariance matrix, which is O(N 3). There is a vast amount of literature that aims to overcome this issue, such as [22], and we use the sparse GP approximation built into the python package gpflow [16]. This approximation is based on work by Titsias [28], which we choose to implement when the total number of data points, N , exceeds 1000. Using this sparse approximation allows our model to scale to large data sets, as will be shown in Section 5. 4 APPLICATION TO SYNTHETIC DATA As an illustrative example, we apply GP-LAPLACE to synthetic data, where the true potential function and its derivatives are known. Importantly, this example allows us to evaluate the performance of our approach. We are then able to build on these results for the real-world data set in Section 5, where we cannot possibly have access to the true potential function. This experiment consists of a multi-agent system of homoge- neous agents, whose dynamics we observe. Our goal is to infer, from trajectories alone, the underlying potential value function. We demonstrate that our approach is able to recover the potential field from a small number of agent trajectories by identifying the true sources and sinks. Therefore, we are able to recover trajectory- influencing locations in an unsupervised manner, with no prior knowledge of the environment. 4.1 Agent model Agents are modelled according to a second order dynamical system, whereby, at each time step t, the acceleration x, velocity x and position x, with η as the update increment, are given by: xt +1 = ∇ϕ(xt , t), xt +1 = xt + η xt +1, xt +1 = xt + η xt +1. 4.2 Agent potential function We define the causal potential function for our illustrative example as a Gaussian mixture model, where each Gaussian has both a time- varying mean µi(t) and covariance Σi(t) that change harmonically according to: αiN(cid:0)µi(t), Σi(t)(cid:1), K i =1 ϕ(x, t) = (19) where K defines the number of Gaussians and α is the weight of each corresponding Gaussian. For the purposes of our synthetic data, we can think of each Gaussian as a time-varying source or sink. In order to validate our results, the first and second order deriva- tives of ϕ(x, t) are calculated using the known derivatives of a multivariate normal distribution p(x) [18]: agent potential functions: (1) stationary attractors ϕ1(x, t) = 4 αiN(cid:0)mi , ci I(cid:1) K i =1 ∂p(x) ∂x = −p(x)Σ−1(x − m), ∂x∂x⊤ = p(x)(cid:16)Σ−1(x − m)(x − m)⊤Σ−1 − Σ−1(cid:17) ∂2p(x) . These derivatives are then used in conjunction with the vector calculus operations in Equations (10) and (12) to calculate the true Laplacian of the utility function, which is used as a ground truth in our experiment. For the purposes of our experiment, we define three different (2) varying-strength attractors ϕ2(x, t) = i =1 (3) rotating attractors ϕ3(x, t) = K αiN(cid:0)mi ,(sin(t + βi) + ci)I(cid:1) (cid:21) K , ci I(cid:1) αiN(cid:0)mi ⊙ (cid:20) cos t sin t i =1 where αi, βi and ci are constants, mi is a two-dimensional constant vector, and ⊙ denotes an element-wise product. Here, we set K = 2 for all three potential functions, along with enforcing αi > 0 to define attractors. Furthermore, we display the derivatives of these agent potential functions in Figure 2 to give an understanding of what they look like in practice. (a) Stationary attractors (b) Varying-strength attractors (c) Rotating attractors Figure 2: We display the three example vector fields intro- duced in Section 4.1. Their time-varing properties are dis- played by including three frames at different time-steps. Along with testing the performance of our model, we can also compare the vector fields in Figure 2 to those occurring in nature. As an example, a study by Sommer et al. [27] tracked fruit bats as they moved between their camps and foraging sites. As nocturnal animals, their foraging sites are time-varying attractors, whereby these sites become the most attractive at night. Therefore, it is important to test our model on the time-varying attractors displayed in Figure 2, where we have access to a ground truth, before applying our model to a real-world data set. 5 4.3 Experimental results For each of the three potential functions, we initialised M agents with a velocity of zero at random starting locations. The experiment stepped through 200 time-steps and our model is used to infer the vector field and Laplacian from the agent trajectories {ζa}M a=1 (see Equation (16)). As a baseline, we took a simple parametric function in x- and y-space ∇U = w0 + w1x + w2y + w3xy + w4x2+ w5y2 + w6x2y + w7xy2 + w8x3 + w9y3 (20) of order three to model the gradient of the potential function, com- parable to the approaches of [4, 5, 20]. Table 1 displays the results for the three experimental set-ups, where we have varied the number of M observed agent trajectories from 4 to 16. We vary the number of agents to demonstrate how the models behave and scale as they observe more agents. Each value in the table is the mean squared error between the inferred Laplacian ∇2 ϕ and the true Laplacian ∇2ϕ, along with their standard deviation. The results show that the parametric model was able to recover the general shape of ∇2 ϕ and often summarised the two attrac- tors as a bowl shape in the three experiments. In comparison, GP- LAPLACE was able to recover the exact shape of the attractors for all three experiments and therefore demonstrated the ability to model the non-stationary characteristics of the potential func- tions. A further result is that our model improved its accuracy as it observed more trajectories, which is a desirable property of the model, while the baseline did not consistently. In order to interpret the behaviour of our model, the subset of frames displayed in Figure 3 show how the attractors of the true potential function are tracked across time by both the posterior inferred Laplacian and the signed KL divergence. This figure high- lights how GP-LAPLACE is able to completely characterise rotating attractors from only observing four trajectories over 200 time-steps. In more detail, the signed KL-divergence gives a measure of sig- nificance of the inferred sinks, through taking into account the uncertainty. The black and white markers indicate the location of the rotating attractors given by the ground truth in the first row. For all three rows, regions of attraction are indicated by blue and it can be seen that the model is able to accurately track the true attractors across time. 5 APPLICATION TO REAL-WORLD DATA In this experiment, we investigate how GP-LAPLACE can be used to determine driving forces in the environment that impact an animal's behaviour. We begin with an overview of the data used, present our experimental results, and then conclude with a discussion. Unlike previous techniques for studying how animals interact with the environment, we don't rely on GPS data to build density maps to construct probability distributions of their likely locations [11, 14]. Furthermore, while more recent approaches have incor- porated time into these models [15], current methods do not focus on inferring the driving force behind animal actions and instead simply show where an animal is likely to be found. −202−3−2−10123−202−3−2−10123−202−3−2−10123−202−3−2−10123−202−3−2−10123−202−3−2−10123−202−3−2−10123−202−3−2−10123−202−3−2−10123Table 1: Results displaying the mean squared error between the true Laplacian and the inferred Laplacian. Each experiment consisted of 200 time-steps with the listed means and standard deviations calculated over 10 different random initialisations of the agents. Number of agents 4 6.38 ± 2.84 Stationary attractors - GP-LAPLACE 8.27 ± 0.84 Stationary attractors - parametric Varying-strength attractors - GP-LAPLACE 62.37 ± 99.9 23.59 ± 7.23 Varying-strength attractors - parametric 13.72 ± 20.81 Rotating attractors - GP-LAPLACE 10.75 ± 2.50 Rotating attractors - parametric 8 2.20 ± 2.81 8.22 ± 0.48 0.88 ± 0.37 19.15 ± 3.37 2.05 ± 0.96 8.00 ± 1.19 12 0.55 ± 0.27 8.02 ± 0.13 0.64 ± 0.11 17.67 ± 1.95 0.78 ± 0.16 6.73 ± 0.91 16 0.35 ± 0.39 7.99 ± 0.49 0.58 ± 0.14 19.19 ± 2.59 0.51 ± 0.15 6.69 ± 0.61 Figure 3: Top row: Laplacian of true utility function. Middle row: inferred Laplacian of utility function. Bottom row: signed KL divergence. The location of the global minimum of the true Laplacian is indicated via the black and white markers. The predicted locations of the sinks, given by both the signed KL divergence and the posterior Laplacian, align well with their true locations, therefore demonstrating the success of our model on identifying non-stationary attractors. 5.1 Data We apply our model to a subset of Scolopi's shearwaters (Calonec- tris diomedea) [26] GPS data to infer the location of influential features in the environment of a MAS of pelagic seabirds. We use the same data set from Pollonara et al. (2015) [19], made available in a Movebank data package [30]. In their experiment, shearwaters were released 400 km from the colony in the northern Mediterranean Sea. Using GPS trackers, the birds' trajectories were mapped and inferences were made about the way in which they navigated. Importantly, the birds were split into separate groups depending on which senses were inhibited. From these, we focus on two sets of birds consisting of the control set and the anosmic set, with the latter unable to use their sense of smell. Of note is that we do not constrain each trajectory to be equally spaced in time, nor do we require the GPS readings to be time aligned across the trajectories. 5.2 Experimental Results In Figure 4, we present frames at different time-steps since the four control birds were released (in chronological order from left to right). Following previous work in this area [4, 5, 20], we focus on the divergence of the velocities of the birds. The top row is our inferred Laplacian, pointing out possible sources and sinks. The bottom row is the signed KL divergence, giving a measure of the significance of these features. Put into context, the birds' colony is slightly to the east of the northern part of Corsica, which is the island shown in the right side 6 Frame:56Frame:77Frame:97Frame:113Frame:146-5.63-4.55-3.48-2.40-1.32-0.240.831.91TrueLaplacian-5.63-4.55-3.48-2.40-1.32-0.240.831.91InferredLaplacian-44.84-38.94-33.05-27.15-21.25-15.35-9.46-3.56SignedKLdivergenceof each frame. Therefore, we expect to see our model placing an attractive region in the vicinity of their nest (marker (5) in figure), which is evident from the dark blue region appearing in the top- right of each of the last three frames, which is an attractor we expect to see as the birds start to approach their nest. Furthermore, the first frame shows that the region in which the birds are released is inferred to be a source as the birds fly outwards from this location (marker 1 in figure). Whereas the control birds in Figure 4 tended to fly directly back to their colony, the anosmic birds flew north until they reached the southern coastline of France (marker (2) in figure). As pointed out in Pollonara et al. (2015) [19], these birds are thought to have relied more on visual cues associated with the coastline, rather than flying straight back to their nest. Figure 5 shows the mean of the signed KL divergence across time for both the anosmic and control set of shearwaters, normalised to be on the same scale. In Figure 5, a direct comparison can be seen between the two sets of birds. The left-hand side plot displays the mean across time for the control set of shearwaters, confirming that the bird colony is an attractor along with parts of the coastline of Corsica (marker (3) and (4) in figure). In contrast, the right-hand side plot for the anosmic shearwaters displays a different behaviour in both the distribution of the sources and sinks and in the trajectories themselves. As previously mentioned, the anosmic birds head North immediately after being released and use the coastline to recover their bearings (marker (1) and (2) in figure). Therefore our GP model, when taking the mean across time, clearly assigns the region along the coastline of southern France as an attractive region. Attributing the coastline as an attractive region to the anosmic birds agrees with the original suggestion by Pollonara et al. [19] that they navigate via 'following coastlines as a form of search strategy, or by recognition of land features previously encountered'. 5.3 Discussion A strength of our model is that there is no requirement to incorpo- rate prior information on the location of environmental features. In our application to the shearwater GPS data set, the model had no prior knowledge of the location of the birds' colony or the coastline of Corsica, yet is still able to identify them as significant features. Furthermore, a strength of our technique is the way it is able to display how attractors and repellers vary with time. It would not be desirable to build a model that labels the birds' release point as a source for the entire duration of the birds' flight, although a simple density map would clearly make this error. As shown in the frames of Figure 4, GP-LAPLACE only describes their release point as a source for the first few hours of the birds' flights and the colony only becomes an attractor towards the end of the time period. The results of this experiment mean that we are now able to take animal GPS data and infer which environment features are influ- encing their movements. We are not constrained by a requirement of modelling any details of the environment and therefore point to the generality of this technique as a flexible method that can be applied to many other data sets. The ability to model time-varying drivers in MASs points to further studies such as in Ellwood et al. [9], where we are interested in setting sampling rates for sensors 7 and also making predictions as to which area of the environment will be attractive at different times of the day. 6 RELATED WORK In this section, we refer to relevant work from both the GP literature and also from work relating to modelling GPS data. As already introduced as a baseline in Section 4.3, there exists other works which model agents as acting according to a potential field [4, 5, 20]. However, we show that GP-LAPLACE outperforms these methods due to the added flexibility of being able to model a time-varying potential function. There also exists alternative tech- niques for studying how animals interact with the environment, which often rely on using GPS data to build density maps to con- struct probability distributions of their likely locations [11, 14]. Although more recent approaches have incorporated time into these models [15], they do not focus on inferring the dynamics and the driving forces behind animal actions, which is one of our motivations for introducing GP-LAPLACE. When relating our use of GPs to previous work with vector fields, Wahlström et al. [29] uses GPs to directly model magnetic fields. This work applies divergence-free and curl-free kernels to enforce linear constraints on their models, where the idea of constraining kernels to be divergence-free and curl-free is extended in [12]. Our work extends on previous work in this area with the introduction of a model that takes advantage of the derivative properties of GPs to give distributions over the operations of vector calculus from lower order observations. We have shown that inferring distributions over these operations results in an interpretable methodology that has not been previously employed in the literature. 7 CONCLUSION AND FUTURE WORK In this paper, we present GP-LAPLACE, a technique for locating attractors from trajectories in time-varying fields. Through apply- ing Gaussian processes combined with vector calculus, we provide a model that is able to infer sources and sinks in the presence of multiple agents from their trajectories alone. Additionally, our prob- abilistic technique enables us to utilise the KL divergence to give a measure of significance of environmental features. To demonstrate the generality of our method, we applied it to two data sets. First, on our synthetic data set, we showed that GP-LAPLACE is able to reconstruct non-stationary attractors ef- fectively, with superior performance over the baseline parametric model. Next, in an unsupervised fashion, we applied our model to a real-world example where we were able to infer features of the environment, such as the release point and bird colony, without prior knowledge of the birds' surroundings. In future work, we will extend GP-LAPLACE to incorporate explicit agent interactions. As an example, work by Preisler et al. [20] included an interaction term in their potential function to measure the strength of disturbances as having an effect on elk movement. We will also look into applying our model to additional real-world data sets, in situations where discovering sources and sinks in the presence of multiple agents will offer further insight into the motivating factors behind agent behaviour. Figure 4: Top row: Inferred Laplacian of utility function. Bottom row: Inferred signed KL divergence. Both rows are superim- posed on a map of the northern Mediterranean Sea. Each frame represents a snapshot of the sources and sinks relating to the velocity flow of the four shearwaters for 48 hours starting from 11:00 pm 21st June 2012. Points of note: (1) Release point, (2) South of France, (3) North of Corsica, (4) South of Corsica, (5) Nest. Best viewed in colour. Figure 5: Left: Mean signed KL divergence for control shearwaters. Right: Mean signed KL divergence for anosmic shearwaters. Both plots contain trajectories of four separate shearwaters released at times similar times. Blue areas show attractive regions and the yellow denotes sources. The black outline corresponds to landmass, where the top is part of the southern French coastline and the island on the right is Corsica. Over a similar time period, the difference in behaviour can be seen from both the routes taken by the birds and the average placement of the sources and sinks. Points of note: (1) Release point, (2) South of France, (3) North of Corsica, (4) South of Corsica, (5) Nest. Best viewed in colour. ACKNOWLEDGMENTS Adam D. Cobb is sponsored by the AIMS CDT (http://aims.robots. ox.ac.uk) and the EPSRC (https://www.epsrc.ac.uk). Richard Everett is sponsored by the UK EPSRC CDT in Cyber Security. We also thank Ivan Kiskin for extensive comments and feedback. REFERENCES [1] Pieter Abbeel and Andrew Y Ng. 2004. Apprenticeship Learning via Inverse Reinforcement Learning. In Proceedings of the twenty-first international conference on Machine learning. ACM, 1. [2] Stefano V Albrecht and Peter Stone. 2017. Autonomous Agents Modelling Other Agents: A Comprehensive Survey and Open Problems. arXiv preprint arXiv:1709.08071 (2017). [3] Mauricio A Alvarez, Lorenzo Rosasco, Neil D Lawrence, et al. 2012. Kernels for Vector-Valued Functions: a Review. Foundations and Trends ® in Machine Learning 4, 3 (2012), 195 -- 266. 8 [4] DR Brillinger, HK Preisler, MJ Wisdom, et al. 2011. Modelling particles moving in a potential field with pairwise interactions and an application. Brazilian Journal of Probability and Statistics 25, 3 (2011), 421 -- 436. [5] David R Brillinger, Brent S Stewart, Charles L Littnan, et al. 2008. Three months journeying of a Hawaiian monk seal. In Probability and statistics: Essays in honor of David A. Freedman. Institute of Mathematical Statistics, 246 -- 264. [6] PR Brook, A Karastergiou, S Johnston, M Kerr, RM Shannon, and SJ Roberts. 2016. Emission-rotation correlation in pulsars: new discoveries with optimal techniques. Monthly Notices of the Royal Astronomical Society 456, 2 (2016), 1374 -- 1393. [7] Wei Chu and Zoubin Ghahramani. 2005. Preference Learning with Gaussian Processes. In Proceedings of the 22nd international conference on Machine learning. ACM, 137 -- 144. [8] John Duchi. 2007. Derivations for Linear Algebra and Optimization. Berkeley, California (2007). [9] Stephen A Ellwood, Chris Newman, Robert A Montgomery, Vincenzo Nicosia, Christina D Buesching, Andrew Markham, Cecilia Mascolo, Niki Trigoni, Bence Pasztor, Vladimir Dyo, et al. 2017. An active-radio-frequency-identification system capable of identifying co-locations and social-structure: Validation with -2.80-2.15-1.51-0.87-0.230.421.061.70Inferred Laplacian12345-4.17-3.00-1.82-0.650.531.702.884.06Signed KL divergence12345Mean signed KL divergence for control shearwatersMean signed KL divergence for anosmic shearwaters-5.38-4.25-3.11-1.97-0.840.301.442.571234512345a wild free-ranging animal. Methods in Ecology and Evolution 8, 12 (2017), 1822 -- 1831. [10] Tracy Holsclaw, Bruno Sansó, Herbert KH Lee, Katrin Heitmann, Salman Habib, David Higdon, and Ujjaini Alam. 2013. Gaussian Process Modeling of Derivative Curves. Technometrics 55, 1 (2013), 57 -- 67. [11] Jon S Horne, Edward O Garton, Stephen M Krone, and Jesse S Lewis. 2007. Analyzing animal movements using Brownian bridges. Ecology 88, 9 (2007), 2354 -- 2363. [12] Carl Jidling, Niklas Wahlström, Adrian Wills, and Thomas B Schön. 2017. Linearly constrained Gaussian processes. arXiv preprint arXiv:1703.00787 (2017). [13] Solomon Kullback and Richard A Leibler. 1951. On Information and Sufficiency. The annals of mathematical statistics 22, 1 (1951), 79 -- 86. [14] Peter N Laver and Marcella J Kelly. 2008. A Critical Review of Home Range Studies. Journal of Wildlife Management 72, 1 (2008), 290 -- 298. [15] Andrew J Lyons, Wendy C Turner, and Wayne M Getz. 2013. Home range plus: a space-time characterization of movement over real landscapes. Movement Ecology 1, 1 (2013), 2. [16] Alexander G. de G. Matthews, Mark van der Wilk, Tom Nickson, Keisuke. Fujii, Alexis Boukouvalas, Pablo León-Villagrá, Zoubin Ghahramani, and James Hens- man. 2017. GPflow: A Gaussian process library using TensorFlow. Journal of Machine Learning Research 18, 40 (April 2017), 1 -- 6. http://jmlr.org/papers/v18/ 16-537.html [17] Andrew McHutchon. 2013. Differentiating Gaussian Processes. http://mlg.eng. cam.ac.uk/mchutchon/DifferentiatingGPs.pdf. (2013). [18] Kaare Brandt Petersen, Michael Syskind Pedersen, et al. 2008. The Matrix Cook- book. Technical University of Denmark 7 (2008), 15. [19] Enrica Pollonara, Paolo Luschi, Tim Guilford, Martin Wikelski, Francesco Bonadonna, and Anna Gagliardo. 2015. Olfaction and topography, but not mag- netic cues, control navigation in a pelagic seabird: displacements with shearwa- ters in the Mediterranean Sea. Scientific reports 5 (2015). [20] Haiganoush K Preisler, Alan A Ager, and Michael J Wisdom. 2013. Analyzing animal movement patterns using potential functions. Ecosphere 4, 3 (2013), 1 -- 13. [21] Martin L Puterman. 2014. Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons. [22] Joaquin Quiñonero-Candela and Carl Edward Rasmussen. 2005. A unifying view of sparse approximate Gaussian process regression. Journal of Machine Learning Research 6, Dec (2005), 1939 -- 1959. [23] Carl Edward Rasmussen. 2006. Gaussian Processes for Machine Learning. (2006). [24] Stephen Roberts, M Osborne, M Ebden, Steven Reece, N Gibson, and S Aigrain. 2013. Gaussian processes for time-series modelling. Phil. Trans. R. Soc. A 371, 1984 (2013), 20110550. [25] Stuart Russell. 1998. Learning agents for uncertain environments. In Proceedings of the eleventh annual conference on Computational learning theory. ACM, 101 -- 103. [26] George Sangster, J Martin Collinson, Pierre-André Crochet, Alan G Knox, David T Parkin, and Stephen C Votier. 2012. Taxonomic recommendations for British birds: eighth report. Ibis 154, 4 (2012), 874 -- 883. [27] Philipp Sommer, Jiajun Liu, Kun Zhao, Branislav Kusy, Raja Jurdak, Adam McKe- own, and David Westcott. 2016. Information Bang for the Energy Buck: Towards Energy-and Mobility-Aware Tracking.. In EWSN. 193 -- 204. [28] Michalis K Titsias. 2009. Variational Learning of Inducing Variables in Sparse Gaussian Processes. In AISTATS, Vol. 5. 567 -- 574. [29] Niklas Wahlström, Manon Kok, Thomas B Schön, and Fredrik Gustafsson. 2013. Modeling Magnetic Fields Using Gaussian Processes. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on. IEEE, 3522 -- 3526. [30] M Wikelski and R Kays. 2014. Movebank: archive, analysis and sharing of animal movement data. www.movebank.org. (2014). [31] Alan M Wilson, JC Lowe, K Roskilly, Penny E Hudson, KA Golabek, and JW McNutt. 2013. Locomotion dynamics of hunting in wild cheetahs. Nature 498, 7453 (2013), 185. 9
1203.0714	1	1203	2012-03-04T06:19:48	Towards an intelligence based conceptual framework for e-maintenance	[ "cs.MA" ]	Since the time when concept of e-maintenance was introduced, most of the works insisted on the relevance of the underlying Information and Communication Technologies infrastructure. Through a review of current e-maintenance conceptual approaches and realizations, this paper aims to reconsider the predominance of ICT within e-maintenance projects and literature. The review brings to light the importance of intelligence as a fundamental dimension of e-maintenance that is to be led in a holistic predefined manner rather than isolated efforts within ICT driven approaches. As a contribution towards an intelligence based e-maintenance conceptual framework, a proposal is outlined in this paper to model e-maintenance system as an intelligent system. The proposed frame is based on CogAff architecture for intelligent agents. Within the proposed frame, more importance was reserved to the environment that the system is to be continuously aware of: Plant Environment, Internal and External Enterprise Environment and Human Environment. In addition to the abilities required for internal coherent behavior of the system, requirements for maintenance activities support are also mapped within the same frame according to corresponding levels of management. A case study was detailed in this paper sustaining the applicability of the proposal in relation to the classification of existing e-maintenance platforms. However, more work is needed to enhance exhaustiveness of the frame to serve as a comparison tool of existing e-maintenance systems. At the conceptual level, our future work is to use the proposed frame in an e-maintenance project.	cs.MA	cs	Towards an intelligence based conceptual framework for e-maintenance Abdessamad Mouzoune QSM Laboratory - CEDOC / EMI, Mohammed V University – Agdal, Rabat, Morocco [email protected] Abstract. Since time when concept of e-maintenance was the introduced, most of the works insisted on the relevance of the underlying Information and Communication Technologies infrastructure. Through a review of current e-maintenance conceptual approaches and realizations, this paper aims to reconsider the predominance of ICT within e-maintenance projects and literature. The review brings to light the importance of intelligence as a fundamental dimension of e- maintenance that is to be led in a holistic predefined manner rather than isolated efforts within ICT driven approaches. As a contribution towards an intelligence based e-maintenance conceptual framework, a proposal is outlined in this paper to model e-maintenance system as an intelligent system. The proposed frame is based on CogAff architecture for intelligent agents. The architecture decomposes the agent into three layers (perception, central processing and action) with a hierarchical vision of three levels. W ithin the proposed frame, more importance was reserved to the environment that the system is to be continuously aware of: Plant Environment, Internal and External Enterprise Environment and Human Environment. In addition to the abilities required for internal coherent behavior of the system, requirements for maintenance activities support are also mapped within the same frame according to corresponding levels (strategic, tactical and operational management). A case study was detailed in this paper sustaining the applicability of the proposal in relation to the classification of existing e-maintenance platforms. The case study also underlined two issues of existing e- maintenance solutions regarding real integration of the system into human teams and more integration at the management strategic level. However, more work is needed to enhance exhaustiveness of the frame to serve as a comparison tool of existing e-maintenance systems. At the conceptual level, our future work is to use the proposed frame in an e- maintenance project that will have the two mentioned issues as an objective. Keywords: E-maintenance, Artificial Intelligence , Intelligent System , Framework, CogAff 1. Introduction 2 E-maintenance has been discussed in many maintenance related literature with different perspectives. E-maintenance basically refers to the integration of the information and communication technologies (ICT) within the maintenance strategy or plan (Chowdhury and Akram 2011). While perpetual development in ICT is providing new opportunities to the existing cultures of maintenance, e-maintenance seems to be predominately ICT-driven, meaning that the creation of innovative engineering solutions advanced the practice and research of e-maintenance (Haftor et al.2010). Reconsidering the distributed artificial intelligence environment generating e- maintenance system, this work aims firstly to highlight the importance of intelligence as a fundamental dimension that conceptual frameworks for e-maintenance may gain by being based on. The rest of the paper is organized as follows: Section2 reviews e-maintenance as an ICT based system. Section 3 explores e-maintenance as an intelligent system and a frame is then proposed to map it from an intelligence point of view. In section 4 a case study is detailed using the proposed frame. In the last sections, our proposal is discussed defining future work to be done. 2. Background: ICT driven e-maintenance models Maintenance is defined by the European Committee for Standardization (EN 13306:2001) as the combination of all technical, administrative and managerial actions during the life cycle of an item intended to retain it in, or restore it to, a state in which it can perform the required function (function or a combination of functions of an item which are considered necessary to provide a given service). The same standards defines maintenance management as all the activities of the management that determine the maintenance objectives or priorities, strategies, and responsibilities and implement them by means such as maintenance planning, maintenance control and supervision, and several improving methods including economical aspects in the organization. Maintenance Management includes corrective, preventive and proactive maintenance, inventory and procurement, work order system, computerized maintenance management systems (CMMS), reliability centred maintenance, total productive maintenance, financial optimization, technical training, and continuous improvement. From an organizational point of view, maintenance management must align actions at three levels of business Activities. Referring to (Márquez, Adolfo Crespo. 2007) ), this means: The Strategic level establishes maintenance priorities in accordance with business priorities. This priorities transformation materialized by a generic maintenance plan will establish critical targets in current operations. In addition, maintenance management at this level is responsible to decide on skills and technologies requirements to improve maintenance effectiveness and efficiency. The tactical level of maintenance management is responsible of the correct assignment of maintenance resources to fulfill the maintenance plan. Hence, detailed 3 maintenance requirements planning and scheduling are established at this level. Tactical maintenance policies are to be improved with experience. The operational level ensures that the maintenance tasks are carried out by skilled technicians, in the time scheduled, following the correct procedures, and using the proper tools. This level of maintenance management is also responsible of data to be recorded for diagnosis and/or prognosis purposes. In addition, Information and Communication Technologies (ICT) is considered as one of the maintenance pillars beside maintenance engineering methods and organizational ones (Crespo Marquez, Adolfo, and Jatinder N.D. Gupta.2006). In fact, CMMS are among the first historical steps of maintenance information systems (Rasovska et al. 2007). W ith the increasing development of ICT, other steps followed namely tele-maintenance, e-maintenance and semantic e-maintenance (s- maintenance). The main issue of these steps globally referred to as “e-maintenance ” is integrating different systems relative to maintenance into one information system. Since the time when concept of e-maintenance was introduced, most of the works insisted on the relevance of the underlying ITC infrastructure. Referring to (Arnaiz et al. 2010), technology sources are basically : Devices miniaturization increasing the ways data can be acquired including mobile systems. Extension of communication technologies including wireless and usage of the Internet as a main distributed platform for business operation. Supported by specific information standards for systems interoperability, such as MIMOSA and OSA-CBM, ICT, technologies include: - New Sensor Systems such as smart sensors —MEMS, perfo rming Condition- Based e-maintenance and RFID (Radio Frequency Identification Device), storage of conventional data - PDA (Personal Digital Assistant), SmartPhones, Graphic tablets, harden laptops, and other mobile devices, - W ireless technologies mainly Bluetooth (IEEE 802.15.1) and ZigBee (IEEE 802.15.4) , In short, (Muller et al. 2008) attributes the e-maintenance emergence its self to two main factors that are (1) E-technologies allowing optimization of maintenance related workflow; and (2) Integrating business performance, and dealing with openness, integration, and collaboration with the other services of the e-enterprise. Ample literature exists on design, development, and integration of e-maintenance within enterprise processes , including theories and practical applications, without any consistent definition as shown by a set of different e-maintenance definitions listed in (Arnaiz et al. 2010). In relation to our topic, none of those definitions explicitly mention intelligence as a fundamental characteristic of e-maintenance. Nevertheless, each implicitly recognizes such an importance especially when it comes to real time data processing and automation of some maintenance tasks. When explicit mention of intelligence is needed, an adjective is added to corresponding system as in “Intelligent Maintenance System ” reported in an editorial of “Com puters in Industry” (2006)): intelligence is defined as ‘ ‘the ability to monitor plant floor assets, link the production and maintenance operations systems, collect feedbacks from remote customer sites, and integrate it in upper level enterprise applications. ’ ’ In fact, from tele-maintenance to semantic e-maintenance, the major concern was integration, communication and interoperability in an ICT driven e-maintenance development. ICT is a fundamental dimension that enhances cooperative and 4 collaborative characteristics of maintenance activities. Being so, ICT are at the base of every definition of e-maintenance. The predominance of ICT dimension in textual frameworks represented by different definitions of e-maintenance is replicated on conceptual ground. Nevertheless, this predominance is more and more rectified by some conceptual frameworks to include strategic vision, business processes and organizations as in (Iung et al. 2009): Following the Zachman framework founded as an information system architecture framework (Zachman, John A. 1999) and later presented as an enterprise architecture, a comprehensive conceptual e-Maintenance framework is proposed based on five abstraction levels. These are (1) E-maintenance strategic vision (business and goal) which supports the ‘ ‘scope ’ ’, (2) E-maintenance bus iness processes which support the Business view, (3) E-maintenance organization which supports the Architect ’s view, (4)E-maintenance service and data architecture which supports the designer ’s view, and(5) E-maintenance IT infrastructure supporting the builder ’s view. Another conceptual framework aiming same objective to limit such an ICT predominance is presented in (Haftor et al. 2010) based on Information Logistics as a Driver for Development. Information logistics are defined as the provision of information for the knowledge worker within the correct context, enabling him or her to make the right decision (Logistics - W ikipedia. 2012). In general terms, though known frameworks, recognize the need for intelligence within e-maintenance systems, none of them goes beyond a shallow model of that dimension. Meanwhile, several industrial and academic e-maintenance platforms have been developed and are in use today. W ithin such platforms, artificial intelligence in general and machine learning algorithms in particular are by far important enablers to model maintenance tasks such as condition monitoring, prognostics, prognosis and decision support. For PROTEUS platform, a generic artificial intelligence (AI) template is defined in (Déchamp et al. 2004) to specify how AI tools must be integrated into the platform for a given diagnosis or prognosis task. More generally, (Muller et al. 2008) reports a state of the art of contributions in the field of e-maintenance classified according to required capabilities (i.e. Remote, Collaborative, Predictive maintenance and Knowledge capitalization )and needs intended to be responded ( i.e. Security/reliability of data, Interoperability, Maintenance integration, Collaboration/MAS, Processes formalization and Knowledge management). According to the mentioned report, few platforms such as TELMA and DYNAMITE undertake works on the whole specter of the chosen capabilities with respect to corresponding needs. Interesting case of IMS/D2B ™ is reported to b e focusing specifically on predictive maintenance and maintenance integration. The mosaic of contributions to enhance e-maintenance components with new AI based capabilities within known platforms are then rarely held in a holistic predefined manner. The way intelligence dimension is led bears yet a resemblance to a kind of more or less successful montage upon given ICT infrastructure. 5 3. Proposal: Towards intelligence based e-maintenance models Intelligence is the base of capabilities intended from e-maintenance concept at every business level strategic, tactical or operational levels. Hence, a global view of this ubiquitous dimension enhances mastering its contribution to global and even changing goals designed for an e-maintenance system. In this work, our intention is to contribute to converging efforts to satisfy real capabilities expected from e-maintenance as an intelligent system within a holistic approach towards an intelligence based framework. This can be understood as a framework where strategy aliened business view is immediately used to design a platform independent intelligence view and then, platform specific and ICT views identify technical possibilities to implementation. Such a framework is intending consensus between globalism in generally Zachman based frameworks and specialism of those focused on limited capabilities and needs. In this work, we focus on the platform independent intelligence view. In the general Zachman based framework proposed in (Iung et al. 2009), intelligence was mainly mentioned within the organization layer that must support the business process integration. After the mentioned paper, the organization expected for e-maintenance (based on collaborative –cooperative a spects) should be close to the IMS (Intelligent Manufacturing System) paradigm (Hiroyuki, and Yoshikawa. 1995) stating that the system behavior emerges through the dynamics of the interactions of basic maintenance agents within IMS the maintenance environment. Among weaknesses, modeling is mentioned in some literature as in (Thomas, Philippe, and André Thomas. 2011). After these authors, that weakness is due to the lack of uniformity in criteria to achieve modeling which makes it very difficult to compare different applications proposed in literature. Our proposal must then deal with such modeling weakness and permit comparison between different proposed e- maintenance solutions. In (Jacek Jakie ła, Bartosz Pomianek .2009), authors a rgue that agent orientation may be considered as a powerful paradigm for organization modeling and the reference architecture for Management Information Systems, that if properly applied, would lead to firm’s overall performance improvemen t. A Multi-Agent System is defined by the authors as a set (society) of decentralized software components, that are carrying out tasks collaboratively in order to achieve a goal of the whole society: those characteristics are among number of similarities between structural and behavioral characteristics of modern organizations and multi- agent systems. Consequently, we will think the whole e-maintenance system as an intelligent system –i.e. a system with artificial intelligence (Intelligent system – W ikipedia. 2012)- interacting with other constituents of the enterprise as part of its environment to achieve the global enterprise goals. At this point, our work reconsider in a direct and different manner the suggestion of “distributed art ificial intelligence environment ” when referring to e-maintenance in (Crespo Marquez, Adolfo, and Jatinder N.D. Gupta.2006). Agent metaphor will then be used in this paper to model e-maintenance system as a whole. Agent definition as well as agent modeling differs from an AI searcher to another according to their own objectives, presuppositions and conceptual frameworks. Sloman, Aaron, and Matthias Scheutz. (2002) introduce a framework for comparing 6 agent architectures: the general CogAff architecture is based on superimposing (1) the distinction between perceptual, central and action components, and (2) a distinction between types of components which evolve at different stages and provide increasingly abstract and flexible processing mechanisms. The reactive components generate goal seeking reactive behavior, whereas the middle layer components enable decision making, planning, and deliberative behavior. The modules of the third layer support monitoring, evaluation and control of the internal process in other layers and levels. We retain CogAff architecture as the main candidate to stand our proposal due to (1) its founding intention to serve as comparison tool of different solutions, and(2) similarities between its three level hierarchy and the three level of business activities here above mentioned. Furthermore, according to (Yan et al. 2003), agent roles and role models provide a vocabulary for describing agent systems, with each role describing a position and a set of responsibilities within a certain context or role model. This approach encourages us to think of the problem in terms of roles that need to be played, and the responsibilities associated with each role. Hence, in this section, each component of the CogAff architecture (perception, reactive, deliberative and reflexive layers and action) will be thought as the role intended from that component from a conceptual point of view. Multiple such roles may be assigned to a real agent within particular implementation. This way, the main objective of classification to meet the largest ways of implementations is respected. The proposed frame, referred to as E-CogAff in the figure here after (fig.1), summarizes our vision to e-maintenance system as an intelligent system called to be more aware of its entire -if not extended- environment. The rest of this section gives details of the frame. 7 Fig. 1 E-CogAff : an application of CogAff to e-maintenance system as an intelligent system. 3.1. Perception Layer The role of the perception horizontal layer is responsible of providing central processing layer with information based on data received from different Environment sources. Here, data are to be processed in a hierarchical manner and categorized into different levels of abstraction allowing the central layer to assess the situation of its environment and to make appropriate decisions. The most direct intelligent system abilities that are candidates to be classified within the perception horizontal layer are: - Data acquisition and pre-processing to detect and remove discrepancies such as outliers and missing data. - Data reconciliation - Data fusion that will help the planning layer assess the situation of the external environment and to make appropriate decisions. 8 - Database management system In general, the perception layer must deal with all aspects of data processing and categorizing issues. Data are either manually (e.g. Murphy, Glen D. 2009) or automatically (e.g. Gilabert et al. 2011) collected. Therefore, data quality is of a major concern within the perception subsystem. Several research studies have indicated that most organizations do have data quality problems (Lin et al. 2006). Accuracy, Completeness, Currency, Consistency and reliability are some of the common dimensions in literature dealing with data quality assessment within information systems as in (Piprani, Baba, and Denise Ernst. 2008). In fact, data quality dimensions characterize properties that are inherent to data. Such dimensions concern data values as well as logical schema or data format to satisfy specific set of semantic rules. Source reliability, defined as the credibility of a source organization with respect to provided data quality values, depends on the cooperative context in which data are exchanged. In general, security issues are to be tackled within this layer. 3.2. Central Processing Layer The central processing horizontal layer is composed with three hierarchical levels corresponding to strategic, tactical and operational levels of both internal capabilities of the system and different activities of the e-maintenance processes. Strategic/Reflective Level The Strategic/Reflective Layer is responsible of providing the ability to monitor, evaluate, and control other components of the architecture. Assuring required level of awareness of the system environment, this layer is to notice and categorize suspicious situations, and through deliberation or observation over an extended time period develop a strategy to deal with such situations. Furthermore, this level coordinates other modules so as to make the whole system performance more robust and coherent assessing credibility of each module's behavior by monitoring their internal states. In addition to monitoring, evaluating and control abilities regarding both internal behavior and e-maintenance activities performances , this level is also responsible of assistance in a cooperative manner for maintenance strategic managers. Its responsibility field covers maintenance contribution in an Executive Management Information System as well as interactive search for appropriate strategic goals and plans to maintenance activities within changing internal and external enterprise Environment. Such ability is of great help for maintenance managers dealing with uncertainty- especially within a risk based management approach (S öderholm, Peter, and Ramin Karim. 2011)- enabling a linkage of e-maintenance to the strategic goals of an organization. In such a case, artificial intelligence paradigms in this layer are to address issues concerning some kind of subjective and uncertain aspects of strategic management. Probabilistic methods as well as possibilistic ones such as fuzzy logic based algorithms are among solutions to deal with a certain level of such aspects. Tactical/Deliberative Level 9 Proactive behavior is achieved in the system in its deliberative layer, which is responsible for governing the system's actions in normal and abnormal circumstances. The choice of the appropriate artificial intelligence paradigm is very crucial to the high performance and real-time requirements of this layer to be endowed with appropriate abilities of planning, reasoning, learning and problem solving: Expert systems, case- based reasoning systems, neural networks are among solutions for such a requirement with different strengths and weaknesses to be taken into account in a given context. Such artificial intelligence requirements are to address different issues including exceptional processing regarding time or resources required by certain maintenance policies: reliability centered maintenance, opportunistic maintenance, health/care management, knowledge management and asset life cycle management. In general, in addition to the supervision of the system in coordination with the reflective layer, this layer is also responsible of interactive decision support towards tactical maintenance managers. Operational/Reactive Level The reactive layer provides direct responses to events that occur in the environment. Inspection and condition based maintenance is the major element of this layer. Many artificial intelligence paradigms can coexist in this layer with appropriate mechanism for collaborative problem-solving concept. This layer is also to support maintenance staff in their daily activities providing them with the right information at the right time and the right place, asking for required data and alerting in case of errors or lack of provided data among other interactive functionalities. 3.3. Action Layer The action layer concretizes plans that are decided in the central layer within a hierarchically organized manner. Among responsibilities of this layer are Maintenance Planning and scheduling, different presentation tasks towards human Environment, updating adequate internal and external Environment and eventually acting on the plant Environment. 3.4. Environment The intelligent system is intended to act on its perceived environment in a way to achieve its maintenance goals and thereby enterprise common goals. We decompose such an environment into four categories as follows: Human Environment 10 This includes all human agents that are to interact with the system in relation to maintenance. Be it management, executive or functional and technical staff or human experts, contractor ’s team members, constructor ’s a fter-sale responsible and so on. Some interactive activities are listed within the three levels of the central layer. Interactivity can be understood as simple information exchange as well as the way the system enables supporting individual maintenance team members in completion of their own tasks or the team as a whole. Plant Environment This environment includes mainly: - Sensors regarding physical assets condition (vibration, oil analysis …) as well as operational conditions (temperature, voltage,...) - Actuators in case of integrated control (emergency shutdown, …) or self- maintainable assets. Enterprise Environment In the enterprise environment, a distinction is made between internal and external environment to allow useful differentiation in the system perception regarding those environments. External environment needs more prudence such as evaluation of the credibility of a source organization. Internal Enterprise Environment The internal enterprise environment includes every system fully accessible to the maintenance system in a manner to allow it adjust its goals and its way to achieve them, as well as contributing in the achievement of the common goal of the enterprise. The main department in daily direct relation with maintenance are typically: Human resources (availability, skills, training, salaries …), Production (operational plan …), MRO Inventory and Purchasing (Maintenance Repair Operations requirement), Finances (costs ….) Regardless to their geographical position or the level of their intelligence, internal systems include Management and Executive Information Systems, Asset Management Systems, Enterprise Resource Planning and CMMS External Enterprise Environment External environment include every system not owned by the enterprise offering services to the maintenance system or cooperating occasionally or permanently to achieve predefined goals. Following systems are examples: Inter Enterprise databases such as reliability databases of a common industrial sector, external knowledge and experts systems, external catalogues and machines manufacturer’s ser vices that do not introduce human agents. This includes the case of exchange of reliability data for and from maintenance system especially when redesign is needed. 4. Case study: DINAWEB 11 In this section, we ’ll use our proposed frame to de monstrate its applicability in studying the case of Dynaweb platform: an ICT architecture related to software web services and communication architecture that support the e-maintenance concept within an integrated European project Dynamite (Dynamic Decisions in Maintenance). W ith its 16 partners, Dynamite Project is a real leader promoting e-maintenance concept either in (1)developing condition monitoring sensors and smart tags, (2) introducing mobile & wireless devices and technologies to support maintenance and diagnosis and prognosis based on WebServices or (3)developing training and economical studies related to maintenance strategy. This section is mainly based on information given in (Jantunen et al. 2009) as reported within the 4th World Congress on Engineering Asset Management, 28-30 September 2009, Athens, Greece. In that conference, important results of Dynamite were discussed with reference to project deliverables here after referred to as DWC followed by the number of some of them. The software architecture of DynaWeb is based on software components offering intelligent distributed web services and was demonstrated in architectures with logically structured decision layers based on the open systems architecture for CBM definition, from condition monitoring to decision support, and provides automated extraction of results(Irigaray et al. 2009). As explained in the main source of this section, different software modules are able to communicate with each other in order to perform a specific task. Indeed, in this context of Dynaweb, information processing is understood as a distributed and collaborative system, where there are different levels of entities that can undertake intelligent tasks. Hence, this case presents features of e-maintenance as an intelligence system as required by our proposed frame. The perception layer is materialized within Dynaweb by its MIMOSA database which follows the ISO 13374 standard: Machine Condition Assessment Data Processing Information Flow Blocks. The central processing layer is based on corresponding layers of OSA-CBM standard that are condition monitoring , health assessment, Prognostics and Decision support . W ithin this central we can recognize following constituents: The operational/reactive level of our frame that corresponds to DWC19 is related to state detection: Measurements received from sensors and their signal processing software are compared to expected values, and alerts are generated in case of anomaly detection. The tactical/deliberative level is well developed within that project. In case of an anomaly detected by condition monitoring modules DWC20 generates health records. Faults are then identified based on health history, operational status and maintenance tasks history. Primary focus of the prognostic module DWC21 is to calculate the future health of an asset and report remaining useful life, taking into account results from DWC 19 and DWC20. The strategic/reflexive level is yet focusing on the only maintenance costs aspect instead of a more global vision based on what maintenance provides and what are its global purposes and benefits. DWC24 refers to a set of modules developed within the Dynamite project for enabling economical analysis related to maintenance. Action layer in Dynaweb is mainly presentations, messages and generally information submitted to the destination of different environment constituents. In 12 addition to these functionalities, special attention was given to maintenance scheduling in DWC13 and DWC23. Plant Environment is by far the most important field concentrating efforts within dynamite project since its starting in 2005. A number of new sensors were developed in this project including vibration (DWC12) and on-line oil analysis sensors. In DWC3 , a multi sensors approach for monitoring vibration, pressure and temperature was concretized based on MEMS technology. Also, a wireless sensor network system, including ZigBee sensing nodes and Zigbee to W ifi gateways, was developed within DWC15 and DWC18. For on-line monitoring of lubricating oil five different types of sensors were developed within DWC4 to DWC8. Most of the developed sensors are based on optical methods. In addition, RFID tags was studied within Dynamite project for the identification of machines to be maintained (DWC1, DWC10, DWC2 and DWC11) and the report mentioned that more research are still needed for commercial use. In relation to the Human Environment, PDAs play a central role in the Dynamite project. The report explains that the PDA are used for accessing measurements data, studying the measurement results, communicating with the diagnosis and prognosis modules, handling of work orders, studying instructions for maintenance work and so on. Furthermore, in DWC9 it is assumed that it must be possible to use the PDA to support maintenance even if wireless connection is not available. Integrating e-maintenance system with its Enterprise Environment is considered within Dynamite project as a key to the success of current e-Maintenance solutions. Referring to the mentioned report, information integration has been pursued by adopting interoperable XML based data exchange formats, operational level data exchange protocols and maintenance-oriented data format definitions. Finally, networking integration has been also pursued through the combined use of wired and wireless communications. However, issues related to performance degradation due to wireless networking are mentioned in the report. Furthermore, CMMS and ERP as essential constituents of e-maintenance system enterprise environment are reported still not fully addressed. In deed, many CMMS and ERP solutions do not necessarily support required data interoperable formats. Finally, it’s useful to end this case study with so me results of industrial demonstrations related to DynaWeb solution. The aim of such industrial demonstrations was to integrate the technological and information technology strands in a business sense and to verify the effectiveness of the complete DynaWeb solution as reported in (Mascolo et al. 2010): The overall result is reported to be globally positive, with technical and economical feasibility proven. However, in an interesting case where the feasibility of integrating the DynaWeb components to form the e-maintenance architecture has been tested on the TELMA Platform, integration issues were still reported. The TELMA was developed mainly for supporting e-maintenance purposes, from both educational and research points of view. According to the same report, tests of feasibility and architecture aspect were held into (1)Internal test, to test the function of each ICT component of the DynaWeb platform and validate its ability to tackle real data, and (2) Integration test, to integrate all components and validate the ability of the Dynaweb platform to tackle an e-maintenance problem from sensor detection, to high level decision and scheduling. Concerning the internal tests, the results are reported to be positive while the results of the integration tests are less positive. According to the same report, this is mainly due to (1) MIMOSA compliancy was not ensured for all components, and (2) Missing functionalities according to use case requirements. Adjustments were required from the developer ’s or the end-user’s point of vie w to assign the components with the 13 right capacities needed for the final implementation on site. This last issue gives an idea of the further long way ahead to a real integration of e-maintenance as an intelligent system into human teams. 5. Discussion and future work The ISO 13374 is a condition monitoring architecture that consists of 8 functional blocks: The first two blocks are considered within the perception layer of our proposed frame, namely: - Data acquisition: Produces digital records with values and descriptive meta- information like time stamps and quality for the sensor/transducer output. - Data manipulation: Performs any kind of processing (e.g., signal processing) necessary in order for the sensor values to be meaningful with respect to state and health assessment. Follows, a reactive layer component: - State detection: Detects whether the asset is in a normal or abnormal state. The next three blocks implement the diagnosis and remediation intelligence of the system. These blocks are deliberative component: - Health assessment: Rates the current state as asset health and diagnoses any faults. - Prognostic assessment: Predicts the remaining lifetime until the next significant state change. - Advisory generation: Provides the user with recommendations on maintenance actions or changes of operational settings to optimize asset performance in the given state. The last blocks are action responsibilities towards some of the environment composers: - Information presentation: Presents the output of any of the previous stages to the user in easily understandable, textual, and graphical form. This block is among responsibilities of the action layer towards human environment. - External systems: Any external system the asset management application is connected to such as archiving, configuration/engineering and history. This block belongs to responsibilities of the action layer towards internal enterprise environment. External enterprise environment is not necessary excluded. (Naedele, M., P. Sager, and C. Frei.2004) As it is the case of Dynaweb here above studied in detail, this standard is at the base of several condition based maintenance solutions in the e-maintenance field. Thus, the first utility of the proposed frame is to provide some kind of classification - as inherited from the founding intention of its CogAff base - of at least architectures based on the mentioned standard. Focusing on condition monitoring, the standard has evident influence on the intelligence dimension of the corresponding solutions. Indeed, this dimension hardly goes beyond the deliberative layer neither deals with actions toward human environment within a real integration of the intelligent system into human teams. In fact, the predominance of the definition of e-maintenance as a part of e-manufacturing 14 and e-business pushed such architectures to focus on more automation and less human intervention. However, to improve the classification ability of the proposed frame, further work is needed to enhance exhaustiveness regarding responsibilities assigned to each role in the proposed frame. The second usefulness of the frame is to map existing intelligence abilities within an e-maintenance system in a given plant or platform and have an idea about intelligence requirements not satisfied. The projected intelligence level deduced from a requirement study can as well be mapped within that frame. This utility at the conceptual level of an e-maintenance project is at the core of our future work towards an intelligence based framework for e-maintenance.Related to this conceptual helpfulness investigation, we note (Sayda, Atalla F., 2008): one of the rare direct applications of the CogAff approach is described in that paper with an intelligent industrial production control system (ICAM). The proposed architecture of the conceptual system consists of four information processing layers and three vertical subsystems, namely, perception, central processing, and action. The lowest horizontal layer above the distributed control system (DCS) contains semi-autonomous agents that represent different levels of data abstraction and information processing mechanisms of the system. The middle two layers interact with the external environment via the DCS by acquiring perceptual inputs and generating actions. The perceptual and action subsystems are divided into several layers of abstraction to function effectively. While ICAM case comfort the feasibility of using the proposed E- CogAff frame in the conceptual stage of an e-maintenance project, in our future work we will attempt to tackle with the two main issues here above underlined: (1) more integration of the e-maintenance system at the strategic level of the enterprise and (2) more integration of e-maintenance as an intelligent system within human teams. 6. Conclusion This paper underlined the importance of intelligence as a fundamental dimension of an e-maintenance system. First, a review of the three level of maintenance management aligned with European standards is given. Then, e-maintenance is presented as an ICT supported information system. W ith a review of literature and known e-maintenance platforms and framework, we discussed the importance explicitly given to the intelligence of e-maintenance systems. According to our findings, this dimension suffers from the lack of frameworks explicitly dealing with the intelligence of e-maintenance systems in a holistic approach. As our contribution towards intelligence based conceptual framework for e- maintenance, we proposed a frame sustaining the eligibility of CogAff architecture as candidate for such a task. Our proposed frame aims to serve as (1) a classification means of existing e-maintenance solutions and (2) a platform independent conceptual tool within an e-maintenance project. To illustrate classification ability of the proposed frame, an application to the case of an e-maintenance leader –Dynaweb- was detailed. Finally, we discussed opportunities and limitations of our proposal. A brief application to ISO 13374 based e-maintenance solutions allowed us to highlight issues regarding (1) real integration of e-maintenance as an intelligent system into human teams, and (2) more mature implication at the strategic level of management. To serve as an accurate tool for e-maintenance solutions classification, our proposal needs further work to enhance its exhaustiveness. At the conceptual level, in our future work within 15 an e-maintenance project we will attempt to tackle with the two main integration issues here above mentioned (ie- at the strategic level and within human teams). 7. References 1) Arnaiz, Aitor, Beno ît Iung, Adam Adgar, Tonu Naks, Av o Tohver, Toomas Tommingas, and Eric Levrat. Information and Communication Technologies W ithin E-maintenance. In E-Maintenance, edited by S. Mekid K.Holmberg, E. Jantunen, A. Adgar, J. Mascolo, A. Arnaiz, 39 –60. Springer. (2010) 2) Chowdhury, S. and Akram, A. E-Maintenance: Opportunities and Challenges. 68 – 81. Turku Centre for Computer Science. (2011) 3) Crespo Marquez, Adolfo, and Jatinder N.D. Gupta. Contemporary Maintenance Management: Process, Framework and Supporting Pillars. In Omega 34, no. 3: 313 –326.(2006) 4) Déchamp L., Dutech A., Montroig T., Qian X., Racoceanu D., Rasovska B., Charpillet F., Jaffray J.-Y., Moine B., Müller G., Pa lluat N., Pelissier L.. On the Use of Artificial Intelligence for Prognosis and Diagnosis in the PROTEUS E- maintenance platform ”, in: IEEE Mechatronics & Robotics Conference (MECHROB). (2004) 5) Editorial Computers in Industry Special issue: E-maintenance 57, no. 6. 6) EN 13306:2001, (2001) Maintenance Terminology. European Standard. CEN (European Committee for Standardization), Brussels. (2006) 7) Gilabert, E., Jantunen, E, Emmanouilidis, C, Starr, A. and Arnaiz, A. Optimizing e- maintenance through intelligent data processing systems. In WCEAM 2011, the 6th annual World Congress on Engineering Asset Management, 2-5/10/2011, Cincinnati, Ohio, USA (SPRINGER). (2011) 8) Haftor, Darek, Miranda Kajtazi, and Anita Mirijamdotter. Research and Practice Agenda of Industrial e-Maintenance: Information Logistics as a Driver for Development. In Proceedings of the 1st International Congress on eMaintenance, 56 –61. Linnaeus University, School of Computer Scienc e, Physics and Mathematics. (2010) 9) Hiroyuki, and Yoshikawa. Manufacturing and the 21st Century — Intelligent Manufacturing Systems and the Renaissance of the Manufacturing Industry. Technological Forecasting and Social Change 49, no. 2 : 195 – 213. (1995) 10) Jacek Jakie ła, Bartosz Pomianek Agent Orientation as a Toolbox for Organizational Modeling and Performance Improvement. In International Book Series "Information Science and Computing" Book 13 - Intelligent Information and Engineering Systems 113-124. (2009) 11) Jantunen, E., E. Gilabert, C. Emmanoulidis, and A. Adgar. E-Maintenance: a Means to High Overall Efficiency. In Proc. 4th World Congress on Engineering Asset Management (WCEAM 2009), Athens, 28 –30. (2009) 12) Intelligent system - W ikipedia, the free encyclopedia.. Intelligent system - W ikipedia, the free encyclopedia. [ONLINE] Available at:http://en.wikipedia.org/wiki/Intelligent_system . [Accessed 01 March 2012]. 13) Information logistics - Wikipedia, the free encyclopedia. Information logistics - W ikipedia, the free encyclopedia. [ONLINE] Available at:http://en.wikipedia.org/wiki/Information_logistics . [Accessed 01 March 2012]. 16 14)Irigaray, Aitor Arnaiz, Eduardo Gilabert, Erkki Jantunen, and Adam Adgar. Ubiquitous Computing for Dynamic Condition-based Maintenance. In Journal of Quality in Maintenance Engineering 15, no. 2: 151 –16 6. (2009) 15) Iung, B., E. Levrat, A.C. Marquez, and H. Erbe. “Co nceptual Framework for e- Maintenance: Illustration by e-Maintenance Technologies and Platforms. ” Annual Reviews in Control 33, no. 2 : 220 –229. (2009) 16) Lin, Shien, Jing Gao, and Andy Koronios. A Data Quality Framework for Engineering Asset Management. In Engineering Asset Management, edited by Joseph Mathew, Jim Kennedy, Lin Ma, Andy Tan, and Deryk Anderson, 473 –482. London: Springer London (2006) 17) Márquez, Adolfo Crespo. On the Definition of Maintenance Management. In The Maintenance Management Framework, 3 –10. Springer Seri es in Reliability Engineering. Springer London. (2007) 18) Mascolo, Julien, Per Nilsson, Beno ît Iung, Eric Le vrat, Alexandre Voisin, Fernando Garramiola, and Jim Bellew. Industrial Demonstrations of E-maintenance Solutions. In E-Maintenance, edited by S. Mekid K.Holmberg, E. Jantunen, A. Adgar, J. Mascolo, A. Arnaiz, 391 –474. Springer. (2010 ) 19) Muller, A., A. Crespo Marquez, and B. Iung. On the Concept of E-maintenance: Review and Current Research. In Reliability Engineering & System Safety 93, no. 8 : 1165 –1187. (2008) 20) Murphy, Glen D. Improving the Quality of Manually Acquired Data: Applying the Theory of Planned Behaviour to Data Quality. In Reliability Engineering & System Safety 94, no. 12: 1881–1886. (2009) 21) Naedele, M., P. Sager, and C. Frei. Using Multi-agent Systems for Intelligent Plant Maintenance Functionality. 4:3059 –3063. IEEE. (2004) 22) Piprani, Baba, and Denise Ernst. A Model for Data Quality Assessment. In On the Move to Meaningful Internet Systems: OTM 2008 Workshops, edited by Robert Meersman, Zahir Tari, and Pilar Herrero, 5333:750 –7 59. Berlin, Heidelberg: Springer Berlin Heidelberg. (2008) 23) Rasovska, Ivana, Brigitte Chebel-Morello, and Noureddine Zerhouni, Classification Des Différentes Architectures En Maintenance . In proc. 7e Congrès international de génie industriel – 5-8 juin 2007 – Trois-Rivières, Québec (CANADA). (2007) 24) Sayda, Atalla F., Intelligent Control and Asset Management of Oil and Gas Production Facilities. University of New Brunswick, Dept. of Electrical and Computer Engineering. (2008) 25) Sloman, Aaron, and Matthias Scheutz. A Framework for Comparing Agent Architectures. doi=10.1.1.28.1945. (2002) 26) Söderholm, Peter, and Ramin Karim. An Enterprise Ris k Management Framework for Evaluation of eMaintenance. In International Journal of System Assurance Engineering and Management 1, no. 3: 219 –228. (2011) 27) Thomas, Philippe, and André Thomas. De La Nécessité Des Bonnes Informations Dans Les Systèmes Contrôlés Par Les Produits. In 7ème Conférence Internationale Conception Et Production Intégrées, CPI’2011, CDROM. Oujda, Maroc. (2011) 28) Yan, Qi, Li-Jun Shan, Xin-Jun Mao, and Zhi-Chang Qi. RoMAS: A Role-. Based Modeling Method for Multi-Agent. System.. 156 –161. Wo rld Scientific Publishing Co. Pte. Ltd. (2003) 29) Zachman, John A. A Framework for Information Systems Architecture. IBM Systems Journal 38, no. 2/3 : 454 –470. (1999)
1710.10753	2	1710	2017-11-21T19:36:27	Computational Social Choice and Computational Complexity: BFFs?	[ "cs.MA", "cs.AI", "cs.CC", "cs.GT" ]	We discuss the connection between computational social choice (comsoc) and computational complexity. We stress the work so far on, and urge continued focus on, two less-recognized aspects of this connection. Firstly, this is very much a two-way street: Everyone knows complexity classification is used in comsoc, but we also highlight benefits to complexity that have arisen from its use in comsoc. Secondly, more subtle, less-known complexity tools often can be very productively used in comsoc.	cs.MA	cs	Computational Social Choice and Computational Complexity: BFFs?∗ Lane A. Hemaspaandra Department of Computer Science University of Rochester Rochester, NY 14627, USA October 30, 2017; revised November 21, 2017 Abstract We discuss the connection between computational social choice (comsoc) and com- putational complexity. We stress the work so far on, and urge continued focus on, two less-recognized aspects of this connection. Firstly, this is very much a two-way street: Everyone knows complexity classification is used in comsoc, but we also highlight ben- efits to complexity that have arisen from its use in comsoc. Secondly, more subtle, less-known complexity tools often can be very productively used in comsoc. 1 Introduction Even in a work context, friendship is amazingly powerful. We all know how valuable it is to have collaborators who have each other's backs: working closely toward a shared project goal yet each bringing their own skills and perspective to the project, to seek to reach insights and advances that neither researcher could have achieved alone. Research areas can be like that too. Computational social choice (henceforward, comsoc)-which brings a computational-cost lens to such social-choice issues as preference aggregation/elections/manipulation/fair division-certainly has roots tracing far back; but comsoc's existence as a recognized, distinctive research area within AI/multiagent-systems is quite recent. Despite that, its growth as an area during these recent years has been ex- plosive, and we suggest that a synergy between comsoc and complexity has been partially responsible for that growth. At the 2017 AAMAS conference [LWDD17], for example, there were four sessions devoted to Computational Social Choice; no other topic had that many sessions. Those four sessions included 21 papers, and of those 21, 7 had the word "com- plexity" in their titles although in the entire conference only two other papers had the word complexity in their titles. In contrast, at the 2003 AAMAS conference [RWSY03], the string ∗A version of this paper will appear in AAAI-2018 [Hem18]. 1 "social choice" does not even appear in the ACM Digital Library online table of contents; neither does the string "election" or any form of the word "vote," and only three papers in the entire conference have the word "complexity" in their titles. One thing that those numbers are reflecting is the fact that comsoc and complexity have worked together very well. Comsoc research, from the seminal work of Bartholdi, Orlin, Tovey, and Trick [BTT89b, BTT89a, BO91, BTT92] onward, has done an enviably skilled job of interestingly employing complexity classification-which admittedly has strengths and weaknesses, both in its theory and in its application-to reveal a subtle and varied landscape as to what is difficult and what is hard, and how small changes in problems and models can induce seismic changes in complexity. Yet this article's primary goal is not to praise that praiseworthy achievement. It is to recognize two more subtle things that have already happened, but that we think need to be recognized for what they are, and what more can be done. Firstly, the comsoc/computational complexity friendship is truly a two-way street. Al- though everyone knows that complexity classification is widely used in comsoc research, we will stress a direction that is not well-recognized within the AI community or the complexity community: Research in comsoc has often been of great benefit to complexity theory. In par- ticular, complexity classification within comsoc has populated key complexity classes with problems that are undeniably natural and compelling. Why is that important? Complexity theorists like creating complexity classes that capture computational capabilities/resources that seem natural and compelling to the complexity theorists. But if those classes then turn out, for example, to have no natural complete problems, it is like hosting a party and having no one show up. It could even make one wonder whether the classes really were important at all. So complexity owes quite a debt to comsoc, for bringing parties to life. In addition to stressing the two-way street aspect of the comsoc/computational complex- ity friendship, we will also point out how rather subtle, lesser-known complexity notions and machinery have in the comsoc world found fertile ground to actually do important things. Complexity notions ranging from search-versus-decision issues to one-to-one reductions to the join operator have arguably been used in the comsoc world in more natural and more satisfying ways than they have been used within complexity theory itself. We present four different issues where the work in comsoc exemplifies one or both of these themes: comsoc work helping complexity, and lesser-known complexity notions and machinery being used to address comsoc issues. We will particularly stress examples that fall in the area of the complexity of elections/voting, probably the most intense current focus of comsoc and a remarkably challenging, interesting, nuanced area (for a general survey of comsoc see the excellent articles by Chevaleyre et al. [CELM07] and Brandt, Conitzer, and Endriss [BCE13], and for surveys of elections/voting within comsoc see the surveys by Faliszewski et al. [FHHR09], Faliszewski, Hemaspaandra, and Hemaspaandra [FHH10], and Faliszewski and Procaccia [FP10]; and for additional accessible surveys, see the relevant chapters in the important recent textbook edited by Rothe [Rot16]). This article gives just a few examples of these themes, but we try to discuss each of those with some context and care, and in two places our examples will connect complexity with other important AI areas, 2 planning and SAT solvers, once with the comsoc/complexity connection growing from the other AI area and once with it creating an opportunity within the other AI area. Then in the conclusion we will suggest that the two themes we have been stressing and the work to date on them make clear that comsoc experts and complexity experts working more closely and extendedly together will be to the benefit of both fields. 2 Make It a Party: Populating Lonely Complexity Classes As mentioned above, complexity classes are typically defined to capture a certain mode/power/model of computing. For example, the important probabilistic complexity class BPP [Gil77], bounded probabilistic polynomial time, is the class of those sets A for which there is a probabilistic polynomial-time Turing machine that on each input agrees with A with probability at least 75% (or 51%, or 99%; all three of those turn out to de- fine the same class). The class NP is the class of those sets accepted by nondeterministic polynomial-time Turing machines. The class UP [Val76], unambiguous polynomial time, is the class of those sets A accepted by nondeterministic polynomial-time Turing machines whose acceptance is always unambiguous; for inputs not in A the machine has no accept- ing computation paths, and for each input in A, the machine has exactly one accepting computation path (in the lingo, "has a unique succinct certificate"). This section will present a few examples of classes that were not wildly filled with natural sets that seem to capture their nature, yet comsoc provided new, or even the first, such examples. By "capture the nature" of a class, in the dream case we would mean "is complete for the class," i.e., our set is in the class and has the property that each set in the class many-one reduces to our set. But if that can't be achieved, finding a set that falls into a given class and does not seem to fall into any smaller class is the natural fallback step to take. Even readers not very familiar with complexity theory may know that there are thou- sands of known NP-complete problems and indeed a whole book is devoted to such prob- lems [GJ79]; and that there is an entire book devoted even to problems complete for P, with respect to the natural completeness type there [GHR95]; and that even for the levels of the polynomial hierarchy beyond NP and coNP, such as NPNP, coNPNP, NPNPNP , and coNPNPNP , there exists a detailed compendium [SU02]. And from that it would be easy to conclude that it is boring and easy to find natural problems complete for virtually any natural complexity class. But that seems not to be so. For example, NP ∩ coNP (the class of sets that have succinct, easily-checkable proofs both of membership and nonmembership) and two of the classes mentioned above, BPP and UP, are not known to have even one complete set-not a natural one, and not even an unnatural one [Sip82, HH88]. In fact, the just-cited papers and the work of Regan [Reg89] in effect are showing that the existence of a complete set for one of these classes would have sweeping consequences for our understanding of the class: that the issue of whether each of these classes has a complete set is in fact a disguised version of the question of whether there exists a nice enumeration of machines that precisely covers 3 the class, i.e., that provides what Regan calls a "constructive programming system." But things are worse still. It is now know that not only do these classes seem to lack complete sets, but in fact, it is plausible that-indeed there are black boxes (aka oracles) relative to which-even far larger classes than the given class do not contain even a single set that is many-one hard, or indeed even Turing hard, for the class [HJV93]. So it is not at all a sure thing that important classes-and BPP is undeniably important, since some would even argue that BPP rather than P should be the working definition of feasibility-have complete sets at all, much less natural complete sets. And yet some classes that had remained largely or wholly unpopulated by natural sets needing the class's power have had natural examples provided by problems from comsoc, and other less empty but far from crowded classes have also turned out to be what pinpoints the complexity of important comsoc problems. 2.1 Θp 2: Parallel Access to NP 2 ⊆ PNP. Perhaps more naturally, Θp Probably the most striking such example is the Θp 2 level of the polynomial hierarchy. This class was introduced by Papadimitriou and Zachos [PZ83], and is the class of sets that can be accepted by polynomial-time machines allowed to make O(log n) sequential queries to an NP oracle; NP ∪ coNP ⊆ Θp 2 is known to be the class of sets that can be accepted by polynomial-time machines allowed to make (an unlimited number of) nonadaptive (i.e., parallel) queries to an NP oracle [Hem89]. For example, consider the language, ParitySAT, that accepts its input exactly if the input is a list of Boolean formulas and the number of them that is satisfiable is itself an odd number. ParitySAT is clearly in Θp 2, since the polynomial-time machine can in parallel ask its oracle about the membership in SAT of each of its input formulas, and then can see whether the number that are satisfiable is odd. We will come back in a moment to the unnaturalness of this example and the search for natural examples. Θp 2 is a very important class in complexity theory. Briefly put, it has a large number of equivalent characterizations that are quite natural, e.g., it is the class of sets that can be accepted by logspace machines allowed to make an unlimited number of adaptive queries to an NP oracle [Wag90]; it is the class the polynomial hierarchy naturally is shown to collapse to if there are sparse (i.e., having at most a polynomial number of strings at each length) NP-Turing-complete sets [Kad89]; and its relationship to PNP is known to com- pletely characterize whether conversations with NP oracles can manufacture time-bounded randomness [HW91]. This at first sounds like quite a party, if one focuses on the results Θp 2 is central to. But wait. Even when all these results were known, the class was not known to have even one natural complete set. It was known to have a large number of rather unnatural complete sets all having to do with counting the parity of items, e.g., the ParitySAT example given above about telling whether out of a collection of Boolean formulas an odd number of them are satisfiable is known to be complete for Θp 2 [Wag87]. So in fact, as to known natural complete sets, the party was at that time a (complexity- theoretically important but) empty room. However, comsoc put first one and then many 4 people into that room. In particular, Charles Lutwidge Dodgson (aka Lewis Carroll) in the year 1876 had defined a fascinating election system [Dod76]. He was motivated by the idea that it would be nice for an election system to ensure that (a) if (relative to the votes, which are assumed to each be a tie-free total ordering of the candidates) there is a candidate- what is known as a Condorcet winner [Con85]-who is preferred in head-on-head contests against each other candidate, then that candidate is chosen the winner, and (b) otherwise, each candidate who is "closest" to being such a winner (in the sense that the number of adjacent exchanges in preference orders needed to make it become a Condorcet winner is the lowest among the candidates) is named a winner. And Dodgson's system then does just that: It counts distance from being a Condorcet winner, and the candidate(s) with the lowest distance wins. In doing that, Dodgson was in the 1800s already using the notation that in modern computer science is central and known as an edit distance. Although Dodgson's system is mathematically very well defined, one of the seminal papers of comsoc showed that the computational complexity of implementing his system is not low. In particular, Bartholdi, Tovey, and Trick [BTT89b] showed that it is NP-hard to tell if a given candidate wins such an election. They left open whether the problem is NP-complete, but eight years later the problem was proven to be complete for Θp 2 [HHR97a]. Θp 2-complete problems cannot be NP-complete unless the polynomial hierarchy collapses to NP ∩ coNP, and thus the winner problem for Dodgson elections is highly unlikely to be NP-complete. But the most important thing to note here is that the complexity of winner-testing for Dodgson elections is not a problem that was rigged to provide a Θp 2-complete set. The election system was defined in the 1800s for its own interest and importance, long before NP was ever dreamed of, much less Θp 2. This problem-from comsoc-thus provides an extremely natural Θp 2-complete set. This result sparked interest in whether other problems in comsoc and beyond might also be Θp 2-complete. And the floodgates opened. Other important election systems such as those of Kemeny and Young were proven to also be Θp 2-complete [HSV05, RSV03]. General tools were extracted from this approach to try to make it easier to prove such results [SV00], Θp 2-completeness was shown to capture the complexity of how well certain greedy algorithms do [HR98], and a survey was written looking at the meaning of improving from NP-hardness results to Θp 2-completeness results [HHR97b]. Even in the quite different field of automata theory, Θp 2-complete problems were found to capture important, natural notions [HM00]. Briefly put, the party was very much on! 2.2 NPPP and Other Classes There are many other classes that comsoc and other work within AI have helped pop- ulate. One of the most striking examples is the class NPPP, i.e., the sets solvable by nondeterministic polynomial-time Turing machines given access to (their choice of a) set from PP (probabilistic polynomial time, the class of sets for which there is a probabilistic polynomial-time Turing machine that on each input is correct with probability greater than 50%). NPPP is a class of great complexity, and it is not always easy to work with. In terms 5 of descriptive generality and its "location," Toda's [Tod91] Theorem says that PPP contains not just the polynomial hierarchy (PH) but even PPPH. So NPPP contains both of those and even NPPPPH , while being itself contained in PSPACE. It was only in the late 1990s that people started finding natural complete sets for NPPP, and that work came not from complexity issues but from an AI domain, the study of planning and decision processes. There are a number of papers on this and we won't here cite them all, but we mention as a pointer the very impressive work of Goldsmith, Littman, and Mundhenk [LGM98], which shows versions of both the plan-existence problem and the plan-evaluation problem are, for partially ordered plans, NPPP-complete, and of Mundhenk et al. [MGLA00] on Markov decision processes, which also yields natural NPPP-complete problems. Comsoc also has helped populate that party. Although currently its NPPP results are upper bounds-and we commend to the reader the open issue of either proving completeness or of lowering the upper bound-the beautiful work of Mattei, Goldsmith, Klapper, and Mundhenk [MGKM15] shows that NPPP plays an important role in the study of bribery and manipulation in tournaments of uncertain information, by showing that many such problems are in NPPP.1 There are many other parties, some admittedly not so previously poorly attended as 2 and NPPP, where comsoc has provided attendees. Here are just a those thrown by Θp few examples. Nguyen et al. [NNRR14] gave the class DP (all sets that are the symmetric difference of two NP sets) a wide range of new complete sets having to do with social welfare optimization in multiagent resource allocation. The first PNP[1]-completeness result in social choice (which also is perhaps the first natural completeness result anywhere for PNP[1]) and the first PNP-completeness result in social choice both come from a study of the complexity of weighted coalitional manipulation of elections, under the voting system, Veto, where each voter vetos one candidate and whoever has the fewest vetos wins; in an online setting (what in political science is called a roll-call vote), the winner problem here is PNP[1]-complete for 3 candidates and is PNP-complete for 4 or more candidates [HHR14]. Filos-Ratsikas and Goldberg [FG17] have recently proven a natural cake-cutting-related problem to be complete for the lonely class PPA. And a huge number of papers in comsoc contribute 1After doing that, that paper comments that "we have actually shown these problems are in NPPP[1]," which is the class in which the NP machine is allowed at most one query per path to the PP oracle. Simply so that comment in the paper doesn't lead any reader of the present paper to think that NPPP-completeness for that paper's problems is unlikely due to that paper having achieved the seemingly better upper bound of NPPP[1], we claim that in fact NPPP[1] = NPPP. Let us prove that claim. NPPP[1] = NPPP holds via altering the NP machine of NPPP to instead guess the answer to each PP query (without asking the query) and then via a single truth-table (parallel) query round to PP checking that all its guesses were correct, and if a given path would have accepted if its guesses of the answers were correct and that path finds via its truth-table-round access to PP that it guessed correctly, then that path accepts. That as just stated is illegally overly using the PP oracle. But since PP is closed under truth-table reductions [FR96], the truth-table round can be replaced with a single query, and thus we indeed have establishing the claimed class equality. Indeed, since the truth-table yields True for only one known-at-truth-table-query-time setting of answers, even Beigel, Reingold, and Spielman [BRS95, p. 195] would suffice to allow a single query per path. 6 natural problems-often related to manipulative attacks on elections-to other classes such as NPNP, coNPNP, and PSPACE. Also, many counting-based issues in comsoc have been shown to be complete for the counting analogue of NP, the class #P. In summary, there is a two-way street of friendship between comsoc and computational complexity. Complexity benefits since its classes-some of which played central roles in key abstract results yet had few or no natural complete problems-are given quite compelling natural problems based in comsoc. Comsoc in turn has benefited since the machinery of complexity-class classification helps clarify how hard or easy many of its problems are. 3 Complexity Machinery and Notions Find Fertile Ground in Comsoc This section provides a tour of two cases where complexity techniques and notions have found very fertile ground in comsoc research. Although one-to-one (Section 3.2) reductions have been used centrally in complexity theory, in particular in the 1970s, their new use in comsoc is both powerful and quite different than the use they had in complexity. And in another case (Section 3.1), comsoc has given new life to lovely complexity machinery that had been developed in 1976, and yet that even 36 years later had not found a single application in a real-world domain. We will cover each topic but briefly, though we will try to give a sense of how each supports this section's thesis that comsoc has been a fertile ground for complexity techniques and notions, helping comsoc of course, but also in the case of little-used or narrowly used complexity techniques helping establish the value or breadth of the techniques. 3.1 Search versus Decision for NP Problems Everyone knows that SAT is a decision problem, and indeed that the vast majority of complexity classes are defined in terms of decision problems, not search problems. This has been the case for so long that it is easy to forget why this is so. The reason why complexity is largely focused on decision problems is because for almost all natural problems the search problem clearly polynomial-time Turing reduces (and of- ten, for example for SAT, even "polynomial-time 2-disjunctive truth-table" reduces) to the problem's decision version. And so the two problems are inherently of the same complex- ity, give or take composition with a polynomial; it is impossible in these cases for decision to be easy yet for search to be hard. SAT is the classic case: Given a Boolean formula, F (v1, v2, . . .), it is satisfiable exactly if at least one of F (True, v2, . . .) and F (False, v2, . . .) is satisfiable, so given a decision procedure for SAT, one can ask those two questions to it, and thus find a good value for the first variable (if any exists), and then can redo the pro- cess, on the "reduced" formula, to similarly set the second variable and so on. By the end, using as a black-box a decision procedure for SAT, one has found a satisfying assignment in polynomial time, when one exists. 7 Are all natural NP problems so polite as to allow themselves to be fitted into this straitjacket of linked complexities? For decades there have been hints that the answer might be "no." Most powerfully, in 1976 Borodin and Demers [BD76] built lovely, clever machinery that showed that if P 6= NP∩coNP, then there exists an infinite polynomial-time recognizable subset of SAT that has no polynomial-time search algorithm finding solutions for its members. In brief, one could quickly determine with certainty that its members were satisfiable, but one has no idea-it turns out not even a good setting for the first variable of the formula-of what a satisfying assignment would be. (The reason this claim is not precluded by the above discussion of self-reducibility is that the Borodin–Demers subset is so pathological that it has infinitely many members such that the formulas generated at some stage of the above self-reduction process are all not in the subset, and so the above process is rendered invalid on the subset.) Stunning though that result was, the fact that it was achieving its power not on all of SAT (which is impossible), but rather on a somewhat pathological subset of SAT, kept the result and its lovely machinery from gaining traction; it remained only as the above 1976 technical report, and did not have a conference or a journal version. The work was likely-and in the first and second of these four examples the work is explicitly discussed- the inspiration behind some later related work, e.g., that for unambiguous computation the implication mentioned above becomes an "if and only if" characterization [HH88], that under the extremely strong assumption that deterministic and nondeterministic double exponential time differ there exists an (artificial) language in NP for which search does not reduce to decision [BG94], that if P 6= NP then there exist NP-complete languages that are not self-reducible [FO10], and that relative to some oracle there is a search versus decision separation for exponential-time classes [IT89]. Nonetheless, the work found not a single application on natural problems for 36 years. And it was comsoc that finally provided that application. In particular, Hemaspaandra, Hemaspaandra, and Menton [HHM13] proved that of the standard types of attacks on elections, for about half it holds that if P 6= NP ∩ coNP then there is an election system having a polynomial-time winner problem for which the decision problem for that type of attack is in P but the search problem cannot be solved in polynomial time. (For the other half of the attack types, that paper proved unconditionally that search does reduce to decision.) Note that the attack types were not created by that paper; they are the long-standard attack types. And so the problems themselves are not tricks, but are the natural, standard ones, and on these, search and decision are being separated under the given complexity-theoretic assumption, which is widely believed to be true. There are four comments which pretty compellingly need to be made at this point. First, if P = NP ∩ coNP, then integer factoring can be done in polynomial time, and so RSA and much of the foundation of modern computer security falls. So the above results say that either the foundations of cryptography are deeply flawed, or the approach taken within comsoc to defining election-attack problems-namely, they are defined in their decision versions-is defining as tractable some problems whose search version (the issue of finding the attack that works) is not tractable. Second, the reason this separation of search 8 and decision is possible is that, quite unexpectedly, about half of the natural election-attack problems do not seem to be nicely self-reducible. Third, natural is in the eye of the beholder; although the election-attack problems used in the result are the standard, natural ones, the election systems used are indirectly specified by the hypothesized sets in (NP ∩ coNP) − P, and so are likely not very natural. And fourth and finally, this entire approach then found application in a different area of AI, namely, the study of backbones and backdoors of Boolean formulas, where it was shown for example that if P 6= NP ∩ coNP, then search versus decision separations occur for backbones [HN17a, HN17b]. 3.2 Density-of-Hardness Transfer and One-to-One Reductions When one says a problem is NP-complete, that doesn't prove that it is hard. It just proves that if anything in NP is hard, then that problem is hard. This is generally viewed as much better than having nothing to say about the difficulty of the problem. Wouldn't it be lovely to have an analogous type of result for frequency of hardness? For example, in Section 3.1, we discussed work showing that if P 6= NP ∩ coNP, then search and decision separate for many key issues regarding manipulative attacks on elections and backbones of Boolean formulas. But maybe in those examples, for which we know decision is easy, the hardness of search may be a con: Perhaps search is only hard on a very, very, very small portion of the inputs, asymptotically? That is, perhaps it is a worst-case separation that doesn't hold in the typical case. To try to argue against that possibility, it would be wonderful if we could argue that if even one problem A in NP ∩ coNP is hard with a certain frequency h (i.e., each polynomial- time heuristic is wrong on A's membership problem on Ω(h(n)) strings up to length n) then every polynomial-time heuristic for our search problem fails with almost that same frequency, namely, for each polynomial-time heuristic there is an ǫ > 0 such that on our given search problem it fails on Ω(h(nǫ)) of the strings up to length n). This would say that if any set in P 6= NP ∩ coNP is frequently hard (relative to heuristic attacks), then our search versus decision separations are not cons. After all, it is widely be- lieved, most crucially in the cryptography community (due in part to the issue of factoring), that there are sets in NP ∩ coNP that are frequently hard with respect to polynomial-time heuristics. And if one believes that, then the hypothetical machinery discussed above would immediately convert that into a claim of the frequent hardness of the created search prob- lems whose decision problems are easy. In fact, machinery that can do this exists, although it was developed for a completely different purpose in the 1970s. So this is an example where comsoc-and the study of SAT solvers-is giving a fresh use to a complexity-theoretic notion. In particular, every CS undergraduate learns why claims of NP-completeness are cru- cially tied to the tool of many-one polynomial-time reductions; it is how we establish them. In the 1970s, though, there was a sudden moment of focus in theoretical computer science on the notion of one-to-one (aka injective) polynomial-time reductions, i.e., polynomial-time reductions that have no collisions. 9 This moment of intense focus on one-to-one reductions came about to support one of the great complexity goals of the 1970s: to prove that all NP-complete sets were the same set in a transparent disguise, namely, that all NP-complete sets were polynomial-time isomorphic. This is known as the Isomorphism Conjecture (aka the Berman–Hartmanis Isomorphism Conjecture). To this day it remains open. However, using one-to-one reductions, in the 1970s Berman and Hartmanis [BH77] proved that all familiar NP-complete sets indeed were polynomial-time isomorphic, which was a huge revelation. (Their insight in using one-to-one reductions was motivated by the fact that in recursive function theory, recursive one-to-one reductions are central to the proof that all RE-complete sets are recursively isomorphic, a result that follows from the so-called Myhill [Myh55] Isomorphism Theorem.) However, due to challenges such as the potential existence of one-way functions, that 1970s result has not only never been expanded to (under the inherently required assumption that P 6= NP) include all NP-complete sets, but indeed there is evidence that there may well exist nonisomorphic NP-complete sets [JY85, KMR95], i.e., that the Isomorphism Conjecture may fail. These days in complexity theory, many-one reductions remain the standard, and one- to-one reductions are not often discussed. However, one thing that one-to-one reductions do fiendishly well is preserve density. If one has a bunch of strings with a certain behavior, their image when pushed through the reduction cannot possibly jumble them on top of each other, because one-to-one reductions don't jumble anything on top of anything else. It is admittedly true that such reductions can leave "holes" (though injective, they need not be surjective), and can polynomially stretch (and superpolynomially contract, though inherently not too frequently) lengths; and that in fact is what is behind the "ǫ" mentioned above. But one-to-one reductions by definition never have collisions. Exploiting that, the "if anything in NP ∩ coNP is frequently hard relative to polynomial- time heuristics then our search problems are almost as frequently hard relative to polynomial-time heuristics as those" results mentioned above are achieved by ensuring that the entire proof structure in those papers can be made to yield polynomial-time one-to-one reductions. (And the additional related paper mentioned earlier, which adds in results on the case of backdoors of Boolean formulas, and has some related results about backbones of Boolean formulas, in both cases under the weaker assumption P 6= NP, also works by achieving polynomial-time one-to-one reductions.) Thus, to summarize, one-to-one reductions are a tool that was briefly extremely impor- tant in complexity in the 1970s in a brilliant but even forty years later not yet successful attempt to show that there is in effect just one NP-complete set. But comsoc-and also, related to SAT solvers, the study of backbones and backdoors of Boolean formulas-has brought the notion of one-to-one reductions again front-and-center, providing an "if any- thing in the class is frequently hard then this is frequently hard" analogue of the "if anything in the class is hard then this is hard" argument that many-reductions have provided for almost half a century in the case of NP-complete sets. This case makes very clear the two-way street that exists between complexity and comsoc (and also the study of SAT solvers). Complexity benefits in that one of its notions that 10 was a bit covered in cobwebs turned out to yield important frequency-of-hardness results in comsoc, and also relatedly in the study of SAT solvers. And comsoc benefits in that it now has these results, which are quite powerful evidence that the hardness of search that these speak of is in fact not some con job that happens only extremely infrequently. 4 Conclusions We have discussed just a few of the rich collection of interactions between comsoc and com- putational complexity. Many others-from how the complexity-theoretic join operation al- lows proofs in comsoc of the impossibility of proving certain impossibility theorems [HHR09] to how work on online control gives an unexpected new quantifier-alternation characteri- zation of the class coNP [HHR17] to the rich interactions with approximation (see as just one of many examples [FST17]) to the power of dichotomy results to the insights given by parameterized complexity (both these last are on view simultaneously, for example, in [DMN17]) to much more-are not even touched on here. What are the take-aways? The benefits of the friendship between comsoc and com- putational complexity are a two-way street; both comsoc and complexity have benefited. Classes from complexity have been populated, cobweb-covered techniques have been given surprising new life, and interesting results have been obtained in comsoc from the use of those classes and techniques. In light of this, we urge researchers in comsoc and complexity to reach out and work with each other more and more, for the benefit of both fields. And we hope that more Ph.D. programs will encourage and create young researchers who are trained in and simultaneously expert in both areas; that will allow advances far beyond even those that so far have come from this wonderful synergy between areas. Let us hope that the areas become best friends forever. Acknowledgments Warm thanks to the anonymous AAAI-18 reviewers, Aris Filos-Ratsikas, and Paul Goldberg for helpful comments. References [BCE13] [BD76] F. Brandt, V. Conitzer, and U. Endriss. Computational social choice. In G. Weiss, editor, Multiagent Systems, pages 213–284. MIT Press, 2nd edition, 2013. A. Borodin and A. Demers. Some comments on functional self-reducibility and the NP hierarchy. Technical Report TR 76-284, Dept. of Computer Science, Cornell University, July 1976. [BG94] M. Bellare and S. Goldwasser. The complexity of decision versus search. SIAM Journal on Computing, 23(1):97–119, 1994. 11 [BH77] [BO91] L. Berman and J. Hartmanis. On isomorphisms and density of NP and other complete sets. SIAM Journal on Computing, 6(2):305–322, 1977. J. Bartholdi, III and J. Orlin. Single transferable vote resists strategic voting. Social Choice and Welfare, 8(4):341–354, 1991. [BRS95] R. Beigel, N. Reingold, and D. Spielman. PP is closed under intersection. Journal of Computer and System Sciences, 50(2):191–202, 1995. [BTT89a] J. Bartholdi, III, C. Tovey, and M. Trick. The computational difficulty of manipulating an election. Social Choice and Welfare, 6(3):227–241, 1989. [BTT89b] J. Bartholdi, III, C. Tovey, and M. Trick. Voting schemes for which it can be difficult to tell who won the election. Social Choice and Welfare, 6(2):157–165, 1989. [BTT92] J. Bartholdi, III, C. Tovey, and M. Trick. How hard is it to control an election? Mathematical and Computer Modeling, 16(8/9):27–40, 1992. [CELM07] Y. Chevaleyre, U. Endriss, J. Lang, and N. Maudet. A short introduction to computational social choice. In Proceedings of the 33rd International Confer- ence on Current Trends in Theory and Practice of Computer Science, pages 51–69. Springer-Verlag Lecture Notes in Computer Science #4362, January 2007. [Con85] [DMN17] [Dod76] [FG17] J.-A.-N. de Caritat, Marquis de Condorcet. Essai sur l'Application de L'Analyse `a la Probabilit´e des D´ecisions Rendues `a la Pluralit´e des Voix. 1785. Facsimile reprint of original published in Paris, 1972, by the Imprimerie Royale. P. Dey, N. Misra, and Y. Narahari. Parameterized dichotomy of choosing committees based on approval votes in the presence of outliers. In Proceedings of the 16th International Conference on Autonomous Agents and Multiagent Systems, pages 42–50, May 2017. C. Dodgson. A method of taking votes on more than two issues. Pamphlet printed by the Clarendon Press, Oxford, and headed "not yet published", 1876. A. Filos-Ratsikas and P. Goldberg. Consensus Halving is PPA-complete. Technical Report arXiv:1711.04503 [cs.CC], Computing Research Repository, arXiv.org/corr/, November 2017. [FHH10] P. Faliszewski, E. Hemaspaandra, and L. Hemaspaandra. Using complexity to protect elections. Communications of the ACM, 53(11):74–82, 2010. [FHHR09] P. Faliszewski, E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. A richer understanding of the complexity of election systems. In S. Ravi and S. Shukla, editors, Fundamental Problems in Computing, pages 375–406. Springer, 2009. 12 [FO10] [FP10] [FR96] [FST17] P. Faliszewski and M. Ogihara. On the autoreducibility of functions. Theory of Computing Systems, 46(2):222–245, 2010. P. Faliszewski and A. Procaccia. AI's war on manipulation: Are we winning? AI Magazine, 31(4):53–64, 2010. L. Fortnow and N. Reingold. PP is closed under truth-table reductions. Infor- mation and Computation, 124(1):1–6, 1996. P. Faliszewski, P. Skowron, and N. Talmon. Bribery as a measure of can- didate success: Complexity results for approval-based multiwinner rules. In Proceedings of the 16th International Conference on Autonomous Agents and Multiagent Systems, pages 6–14, May 2017. [GHR95] R. Greenlaw, H. Hoover, and W. Ruzzo. Limits to Parallel Computation: P- Completeness Theory. Oxford University Press, 1995. [Gil77] [GJ79] [Hem89] [Hem18] [HH88] J. Gill. Computational complexity of probabilistic Turing machines. SIAM Journal on Computing, 6(4):675–695, 1977. M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman, 1979. L. Hemachandra. The strong exponential hierarchy collapses. Journal of Com- puter and System Sciences, 39(3):299–322, 1989. L. Hemaspaandra. Computational social choice and computational complexity: BFFs? In Proceedings of the 32nd AAAI Conference on Artificial Intelligence. AAAI Press, February 2018. To appear. J. Hartmanis and L. Hemachandra. Complexity classes without machines: On complete languages for UP. Theoretical Computer Science, 58(1–3):129–142, 1988. [HHM13] E. Hemaspaandra, L. Hemaspaandra, and C. Menton. Search versus decision for election manipulation problems. In Proceedings of the 30th Annual Symposium on Theoretical Aspects of Computer Science, pages 377–388, 2013. [HHR97a] E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. Exact analysis of Dodgson elections: Lewis Carroll's 1876 voting system is complete for parallel access to NP. Journal of the ACM, 44(6):806–825, 1997. [HHR97b] E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. Raising NP lower bounds to parallel NP lower bounds. SIGACT News, 28(2):2–13, 1997. [HHR09] E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. Hybrid elections broaden complexity-theoretic resistance to control. Mathematical Logic Quarterly, 55(4):397–424, 2009. 13 [HHR14] [HHR17] [HJV93] [HM00] [HN17a] [HN17b] [HR98] E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. The complexity of online manipulation of sequential elections. Journal of Computer and System Sciences, 80(4):697–710, 2014. E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. Online voter control in sequential elections. Autonomous Agents and Multi-Agent Systems, 31(5):1055– 1076, May 2017. L. Hemaspaandra, S. Jain, and N. Vereshchagin. Banishing robust Turing com- pleteness. International Journal of Foundations of Computer Science, 4(3):245– 265, 1993. M. Holzer and P. McKenzie. Alternating and empty alternating auxiliary stack automata. In Proceedings of the 25th International Symposium on Mathemati- cal Foundations of Computer Science, pages 415–425. Springer-Verlag Lecture Notes in Computer Science #1893, August/September 2000. L. Hemaspaandra and D. Narv´aez. The opacity of backbones. In Proceedings of the 31st AAAI Conference on Artificial Intelligence, pages 3900–3906. AAAI Press, February 2017. L. Hemaspaandra and D. Narv´aez. The opacity of backbones and backdoors un- der a weak assumption. Technical Report arXiv:1706.04582 [cs.AI], Computing Research Repository, arXiv.org/corr/, June 2017. E. Hemaspaandra and J. Rothe. Recognizing when greed can approximate maximum independent sets is complete for parallel access to NP. Information Processing Letters, 65(3):151–156, 1998. [HSV05] E. Hemaspaandra, H. Spakowski, and J. Vogel. The complexity of Kemeny elections. Theoretical Computer Science, 349(3):382–391, 2005. [HW91] [IT89] L. Hemachandra and G. Wechsung. Kolmogorov characterizations of complex- ity classes. Theoretical Computer Science, 83:313–322, 1991. R. Impagliazzo and G. Tardos. Decision versus search problems in super- polynomial time. In Proceedings of the 30th IEEE Symposium on Foundations of Computer Science, pages 222–227. IEEE Computer Society Press, Octo- ber/November 1989. [JY85] D. Joseph and P. Young. Some remarks on witness functions for non-polynomial and non-complete sets in NP. Theoretical Computer Science, 39(2–3):225–237, 1985. [Kad89] J. Kadin. PNP[log n] and sparse Turing-complete sets for NP. Journal of Com- puter and System Sciences, 39(3):282–298, 1989. 14 [KMR95] S. Kurtz, S. Mahaney, and J. Royer. The Isomorphism Conjecture fails relative to a random oracle. Journal of the ACM, 42(2):401–420, 1995. [LGM98] M. Littman, J. Goldsmith, and M. Mundhenk. The computational complexity of probabilistic planning. Journal of Artificial Intelligence Research, 9:1–36, 1998. [LWDD17] K. Larson, M. Winikoff, S. Das, and E. Durfee, editors. Proceedings of the 16th International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2017, Sao Paulo, Brazil, May 8–12, 2017. International Foundation for Autonomous Agents and Multiagent Systems, 2017. [MGKM15] N. Mattei, J. Goldsmith, A. Klapper, and M. Mundhenk. On the complexity of bribery and manipulation in tournaments with uncertain information. Journal of Applied Logic, 13(4, Part 2):557–581, 2015. [MGLA00] M. Mundhenk, J. Goldsmith, C. Lusena, and E. Allender. Complexity of finite- horizon Markov decision process problems. Journal of the ACM, 47(4):681–720, July 2000. [Myh55] J. Myhill. Creative sets. Zeitschrift fur Mathematische Logik und Grundlagen der Mathematik, 1:97–108, 1955. [NNRR14] N. Nguyen, T. Nguyen, M. Roos, and J. Rothe. Computational complexity and approximability of social welfare optimization in multiagent resource allocation. Autonomous Agents and Multi-Agent Systems, 28(2):256–289, 2014. [PZ83] [Reg89] [Rot16] C. Papadimitriou and S. Zachos. Two remarks on the power of counting. In Proceedings of the 6th GI Conference on Theoretical Computer Science, pages 269–276. Springer-Verlag Lecture Notes in Computer Science #145, January 1983. K. Regan. Manuscript, April 1989. Provable complexity properties and constructive reasoning. J. Rothe, editor. Economics and Computation: An Introduction to Algorithmic Game Theory, Computational Social Choice, and Fair Division. Springer, 2016. [RSV03] J. Rothe, H. Spakowski, and J. Vogel. Exact complexity of the winner problem for Young elections. Theory of Computing Systems, 36(4):375–386, 2003. [RWSY03] J. Rosenschein, M. Wooldridge, T. Sandholm, and M. Yokoo, editors. Pro- ceedings of the 2nd International Joint Conference on Autonomous Agents and Multiagent Systems, AAMAS 2003, Melbourne, Australia, June 14–18, 2003. International Foundation for Autonomous Agents and Multiagent Sys- tems, 2003. 15 [Sip82] [SU02] [SV00] [Tod91] [Val76] [Wag87] [Wag90] M. Sipser. On relativization and the existence of complete sets. In Proceed- ings of the 9th International Colloquium on Automata, Languages, and Pro- gramming, pages 523–531. Springer-Verlag Lecture Notes in Computer Science #140, July 1982. M. Schaefer and C. Umans. Completeness in the polynomial-time hierarchy: Part I: A compendium. SIGACT News, 33(3):32–49, 2002. Updated version available online at ovid.cs.depaul.edu/documents/phcom.pdf. H. Spakowski and J. Vogel. Θp 2-completeness: A classical approach for new results. In Proceedings of the 20th Conference on Foundations of Software Technology and Theoretical Computer Science, pages 348–360. Springer-Verlag Lecture Notes in Computer Science #1974, December 2000. S. Toda. PP is as hard as the polynomial-time hierarchy. SIAM Journal on Computing, 20(5):865–877, 1991. L. Valiant. The relative complexity of checking and evaluating. Information Processing Letters, 5(1):20–23, 1976. K. Wagner. More complicated questions about maxima and minima, and some closures of NP. Theoretical Computer Science, 51(1–2):53–80, 1987. K. Wagner. Bounded query classes. SIAM Journal on Computing, 19(5):833– 846, 1990. 16
1912.04451	1	1912	2019-12-10T02:03:39	ColosseumRL: A Framework for Multiagent Reinforcement Learning in $N$-Player Games	[ "cs.MA" ]	Much of recent success in multiagent reinforcement learning has been in two-player zero-sum games. In these games, algorithms such as fictitious self-play and minimax tree search can converge to an approximate Nash equilibrium. While playing a Nash equilibrium strategy in a two-player zero-sum game is optimal, in an $n$-player general sum game, it becomes a much less informative solution concept. Despite the lack of a satisfying solution concept, $n$-player games form the vast majority of real-world multiagent situations. In this paper we present a new framework for research in reinforcement learning in $n$-player games. We hope that by analyzing behavior learned by agents in these environments the community can better understand this important research area and move toward meaningful solution concepts and research directions. The implementation and additional information about this framework can be found at https://colosseumrl.igb.uci.edu/.	cs.MA	cs	ColosseumRL: A Framework for Multiagent Reinforcement Learning in N -Player Games Alexander Shmakov, John Lanier, Stephen McAleer, Rohan Achar, Cristina Lopes, and Pierre Baldi University of California Irvine, Donald Bren School of Information and Computer Sciences {ashmakov, jblanier, smcaleer, rachar}@uci.edu, {lopes, pfbaldi}@ics.uci.edu Abstract Much of recent success in multiagent reinforcement learning has been in two-player zero-sum games. In these games, algorithms such as fictitious self-play and minimax tree search can converge to an approx- imate Nash equilibrium. While playing a Nash equi- librium strategy in a two-player zero-sum game is op- timal, in an n-player general sum game, it becomes a much less informative solution concept. Despite the lack of a satisfying solution concept, n-player games form the vast majority of real-world multiagent sit- uations. In this paper we present a new frame- work for research in reinforcement learning in n-player games. We hope that by analyzing behavior learned by agents in these environments the community can better understand this important research area and move toward meaningful solution concepts and re- search directions. The implementation and additional information about this framework can be found at https://colosseumrl.igb.uci.edu. Introduction Recently, reinforcement learning has had success beat- ing humans in 2-player zero-sum games such as Go (Sil- ver et al. 2018), Starcraft (Vinyals et al. 2019), and Poker (Brown and Sandholm 2018). Two-player zero sum games have many nice properties that make train- ing a reinforcement learning agent for them relatively straightforward. Mainly, playing a Nash equilibrium strategy in zero-sum two-player games is optimal, and in the worst case results in a tie. This fact forms the basis of most current algorithms in this domain, such as Monte Carlo tree search (Browne et al. 2012) and coun- terfactual regret minimization (Zinkevich et al. 2008). While most empirical and theoretical success has come in zero-sum two-player games, we rarely find ex- amples of these games in practical problems. Much more often we have scenarios with more than two agents, and where one agents success is not necessar- ily another agents failure. These n-player general sum games include scenarios such as self-driving cars, mar- ketplaces, robotic teams, and the global economy. Un- fortunately, the Nash equilibrium solution concept does not provide a satisfying solution to these n-player zero sum games. Besides being intractable to compute, a Nash equilibrium strategy has each player playing a best response to all the other players in that current strategy. Another way to put this is that no player has incentive to unilaterally deviate. However, if two play- ers deviate at the same time, they could be better off than if they hadn't. For example, in three-player Kuhn poker, one Nash equilibrium has the first two players getting negative utility and the third player getting pos- itive utility. Neither the first or second player is incen- tivized to unilaterally deviate, but if they both deviate they could end up with positive utility. Furthermore, even if every player plays a Nash equilibrium strategy, their combined strategy might not be a Nash equilib- rium strategy. Currently, there is no satisfying theoret- ical solution concept in n-player general sum games. Although the theory behind n-player general sum games is lacking, in practice many of the same algo- rithms that find success in two-player zero-sum games such as CFR, tree search and self-play can find very good strategies in n-player general sum games. Per- haps the most well-known example of this is Pluribus (Brown and Sandholm 2019), the algorithm that beat the top human players in multiplayer poker using the same techniques from the two-player superhuman algo- rithm Libratus. Pluribus is conjectured to be highly exploitable, but it is still able to beat the best humans at multiplayer poker. How is this possible? What can we learn from solutions found from traditional meth- ods when used in the n-player general sum setting? We believe that these questions need to be emphasized in the multiagent reinforcement learning community. To that end, we have created a framework for reinforce- ment learning in n-player general sum games. This framework includes a number of new n-player game en- vironments such as Blokus, Tron, and 3/4 player tic- tac-toe as well as traditional normal form games and Kuhn poker. Our hope is that by comparing learning approaches and algorithms against each other in these environments we can better understand the dynamics of n-player games and start to discover relevant solution concepts. one player has no possibility of winning but they may decide which of the other two players would win. This challenges tree search algorithms because it undermines the notion of optimal play in all scenarios. Blokus ColosseumRL We introduce a multi agent framework designed to allow researchers to tackle the challenges presented. Colos- seum is a reinforcement learning framework that in- cludes several environments, a unified format for adding new environments, and a distributed competition server for testing and evaluation. The matchmaking system allows for multi agent competitions to be executed in a distributed manner, allowing the evaluation of large populations of agents. Environments Currently, Colosseum has three implemented compet- itive n-player environments. We also define a generic framework for defining n-player games with potential imperfect information that will automatically integrate with the evaluation features. Tic-Tac-Toe Blokus is a strategy board game where the players' goal is to gain maximum control over the board. Blokus has a maximum of 4 players, each receiving one of every possible polyominos ranging one to five sections each. Players may place these polyominos in any diagonally adjacent location to the pieces they already have. Every time a player places a block, they gain control over the number of sections that block contained; for example, large pentonimos give five control whereas the smaller dominos only give two. Rewards for this game are equal to the amount of control received for each move, or 0 if no more moves are possible. At the end of the game, players are given a ranking from first to fourth based on their total control. The observations for this environ- ment are the entire board and the pieces each player has remaining; and, since it is a sequential board game, it is again fully observable with perfect information. Al- though Blokus only has a maximum of four players, it posses interesting challenges for reinforcement learning due to the area control aspect of the game. It is easy for players to form short term alliances by focusing their control efforts on a single player and avoiding blocking each other's possible moves. Also, although the game is perfect information and sequential, the large action spaces makes simple search algorithm infeasible for find- ing optimal agents. Tron Tron is simple n-player grid based video game. The game consists of an K×M arena with surrounding walls We extend the notion of Tic-Tac-Toe to n-player games by using identical rules to regular tic-tac-toe and introducing additional symbols for players. We create an n × m board use the symbols X, O, Y, Z, . . . in or- der to label the players. The goal for every player is to attain three of their symbols sequentially in any cardi- nal or inter-cardinal direction. The winner of the game is the player that achieves this configuration first, and all of the other players are equal losers. Rewards for this game are r = 0 for a regular action, 1 for win- ning, and −1 for losing. The observation for this game is the entire game board, with no obscured informa- tion. This game is a perfect information game with a relatively small game tree, meaning that agents may search essentially all possibilities given enough compu- tation time. Tic-Tac-Toe represents a good challenge for multi-agent reinforcement learning due to the pres- ence of king-maker scenarios: states of the game where Rock Paper Scissors A classical environment from game theory is Rock Paper Scissors, which servers as a common baselines for evolutionary mechanics. We pro- vide both the traditional two player rock paper scissors as well as a generalization of the game to three players. This generalization, created by Bernard Koven (Koven 2014), extends the rules so that the winner of any game is the player with the single unique action out of the three players. If all three players have the same action or all three are unique, then the game is a tie. We allow for user-defined payoff values for winning, losing, and tying the game. The observation for this environment is the normal form matrix given the payoff values, and the actions for each players are the traditional Rock, Paper, and Scissors. Kuhn Poker A classical example of a partial infor- mation environment from game theory is Kuhn Poker, a simplified version of Texas Hold'em. The original ver- sion of the game designed by (Kuhn 1951) and involves two players with a maximum of three actions in a game. A generalization of the mechanics to three players was defined and studied by (Szafron, Gibson, and Sturte- vant 2013). We further generalize the mechanics to include an arbitrary N players, although the analogy to Texas Hold'em breaks down with enormous player counts. The deck of N player Kuhn Poker consists of N + 1 cards labeled 0 to N . The game consists of two rounds: If there is no outstanding bet, then a player may either bet or check. If there is an outstanding bet, then a player may either call or fold. If all players have checked or a player has bet and at least one other has called, then the participating player with the highest card wins the entire pot. Reward is 0 for each action and the change in net worth at the end of a game. Three player Kuhn poker is one of the few non-trivial multi- player games with known Nash equilibria. Baselines One major difficulty in multi agent reinforcement learn- ing is finding good opponents to collect episodes from. We may only train a competent agent if we have good opponents to train on. We apply several standard mul- tiplayer reinforcement learning techniques to examine the difficulties of training in these MARL environments. All of the agents are optimized using Proximal Policy Optimization (Schulman et al. 2017) using the same neural network, and we only vary the training strategy. We focus on the simultaneous-action fully-observable Tron environment for these test because it has a small enough action and observation space for a network to reliably learn but that is not trivial enough for tree search. We use a tron board with size 15 × 15 and 4 players. Fixed Opponent We first train an agent using single player reinforcement learning against a simple, hand-coded agent. This agent moves forward if the path is clear, or randomly moves and N players initially starting in an circle around the arena. Players may move forward, left, or right and when doing so, leave behind a trailing wall. Players are eliminated if they crash into either the arena wall, another player's wall, or their own wall. The goal of the game is to force your opponents to crash into a wall by limiting their movement options. The game supports arbitrarily high number of players as well as teams of multiple players. The rewards are 1 for sur- viving an action, −1 for crashing, and 10 for achieving first place without a tie. At the end of the game, players are ranked (with possible ties) based on the order that they crashed. The observation for Tron may either be the entire game board or a small window around only the given player. Tron may also either be played se- quentially for a perfect information variant of the game or it can be played simultaneously to mimic the real game. Tron is an interesting game to study MARL on because it is easily extendable to many players, features an area control mechanic that allows players to form short-term alliances, and allows for a large of amount of simple variations. Normal Form Games Random Payoff Matrix For theoretical explo- ration, we implement games with randomly generated payoff matrices. We define each environment with a payoff matrix M ∈ RS1,S2,...,SP ,P , where P is the num- ber of players and Si is the number of possible actions for player i. The payouts for each action are randomly generated from a uniform distribution on [0, 1), and we include an option to ensure that the game will always be zero-sum. The observation for this environment is the entire normal form game matrix, and the action for each agent is a number from 0 to Si. either left or right if the path is clear in either direction. We train a neural network on this fixed opponent to en- sure that our PPO and network are capable of learning the environment and task given a stationary opponent. We test three different variants of these fixed agents. In each variant, we change the probability that the agent will perform a random action instead of the hand-coded action. We set this probability to 5% (Sα), 25% (Sβ), and 100% (Sγ). Self Play We also apply the common two-player tactic of de- layed update self play (SP). In this technique, we train against fixed copies of ourselves until the current agent reaches a specified average win percentage. We then update all of the opponent agents and repeat this train- ing procedure on the new agents. This will allow the agent to learn against better opponents than our hand- coded option, but it also provides stability since we do not update the opponents very often. We find that self play converges to a stable agent within 1 to 2 million time-steps. We try two different minimum win per- centages for the update: 50% (SP α) and 80% (SP β). Lower update percentages produces quicker updates to the opponents. Fictitious Self Play We also use an extended version of self play known as fictitious self play (Heinrich, Lanctot, and Silver 2015). In this variant, instead of only keeping track of the last iteration of the target network, we keep the last K it- erations and play against all of them. In each game, each opponent has an 80% chance of being the latest iteration of the policy and a 20% change of being one of the other K − 1 iterations. This encourages diver- sity in the opponent population and ensures that our agent does not over-fit to itself. We again try update percentages of 50% and 80% and population sizes of K = 4 (F SP α, F SP β) and K = 16 (F SP γ, F SP δ). Higher populations should create more robust agents since they have to maintain dominance against a wider variety of opponents. Evaluation Another major difficulty in multi agent reinforcement learning is the task of ranking a population of trained agents. Simple round robin tournaments may results in non-transitive relations between agents. It may be difficult to find a single agent who beats all others, but simply looking at each agents win count may not always find the intuitively best agent either. We provide a simple measure of agent performance based on the probability of an agent finishing in a given position. We run a tournament of 25, 000 games, in which we uniformly sample the opponents from our 12 variations without replacement. From every game, we receive a ranking for each agent based on when they crashed into a wall. We collect the number of times each Distribution of Rankings Figure 1: The distribution of rankings each agent re- ceived throughout tournament. agent finished first, second, third, and fourth. We then sort these agents based on their first place performance and plot the times each agent achieved each ranking. For model design, we flatten our board representa- tion and stack the last three observations so the agent has access to temporal information about the environ- ment. We then feed this flattened observation into a two layer feed forward network with hidden sizes of 512 and 256. Finally, we split this hidden vector into a policy and value layer. We train each agent for 2.72 million timesteps with 5, 440 timesteps for each train- ing iteration. Our network runs 16 iterations of of the Adam Optimizer with an SGD batch size of 1024 for each training batch. We run a tournament featuring the three agents trained on those fixed opponents, the two self-play agents, and the four variations of fictitious self play agents. We also include three variants of our fixed agent to act as a sanity check for the other agents (F α, F β, F γ). We note that if two players tied, for example if they both crashed into a wall at the same time, their rank is rounded down to the lowest nearest rank. The results of the competition are showing in Figure 1. We see an interesting result that the agents trained against only the fixed opponent with single player re- inforcement learning performed the best. The self play and fictitious self play policies performed less good, and the fixed opponents were always last. This is a prelim- inary analysis, and raises more questions than it an- swers. Koven, B. L. D. 2014. Three-player r/p/s • a playful path. A Playful Path. Kuhn, H. W. 1951. A simplified two-person poker. Con- tributions to the Theory of Games (AM-24), Volume I 97 -- 104. Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Silver, D.; Hubert, T.; Schrittwieser, J.; Antonoglou, I.; Lai, M.; Guez, A.; Lanctot, M.; Sifre, L.; Kumaran, D.; Graepel, T.; et al. 2018. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science 362(6419):1140 -- 1144. Szafron, D.; Gibson, R.; and Sturtevant, N. 2013. A parameterized family of equilibrium profiles for three- player kuhn poker. volume 1, 247 -- 254. Tideman, T. N. 1987. Independence of clones as a criterion for voting rules. Social Choice and Welfare 4(3):185 -- 206. Vinyals, O.; Babuschkin, I.; Czarnecki, W. M.; Math- ieu, M.; Dudzik, A.; Chung, J.; Choi, D. H.; Powell, R.; Ewalds, T.; Georgiev, P.; et al. 2019. Grandmas- ter level in starcraft ii using multi-agent reinforcement learning. Nature 1 -- 5. Zinkevich, M.; Johanson, M.; Bowling, M.; and Pic- cione, C. 2008. Regret minimization in games with incomplete information. In Advances in neural infor- mation processing systems, 1729 -- 1736. As an additional test, we ranked the agents using the ranked pairs voting scheme (Tideman 1987). Ranked pairs constructs an acyclic directed graph by sequen- tially placing directed edges between agents in order of their win percentage, checking that each edge does not create a cycle. The rankings generated by this voting scheme, in order from highest to lowest, are as follows: {Sα, Sβ, F SP α, SP α, F SP δ, SP β, F SP γ, Sγ, F SP β, F β, F α, F γ}. Discussion It is striking that even in the simplest possible n-player games such as three-player tic-tac-toe, we don't know what an optimal strategy for an individual player is. This is concerning because n-player games form the vast majority of real-world multiagent systems. We believe that the multiagent reinforcement learning community needs to put much more emphasis on studying these games and designing algorithms that learn how to play optimally in these games. We also think that much greater emphasis needs to be placed in this area on the theoretical side from game theorists, both in defining ranking and optimality concepts as well as developing algorithms that create agents which achieve them. In this paper we have mainly focused on the question of how to design an optimal agent or rank agents in general sum n-player games. Perhaps this is the wrong question though. Maybe in real-world n-player scenar- ios, finding the optimal strategy for each agent doesn't make sense. For example, in self-driving cars, we don't want to have each car get to their location the fastest by endangering other cars. There is a large literature in economics on social choice theory, which approaches this problem from the perspective of a social planner. It is possible that this is a more useful direction be- cause the price of anarchy can be quite large in n-player games. There is also much work done on cooperative n-player games which might be helpful in this area. Fi- nally, it is interesting to examine connections to other n- player settings that are well-studied in economics such as market design and mechanism design. References Brown, N., and Sandholm, T. 2018. Superhuman ai for heads-up no-limit poker: Libratus beats top profession- als. Science 359(6374):418 -- 424. Brown, N., and Sandholm, T. 2019. Superhuman ai for multiplayer poker. Science 365(6456):885 -- 890. Browne, C. B.; Powley, E.; Whitehouse, D.; Lucas, S. M.; Cowling, P. I.; Rohlfshagen, P.; Tavener, S.; Perez, D.; Samothrakis, S.; and Colton, S. 2012. A survey of monte carlo tree search methods. IEEE Trans- actions on Computational Intelligence and AI in games 4(1):1 -- 43. Heinrich, J.; Lanctot, M.; and Silver, D. 2015. Ficti- tious self-play in extensive-form games. In International Conference on Machine Learning, 805 -- 813.
1903.08228	1	1903	2019-03-19T19:34:49	How to Make Swarms Open-Ended? Evolving Collective Intelligence Through a Constricted Exploration of Adjacent Possibles	[ "cs.MA", "cs.DC", "cs.NE", "nlin.AO" ]	We propose an approach of open-ended evolution via the simulation of swarm dynamics. In nature, swarms possess remarkable properties, which allow many organisms, from swarming bacteria to ants and flocking birds, to form higher-order structures that enhance their behavior as a group. Swarm simulations highlight three important factors to create novelty and diversity: (a) communication generates combinatorial cooperative dynamics, (b) concurrency allows for separation of timescales, and (c) complexity and size increases push the system towards transitions in innovation. We illustrate these three components in a model computing the continuous evolution of a swarm of agents. The results, divided in three distinct applications, show how emergent structures are capable of filtering information through the bottleneck of their memory, to produce meaningful novelty and diversity within their simulated environment.	cs.MA	cs	How to Make Swarms Open-Ended? Evolving Collective Intelligence Through a Constricted Exploration of Adjacent Possibles Olaf Witkowski1,2,∗ and Takashi Ikegami3 1 Earth-Life Science Institute, Tokyo Institute of Technology, Tokyo, Japan 2 Institute for Advanced Study, Princeton, USA 3 University of Tokyo, Tokyo, Japan * [email protected] Abstract. We propose an approach of open-ended evolution via the simulation of swarm dynamics. In nature, swarms possess remarkable properties, which allow many organisms, from swarming bacteria to ants and flocking birds, to form higher-order structures that enhance their behavior as a group. Swarm simulations highlight three important fac- tors to create novelty and diversity: (a) communication generates com- binatorial cooperative dynamics, (b) concurrency allows for separation of timescales, and (c) complexity and size increases push the system to- wards transitions in innovation. We illustrate these three components in a model computing the continuous evolution of a swarm of agents. The results, divided in three distinct applications, show how emergent structures are capable of filtering information through the bottleneck of their memory, to produce meaningful novelty and diversity within their simulated environment. Keywords: Open-ended evolution · continuous swarm evolution · col- lective intelligence · artificial neural networks · evolution of cooperation · intrinsic novelty 2 O. Witkowski and T. Ikegami 1 OEE Life has been evolving on our planet over billions of years, undergoing several major transitions which transformed the way it stored, processed and transmit- ted information. All these transitions, from multicellularity to the formation of eusocial systems and the development of complex brains, seem to lead to the idea that the evolution of living systems is open-ended. In other words, life appears to be capable of increasing its complexity indefinitely Another formulation of open-endedness, echoed by Standish [42] and Soros [41], is that open-endedness depends fundamentally on the continual production of novelty. Life keeps un- covering new inventions, in a process which never seems to stop. Since the 1950's, open-ended evolution (OEE) has been a central topic of research for artificial life approaches to the fundamental principles of life. Soon after, John Von Neumann [51] has contributed to the issue as well, with his early model of self-reproducing automata. Since 2015, a series of workshops have been taking place at Artificial Life conferences [45], the last of which4 was a launchpad for the present special issue. In general, an evolutionary system is considered to be open-ended when it is able to endlessly generate diverse novel entities of growing complexity. Engineering open-ended systems in the lab is not easy, and the main obstacle is that the designed evolutionary systems are subject to a thermodynamic drift making them collapse into equilibrium states. Once local optima are reached, they do not produce novelty anymore, which bounds their complexity and diversity. Innovation seems to emerge from collective intelligence, a phenomenon which refers to groups or networks of agents that develop together the ability to en- hance the group's cognitive capacity or creativity. This is reminiscent of the ongoing innovative process of science, which does not have any other fixed ob- 4 at ALIFE 2018, in Tokyo How to Make Swarms Open-Ended? 3 jective than the production of new knowledge, but makes discoveries mostly through accidents. Stuart Kauffman advocated for the idea of the adjacent pos- sible, claiming that a biosphere can be viewed as a secular or long-term trend and it can maximize the rate of exploration of the adjacent possible of an exist- ing organization [20, 19]. Ikegami et al. (2017) [18] builds on that idea to explain how, in terms of evolutionary transitions [28], a new stage (e.g. multicellular oranism) of evolution may be produced without any information being passed on from the previous stage (e.g. from single cells). Rather, structural properties are assembled, producing a stepping stone to the next level of innovation. These structural properties of a collective group can be compared to a bot- tleneck that acts as a filter on several levels of the system, implementing compu- tation that is not present in any of its parts. Part of this idea is analog to Tishby and Polani (1999) [46], where the information is squeezed through a bottleneck, to extract relevant or meaningful information from an environment. The result- ing "filtered" information through the bottleneck, retains only the features most relevant to general concepts. In this paper, we present three "C" factors that we deem important for nov- elty and diversity. We then introduce a swarm model to study these factors, ap- plied in three different studies. We conclude with a discussion on open-endedness at large, framing the three factors in terms of the emergence of collective intel- ligence in swarm simulations. 2 Conditions for OEE The OEE literature has proposed various conditions that are supposed to lead to the successful production of OEE in a system. Number of studies have attempted to formalize necessary conditions for OEE [16, 6]. Taking a recent example, Soros and Stanley (2014) [41] suggest four conditions at the scale of single reproducing 4 O. Witkowski and T. Ikegami individuals in the system, which should each fulfill some nontrivial minimal criterion, be able to create novelty, act autonomously, and dispose of access to unbounded memory. Such papers have typically been proposing their own model, to demonstrate the importance of the hypothesized conditions for the emergence of OEE within it. However, most evolutionary algorithms seem to either converge very quickly to a solution, or get stuck in a confined area of the search space. Either way, they don't seem to be able to intrinsically generate the amounts of complexity and novelty we find in nature, even at a scale. Although this failure of simulated evolution to match open-ended properties found in natural evolution may be explained by shorter timescales, in principle one would have expected decades of efforts, and increasingly larger resources poured into research in evolutionary computation, to have unlocked more of its potential to create novelty. However, even the latest technologies don't seem to keep their inventivity. In general, promising models [24, 23, 12, 13, 40] that manage to demonstrate at least a few phases transitions or creative leaps -- not necessarily with evolutionary computation -- seem to have one common denom- inator of containing several structural bottlenecks which filter relevant informa- tion through them, as a catalyst of creativity. What seems to be missing to achieve general OEE? We choose to empha- size three "C" candidates which we see as worth pursuing -- Communication, Concurrency, and Complexity: (a) Communication: constricted information flows between parts of the systems allow for synergy and cooperation effects. (b) Concurrency: the creation of separate space and time scales requires concur- rent, nondeterministic, asynchronous models. (c) Complexity: mere system growth can boost novelty and diversity. How to Make Swarms Open-Ended? 5 These are the three C-factors on which we choose to concentrate, in this paper. Next, will expand on each of them a little further, before proposing how to apply them in concrete models. We will now expand on these three points, before presenting concrete exam- ples, with Study 1, 2 and 3. Communication This first point addresses synergies and coordination between components of the systems, using information transfers. One well known example of OEE is combinatorial creativity in human language, where syntactical rules are capable of producing infinite well-formed structures using recursion, thus making the number of potential sentences unbounded [15]. Although these may seem slightly dated remarks, at the advent of language studies based on artificial life systems [22], it is promising to focus on the cultural layer of dynamics, that lives on top of the main layer of entities. For example, in the case of web services (social tagging networks), we can analyze how combinatioral complexity is effective in evoking OEE. In cellular automata, one may want to study the interactions between gliders or other patterns. In artificial chemistry, one may want to look at information flows between types of molecules or replicators. In agent-based modeling, perhaps establishments of protocols between agents or groups of agents can become a factor to focus on. Communication naturally adds relevant computing filters on unexploited in- formation flows, effectively increasing the bandwidth of useful information flows within the system per clock cycle, communication offers a layer for metadynamics at a different timescale from the first-order dynamics. This induces a separation of timescales, thus doubling the system's capacity to implement learning mech- anisms. Designing information to circulate between a sub-entities of the system forces the creation of more structural bottlenecks. 6 O. Witkowski and T. Ikegami We propose information exchanges as a central mechanism promoting open- ended evolution. From information flows in groups of individuals, a system can boost its own production of creativity to achieve indefinite complexity. Examples are detailed in Study 1 and 2, below. Concurrency In many situations, a system cannot scale up to larger space-time scales as it is. We need to add some ingredient to make it work in the larger scales. One such remedy is to give it asynchronous updates. Removing the global clock is needed to make larger systems function consistently without constantly checking local consistency. On the other hand, we know that cellular automata (CA) tend to lose their complexity by adopting asynchronous dynamics. Yet asynchrony is an original natural phenomena difficult to bring to artificial systems. According to Dave Ackley [1], models should be indefinitely scalable, ruling out deterministic, synchronous models (such as simple cellular automata), and suggesting nondeterministic, asynchronous ones. Bersini et al. (1994) [4] pro- posed that asynchronous rather than synchronous updating may be key factor in inducing stability in simulations. Although they were examining a variant of cellular automata, their results, based on an analysis of the Lyapunov exponent, indicated the responsibility of asynchrony for sensitivity of the update function. Ackley and Ackley (2015) [1] propose to use asynchrony. The concurrency is also closely related to the ability a system has to evolve separate timescales. Although highly contingent on other properties of a simula- tion, the capacity to develop heterogeneous timescales often constitutes a barrier to producing intrinsic novelty. Researchers used to separate lifetime learning and evolutionary learning, as two distinct mechanisms [31]. However, the effects of accumulating and filtering information into and out of one system's memory occurs at a much more continuous range of different timescales. In nature, from How to Make Swarms Open-Ended? 7 phenotypic plasticity, to maternal effects, to sexual selection, or to gene flow, many events have their time scales intricately interlaced. We will address this in particular in Study 1 and 2. Complexity We have no grounds to argue that nature is its own only possible realization, since there would be no satisfactory explanation for that. One important feature of nature though, is its complexity, which can translate into both system size and landscape complexity. In the simplest of all cases, complexity can be reached merely with large population sizes. Ikegami et al. (2017) [18] proposed that large groups of individuals, given the right set of structural characteristics, may be the main driver for emergence. They discussed this hypothesis in relation to large-scale boid swarm simulations [37], in which the nucleation, organization and collapse dynamics were found to be more diverse in larger flocks than in smaller flocks. Collective behaviors can be qualitatively different by increasing the number of agents, i.e. a colony or group size. In the actual observation, e.g., the individual bees change their behaviors depending on the colony size. Also the fish change their performance of sensing the environmental gradient depending on the school size. In previous works [30], we simulated half a million birds flock using a boids model [37, 49] and found that qualitatively different behavior emerges when the total number of individuals exceeds a few thousands or so. Flocks of different sizes and different shapes interact to diminish some flocks but to generate new ones. Different types of fluctuation become dominant in different size of flocks. A correlation of the local density fluctuation becomes dominant in the larger size flocks and that of the velocity fluctuation dominates in the smaller size flocks. An example of that is offered in Study 3. 8 O. Witkowski and T. Ikegami Stretching the argument on size, environmental complexity is definitely nec- essary to a certain extent to create complex behaviors. Only with richer envi- ronments, encompassing complex distributions of energy resources and ways for systems to survive, can emerging individuals explore a rich set of strategies and increasingly increasingly complex solutions. As mentioned earlier, evolutionary landscapes have become an important concept in biology to analyze the dynam- ics at play in an ecosystem. The picture to have is the one of a unit of selection (e.g. a gene, among many other options) being represented by a point in a multidimensional search space. That space is typically given as many dimensions as there are degrees of freedom for the entity to vary and evolve in the space (e.g. combination of nucleotide sequences). The search space is mapped onto an additional dimension, which is usually the reproductive success, or fitness. The shape of the fitness across all degrees of freedom of the system have a strong impact on the dynamics that it can achieve. Malan et al. (2013) [26] identifies eleven characteristics of fitness, which make them more or less difficult to solve. These characteristics include the degree of variable inter-dependency, noise, fitness distribution, fitness distribution in search space, modality, information to guide search and deception, global landscape structure, ruggedness, neutrality, symmetry, and searchability. In evolutionary systems, richer environments, benefiting of a complex distri- bution of energy sources and ways for systems to survive, give rise to richer sets of pathways. The larger the search spaces, the more complex fitness functions are potentially evolved. Another way is to make the environment a more com- plex function of time, which the agents will need to learn in order to extract more energy from it. In Study 1 and 2, we present results suggesting that sim- ulations should be ensured to provide sufficient system complexity in terms of the environment of agents. How to Make Swarms Open-Ended? 9 Simulation time and memory, though not mentioned yet, are important com- ponents to consider. Computationally, the whole course of evolution on Earth is like a single run of a single algorithm that invented all of nature, and seems like it will never end. One obvious difference is the size of the systems, which might be the missing element to get ever-greater emergent complexity and nov- elty through very long time. However, we do not insist on this component in this paper, as we consider trivial that a system with too low computational power will not be able to achieve OEE to any extent. Similarly, there is point to be made about endo-OEE, producing novelty from within, against exo-OEE, which makes use of input from outside the sys- tem. Picbreeder [39], for example, explicitly requires external human input to function, which makes it a debatable generator of OEE. Nevertheless, OEE is not about new information, but rather inventions achieved by the system. In that respect, swarms are a promising model: without increasing ensemble size, they let us focus on how coordination patterns self-organize, generating intrin- sic novelty. To give another example, even increasing the number of neurons in a neural network still requires neurons to differentiate themselves, and create coordinated networks before they get to foster innovative ideas. We exemplify the importance of size and complexity in Study 1 and 3, while discussing how to make simulations parallelizable, to save considerable amounts of time and memory by distributing them over many machines. 3 Concurrent evolutionary neural boids model The model we choose to present puts together the abovementioned series of features, as a means to promote the system's open-endedness. We give some details here, and will go over the details of several applications of it in the 10 O. Witkowski and T. Ikegami next section. The evolutionary system is an agent-based simulation, based off Reynolds' boids model [37]. The boids model was based on simple rules computed locally, allowing to simulate flocks of agents moving through artificial environments. As with most artificial life simulations, boids showed emergent behavior, that is, the complexity of boids arises from the interaction of individual agents adhering to a set of simple rules of separation (steer to avoid crowding local flockmates), alignment (steer towards the average heading of local flockmates, and cohesion (steer to move toward the center of mass of local flockmates). In our model, like in Reynolds' model, the population of agents moves around in a continuous three-dimensional space, with periodic boundary conditions (Fig- ure 1). However, instead of using fixed rules to control the boids' motion, we allow agents to evolve their own controllers through concurrent evolutionary computa- tion. Each agent, instead of responding to simple rules, is controlled by its own neural network. The parameters of the neural network are encoded in a genotype vector, which determines the individual's sensorimotor choices at each moment in time. This corresponds to standard evolutionary robotics methodologies [32], although we introduce the following variant. The genotype is evolved through the course of the simulation, via a continuous variant of an evolutionary algo- rithm [54], that is, agents with high level of fitness are allowed at any point to replicate with mutation in the middle of the running simulation. This model also builds up on prior work on the effect of self-organized inter- agent communication and cooperative behavior on the performance of agents to solve tasks [35, 34]. Previous research has shown the difficulty of using commu- nication channels [36, 29] but showed cooperative value of information trans- fers [55]. This will be complemented by the results from previous information- How to Make Swarms Open-Ended? 11 theoretic analyses of learning systems, which managed to shed light on the role of information flows in learning [48, 25, 47]. Fig. 1: Graphical representation of the world in a neural boids simulation. Each agent is represented as an arrow indicating its current direction. The color of an agent indicates the average value of its internal nodes. The green spheres represent the centers for energy sources. Although variants presented later in the paper display slightly different graphics, the backbone is the same. Agents are given a certain energy, that also acts as their fitness. This will be specific to the study cases. Each agent comes with a set of 12 different sensors. The neural network (represented on Figure 2) takes the information from those sensors as inputs, in order to decide the agent's actions at every time step. The possible actions amount to the agent's motion, and in the specific variant shown here, a Prisoner's Dilemma action (cooperate or defect), as well as two output 12 O. Witkowski and T. Ikegami signals. The architecture is composed of a 12 input, 10 hidden, 5 output, and 10 context neurons connected to the hidden layer (see Figure 2). The agents' motion is controlled by M1 and M2, outputting two Euler rota- tion angles: ψ for pitch (i.e. elevation) and θ for yaw (i.e. heading), with floating point values between 0 and π. Even though the agents' speed is fixed, the rota- tion angles still allow the agent to control its average speed (for example, if ψ is constant and theta equals zero, the agents will continuously loop on a circular trajectory, which results in an almost-zero average speed over 100 steps). The outputs S(1) out and S(2) out control the signals emitted on two distinct chan- nels, which are propagated through the environment to the agents within a neighboring radius set to 50. The choice for two channels was made to allow for signals of higher complexity, and possibly more interesting dynamics than greenbeard studies [11]. The received signals are summed separately for each direction (front, back, right, left, up, down), and weighted by the squared inverse of the emitters dis- tance. This way, agents further away have much less impact on the sensors than closer ones do. Every agent is able to receive signals on the two emission chan- nels, from 6 different directions, totalling 12 different values sensed per time step. For example, the input S(6,1) in corresponds to the signals reaching the agent from the neighbors below. The evolution is performed continuously over the population. Agents with negative or zero energy are removed, while agents with energy above a thresh- old are forced to reproduce, within the limits of one infant per time step. The reproduction cost is low enough, considering the threshold, to not put the life of the agent at risk. How to Make Swarms Open-Ended? 13 Fig. 2: Architecture of the agent's controller. The network is composed of 12 input neurons, 10 hidden neurons, 10 context neurons and 5 output neurons. 4 Study cases We go over the application of this model in three selected examples of studies. Each of them highlights a specific property for OEE. The first model shows how agents can form patterns to accelerate their search for energy, distributed over an n-dimensional space, collaborating via local signaling with their neighbors. The second study shows the invention of dynamical group strategies in a spatial prisoner's dilemma, allowing for specific cooperation effects. The third example shows the impact of growth on the emergence of noise-canceling effects. 4.1 OEE via collective search based on communication Since Reynold's boids, coordinated motion has often been reproduced in number of artificial models, but the conditions leading to its emergence are still subject to research, with candidates ranging from obstacle avoidance to virtual leaders. The relation of spatial coordination and group cooperation has long been studied in game theory and evolutionary biology. 14 O. Witkowski and T. Ikegami We here apply our model of agents exchanging signals and moving in a three- dimensional environment, to a task of dynamical search for free energy in space [54, 55]. Each agent's movements are controlled by artificial networks, evolved through generations of an asynchronous selection algorithm. At the term of the evolution, the agents are able to communicate to produce cooperative, coordi- nated behavior. Individuals develop swarming using only their ability to listen to each other's signals. The agents are selected based on their performance at finding invisible resources in space giving them fitness. The agents are shown to use the informa- tion exchanged between them via signaling to form temporary leader-follower relations allowing them to flock together. The swarmers outperform the non- swarmers at finding the resource, thus reaching a neutral evolutionary space which leads to a genetic drift. This work constructs an adaptive system to evolve swarming based only on individual sensory information and local communication with close neighbors. This addresses directly the problem of group coordination without central con- trol, without being aware of the position direct neighbors, nor any use of the substrate where to deposit information (stigmergy) [14]. The approach has also the advantage of yielding original and efficient swarming strategies. A detailed behavioral analysis is then performed on the fittest swarm to gain insight as to the behavior of the individual agents. The results show that agents progressively evolve the ability to flock through communication to perform a foraging task. We observe a dynamical swarming behavior, including coupling/decoupling phases between agents, allowed by the only interaction at their disposal, that is signaling. Eventually, agents come to react to their neighbors' signals, which is the only information they can use to improve their foraging. This can lead them to either head towards or move How to Make Swarms Open-Ended? 15 away from each other. While moving away from each other has no special effect, moving towards each other, on the contrary, leads to swarming. Flocking with each other may lead agents to slow down their pace, which for some of them may keep them closer to a food resource. This creates a beneficial feedback loop, since the fitness brought to the agents will allow them to reproduce faster, and eventually multiply this type of behavior within the total population. The algorithm converges to build a heterogeneous population, as shown on Figure 3. The phylogeny is represented horizontally in order to compare it to the average number of neighbors throughout the simulation. The neighborhood becomes denser around iteration 400k, showing a higher portion of swarming agents. This leads to a firstly strong selection of the agents able to swarm together over the other individuals, a selection that is soon relaxed due to the signaling pattern being largely spread, resulting in a heterogeneous population, as we can see on the upper plot, with numerous branches towards the end of the simulation. In this scenario, agents do not need extremely complex learning to swarm and eventually get more easily to the resource, but rather rely on dynamics emerging from their communication system to create inertia and remain close to goal areas. The simulated population displays strong heterogeneity due to the asyn- chronous reproduction schema, which can be seen in the phylogenetic tree (Fig- ure 3). The principal component analysis plotted on Figure 4 shows a large cluster (left side) in addition to a series of smaller ones (right side). The geno- types in the early stages of the simulation belong to the right clusters, but get to the left cluster later on, reaching a higher number of neighbors. The plot shows a diverse set of late clusters, which translates to numerous distinct behaviors in the late stage of the simulation. 16 O. Witkowski and T. Ikegami Fig. 3: Top: average number of neighbors during a single run. Bottom: agents phylogeny for the same run. The roots are on the left, and each bifurcation represents a newborn agent. Such heterogeneity may suppress swarming but the evolved signaling helps the population to form and keep swarming. The simulations do not exhibit strong selection pressures to adopt specific behavior apart from the use of the signaling. Without high homogeneity in the population, the signaling alone allows for inter- action dynamics sufficient to form swarms, which proves in turn to be beneficial to get extra fitness. These results represent an improvement on previous models using hard-coded rules to simulate swarming behavior, as they are evolved from very simple con- ditions. Our model also does not rely on any explicit information from leaders, How to Make Swarms Open-Ended? 17 Fig. 4: Two principal components of a PCA on the genotypes of all agents of a typical run, over one million iterations. Each circle represents one agent's genotype, the diameter representing the average number of neighbors around the agent over its lifetime, and the color showing its time of death ranging from bright green (at time step 0, early in the simulation) to red (at time step 106, when the simulation approaches one million iterations). like previously used in part of the literature [8, 44]. It does not impose any ex- plicit leader-follower relationship beforehand, letting simply the leader-follower dynamics emerge and self-organize. In spite of being theoretical, the swarming model presented in this paper offers a simple, general approach to the emergence of swarming behavior once approached via the boids rules. This simulation im- proves on previous work because agents naturally switch leadership and follow- ership by exchanging information over a very limited channel of communication. Finally, our results also show the advantage of swarming for resource finding. It's only through swarming, enabled by signaling behavior, that agents are able to reach and remain around the goal areas. 18 O. Witkowski and T. Ikegami In terms of cooperation, this model exemplifies a case of multilevel selection theory [52, 50], which models the layers of competition and evolution, within an ecological system. Our system shows the emergence of different levels which function cohesively to maximize reproductive success. The fitness value of the group level dynamics outweighs the competitive costs, resulting in individuals constantly innovating in ways they are cooperating in a non-trivial way, to create behaviors which are not centrally coded for. This study shows swarming dynamics emerging from a communication sys- tem between agents, immersed in a simulated environment with spatial distribu- tion of energy resource. The concurrent evolution scheme, running at the same time as the simulation itself, led to decentralized leader-follower interactions, which allowed for collective motion patterns, which in turn significicantly im- proved the groups' fitnesses. This model encodes the stochastic evolution of a controller that maps sensory input onto motor output, to optimize the performance on a task, as framed broadly by Nolfi and Floreano (2000) [32]. We capture the fight against a difficult wall [38], which simulations typically fail at because they suddenly hit a so-called "wall of complexity": trivial tasks are solved easily, but it's hard to jump to solving difficult ones. If we take the no-free-lunch argument from Wolpert and Macready (1997) [56] that no optimisation algorithm is at all times superior to others, it is natural that the more specific the algorithm, the more it is likely to fail with new problems. Our results suggest that novelty can be produced by the asynchronous evo- lution of a heterogeneous community of agents, which through their mixture of strategies, may achieve open-ended, uninformed learning. The heterogeneity present in the model also offers an extension to the advantages of particle swarm optimization (PSO) [10]. While PSO only offers one unique objective function How to Make Swarms Open-Ended? 19 to optimize, each agent in the swarm effectively runs its own function, which are combined into a swarm behavior. Although these results suggest open-endedness, it is worth noting we do not bring a proof that the phenomenon is truly open- ended, which may require the emergence of ever-complexifying communication, or an uninterrupted sequence of evolutionary innovations. The information flows were a focus of the original work [55]. From these flows, one can notice three main bottlenecks. The evolutionary computation contains a bottleneck effect, as a result from the selection based on the agents' performance on the task. Another bottleneck can be found between the sensory inputs of each agent, and its motor outputs, as the neural controller acts as a filter for the information. The agents' signaling also naturally contains a bottleneck effect, as the information transmitted from agent to agent is constrained by the physical communication bandwidth. The combination of these three bottlenecks allows for relevant information to be filtered into the swarm, which is able learn certain behaviors (see also next section). 4.2 OEE via cooperative flocking The evolution of cooperation is studied in game theory, and stretches have been made to include spatial dimensions. This problem is often tackled by using simple models, such as considering interactions to be a game of Prisoner's Dilemma (PD). We examined a variation of the model with a distinct fitness function in a separate study, based this time on the agents playing a spatial version of the Pris- oner's Dilemma [53]. We study the impact of the movement control on optimal strategies, and show that cooperators rapidly join into static clusters, creating favorable niches for fast replications. It is also noted that, while remaining inside those clusters, cooperators still keep moving faster than defectors. The system dynamics are analyzed further to explain the stability of this behavior. 20 O. Witkowski and T. Ikegami This work presents, in an even more explicit fashion than the previous study, a model aimed at showing emergent levels of selection for cooperative behavior [52]. At every time step, agents are playing a N-player version of the prisoner's dilemma with their surrounding, meaning that they make a single decision that affects all agents around them. They get reward and/or punishment based on the number of cooperator around them. Their decision is one of the outputs of their neural network. Effectively, the payoff matrix we used is an extension of Chiong and Kirley (2012) [5], where we added the distance to take into account the spatial continuity. Based on the outcome of the match, agents can choose a new direction, which is similar to leaving the group in the walk away strategy [2], the main difference being that, in our case, it is also possible for groups to split. It is also similar in another aspect: there is a cost to leaving a group, as a lone agent may need time to meet others. At the beginning of each run, the environment is seeded with random agents. Since all weights in their neural network are set at random, roughly half of the agents initially choose to cooperate while the other half choose to defect. This leads to a fast extinction of cooperators, until approximately 50000 time steps), until a group emerges strong enough to survive. The second phase follows, in which cooperators are quickly increasing in number due to the autocatalytic na- ture of this strategy. A third step happens eventually, where defectors invade the cluster, followed either by the survival of the cluster due to cooperators running away or a reboot of the cycle. In case of survival, oscillations in the proportion of cooperators can be observed. However, this phenomenon is averaged away over multiple runs, since period and phase of the oscillations are not correlated from one experiment to the other. Were a defector to appear near a cluster of cooper- How to Make Swarms Open-Ended? 21 Fig. 5: Graphical representation of the world in a simulation. Each agent is repre- sented as an arrow indicating its current direction. The color of an agent indicates its current action, either cooperation (blue) or defection (red). Note the cluster of cooperators being invaded by defectors. ators, the cluster would react by "reproducing away". However, the chances to be overtaken by the defectors is much higher than in the dynamic case. From this three-dimensional model of agents playing the Prisoner's Dilemma, the first result is that cooperators, when they are present, quickly evolve to form clusters as they represent a favorable pattern. The clustering behavior can be in- terpreted as a degenerated version of the simulations presented above, since the cooperating agents present the same capacities of information exchange as that model. We note that this solution is evolved through a longer time scale, as it is not always viable locally, depending on the distribution and behavioral thresh- olds of defectors. While the clustering itself can be expected, it is interesting to 22 O. Witkowski and T. Ikegami observe that their overall movement rate is still higher than defectors. This is even more surprising considering that those clusters do not seem to move fast. Instead, analysis shows that cooperators are moving quickly inside the cluster, which may be a way to adapt to an aggressive environment. In addition, comparison with the static case showed that movement made the emergence of cooperators harder, but more stable in the long run. Since it is harder for defectors to overtake a cluster of cooperators, our systems often show a soft bistability, meaning that they will eventually switch from one state to the other. It is even possible to observe a sort of symbiosis, where cooperators are generating more energy than necessary, which is in turn used by peripheral defectors. In this case, replacement rates allow cooperators to stay ahead, keep- ing this small ecosystem stable. This cohesion among cooperators seems to be enhanced by signaling, even though signals might attract defectors. Additional investigation on the transfer entropy, for instance, could be a promising next step. Another result is found in the choice of actions, generated by the neural networks without consideration of the past actions. We notice the emergence of a dynamical memory effect, that otherwise requires to be encoded in each agent, here emerging from the agents' motion in space. Since the Prisoner's Dilemma game has become a common model used in evolutionary biology to study the outcomes depending on the costs that char- acterize an ecosystem, this model, with a fitness based on the results of such game, showed the emergence of spatial coordination based on a the exchange of signals between agents. The signals remained very simple, and the environment was fixed in time. This model's evolutionary computation reached solutions composed of differ- ent parts, including soft bistable strategies, different radiuses of clusters, as well How to Make Swarms Open-Ended? 23 as the use of dynamical patterns to improve their fitness. The solutions were also distributed over different timescales. The communication between agents also allowed them to converge on these behaviors more quickly. These elements refer back to our 3-C arguments earlier, for the discovery of novel solutions to a simple PD game. Lastly, we note that many different neural architectures may coexist as only a part of the neural architecture is used to implement flocking. This neural heterogeneity is something we'd like to insist in the context of OEE. Additionally, communication is important to filter out the neural architecture heterogeneity, which potentially holds the heterogeneity in a community (i.e. agents can stay in a community if they can communicate to each other). The communication may therefore indirectly help preserving the heterogeneity. 4.3 OEE via large scale swarms Studying flocking models can also lead to the emergence of OEE, by focusing on emergent phenomena as macroscopic layers of patterns and structures that appear as a result of cooperative phenomena between autonomously behaving elements. A group of elements creates a self-organizing structure, which gov- erns the individual micro rules and creates a new macro structure. Therefore, consecutive micro -- macro recurrent self-organization is defined as an emergent phenomenon. Here, we describe the contribution of the same swarming simulation, scaled up, showcasing the effect of size on emergence of open-endedness [9]. Before that, we start by presenting a degenerated version of that model, which shows large-scale dynamics in the less computationally costly case of agents that don't preserve any internal state other than position and velocity [18]. Starting with this simpler stateless model, we observe a noticeable change when the total number of boids increases from 2048 to 524,288, while the density 24 O. Witkowski and T. Ikegami is kept constant (Figure 6). In order to compute large swarming behaviour, we parallelized the computational steps using the general-purpose computing on graphics processing units (GPGPU) method. The next step was to extend it to stateful agents. Fig. 6: Visualization of swarming behavior, simulated by a large scale stateless boids model [18]. The total number of boids in each panel is (a) 2048, (b) 16 384, (c) 131 072 and (d) 524 288, respectively. Some flocks are composed of a very large number of boids with narrow filament patterns. The initial velocity of each boid is set at random, and the density of the total number of boids is kept constant at 16,384 (number per cubic unit). The minimum and the maximum speed are set at 0.001 and 0.005 (unit per step), respectively. We explore the effect of reaching a critical mass, and how it impact the ef- ficiency of the swarm's foraging behavior. In particular, we study the problem of maintaining the swarm's resilience to noisy signals from the environment. To do so, we look at stateful boids, i.e. moving agents controlled by neural network How to Make Swarms Open-Ended? 25 controllers, which we evolve through time in order to explore further the emer- gence of swarming, like in the previous two sections. However, we now ground our model in a more realistic setting where information about the resource location made partly accessible to the agents, but only through a highly noisy channel. The swarming is shown to critically improve the efficiency of group foraging, by allowing agents to reach resource areas much more easily by correct individual mistakes in group dynamics. As high levels of noise may make the emergence of collective behavior depend on a critical mass of agents, it is crucial to reach in simulation sufficient computing power to allow for the evolution of the whole set of dynamics. Because this type of simulations based on neural controllers and information exchanges between agents is computationally intensive, it is critical to optimize the implementation in order to be able to analyze critical masses of individ- uals. In this work, we address implementation challenges, by showing how to apply techniques from astrophysics known as treecodes to compute the signal propagation, and efficiently parallelize for multi-core architectures. The results confirm that signal-driven swarming improves foraging performance. The agents overcome their noisy individual channels by forming dynamic swarms. The mea- sured fitness is found to depend on the population size, which suggests that large scale swarms may behave qualitatively differently. The minimalist study presented in this paper together with crucial compu- tational optimizations opens the way to future research on the emergence of signal-based swarming as an efficient collective strategy for uninformed search. Future work will focus on further information analysis of the swarming phe- nomenon and how swarm sizes can affect foraging efficiency. In this model, we specifically focus on the addition of noise to the food detection sense that the agents possess, and hypothesize that it can be overcome 26 O. Witkowski and T. Ikegami by the emergence of a collective behavior involving sufficiently large groups of agents. Many systems, from atomic piles to swarms, seem to work towards preserving a precarious balance right at their critical point [3]. An atomic pile is said to be "critical" when a chain reaction of nuclear fission becomes self-sustaining. A minimal amount of fissionable material has to be compacted together to keep the dynamics from fading away. The notion of critical mass as a crucial factor in collective behavior has been studied in various areas of application [27, 33]. Similarly, the size of the formed groups of agents may be crucial, in order to reach a critical mass in swarms, enough to overcome very noisy environments. Part of the focus will therefore be on the optimization of the computer simulation itself, as large-scale swarms may qualitatively differ in behavior from regular- sized ones. The model extends the original setup described before, which proposed an asynchronous simulation evolving a swarming behavior based on signaling be- tween individuals. However, unlike the original model, where the individuals don't perceive directly either the food patches or the other agents around them, here we give a sense of vision to every agent, allowing them to detect nearby resources. However, we add a high level of noise to make this information highly imperfect. We used an agent-based simulation to show how signal-driven swarming, emerging in an evolutionary simulation such as in Witkowski and Ikegami (2014) [54], allow agents to overcome noisy information channels an improve their per- formance in a resource finding task. Our first contribution is the very intro- duction of noise, demonstrating that the algorithm performs well against noises filling up channels of information almost up to their full capacity, in the inputs of agents. The individuals, by means of a swarming behavior helped by basic How to Make Swarms Open-Ended? 27 signaling, manage to globally filter out the noise present in the information from their sensory inputs, to reach the food sites. We proposed a hierarchical method based on the Barnes-Hut simulation in computational physics and its parallel implementation. We achieved a perfor- mance improvement of a few orders of magnitude over the previous implementa- tion [54]. This implementation is crucial to achieve the simulation of a sufficient number of agents to test for large-scale swarms (i.e. involving a very large num- ber of individuals), which have been suggested to generate qualitatively different dynamics. The optimization of the fitness acquired by phenotypes using efficient pat- terns of behavior (motion and signaling), which themselves are encoded in the weights of agents' neural networks. The real optimization therefore occurs at the higher level of the darwinian-like process in the genotypic search space. Effi- cient genotypes are selected by the asynchronous genetic algorithm throughout a simulation run. We observed that signaling improves the foraging of agents (see Figure 7 for plots from Drozd et al. [9] of efficiency or fitness against simulation time), the average resource retrieved per agent per iteration as a measure of the popula- tion's fitness. Without noise, the agents using signaling are less efficient than their silent counterpart, which we found is not due to the cost of signaling, but rather because of the excess of noise brought by the signal inputs. The difference remains very small between signaling and non-signaling agents. We find however that from a certain noise level, the cost to signal is fully compensated by the benefits of signaling, as it helps the foraging of agents. The average fitness becomes even higher as we increase the noise level, which suggest that the signaling behavior increases in efficiency for high levels of noise, allowing the agents to overcome imperfect information by forming swarms. 28 O. Witkowski and T. Ikegami Fig. 7: Agents' efficiency plot with and without signal, from a stateful (neural- network-controlled) boids model in the original work [9], with mean (central line) and standard deviation range (area plot) over 10 runs. The plots correspond to noiseless (top), constant noise 20 (middle), and constant noise 40 (bottom), respectively. We also observe scale effects in the influence of the signal propagation on the average fitness of the population. For a smaller population, only middle values of signal propagation seem to bring about fitter behaviors, whereas this is not the case for larger sizes of population. On the contrary, larger populations are most efficient for lower levels of signal propagation. This may suggest a phase How to Make Swarms Open-Ended? 29 transition in the agents' behavior for large populations, eventually in the way the swarming itself helps foraging. Understanding criticality seems strongly related to a broad, fundamental theory for the physics of life as it could be, which still lacks a clear description of how it can arise and maintain itself in complex systems. The effects of criticality have recently been investigated futher by one of the authors, using a similar setup [21]. The results showed exploratory dynamics at criticality in the evolution of foraging swarms, and the tension between local and global scale adaptation. Through this work, by increasing the number of simulated boids that main- tain their own states, we may introduce more than the mere number. By allowing for many information exchanges between computing agents, the simulation can effectively take leaps of creativity. In Stanley and Lehman's 2015 book [43], ob- jective functions are presented as a distraction, as novelty and diversity might not be achieved by hard-coding the arrival point. Here, in contrast, we have many evolvable objective functions cooperate in reaching a solution, as a step- ping stone to reach the search for novelty. By letting the swarm grow, we see the emergence of collective intelligence, which corresponds to the invention of signal-based error correction. By exchang- ing signals, the agents are able to correct the error induced by the noise we in- jected in the simulation. Like for the large scale boids simulation, the invention happens after a critical mass of agents is reached, suggesting similar dynamics with stateful agents. 5 Discussion OEE is the everlasting innovative processes found in human technologies and biological evolutions, and we barely observe open-endedness outside these exam- 30 O. Witkowski and T. Ikegami ples. Yet, some artificial systems demonstrate close-to-OEE phenomena, which we have discussed earlier in this paper. Achieving real OEE remains an open challenge, and at this point all works in the literature works fall short of that objective. Although that may be the case with the swarm models presented above, it was one of our goals to emphasize the importance of maintaining the evolvability of a system. In an adversarial game theoretical setup for example, reaching an ESS or a local attractor may keep the system from inventing new solutions. In this sense, explicitly stopping the system from learning too much may allow the system to avoid being stuck in such attractor, and possibly keep innovating forever. In this paper, we propose collective intelligence as a driving force towards open-ended evolution, suggesting that collective groups can develop the ability to be more innovative. Instead of aiming at optimizing one fixed objective function, a collective swarm of agents works with as many competing objectives as there are agents in the swarm. Through information exchanges between a certain number of agents, these objectives, embodied in the agent's behaviors, can collaborate to implement a search for novelty. All agents contribute to the search in behavioral space, as one whole organization, by exploring the adjacent possible. Each novel discovery in the system, or emergent level of organization, can be reached from an adjacent state where the system was previously. The way one moves from one point to the next, which should retain information accumulated in the past, is constrained by the structure of the swarm, in a bottleneck effect. We discuss several instances of bottlenecks in this paper. One is a task or environmental condition which each agent must overcome. In the case of foraging environment, organizing swarming turned out to be a critical step. So it became a major transition from non-swarming to swarming agents. Swarming behavior How to Make Swarms Open-Ended? 31 was obtained by organizing a hill side function by the neural controller. After the swarming behavior has been achieved, other properties (e.g. individual pattern) start to evolve. So for the task, swarming was a necessary behavior to organize and was a bottleneck for the entire evolutionary process. In other words, OEE emerges by setting up a right environmental condition. In case of a game-theoretic situation, such as Study 2, the communication system among all agents constituted a bottleneck to achieve mutual coopera- tion. With the emergence of niche construction, the door opens for regulating mechanisms such as cooperation, reciprocal altruism or social punishment, to get implemented. In this example, the OEE in terms of the invention of cooper- ation mechanisms, can only evolve as a secondary structure once the swarming structure is already established. In the case of large swarm models, the bottleneck is twofold. One is scale itself and also its CPU resource. We discussed the evolution of swarm by increasing its size, showing there is a critical size where the different kinds of fluctuation dominates in larger flocks (i.e. heading direction fluctuation to the density fluc- tuation). If such a transition occurs at the larger size simulation (we expect it can happen in each 3-4 order difference in sizes), we say that OEE is caused by increasing the size. In addition to this point, 3D swarm models require a huge computational power and we need to elaborate programming for a large scale systems. In Study 3, each boid has a list of neighbors and it is updated periodically. This speeds up the calculation of the distance from the one to its neighboring boids. In study 1, each bird can listen to the sound sent unidirectionally from the other boids, so that we don't have to calculate the exact distance. Real birds will never measure the distance to the other individuals. So measuring the distance is an 32 O. Witkowski and T. Ikegami unnecessary bottleneck due to the computational model. Here the OEE is the new computational techniques to overcome this computational bottlenecks. The computation of a swarm displays a bottleneck effect, in the sense that the emergent properties of the swarm and its embodiment in a simulated envi- ronment may constrain the way the information (communication, lineage infor- mation) flows within the system, and the way relevant information (strategies, motion patterns) is progressively retained5 through time in its structure. Never- theless, the simplicity of such information flows may be limiting, more complex information transfer protocols may need to emerge from bottlenecks to bootstrap OEE. For open-endedness, bottlenecks are crucial to, perhaps counterintuitively, act against learning. We observe examples of such bottlenecks in systems like Picbreeder [39], where one must find a way to avoid the system from assuming that the current apparent goal is the ultimate goal, as this would preclude further innovations. Picbreeder-like systems present similarities with our signal-based swarms, as they have communication between many agents filter information to let innovations come about. As suggested in the beginning of the paper, bottlenecks can be caused by dif- ferent components: an explicit communication system, a concurrent evolutionary system, and a greater complexity. These three components are highlighted in the studies described above, and we propose them as the principal ones to create novelty and heterogeneity in solutions. First, the communication between agents is shown to catalyze swarming and cooperation strategies. In previous work of turn-taking interaction between two agents installed with neural networks [17], we noticed that performing demo- cratic turn-taking offers novel styles of motion evolutionarily. Accordingly, here, 5 A swarm can be shown to act as a collective memory, either explicitly/statefully [55] or dynamically/statelessly [7]. How to Make Swarms Open-Ended? 33 the local interactions between agents in a flock allow for the swarm to take par- ticular shapes (Study 1), invent an explicit cooperative protocol (Study 2), and implement a noise-canceling policy (Study 3). To reach OEE, perhaps more than mere signaling, higher complexity levels of language may need to emerge. Second, the concurrent evolution algorithm essentially selects for meaningful information in behavioral space, by squeezing noisy behavior through a selective bottleneck. However, instead of using one unique objective function, the selection is distributed asynchronously in space and time. Differential timescales also helps accelerating the learning, which should happen as fast as possible, while retaining the way to generate the best found patterns discovered in the past. Lastly, once past the selection bottleneck, heterogeneity seems to increase considerably in genotypic space. Third, in terms of complexity, given the population size is large enough, with a consequently large number of degrees of freedom, we notice the swarm dynam- ics significantly change, in various ways. The flock's surface curvature may vary for large or small flocks, as well as the attraction and repulsion induced by the exchange of different signals. The motor responses may be amplified, since the input signals may significantly increase, given a higher density of neighboring agents, as seen in Study 1. Similarly, smaller flocks may display a more ordered behavior, with the trade-off however of being more sensitive to noise, since the critical mass is not reached to implement noise-canceling effects, as demonstrated in Study 3. Larger flocks can also be a source of individual behavioral differenti- ation, when a higher order of organization emerges. The key is not the size nor the amount of new information, but rather the system promoting the invention of new coordination patterns within itself. We have shown how collective intelligence has the ability to augment the creation of new and diverse solutions in a swarm, when given limited channels 34 O. Witkowski and T. Ikegami of communication, a concurrent evolution bottleneck and a large number of constrainted degrees of freedom. It come as an inspiration for scientists: a good way to build an open-ended system, able to indefinitely discover new inventions, seems not to reside in centralized computation, but rather in distributed systems, composed of large collectives of communicating agents. 6 Acknowledgements The authors would like to thank their collaborators who contributed partially to this work: Nathanael Aubert-Kato, Aleksandr Drozd, Yasuhiro Hashimoto, Norihiro Maruyama, Yoh-ichi Mototake and Mizuki Oka. Bibliography [1] Ackley, D. H. and Ackley, E. S. (2015). Artificial life programming in the robust-first attractor. In Andrews, P., Caves, L., Doursat, R., Hickinbotham, S., Polack, F., Stepney, S., Taylor, T., and Timmis, J., editors, Artificial Life Conference Proceedings 13, pages 554 -- 561. Cambridge, MA: MIT Press. [2] Aktipis, C. (2004). Know when to walk away: contingent movement and the evolution of cooperation. Journal of Theoretical Biology, 231(2):249 -- 260. [3] Bak, P. (2013). How nature works: the science of self-organized criticality. Springer Science & Business Media. [4] Bersini, H. and Detours, V. (1994). Asynchrony induces stability in cellular automata based models. In Brooks, R. A. and Maes, P., editors, Artificial Life IV, pages 382 -- 387. Cambridge, MA: MIT Press. [5] Chiong, R. and Kirley, M. (2012). Random mobility and the evolution of cooperation in spatial n-player iterated prisoner's dilemma games. Physica A: Statistical Mechanics and its Applications, 391(15):3915 -- 3923. [6] Conrad, M. and Pattee, H. (1970). Evolution experiments with an artificial ecosystem. Journal of Theoretical Biology, 28(3):393 -- 409. [7] Couzin, I. D., Krause, J., James, R., Ruxton, G. D., and Franks, N. R. (2002). Collective memory and spatial sorting in animal groups. Journal of theoretical biology, 218(1):1 -- 11. [8] Cucker, F. and Huepe, C. (2008). Flocking with informed agents. Mathe- matics in Action, 1(1):1 -- 25. [9] Drozd, A., Witkowski, O., Matsuoka, S., and Ikegami, T. (2016). Critical mass in the emergence of collective intelligence: a parallelized simulation of swarms in noisy environments. Artificial Life and Robotics, 21(3):317 -- 323. 36 O. Witkowski and T. Ikegami [10] Eberhart, R. C. and Kennedy, J. (1995). A new optimizer using particle swarm theory. In Proceedings of the sixth international symposium on micro machine and human science, volume 1, pages 39 -- 43. New York, NY: Institute of Electrical and Electronics Engineers. [11] Gardner, A. and West, S. A. (2010). Greenbeards. Evolution, 64(1):25 -- 38. [12] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial nets. In Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., and Weinberger, K. Q., editors, Advances in neural information processing systems, pages 2672 -- 2680. [13] Greenbaum, B. and Pargellis, A. (2016). Digital replicators emerge from a self-organizing prebiotic world. In Gershenson, C., Froese, T., Siqueiros, J. M., Aguilar, W., Izquierdo, E. J., and Sayama, H., editors, Proceedings of the European Conference on Artificial Life 13, pages 60 -- 67. Cambridge, MA: MIT Press. [14] Hauert, S., Zufferey, J.-C., and Floreano, D. (2009). Evolved swarming without positioning information: an application in aerial communication relay. Autonomous Robots, 26(1):21 -- 32. [15] Hauser, M. D., Chomsky, N., and Fitch, W. T. (2002). The faculty of lan- guage: What is it, who has it, and how did it evolve? science, 298(5598):1569 -- 1579. [16] Holland, J. H. (1999). Echoing emergence: Objectives, rough definitions, and speculations for echo-class models. In Cowan, G. A., Pines, D., and Meltzer, D., editors, Complexity, pages 309 -- 342. Cambridge, MA: Perseus Books. [17] Iizuka, H. and Ikegami, T. (2003). Adaptive coupling and intersubjectivity in simulated turn-taking behaviour. In Banzhaf, W., Ziegler, J., Christaller, T., Dittrich, P., and Kim, J. T., editors, Advances in Artificial Life, pages 336 -- 345, Berlin, Heidelberg. Springer Berlin Heidelberg. How to Make Swarms Open-Ended? 37 [18] Ikegami, T., Mototake, Y.-i., Kobori, S., Oka, M., and Hashimoto, Y. (2017). Life as an emergent phenomenon: studies from a large-scale boid simulation and web data. Phil. Trans. R. Soc. A, 375(2109):20160351. [19] Kauffman, S. (2003). The adjacent possible: A talk with stuart kauffman. Retrieved Nov, 30. [20] Kauffman, S. A. (2000). Investigations. Oxford University Press. [21] Khajehabdollahi, S. and Witkowski, O. (2018). Critical learning vs. evolu- tion: Evolutionary simulation of a population of ising-embodied neural net- works. In Ikegami, T., Virgo, N., Witkowski, O., Oka, M., Suzuki, R., and Iizuka, H., editors, Artificial Life Conference Proceedings, pages 47 -- 54. Cam- bridge, MA: MIT Press. [22] Kirby, S. (2002). Natural language from artificial life. Artificial life, 8(2):185 -- 215. [23] Lehman, J. and Stanley, K. O. (2011). Evolving a diversity of virtual crea- tures through novelty search and local competition. In Krasnogor, N., editor, Proceedings of the 13th Annual Conference on Genetic and Evolutionary Com- putation, GECCO '11, pages 211 -- 218, New York, NY, USA. ACM. [24] Lenski, R. E., Ofria, C., Pennock, R. T., and Adami, C. (2003). The evo- lutionary origin of complex features. Nature, 423(6936):139. [25] Lin, H. W., Tegmark, M., and Rolnick, D. (2017). Why does deep and cheap learning work so well? Journal of Statistical Physics, 168(6):1223 -- 1247. [26] Malan, K. M. and Engelbrecht, A. P. (2013). A survey of techniques for characterising fitness landscapes and some possible ways forward. Information Sciences, 241:148 -- 163. [27] Marwell, G. and Oliver, P. (1993). The critical mass in collective action. Cambridge University Press. [28] Maynard-Smith, J. and Sz´athmary, E. (1997). The Major Transitions in Evolution. Oxford University Press, Oxford, United Kingdom. 38 O. Witkowski and T. Ikegami [29] Mitri, S., Wischmann, S., Floreano, D., and Keller, L. (2013). Using robots to understand social behaviour. Biological Reviews, 88(1):31 -- 39. [30] Mototake, Y. and Ikegami, T. (2015). A simulation study of large scale swarms. pages 446 -- 450. [31] Nolfi, S. (1990). Learning and evolution in neural networks. [32] Nolfi, S., Floreano, D., and Floreano, D. D. (2000). Evolutionary robotics: The biology, intelligence, and technology of self-organizing machines. Cam- bridge, MA: MIT press. [33] Oliver, P. E. and Marwell, G. (2001). Whatever happened to critical mass theory? a retrospective and assessment. Sociological Theory, 19(3):292 -- 311. [34] Olson, R. S., Hintze, A., Dyer, F. C., Knoester, D. B., and Adami, C. (2013). Predator confusion is sufficient to evolve swarming behaviour. Journal of The Royal Society Interface, 10(85):20130305. [35] Prokopenko, M., Gerasimov, V., and Tanev, I. (2006). Evolving spatiotem- poral coordination in a modular robotic system. In International Conference on Simulation of Adaptive Behavior, pages 558 -- 569. Springer. [36] Rasmusen, E. and Blackwell, B. (1994). Games and information. Cam- bridge, MA, 15. [37] Reynolds, C. W. (1987). Flocks, herds and schools: A distributed behav- ioral model. In Stone, M. C., editor, ACM SIGGRAPH Computer Graphics, volume 21, pages 25 -- 34. ACM. [38] Schmickl, T., Zahadat, P., and Hamann, H. (2016). Sooner than expected: Hitting the wall of complexity in evolution. arXiv preprint arXiv:1609.07722. [39] Secretan, J., Beato, N., D Ambrosio, D. B., Rodriguez, A., Campbell, A., and Stanley, K. O. (2008). Picbreeder: Evolving pictures collaboratively on- line. In Czerwinski, M., Lund, A., and Tan, D., editors, Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI '08, pages 1759 -- 1768, New York, NY, USA. ACM. How to Make Swarms Open-Ended? 39 [40] Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al. (2017). Mastering the game of go without human knowledge. Nature, 550(7676):354. [41] Soros, L. B. and Stanley, K. O. (2014). Identifying necessary conditions for open-ended evolution through the artificial life world of chromaria. In Sayama, H., Rieffel, J., Risi, S., Doursat, R., and Lipson, H., editors, Proceedings of the Fourteenth International Conference on the Simulation and Synthesis of Living Systems (Artificial Life 14), pages 793 -- 800. Citeseer. [42] Standish, R. K. (2003). Open-ended artificial evolution. International Jour- nal of Computational Intelligence and Applications, 3(02):167 -- 175. [43] Stanley, K. O. and Lehman, J. (2015). Why greatness cannot be planned: The myth of the objective. London: Springer. [44] Su, H., Wang, X., and Lin, Z. (2009). Flocking of multi-agents with a virtual leader. Automatic Control, IEEE Transactions on, 54(2):293 -- 307. [45] Taylor, T., Bedau, M., Channon, A., Ackley, D., Banzhaf, W., Beslon, G., Dolson, E., Froese, T., Hickinbotham, S., Ikegami, T., McMullin, B., Packard, N., Rasmussen, S., Virgo, N., Agmon, E., Clark, E., McGregor, S., Ofria, C., Ropella, G., Spector, L., Stanley, K. O., Stanton, A., Timperley, C., Vostinar, A., and Wiser, M. (2016). Open-ended evolution: perspectives from the oee workshop in york. Artificial life, 22(3):408 -- 423. [46] Tishby, N., Pereira, F. C., and Bialek, W. (1999). The information bottle- neck method. In Hajek, B. and Sreenivas, R. S., editors, Proceedings of the 37th Annual Allerton Conference on Communication, Control and Computing, pages 368 -- 377. University of Illinois, Urbana-Champaign. [47] Tishby, N. and Polani, D. (2011). Information theory of decisions and ac- tions. In Cutsuridis, V., Hussain, A., and Taylor, J. G., editors, Perception- action cycle, pages 601 -- 636. London: Springer. 40 O. Witkowski and T. Ikegami [48] Tishby, N. and Zaslavsky, N. (2015). Deep learning and the information bottleneck principle. In 2015 IEEE Information Theory Workshop (ITW), pages 1 -- 5. [49] Toner, J. and Tu, Y. (1998). Flocks, herds, and schools: A quantitative theory of flocking. Physical review E, 58(4):4828. [50] Traulsen, A. and Nowak, M. A. (2006). Evolution of cooperation by multilevel selection. Proceedings of the National Academy of Sciences, 103(29):10952 -- 10955. [51] Von Neumann, J., Burks, A. W., et al. (1966). Theory of self-reproducing automata. IEEE Transactions on Neural Networks, 5(1):3 -- 14. [52] Wilson, D. S. and Sober, E. (1994). Reintroducing group selection to the human behavioral sciences. Behavioral and brain sciences, 17(4):585 -- 608. [53] Witkowski, O. and Aubert, N. (2014). Pseudo-static cooperators: Moving isn't always about going somewhere. In Sayama, H., Rieffel, J., Risi, S., Dour- sat, R., and Lipson, H., editors, Proceedings of the Fourteenth International Conference on the Simulation and Synthesis of Living Systems (Artificial Life 14), volume 14, pages 392 -- 397. [54] Witkowski, O. and Ikegami, T. (2014). Asynchronous evolution: Emergence of signal-based swarming. In Sayama, H., Rieffel, J., Risi, S., Doursat, R., and Lipson, H., editors, Proceedings of the Fourteenth International Conference on the Simulation and Synthesis of Living Systems (Artificial Life 14), volume 14, pages 302 -- 309. [55] Witkowski, O. and Ikegami, T. (2016). Emergence of swarming behav- ior: foraging agents evolve collective motion based on signaling. PloS one, 11(4):e0152756. [56] Wolpert, D. H. and Macready, W. G. (1997). No free lunch theorems for optimization. IEEE transactions on evolutionary computation, 1(1):67 -- 82.
1903.05431	1	1903	2019-03-13T11:54:04	Resource Abstraction for Reinforcement Learning in Multiagent Congestion Problems	[ "cs.MA", "cs.LG" ]	Real-world congestion problems (e.g. traffic congestion) are typically very complex and large-scale. Multiagent reinforcement learning (MARL) is a promising candidate for dealing with this emerging complexity by providing an autonomous and distributed solution to these problems. However, there are three limiting factors that affect the deployability of MARL approaches to congestion problems. These are learning time, scalability and decentralised coordination i.e. no communication between the learning agents. In this paper we introduce Resource Abstraction, an approach that addresses these challenges by allocating the available resources into abstract groups. This abstraction creates new reward functions that provide a more informative signal to the learning agents and aid the coordination amongst them. Experimental work is conducted on two benchmark domains from the literature, an abstract congestion problem and a realistic traffic congestion problem. The current state-of-the-art for solving multiagent congestion problems is a form of reward shaping called difference rewards. We show that the system using Resource Abstraction significantly improves the learning speed and scalability, and achieves the highest possible or near-highest joint performance/social welfare for both congestion problems in large-scale scenarios involving up to 1000 reinforcement learning agents.	cs.MA	cs	Resource Abstraction for Reinforcement Learning in Multiagent Congestion Problems ∗ Kleanthis Malialis Dept. of Computer Science University College London, UK [email protected] Sam Devlin Dept. of Computer Science University of York, UK [email protected] Daniel Kudenko Dept. of Computer Science University of York, UK [email protected] 9 1 0 2 r a M 3 1 ] A M . s c [ 1 v 1 3 4 5 0 . 3 0 9 1 : v i X r a ABSTRACT Real-world congestion problems (e.g. traffic congestion) are typically very complex and large-scale. Multiagent rein- forcement learning (MARL) is a promising candidate for dealing with this emerging complexity by providing an au- tonomous and distributed solution to these problems. How- ever, there are three limiting factors that affect the deploy- ability of MARL approaches to congestion problems. These are learning time, scalability and decentralised coordination i.e. no communication between the learning agents. In this paper we introduce Resource Abstraction, an approach that addresses these challenges by allocating the available re- sources into abstract groups. This abstraction creates new reward functions that provide a more informative signal to the learning agents and aid the coordination amongst them. Experimental work is conducted on two benchmark domains from the literature, an abstract congestion problem and a re- alistic traffic congestion problem. The current state-of-the- art for solving multiagent congestion problems is a form of reward shaping called difference rewards. We show that the system using Resource Abstraction significantly improves the learning speed and scalability, and achieves the highest possible or near-highest joint performance/social welfare for both congestion problems in large-scale scenarios involving up to 1000 reinforcement learning agents. CCS Concepts •Computing methodologies → Multi-agent systems; Multi-agent reinforcement learning; Keywords Multiagent learning; Resource abstraction; Congestion prob- lems 1. INTRODUCTION There has been in recent years an explosion of interest in multiagent systems (MAS) where multiple interacting agents ∗Most of this work was carried out when the author was at the University of York, UK. Appears in: Proceedings of the 15th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2016), J. Thangarajah, K. Tuyls, C. Jonker, S. Marsella (eds.), May 9 -- 13, 2016, Singapore. are situated in a common environment. A larger set of prob- lem domains can be practically modelled using MAS for ex- ample by taking advantage of the geographic distribution and sharing experiences for faster and better learning [8]. MAS often need to be very complex, and MARL is a promis- ing candidate for dealing with this emerging complexity [7]. MARL approaches provide adaptive and autonomous agents that can improve from experience. Congestion problems have been formulated in the past as MARL problems [10, 11, 6, 3]. Congestion problems are typ- ically very complex and large-scale. Real-life examples in- clude traffic congestion, air traffic management and network routing. Despite the benefits offered by MARL approaches there are three limiting factors that affect their deployability to real-world congestion problems; these are, learning time, scalability and decentralised coordination i.e. no communi- cation between learning agents. Firstly, learning time is of particular importance to offline learning and critical to online learning. If a MARL system is slow it may be practically useless even if it eventually achieves a great performance. Furthermore, in real-life com- plex congestion problems, learning time is typically limited. Secondly, the curse of dimensionality [9] refers to the ex- ponential growth of the search space, in terms of the number of states and actions. If a MARL solution is not scalable, it will never be considered, let alone adopted, for deployment by a company or organisation. Lastly, while sharing experiences may be beneficial and even necessary in specific applications, allowing communica- tion between the learning agents is not always desirable or even possible. This is because agents may have been devel- oped by different designers or controlled by different owners. Even if this is not the case, communication messages can be costly, noisy, corrupted, dropped or received out of order [5]. In this paper we introduce Resource Abstraction, an ap- proach that addresses the aforementioned challenges. Ex- periments are conducted on two benchmark domains from the literature; an abstract congestion problem and a realis- tic traffic congestion problem where agents or cars "compete" for traffic lanes. Our contributions are the following. We propose the novel Resource Abstraction approach where its main idea is to allocate the available resources into groups called abstract groups. This grouping provides an abstract (A) reward function that is more informative and aids the coordination among the learning agents to significantly im- prove the learning speed, scalability and agents' final joint performance or social welfare. We compare our approach against systems which learn from local (L), global (G) and difference (D) rewards. The first two are the common approaches while the third is a form of reward shaping that constitutes the state-of-the-art for solving multiagent congestion problems. Like D, the system using A provides decentralised coor- dination. It is further shown that the system using A re- quires significantly fewer time steps per learning episode to achieve the highest system performance or social welfare in large-scale scenarios. The proposed approach scales to a large number of agents and is successfully demonstrated in experiments involving up to 1000 learning agents. The organisation of this paper is as follows. Section 2 describes the background material necessary to understand the contributions made by this work. Section 3 presents a formal definition of congestion problems as multiagent learn- ing and coordination problems. Section 4 introduces in de- tail our proposed Resource Abstraction approach. Section 5 and Section 6 provide the experimental work and results for the abstract and traffic congestion problems respectively. A discussion and conclusion is presented in Section 7. 2. BACKGROUND 2.1 Reinforcement Learning Reinforcement learning is a paradigm in which an active decision-making agent interacts with its environment and learns from reinforcement, that is, a numeric feedback in the form of reward or punishment [9]. The feedback re- ceived is used to improve the agent's actions. The problem of solving a reinforcement learning task is to find a policy (i.e. a mapping from states to actions) which maximises the accumulated reward. The concept of an iterative approach constitutes the back- bone of the majority of reinforcement learning algorithms. These algorithms apply so called temporal-difference up- dates to propagate information about values of states, V (s), or state-action, Q(s, a), pairs. These updates are based on the difference of the two temporally different estimates of a particular state or state-action value. The Q-learning al- gorithm is such a method [12]. After each real transition, (s, a) → (s(cid:48), r), in the environment, it updates state-action values by the formula: Q(s, a) ← Q(s, a) + α[r + γ max (cid:48) a(cid:48) Q(s (cid:48) ) − Q(s, a)] , a (1) where α is the rate of learning and γ is the discount factor. The exploration-exploitation trade-off constitutes a crit- ical issue in the design of a reinforcement learning agent. It aims to offer a balance between the exploitation of the agent's knowledge and the exploration through which the agent's knowledge is enriched. A common method of doing so is -greedy, where the agent behaves greedily most of the time, but with a probability it selects an action randomly. To get the best of both exploration and exploitation, it is advised to reduce over time [9]. Applications of MARL typically take one of two approaches; multiple individual learners (ILs) or joint action learners (JALs) [1]. Multiple ILs assume any other agents to be part of the environment and so, as the others simultaneously learn, the environment appears to be dynamic as the prob- ability of transition when taking action a in state s changes over time. To overcome the appearance of a dynamic envi- ronment, JALs were developed that extend their value func- tion to consider for each state the value of each possible combination of actions by all agents. The consideration of the joint action causes an exponential increase in the num- ber of values that must be calculated with each additional agent added to the system. Therefore, as we are interested in scalability and decentralised coordination (i.e. no com- munication between agents), this work focuses on multiple individual learners and not joint action learners. In MARL the typical approach provides agents with one of two types of reward. A local reward (L) is unique to each agent and often promotes "selfish" behaviours since each agent attempts to increase its own reward, potentially at the cost of the system performance. The global reward (G) is in fact the system performance used as a learning signal1. This allows each agent to act in the system's interest and typically, the learnt behaviour is better compared to the case where agents are learning using L. G, however, includes a substantial amount of noise due to other agents acting in the system. This is because an agent may be rewarded for taking a bad action (if the other agents executed a good action), or punished for taking a good ac- tion (if the others performed a bad action). This is known as the multiagent credit assignment problem. The reward signal should reward an agent depending on its individual contribution to the system objective. Difference rewards [13], a form of reward shaping, address this issue and are discussed in the next section. 2.2 Difference Rewards Difference rewards [13] were introduced to address the multiagent credit assignment problem encountered in re- inforcement learning. The difference rewards (Di) is a shaped reward signal that helps an agent i to learn the con- sequences of its actions on the system objective by removing a large amount of the noise created by the actions of other agents active in the system. It is defined as: Di(z) = G(z) − G(z−i) (2) where z is a general term representative of either states or state-action pairs depending on the application, G(z) is the global reward used, and G(z−i) is G(z) for a theoretical system without the contribution of agent i. Difference rewards exhibit the following two properties. Firstly, any action taken that increases Di simultaneously increases G. Secondly, since difference rewards only depend on the actions of agent i, Di provides a cleaner signal with reduced noise. These properties allow for the difference re- wards to significantly boost learning performance in a MAS. The challenge for deriving the difference rewards signal is obviously how to calculate the second term of the equation G(z−i) which is called the counterfactual. In particular do- mains, the counterfactual G(z−i) is possible to be directly calculated [11, 3]. Alternatively, in cases where this is not possible, it has been demonstrated that difference rewards can be estimated [10, 6]. Difference rewards have been successfully applied in dif- ferent domains such as congestion problems [10, 11, 6, 3], network security [4] and distributed sensor networks [2]. 1The terms global reward, system performance and social welfare are used interchangeably. 3. CONGESTION PROBLEMS AS MULTI- AGENT LEARNING PROBLEMS This section provides a formal definition of congestion problems as multiagent learning problems. A congestion problem occurs when multiple entities are competing for a limited amount of available resources. Different congestion problems (traffic congestion, air traffic management, Beach problem, El Farol Bar problem) have been formulated in the past as multiagent learning problems [10, 11, 6, 3]. Formally, there exists a set S of n resources S = {s1, s2, ..., sn}. Each resource s is a 3-tuple s =< ws, ψs, xs,t > where ws ≥ 0 is the weighting or importance of resource s, ψs > 0 is the capacity of resource s, and xs,t ≥ 0 is the consumption of resource s at time t. A resource s at time t is said to be congested if its consumption is higher than its capacity i.e. xs,t > ψs. The local reward (L) of an entity located in resource s at time t depends on the resource's weighting, capacity and consumption as shown in Equation 3: L(s, t) = f (ws, ψs, xs,t) (3) The global reward (G) is defined as the summation over all local rewards and is given by Equation 4. G(t) = L(s, t) (4) (cid:88) s∈S Recall from Section 2.2 that the difference rewards sig- nal is calculated by subtracting the counterfactual from the global reward. Therefore, the difference rewards (D) for agent i located in section s at time t can be calculated using Equation 5. Di(t) = G(t) − G−i(t) = L(s, t) − L−i(s, t) (5) The straightforward approach is to consider the entities which compete for the resources as the learning agents. For instance, in a traffic congestion domain [11] where the enti- ties are cars which compete for traffic lanes, we can install a learning agent on each car. In such domains, there is no need to estimate the counterfactual as it can simply be calculated by decreasing the resource's consumption or attendance by 1 as shown in Equation 6 Di(t) = L(s, t) − L−i(s, t) = f (ws, ψs, xs,t) − f (ws, ψs, xs,t − 1) (6) Alternatively, there exist congestion problems where learn- ing agents are not installed on the entities competing for the resources and in such cases the counterfactual needs to be estimated [10]. Although estimating Di has been success- fully demonstrated in a number of applications this topic is under ongoing research [6]. 4. RESOURCE ABSTRACTION The idea behind our approach is to allocate the available resources into groups called abstract groups which provide a more informative signal to the learning agents. {b1, b2, ..., bp}. Each abstract group b is a 4-tuple: b =< Mb, Wb, Ψb, Xb,t >. Formally, there exists a set B of p abstract groups: B = The set Mb = {m1, m2, ..., mk} is the member set of ab- stract group b that consists of k resources. A resource m can belong only to one abstract group, that is, m ∈ Mi and m ∈ Mj only if i = j. The number of abstract groups (and their members) depends on the application domain. The weight Wb of abstract group b is defined as the average weight of its members as shown in Equation 7. wm (7) The capacity Ψb of abstract group b is the total capacity of its members. Similarly, the consumption Xb,t of abstract group b at time t is the total consumption of its members. These are calculated using Equations 8 and 9. (cid:88) m∈Mb Wb = 1 k (cid:88) (cid:88) m∈Mb m∈Mb Ψb = Xb,t = ψm xm,t (8) (9) As discussed, the straightforward approach is to consider the entities which compete for the resources as the learning agents. However, depending on the application, this may not necessarily be the case, and our proposed approach does not impose any limitations on the agent selection. Let us now introduce the proposed reward functions. We start by extending the local reward L(s, t), and define the signal H(b, t) for an abstract group b at time t. It is cal- culated by taking into account the abstract group's weight, capacity and consumption as shown in Equation 10. It is important to note that the function f (·) is the same which is used to calculate the local reward L (see Equation 3). H(b, t) = −f (Wb, Ψb, Xb,t) (10) The abstract reward (A) is defined as follows. When re- source s which belongs to abstract group b is not congested, we provide a reward from L(s, t). Alternatively, we provide a reward from the abstraction configuration i.e. from H(b, t). This is shown in Equation 11. (cid:40) L(s, t) H(b, t) if xs,t ≤ ψs if xs,t > ψs (11) A(b, s, t) = where s ∈ Mb. The intuition behind this reward function is to facilitate the coordination among the learning agents by providing them with a more informative signal. Assuming a "good" abstraction configuration (as mentioned, this is application domain-specific), the abstract reward (A) will provide a big- ger (than L) punishment to the agents to encourage them to keep resources decongested. Furthermore, as we will later demonstrate, in domains where the number of entities com- peting for the resources is greater than the number of avail- able resources, the desired solution is to congest only a min- imal amount of them (e.g. only one) and leave the rest decongested. The approach will encourage some agents to consume a congested resource. In such cases the agents al- ready receive a punishment, but they will be provided with a bigger punishment if the agents attempt to switch to a decongested abstract group and congest it. 5. BEACH PROBLEM DOMAIN (BPD) 5.1 Domain Description The Beach Problem Domain (BPD) [3] is an abstract con- gestion problem that relates to many real-life congestion problems and allows us to perform a thorough analysis of our proposed approach. In this domain, each tourist (learn- ing agent) gets to choose the beach section it will visit. At each time step each agent knows which beach section it is currently on and must choose to either stay still or move to an adjacent beach section (i.e. left or right). When all the agents have performed their actions, they receive a reward. The highest reward for a beach section is received when the attendance of agents is equal to the capacity of the beach section. If a beach section gets con- gested or overcrowded it is undesirable. Beach sections with low attendance are also undesirable. This constitutes a con- gestion problem when the number of agents is much greater than the total capacity of the beach sections. In this abstract congestion problem it is assumed that all beach sections have an equal weighting of 1 and they all have the same capacity ψ. The local reward function (L) is given by Equation 12: L(s, t) = xs,te −xs,t ψ (12) for timestep = 1 : num timesteps do for agent = 1 : num agents do move to nearest section s(cid:48) ∈ S perceive current beach section s ∈ S choose action a = {−1, 0, +1} using -greedy move to section s(cid:48) = {s − 1, 0, s + 1} if s(cid:48) /∈ S then Algorithm 1 Beach Problem with Resource Abstraction 1: define abstraction configuration B 2: initialise Q-values: ∀s, aQ(s, a) = −1 3: set agents to initial locations 4: for episode = 1 : num episodes do 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: end for end if end for for section = 1 : num sections do calculate A reward (Equation 16) end for reduce α using alpha decay rate reduce using epsilon decay rate end for for agent = 1 : num agents do end for reset agents to initial locations update Q(s,a) using A (Equation 1) where s is the beach section and xs,t is the attendance of beach section s at time step t. 5.2 Experimental Setup The global reward (G) is a summation over all local re- wards and is given by Equation 13. G(t) = −xs,t ψ xs,te (13) (cid:88) s∈S The difference rewards (D) signal is calculated by apply- ing Equation 6 to Equation 12. This results in Equation 14. Di(t) = xs,te −xs,t ψ − (xs,t − 1)e −(xs,t−1) ψ (14) Similarly to the local reward function, we define the cor- responding reward function for an abstraction configuration B (defined as described in Section 4). Therefore, the appli- cation of Equation 10 to Equation 12 results in Equation 15: H(b, t) = −Xb,te −Xb,t Ψb (15) where b is the abstract group, Xb,t is the attendance at time step t, and Ψb is the capacity of abstract group b. Lastly, the abstract reward (A) is calculated by applying Equation 11 and is shown in Equation 16. xs,te −Xb,te −xs,t ψ −Xb,t Ψb if xs,t ≤ ψ if xs,t > ψ (16) A(b, s, t) = where s ∈ Mb. Algorithm 1 presents what has been described thus far. Before describing the experimental work, we should note that the best solution for this problem is to overcrowd one (any) beach section, and leave ψ number of agents in the rest of the sections. This will give the highest possible sys- tem performance or global reward. Our experimental setup is as follows. The learning rate is set to α = 0.1 and alpha decay rate = 0.9999. The explo- ration parameter is set to = 0.05 and epsilon decay rate = 0.9999. The discount factor is set to γ = 1.0 and the num- ber of episodes is set to num episodes = 10000. Initially, the agents are uniformly distributed (i.e. it is assumed that there exist hotels for the tourists in each beach section). The rest of the parameters are given later for each experiment. In all experiments, we plot the system performance/global reward G at the last time step of each episode. The values are averaged over 30 statistical runs and error bars showing the standard error around the mean are plotted. In some plots the error bars are very small and hardly visible, but they are present on all plots. 5.3 Experimental Results We have set num agents = 100, num sections = 6, ca- pacity ψ = 6 and num timesteps = 5. The highest social welfare or global reward is 11.04 and occurs when one (any) of the six sections is overcrowded with 70 agents, while each of the remaining five sections have six agents; this is shown with a black dashed line in all figures. The first experimental study investigates how different ab- straction (A) configurations affect the performance, and how they compare to the approaches which use the local (L), global (G) and difference (D) rewards. Figure 1 demon- strates this with abstraction configurations of two and three abstract groups. For example, the plot with label "A-4+2" means that the first four beach sections are grouped into one abstract group, while the remaining two beach sections are grouped into another abstract group. There are four important outcomes from this study: • Abstraction selection does affect the system perfor- mance. Figure 1: Abstraction configurations Figure 2: Impact of episode length on D • However, all the abstractions outperform the system that uses the difference (D) rewards (third plot from the bottom). • The abstraction "A-2+1+3" achieves the highest per- formance while three other achieve a near-highest per- formance. The system using D never achieves the high- est social welfare. • All the abstractions outperform the systems that use the local (L) and global (G) rewards (bottom two plots). The second experimental study examines how the episode length i.e. number of time steps per episode affects the system performance. Recall that at the start of each new episode, that is, after the last time step of the previous episode (Algorithm 1/Line 23) the agents return to their initial location. It is important to note that changing the length of the episode produces a different and independent version of the problem. In Figure 2 we apply D in different versions of the BPD, where we vary the time (episode length) agents are allowed to settle to a beach section within an episode. Note that with five or more time steps per episode, an agent located at the far left (or right) beach section can, in principle, move to the far right (or left) section at the end of an episode. However, D achieves the highest performance only in BPD versions with episode length of 15 time steps or more. We now run the same experiment for the proposed Re- source Abstraction approach. Based on previous findings (Figure 1) we proceed with abstraction "A-2+1+3". Figure 3 depicts the performance of A in different BPD versions. It is observed that the highest social welfare is achieved in all BPD versions except when a single time step is used. Also, A learns faster and achieves a significantly higher final perfor- mance than D in all BPD versions except from versions with episode length of 15 steps or more, where both have a sim- ilar performance. These are important results for potential real-world applications as they show that A is more robust to changes in the problem domain (i.e. episode length) than the current state-of-the-art approach D. The third experimental study of BPD aims at investigat- ing the scalability of the proposed approach. The setup for this study involves num agents = 1000, num sections = Figure 3: Impact of episode length on A 20, capacity ψ = 18 and num timesteps = 5. The highest social welfare is 125.82 and is shown with a black dashed line in all figures. Like before, we have investigated a range of abstraction configurations and many of them are shown in Figure 4. We observe the following. All the abstractions except one (third plot from the bottom) outperform the system that uses the difference (D) rewards (fourth plot from the bottom). All the abstractions outperform the systems that use the lo- cal (L) and global (G) rewards (bottom two plots). Note that no approach manages to achieve the best performance in just five time steps per episode. However, taking into consideration the scale of the scenario (1000 agents, 20 sec- tions), some of the abstractions (see upper plots in Figure 4) perform extremely well. We proceed with abstraction "A-8+1+11" and repeat the same experiment with ten and 20 time steps per episode; results are shown in Figures 5 and 6 respectively. In both figures, the system using A achieves a near-highest perfor- mance. The system using D performs orders of magnitude better than the local (L) and global (G) rewards, but still it is significantly outperformed by the proposed A approach. Figure 4: BPD with 1000 agents, 5 time steps Figure 6: BPD with 1000 agents, 20 time steps The local reward (L) function is given in Equation 17: (cid:40)wse−1 −xs,t ψs wse L(s, t) = if xs,t ≤ ψs if xs,t > ψs (17) where s is the traffic lane, ws is the weighting, xs,t is the attendance at time step t and ψs is the capacity. The global (G) and difference (D) rewards are calculated using Equations 4 and 6 respectively. The application of Equation 10 to Equation 17 is shown in Equation 18: H(b, t) = (cid:40)−Wbe−1 −Wbe −Xb,t Ψb if Xb,t ≤ Ψb if Xb,t > Ψb (18) Figure 5: BPD with 1000 agents, 10 time steps 6. TRAFFIC LANE DOMAIN (TLD) 6.1 Domain Description Despite the BPD being an abstract congestion problem, it is very important as it relates to many real-world congestion problems. One such complex problem is the Traffic Lane Domain (TLD) [11] which is defined as follows. In this domain each car (learning agent) gets to choose which traffic lane to follow. Each agent knows at each time step which lane it currently follows. At each time step, each agent can either stay in the same lane, or move to an ad- jacent lane (i.e. left or right). Once all the agents have executed their actions, they receive a reward. To simulate real-life characteristics there are three major differences with the BPD: 1. traffic lanes have a different weighting (represents the different preference of drivers) 2. lanes have different capacities (represents the fact that lanes are of different sizes, or that there exist restrictions due to construction works, tolls, carpools etc.) and 3. it has a different reward function as it doesn't matter how many cars are on a lane as long as the lane is not congested (as opposed to a beach section which is desirable to have more people as long as it is not overcrowded). where b is the abstract group, Wb is the weighting, Xb,t is the attendance at time step t, and Ψb is the capacity. Lastly, the abstract reward (A) is calculated using Equation 11. 6.2 Experimental Setup We have set α = 0.1, alpha decay rate = 0.9999, = 0.05, epsilon decay rate = 0.9999, γ = 1.0 and num episodes = 10000. Initially, the agents are uniformly distributed. The rest of the parameters are given below. In all experiments, we plot the system performance/global reward G at the last time step of each episode (except where otherwise is stated). The values are averaged over 30 statistical runs and error bars showing the standard error around the mean are plot- ted. In some plots the error bars are very small and hardly visible, but they are present on all plots. 6.3 Experimental Results For the first experimental study in the TLD we adopt the same traffic scenario as in [11]. In this scenario, there are 500 cars or learning agents and 9 traffic lanes each with capacity 167, 83, 33, 17, 9, 17, 33, 83, 167 (i.e. a total of 609). The weightings for each lane are 1, 5, 10, 1, 5, 10, 1, 5, 10. The highest possible system performance or global reward is 17.66 and is achieved when all the lanes are decongested; this is shown with a black dashed line in all figures. The fact that there are different weightings for each re- source or traffic lane gives us the opportunity to try non- contiguous abstractions i.e. abstract groups with non-adjacent Figure 7: TLD with 500 cars, 5 time steps Figure 9: TLD with accidents The second experimental study in the TLD provides an- swers to these questions. To simulate such a behaviour we assume that two accidents occurred in lanes 3 and 9 at episode 2000. As a result, their capacity is reduced to half. Also, each of the two lanes has swapped weighting (pref- erence) with one of its adjacent lanes. Therefore, the new capacities are 167, 83, 17, 17, 9, 17, 33, 83, 83 (i.e. a total of 509) and the new weightings are 1, 10, 5, 1, 5, 10, 1, 10, 5. We keep the same number of cars i.e. 500 and at the time of the accidents the agents' exploration parameter is reset to the initial. The number of time steps per episode is five. Figure 9 shows the results for the scenario with the ac- cidents. The system using L is affected the most and does not manage to recover i.e. to reach the performance it had achieved before the occurrence of the accidents. Interest- ingly the system using G is not affected at all, but its per- formance is poor. The system using A is affected but man- ages to completely recover and achieve the highest possible welfare as it used to be the case without the occurrence of the accidents. The system using D is also affected but only slightly (it is not visible in the graph as it is out-shaded by the decline of the A plots), however, the only system that achieves the best performance is the one using A. The out- come of this study suggests that a system using the proposed approach is robust and does not need to refine the abstrac- tion configuration if something is altered in due course such as the traffic lane capacities and weightings. It has been assumed so far that all learning agents or cars are participating in the learning scheme. The third exper- imental study for the TLD examines the behaviour of the approaches when this assumption is violated. In practise, 100% compliance by drivers is unlikely to be achieved [11]. This is because some drivers may not be convinced to par- ticipate or even if they are all willing to participate, some of them may come across different problems such as informa- tion and sensing limitations. Figure 10 depicts the system performance (num agents = 500, num timesteps = 5) in the presence of 25% non-compliant drivers. Non-compliant drivers are simulated as agents stick- ing to their initial choice/location. It is observed that the performance of L and G is not affected at all, but it is very poor. The performance of D and A is not affected much and, as a result, the system using A still outperforms the Figure 8: TLD with 500 cars, 10 time steps members. The obvious configuration is to have three ab- stract groups each with three member lanes. The first, sec- ond and third abstract groups include lanes with weight- ing 1, 5, and 10 respectively. We will refer to this abstrac- tion configuration as "A (non-contig)". Moreover, like in the BPD, we have experimented with many abstraction configu- rations and found out that the best performance is obtained with "A-1+8", which we will refer to as "A (contig)". Figure 7 shows the performance for all approaches for five time steps per episode. The system using abstract rewards (A) achieves the highest performance while both A and D completely outperform L and G. The same experiment with ten time steps per episode is repeated in Figure 8 where very similar results are obtained. Note that with ten time steps, an agent located at the far left (or right) lane has, in principle, adequate time to reach the far right (or left) end. Despite the increase in the number of time steps, the system performance using D is only slightly improved. An important question is how will the system perform if the lane weightings and capacities change in due course e.g. because of an accident? Consequently, a crucial question that certainly has deployment implications is will a new ab- straction configuration be needed each time the lane weight- ings and capacities change? Figure 10: TLD with 25% non-compliant drivers Figure 12: TLD (1000 cars) with accidents 7. DISCUSSION AND CONCLUSION We have introduced Resource Abstraction for multiagent coordination problems. Our approach allocates available re- sources into abstract groups to provide a more informative signal to learning agents, significantly improving both learn- ing speed and social welfare. Experimental work was conducted on two benchmark do- mains, an abstract congestion problem and a realistic traffic congestion problem. The state-of-the-art solution to solve multiagent congestion problems is a form of reward shaping called difference rewards. We have shown that our novel approach of resource abstraction requires significantly fewer time steps per learning episode to achieve the highest social welfare. This is important for real-world applications where the availability of learning time is limited. Furthermore, it has been empirically demonstrated in both experimental domains that the proposed approach not only performs better than the two baseline and the difference re- wards approaches, but it achieves the highest possible or near-highest system performance even in experiments in- volving 1000 learning agents. Scalability is another strong aspect of the proposed approach. It has been shown that the proposed approach is robust against changes (e.g. car accidents in the traffic congestion domain) and capable of operating without 100% agent compliance. It has been shown that when the number of learning steps is limited, then in almost every case Resource Abstraction outperforms the state-of-the-art irrespective of the abstrac- tion grouping (Figures 1 and 4). When the number of learn- ing steps is not limited, we suggest some guidelines that help us define the abstraction grouping. Some experimentation is still necessary but the following heuristics simplify this pro- cess: a. three abstract groups typically suffice as this allows more resources to be included in an abstract group, hence, a larger penalty can be provided if the group becomes con- gested b. group resources with similar characteristics e.g. the same weightings; the effectiveness of this heuristic is shown by A (non-contig) in Figures 7 - 12. Moreover, it has been shown that if something unexpected occurs (e.g. accidents in the traffic lane domain) it is highly likely that there will be no need to refine the abstraction grouping, as, given some additional exploration time, the system can recover and still achieve the highest performance. Figure 11: TLD with 4 delivery phases system that uses D. The results suggest that the proposed system works well under non-compliant drivers. The fourth experimental study further investigates the po- tential deployment of such a learning scheme. In practise, it is expected that such a system will be rolled out in phases [11]. For instance, consider a scenario of four delivery phases where the participation is 25%, 50%, 75% and 100% of the drivers. To simulate this situation we vary the rate of non- compliant drivers, for example, the first phase with 25% participation corresponds to 75% non-compliant drivers. Figure 11 shows the converged performance (num agents = 500, num timesteps = 5) for the scenario with four delivery phases where the superiority of the proposed approach is clearly demonstrated, suggesting that the proposed system can be introduced to the public in different stages. In the last study we repeat the experiment with 1000 cars. The highest performance is 17.33 and occurs when the first lane (capacity 167, weighting 1) is overcrowded with 558 agents while the remaining eight lines have full capacity. We repeat the case under accident conditions (as in Figure 9), where the highest performance is 17.31 and occurs when the first lane is overcrowded with 658 agents. Figure 12 shows the results (num timesteps = 5) where the superiority and the ability of the proposed approaches to recover is shown. REFERENCES [1] C. Claus and C. Boutilier. The dynamics of reinforcement learning in cooperative multiagent systems. In Proceedings of the AAAI Conference on Artificial Intelligence, 1998. [2] M. Colby and K. Tumer. Multiagent reinforcement learning in a distributed sensor network with indirect feedback. In Proceedings of the 2013 international conference on Autonomous agents and multi-agent systems, pages 941 -- 948, 2013. [3] S. Devlin, L. Yliniemi, K. Tumer, and D. Kudenko. Potential-based difference rewards for multiagent reinforcement learning. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems, 2014. [4] K. Malialis, S. Devlin, and D. Kudenko. Distributed reinforcement learning for adaptive and robust network intrusion response. Connection Science, 27(3):234 -- 252, 2015. [5] M. J. Matari´c. Using communication to reduce locality in distributed multi-agent learning. Journal of Experimental and Theoretical Artificial Intelligence, special issue on Learning in DAI Systems, Gerhard Weiss, ed., 10(3):357 -- 369, 1998. [6] S. Proper and K. Tumer. Multiagent learning with a noisy global reward signal. In Proceedings of the AAAI Conference on Artificial Intelligence, 2013. [7] P. Stone. Learning and multiagent reasoning for autonomous agents. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 12 -- 30, 2007. [8] P. Stone and M. Veloso. Multiagent systems: A survey from a machine learning perspective. Autonomous Robots, 8(3):345 -- 383, 2000. [9] R. S. Sutton and A. G. Barto. Introduction to Reinforcement Learning. MIT Press Cambridge, MA, USA, 1998. [10] K. Tumer and A. Agogino. Distributed agent-based air traffic flow management. In Proceedings of the 6th International Joint Conference on Autonomous Agents and Multiagent Systems, 2007. [11] K. Tumer, Z. Welch, and A. K. Agogino. Traffic congestion management as a learning agent coordination problem. In A. Bazzan and F. Kluegl, editors, Multiagent Architectures for Traffic and Transportation Engineering, pages 261 -- 279. Lecture notes in AI, Springer, 2009. [12] C. J. Watkins and P. Dayan. Q-learning. Machine learning, 8(3-4):279 -- 292, 1992. [13] K. R. W. Wolpert, David H. and K. Tumer. Collective intelligence for control of distributed dynamical systems. Europhysics Letters, 49(6), 2000.
1711.06871	2	1711	2018-07-25T03:30:42	Anonymous Hedonic Game for Task Allocation in a Large-Scale Multiple Agent System	[ "cs.MA", "cs.AI", "cs.GT" ]	This paper proposes a novel game-theoretical autonomous decision-making framework to address a task allocation problem for a swarm of multiple agents. We consider cooperation of self-interested agents, and show that our proposed decentralized algorithm guarantees convergence of agents with social inhibition to a Nash stable partition (i.e., social agreement) within polynomial time. The algorithm is simple and executable based on local interactions with neighbor agents under a strongly-connected communication network and even in asynchronous environments. We analytically present a mathematical formulation for computing the lower bound of suboptimality of the solution, and additionally show that 50% of suboptimality can be at least guaranteed if social utilities are non-decreasing functions with respect to the number of co-working agents. The results of numerical experiments confirm that the proposed framework is scalable, fast adaptable against dynamical environments, and robust even in a realistic situation.	cs.MA	cs	Anonymous Hedonic Game for Task Allocation in a Large-Scale Multiple Agent System Inmo Jang, Hyo-Sang Shin, and Antonios Tsourdos 1 8 1 0 2 l u J 5 2 ] A M . s c [ 2 v 1 7 8 6 0 . 1 1 7 1 : v i X r a Abstract -- This paper proposes a novel game-theoretical au- tonomous decision-making framework to address a task allo- cation problem for a swarm of multiple agents. We consider cooperation of self-interested agents, and show that our proposed decentralized algorithm guarantees convergence of agents with social inhibition to a Nash stable partition (i.e., social agreement) within polynomial time. The algorithm is simple and executable based on local interactions with neighbor agents under a strongly- connected communication network and even in asynchronous environments. We analytically present a mathematical formu- lation for computing the lower bound of suboptimality of the solution, and additionally show that 50% of suboptimality can be at least guaranteed if social utilities are non-decreasing functions with respect to the number of co-working agents. The results of numerical experiments confirm that the proposed framework is scalable, fast adaptable against dynamical environments, and robust even in a realistic situation. Index Terms -- Distributed robot systems, Networked robots, Task allocation, Game theory, Self-organising systems I. INTRODUCTION Cooperation of a large number of possibly small-sized robots, called robotic swarm, will play a significant role in complex missions that existing operational concepts using a few large robots could not deal with [1]. Even if every single robot (or called agent) in a swarm is incapable of accomplishing a task alone, their cooperation will lead to successful outcomes [2] -- [5]. The possible applications include environmental monitoring [6], ad-hoc network relay [7], dis- aster management [8], cooperative radar jamming [9], to name a few. Due to the large cardinality of a swarm robot system, however, it is infeasible for human operators to supervise each agent directly, but needed to entrust the swarm with certain levels of decision-makings (e.g., task allocation, path planning, and individual control). Thereby, what only remains is to provide a high-level mission description, which is manageable for a few or even a single human operator. Nevertheless, there still exist various challenges in the autonomous decision- making of robotic swarms. Among them, this paper addresses a task allocation problem where the number of agents is higher than that of tasks: how to partition a set of agents into subgroups and assign the subgroups to each task. In the problem, it is assumed that each agent can be assigned to at most one task, whereas each task may require multiple agents: Inmo Jang, Hyo-Sang Shin, and Antonios Tsourdos are with Centre for Autonomous and Cyber-Physical Systems, Cranfield University, MK43 0AL, United Kingdom (e-mail: [email protected]; [email protected]; [email protected]). this case falls into ST-MR (single-task robot and multi-robot task) category [10], [11]. According to [4], [5], [12] -- [14], decision-making frame- works for a robotic swarm should be decentralized (i.e., the desired collective behavior can be achieved by individual information), scalable, predictable agents relying on local (e.g., regarding convergence performance and outcome qual- ity), and adaptable to dynamic environments (e.g., unexpected elimination or addition of agents or tasks). Moreover, the frameworks are also desirable to be robust against asyn- chronous environments because, due to the large cardinality of the system and its decentralization, it is very challenging for every agent to behave synchronously. For synchronization in practice, "artificial delays and extra communication must be built into the framework" [14], which may cause considerable inefficiency on the system. In addition, it is also preferred to be capable of accommodating different interests of agents (e.g., different swarms operated by different organizations [15]). In this paper, we propose a novel decision-making frame- work based on hedonic games [16] -- [18]. The task allocation problem considered is modeled as a coalition-formation game where self-interest agents are willing to form coalitions to improve their own interests. The objective of this game is to find a Nash stable partition, which is a social agreement where all the agents agree with the current task assignment. Despite any possible conflicts between the agents, this paper shows that if they have social inhibition, then a Nash stable partition can always be determined within polynomial times in the proposed framework and all the desirable characteristics mentioned above can be achieved. Furthermore, we analyze the lower bound of the outcome's suboptimality and show that 50% is at least guaranteed for a particular case. Various settings of numerical experiments validate that the proposed framework is scalable, adaptable, and robust even in asyn- chronous environments. This paper is organized as follows. Section II reviews exist- ing literature on decentralized task allocation approaches and introduces a recent finding in hedonic games that inspires this study. Section III proposes our decision-making framework, named GRAPE, and analytically proves the existence of and the polynomial-time convergence to a Nash stable partition. Section IV discusses the framework's algorithmic complexity, suboptimality, adaptability, and robustness. Section V shows that the framework can also address a task allocation problem in which each task may need a certain number of agents for completion. Numerical simulations in Section VI confirm that the proposed framework holds all the desirable characteristics. Finally, concluding remarks are followed in Section VII. II. RELATED WORK A. Decentralized Coordination of Robotic Swarms Existing approaches for task allocation problems can be categorized into two branches, depending on how agents even- tually reach a converged outcome: orchestrated and (fully) self- organized approaches [19]. In the former, additional mech- anism such as negotiation and voting model is imposed so that some agents can be worse off if a specific condition is met (e.g., the global utility is better off). Alternatively, in self-organized approaches, each agent simply makes a decision without negotiating with the other agents. The latter generally induce less resource consumption in communication and computation [20], and hence they are preferable in terms of scalability. On the other hand, the former usually provide a better quality of solutions with respect to the global utility, and a certain level of suboptimality could be guaranteed [21] -- [23]. A comparison result between them [20] presents that as the available information to agents becomes local, the latter becomes to outperform the former. In the following, we particularly review existing literature on self-organized approaches because, for large-scale multiple agent systems, scalability is at least essential and it is realistic to regard that the agents only know their local information but instead the global information. Self-organized approaches can be categorized into top-down approaches and bottom-up approaches according to which level (i.e., between an ensemble and individuals) is mainly focused on. Top-down approaches emphasize developing a macroscopic model for the whole system. For instance, popu- lation fractions associated with given tasks are represented as states, and the dynamics of the population fractions is modeled by Markov chains [12], [24] -- [26] or differential equations [27] -- [31]. Given a desired fraction distribution over the tasks, agents can converge to the desired status by following local decision policies (e.g., the associated rows or columns of the current Markov matrix). One advantage of using top-down approaches is predictability of average emergent behavior with regard to convergence speed and the quality of a stable outcome (i.e., how well the agents converge to the desired fraction distribution). However, such prediction, to the best of our knowledge, can be made mainly numerically. Besides, as top-down generated control policies regulate agents, it may be difficult to accommodate each agent's individual preference. Also, each agent may have to physically move around ac- cording to its local policy during the entire decision-making process, which may cause waste of time and energy costs in the transitioning. Bottom-up approaches focus on designing each agent's individual rules (i.e., microscopic models) that eventually lead to a desired emergent behavior. Possible actions of a single agent can be modeled by a finite state machine [32], and a change of behavior occurs according to a probabilistic thresh- old model [33]. A threshold value in the model determines the decision boundary between two motions. This value is adjustable based on an agent's past experiences such as the time spent for working a task [19], [34], the success/failure rates [32], [35], and direct communication from a central 2 unit [33]. This feature can improve system adaptability, and may have a potential to incorporate each agent's individual interest if required. However, it was shown in [35] -- [41] that, to predict or evaluate an emergent performance of a swarm utilizing bottom-up approaches, a macroscopic model for the swarm is eventually required to be developed by abstracting the microscopic models. B. Hedonic Games Hedonic games [16] -- [18] model a conflict situation where self-interest agents are willing to form coalitions to improve their own interests. Nash stability [18] plays a key role since it yields a social agreement among the agents even without having any negotiation. Many researchers have investigated conditions under which a Nash stable partition is guaranteed to exist and to be determined [18], [42] -- [44]. Among them, the works in [43], [44] mainly addressed an anonymous hedonic game, in which each agent considers the size of a coalition to which it belongs instead of the identities of the members. Recently, Darmann [44] showed that selfish agents who have social inhibition (i.e., preference toward a coalition with a fewer number of members) could converge to a Nash stable partition in an anonymous hedonic game. The author also proposed a centralized recursive algorithm that can find a Nash a · nt) of iterations. Here, na is the stable partition within O(n2 number of agents and nt is that of tasks. C. Main Contributions Inspired by the recent breakthrough of [44], we propose a novel decentralized framework that models the task allocation problem considered as an anonymous hedonic game. The proposed framework is a self-organized approach in which agents make decisions according to its local policies (i.e., indi- vidual preferences). Unlike top-down or bottom-up approaches reviewed in the previous section, which primarily concentrate on designing agents' decision-making policies either macro- scopically or microscopically, our work instead focuses on in- vestigating and exploiting advantages from socially-inhibitive agents, while simply letting them greedily behave according to their individual preferences. Explicitly, the main contributions of this paper are as follows: 1) This paper shows that selfish agents with social inhibi- tion, which we refer to as SPAO preference (Definition 4), can reach a Nash stable partition within less algorith- mic complexity compared with [44]: O(n2 a) of iterations are required1. 2) We provide a decentralized algorithm, which is exe- cutable under a strongly-connected communication net- work of agents and even in asynchronous environments. Depending on the network assumed, the algorithmic complexity may be additionally increased by O(dG), where dG < na is the graph diameter of the network. 3) This paper analyzes the suboptimality of a Nash stable partition in term of the global utility. We firstly present a 1Note that the definition of iteration is described in Definition 5. This comparison assumes the fully-connected communication network because the algorithm in [44] is centralized. mathematical formulation to compute the suboptimality lower bound by using the information of a Nash stable partition and agents' individual utilities. Furthermore, we additionally show that 50% of suboptimality can be at least guaranteed if the social utility for each coalition is defined as a non-decreasing function with respect to the number of members in the coalition. 4) Our framework can accommodate different agents with different interests as long as their individual preferences hold SPAO. 5) Through various numerical experiments, it is confirmed that the proposed framework is scalable, fast adaptable to environmental changes, and robust even in a realistic situation where some agents are temporarily unable to proceed a decision-making procedure and communicate with the other agents during a mission. TABLE I NOMENCLATURE Symbol A ai T ∗ tj tφ T (tj , p) X Pi (cid:31)i ∼i (cid:23)i Π Sj Π(i) dG Ni Description a set of na agents the i-th agent a set of nt tasks the j-th task the void task (i.e., not to work any task) a set of tasks, T = T ∗ ∪ {tφ} a task-coalition pair (i.e. to do task tj with p participants) the set of task-coalition pairs, X = X ∗ ∪ {tφ}, where X ∗ = T ∗ × {1, 2, ..., na} agent ai's preference relation over X the strong preference of agent ai the indifferent preference of agent ai the weak preference of agent ai a partition: a disjoint set that partitions the agent set A, Π = {S1, S2, ..., Snt , Sφ} the (task-specific) coalition for tj the index of the task to which agent ai is assigned given Π the graph diameter of the agent communication network The neighbor agent set of agent ai given a network III. GROUP AGENT PARTITIONING AND PLACING EVENT A. Problem Formulation Let us first introduce the multi-robot task allocation problem considered in this paper and underlying assumptions. Problem 1. Suppose that there exist a set of na agents A = {a1, a2, ..., ana} and a set of tasks T = T ∗ ∪ {tφ}, where T ∗ = {t1, t2, ..., tnt} is a set of nt tasks and tφ is the void task (i.e., not to perform any task). Each agent ai has the individual utility ui : T × A → R, which is a function of the task to which the agent is assigned and the number of its co-working agents (including itself) p ∈ {1, 2, ..., na} (called participants). The individual utility for tφ is zero regardless of the participants. Since every agent is considered to have limited capabilities to finish a task alone, the agent can be assigned to at most one task. The objective of this task allocation problem is to find an assignment that maximizes the global utility, which is the sum of individual utilities of the entire agents. The problem described above is defined as follows: (cid:88) subject to 3 (1) (2) (3) (cid:88) (cid:88) ∀tj∈T max{xij} ∀ai∈A ∀tj∈T ui(tj, p)xij, xij ≤ 1, ∀ai ∈ A, xij ∈ {0, 1}, ∀ai ∈ A,∀tj ∈ T , where xij is a binary decision variable that indicates whether or not task tj is assigned to agent ai. The term social utility is defined as the sum of individual utilities within any agent group. Assumption 1 (Homogeneous agents with limited capabili- ties). This paper considers a large-scale multi-robot system of homogeneous agents since the realisation of a swarm can be in general achieved through mass production [4]. Therefore, each individual utility ui is concerned with the cardinality of the agents working for the task. Note that agents in this paper may have different preferences with respect to the given tasks, e.g., for an agent, a spatially closer task is more preferred, whereas this may not be the case for another agent. Besides, noting that "mass production favors robots with fewer and cheaper components, resulting in lower cost but also reduced capabilities [45]", it is also assumed that each agent can be only assigned to perform at most a single task. According to [10], such a robot is called a single-task (ST) robot. Assumption 2 (Agents' communication). The communication network of the entire agents is at least strongly-connected, i.e., there exists a directed communication path between any two arbitrary agents. Given a network, Ni denotes a set of neighbor agents for agent ai. Assumption 3 (Multi-robot-required tasks). Every task is a multi-robot (MR) task, meaning that the task may require multiple robots [10]. For now, we assume that each task can be performed even by a single agent although it may take a long time. However, in Section V, we will also address a particular case in which some tasks need at least a certain number of agents for completion. Assumption 4 (Agents' pre-known information). Every agent ai only knows its own individual utility ui(tj, p) with regard to every task tj, while not being aware of those of the other agents. Through communication, however, they can notice which agent currently choses which task, i.e., partition (Definition 2). Note that the agents do not necessarily have to know the true partition information at all the time. Each agent owns its locally-known partition information. B. Proposed Game-theoretical Approach: GRAPE Let us transform Problem 1 into an anonymous hedonic game event where every agent selfishly tends to join a coalition according to its preference. a tasks; set of (2) T = T ∗ ∪ {tφ}, Definition 1 (GRAPE). An instance of GRoup Agent Par- titioning and placing Event (GRAPE) is a tuple (A,T ,P) (1) A = {a1, a2, ..., ana}, a set of na that consists of agents; and (3) P = (P1,P2, ...,Pna ), an na-tuple of preference relations of the agents. For agent ai, Pi describes its preference relation over the set of task-coalition pairs X = X ∗ ∪ {tφ}, where X ∗ = T ∗ × {1, 2, ..., na}; a task-coalition pair (tj, p) is in- terpreted as "to do task tj with p participants". For any task-coalition pairs x1, x2 ∈ X , x1 (cid:31)i x2 implies that agent ai strongly prefers x1 to x2, and x1 ∼i x2 means that the preference regarding x1 and x2 is indifferent. Likewise, (cid:23)i indicates the weak preference of agent ai. a partition is defined as Note that agent ai's preference relation can be derived from its individual utility ui(tj, p) in Problem 1. For instance, given that ui(t1, p1) > ui(t2, p2), it can be said that (t1, p1) (cid:31)i (t2, p2). Definition 2 (Partition). Given an instance (A,T ,P) of GRAPE, set Π = {S1, S2, ..., Snt, Sφ} that disjointly partitions the agent set A. Here, Sj ⊆ A is the (task-specific) coalition for executing task tj such that ∪nt j=0Sj = A and Sj ∩ Sk = ∅ for j (cid:54)= k. Sφ is the set of agents who choose the void task tφ. Note that this paper interchangeably uses S0 to indicate Sφ. Given a partition Π, Π(i) indicates the index of the task to which agent ai is assigned. For example, SΠ(i) is the coalition that the agent belongs to, i.e., SΠ(i) = {Sj ∈ Π ai ∈ Sj}. a The objective of GRAPE is to determine a stable partition that all the agents agree with. In this paper, we seek for a Nash stable partition, which is defined as follows: Definition 3 (Nash stable). A partition Π is be Nash stable if, for every agent ai ∈ A, (tΠ(i),SΠ(i)) (cid:23)i (tj,Sj ∪ {ai}), ∀Sj ∈ Π. said to it holds that In other words, in a Nash stable partition, every agent prefers its current coalition to joining any of the other coali- tions. Thus, every agent does not have any conflict within this partition, and no agent will not unilaterally deviate from its current decision. Remark 1 (An advantage of Nash stability: low communica- tion burden on agents). The rationale behind the use of Nash stability among various stable solution concepts in hedonic games [16], [46] -- [48] is that it can reduce communication burden between agents required to reach a social agreement. In the process of converging to a Nash stable partition, an agent does not need to get any permission from the other agents when it is willing to deviate. This property may not be the case for the other solution concepts. Therefore, each agent is only required to notify its altered decision without any negotiation. This fact can reduce inter-agent communication in the proposed approach. C. SPAO Preference: Social Inhibition This section introduces the key condition, called SPAO, that enables our proposed approach to provide all the desirable 4 is said that properties described in Section I, and then explains its impli- cations. Definition 4 (SPAO). Given an instance (A,T ,P) of GRAPE, it the preference relation of agent ai with respect it holds that, for every (tj, p) ∈ X ∗, (tj, p1) (cid:23)i (tj, p2) for any p1, p2 ∈ {1, ..., na} such that p1 < p2. Besides, we say that an instance (A,T ,P) of GRAPE is SPAO if the preference relation of every agent in A with respect to every task in T ∗ is SPAO. is SPAO (Single-Peaked-At-One) to task tj if For an example, suppose that Pi is such that (t1, 1) (cid:31)i (t1, 2) (cid:23)i (t1, 3) (cid:31)i (t2, 1) ∼i (t1, 4) (cid:31)i (t2, 2). This preference relation indicates that agent ai has (t1, 1) (cid:31)i (t1, 2) (cid:23)i (t1, 3) (cid:31)i (t1, 4) for task t1, and (t2, 1) (cid:31)i (t2, 2) for task t2. According to Definition 4, the preference relation for each of the tasks holds SPAO. For another example, given that (t1, 1) (cid:31)i (t1, 2) (cid:23)i (t1, 3) (cid:31)i (t2, 2) ∼i (t1, 4) (cid:31)i (t2, 1), the preference relation regarding task t1 holds SPAO, whereas this is not the case for task t2 because of (t2, 2) (cid:31)i (t2, 1). This paper only considers the case in which every agent has SPAO preference relations regarding all the given tasks. Such agents prefer to execute a task with smaller number of collaborators, namely, they have social inhibition. Remark 2 (Implications of SPAO). SPAO implies that an agent's individual utility should be a monotonically decreasing function with respect to the size of a coalition. In practice, SPAO can often emerge. For instance, experimental and simu- lation results in [49, Figures 3 and 4] show that the total work capacity resulted from cooperation of multiple robots does not proportionally increase due to interferences of the robots. In such a non-superadditive environment [50], assuming that an agent's individual work efficiency is considered as its indi- vidual utility, the individual utility monotonically drops as the number of collaborators enlarges even though the social utility is increased. For another example, SPAO also arises when individual utilities are related with shared-resources. As more agents use the same resource simultaneously, their individual productivities become diminished (e.g., traffic affects travel times [51] [52, Example 3]). As the authors in [50] pointed out, a non-superadditive case is more realistic than a superadditive case: agents in a superadditive environment always attempt to form the grand coalition whereas those in a non-superadditive case are willing to reduce unnecessary costs. Note that social utility functions are not restricted so that they can be either monotonic or non-monotonic. Remark 3 (Cooperation of selfish agents with different in- terests). The proposed framework can accommodate selfish agents who greedily follow their individual preferences as long as the preferences hold SPAO. This implies that the framework may be utilized for a combination of swarm systems from different organisations under the condition that the multiple systems satisfy SPAO. D. Existence of and Convergence to a Nash Stable Partition Let us prove that if an instance of GRAPE holds SPAO, there always exists a Nash stable partition and it can be found within polynomial time. Definition 5 (Iteration). This paper uses the term iteration to represent an iterative stage in which an arbitrary agent compares the set of selectable task-coalition pairs given an existing partition, and then determines whether or not to join another coalition including the void task one. Assumption 5 (Mutual exclusion algorithm). We assume that, at each iteration, a single agent exclusively makes a decision and updates the current partition if necessary. This paper refers to this agent as the deciding agent at the iteration. Based on the resultant partition, another deciding agent also performs the same process at the next iteration, and this process continues until every agent does not deviate from a specific partition, which is, in fact, a Nash stable partition. To implement this the agents need a mutual algorithmic process in practice, exclusion (or called mutex) algorithm to choose the deciding agent at each iteration. In this section, for simplicity of description, we assume that all the agents are fully-connected, by which they somehow select and know the deciding agent. However, in Section III-E, we will present a distributed mutex algorithm that enables the proposed approach to be executed under a strongly-connected communication network even in an asynchronous manner. Lemma 1. Given an instance (A,T ,P) of GRAPE that is SPAO, suppose that a new agent ar /∈ A holding a SPAO pref- erence relation with regard to every task in T joins (A,T ,P) in which a Nash stable partition is already established. Then, the new instance ( A,T ,P), where A = A ∪ {ar}, also (1) satisfies SPAO; (2) contains a Nash stable partition; and (3) the maximum number of iterations required to re-converge to a Nash stable partition is A. Proof. Given a partition Π, for agent ai, the number of additional co-workers tolerable in its coalition is defined as: (cid:8)∆ (tΠ(i),SΠ(i) + ∆) (cid:23)i (tj,Sj ∪ {ai})(cid:9). ∆Π(i) := min Sj∈Π\{SΠ(i)} max ∆∈Z (4) Due to the SPAO preference relation, this value satisfies the following characteristics: (a) if Π is Nash stable, for every agent ai, it holds that ∆Π(i) ≥ 0; (b) if ∆Π(i) < 0, then agent ai is willing to deviate to another coalition at a next iteration; and (c) for the agent ai who deviated at the last iteration and updated the partition as Π(cid:48), it holds that ∆Π(cid:48)(i) ≥ 0. From Definition 4, it is clear that the new instance ( A,T ,P) still holds SPAO. Let Π0 denote a Nash stable partition in the original instance (A,T ,P). When a new agent ar /∈ A decides to execute one of the tasks in T and creates a new partition Π1, it holds that ∆Π1(r) ≥ 0, as shown in (c). If there is no existing agent aq ∈ A whose ∆Π1(q) < 0, then the new partition Π1 is Nash stable. Suppose that there exists at least an agent aq whose ∆Π1(q) < 0. Then, the agent must be one of the existing members in the coalition that agent ar selected in the last iteration. As agent aq moves to another coalition and creates 5 a new partition Π2, the previously-deviated agent ar holds ∆Π2(r) ≥ 1. In other words, an agent who deviates to a coalition and expels one of the existing agents in that coalition will not deviate again even if another agent joins the coalition in a next iteration. This implies that at most A of iterations are required to hold ∆ Π(i) ≥ 0 for every agent ai ∈ A, where the partition Π is Nash stable. Lemma 1 is essential not only for the existence of and convergence to a Nash stable partition but also for fast adaptability to dynamic environments. Theorem 1 (Existence). If (A,T ,P) is an instance of GRAPE holding SPAO, then a Nash stable partition always exists. Proof. This theorem will be proved by induction. Let M (n) be the following mathematical statement: for A = n, if an instance (A,T ,P) of GRAPE is SPAO, then there exists a Nash stable partition. Base case: When n = 1, there is only one agent in an instance. This agent is allowed to participate in its most preferred coalition, and the resultant partition is Nash stable. Therefore, M (1) is true. Induction hypothesis: Assume that M (k) is true for a positive integer k such that A = k. Induction step: Suppose that a new agent ai /∈ A whose preference relation regarding every task in T is SPAO joins the instance (A,T ,P). This induces a new instance ( A,T ,P) where A = A ∪ {ai} and A = k + 1. From Lemma 1, it is clear that the new instance also satisfies SPAO and has a Nash stable partition Π. Consequently, M (k + 1) is true. By mathematical induction, M (n) is true for all positive integers n ≥ 1. Theorem 2 (Convergence). If (A,T ,P) is an instance of GRAPE holding SPAO, iterations re- quired to determine a Nash stable partition is at most A · (A + 1)/2. Proof. Suppose that, given a Nash stable partition in an instance where there exists only one agent, we add another arbitrary agent and find a Nash stable partition for this new instance, and repeat the agents in A are included. From Lemma 1, if a new agent joins an instance in which the current partition is Nash stable, then the maximum number of iterations required to find a new Nash stable partition is the number of the existing agents plus one. Therefore, it is trivial that the maximum number of iterations to find a Nash stable partition of an instance (A,T ,P) is given as the procedure until all then the number of k = A · (A + 1)/2. (5) A(cid:88) k=1 Note that this polynomial-time convergence still holds even if the agents are initialized to a random partition. Suppose that we have the following setting: the entire agents A are firstly not movable from the existing partition, except a set of free agents A(cid:48) ⊆ A; whenever the agents A(cid:48) find a Nash stable partition Π(cid:48), one arbitrary agent in ar ∈ A \ A(cid:48) additionally becomes liberated and deviates from the current coalition SΠ(cid:48)(r) to another coalition in Π(cid:48). In this setting, from the viewpoint of the agents in A(cid:48) \ SΠ(cid:48)(r), the newly liberated agent is considered as a new agent as that in Lemma 1. Accordingly, we can still utilize the lemma for the agents in A(cid:48) \ SΠ(cid:48)(r) ∪ {ar}. The agents also can find a Nash stable partition if one of them moves to SΠ(cid:48)(r) during the process, because, due to ar, it became ∆Π(cid:48)(i) ≥ 1 for every agent ai ∈ SΠ(cid:48)(r) \ {ar}. In a nutshell, the agents A(cid:48) ∪ {ar} can converge to a Nash stable partition within A(cid:48)∪{ar}, which is equivalent to Lemma 1. Hence, Theorem 1 and this theorem are also valid for the case when the initial partition of the agents are randomly given. E. Decentralized Algorithm In the previous section, it was assumed that only one agent is somehow chosen to make a decision at each iteration under the fully-connected network. On the contrary, in this section, we propose a decentralized algorithm, as shown in Algorithm 1, in which every agent does decision making based its local information and affects its neighbors simultaneously under a strongly-connected network. Despite that, we show that Theorems 1 and 2 still hold thanks to our proposed distributed mutex subroutine shown in Algorithm 2. The details of the decentralized main algorithm are as follows. // Initialisation Algorithm 1 Decision-making algorithm for each agent ai 1: satisfied ← f alse; ri ← 0; si ← 0 2: Πi ← {Sφ = A, Sj = φ ∀tj ∈ T } // Decision-making process begins 3: while true do // Make a new decision if necessary if satisfied = f alse then (tj∗,Sj∗) ← arg max∀Sj∈Πi(tj,Sj ∪ {ai}) if (tj∗,Sj∗) (cid:31)i (tΠi(i),SΠi(i)) then 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: Join Sj∗ and update Πi ri ← ri + 1 si ∈ unif[0, 1] end if satisfied = true // Broadcast the local information to neighbor agents end if Broadcast M i = {ri, si, Πi} and receive M k from its neighbors ∀ak ∈ Ni Construct Mi {ri, si, Πi}, satisfied ← D-MUTEX(Mi // Select the valid partition from all the received messages rcv = {M i,∀M k} rcv) 14: 15: 16: end while Each agent ai has local variables such as Πi, satisfied, ri, and si (Line 1 -- 2). Here, Πi is the agent's locally-known partition; satisfied is a binary variable that indicates whether or not the agent is satisfied with Πi such that it does not want to deviate from its current coalition; ri ∈ Z+ is an integer variable to represent how many times Πi has evolved (i.e., the number of iterations happened for updating Πi until that moment); and si ∈ [0, 1] is a uniform-random variable that 6 is generated whenever Πi is newly updated (i.e., a random time stamp). Given Πi, agent ai examines which coalition is the most preferred among others, assuming that the other agents remain at the existing coalitions (Line 5). Then, the agent joins the newly found coalition if it is strongly preferred than its current coalition. In this case, the agent updates Πi to reflect its new decision, increases ri, and generates a new random time stamp si (Line 6 -- 10). In any case, since the agent ascertained that the currently-selected coalition is the most preferred, the agent becomes satisfied with Πi (Line 11). Then, agent ai generates and sends a message M i := {ri, si, Πi} to its neighbor agents, and vice versa (Line 13). Since every agent locally updates its locally-known partition simultaneously, one of the partitions should be regarded as if it were the partition updated by a deciding agent at the previous iteration. We refer to this partition as the valid partition at the iteration. The distributed mutex subroutine in Algorithm 2 enables the agents to recognize the valid partition among all the locally-known current partitions even under a strongly- connected network and in asynchronous environments. Before executing this subroutine, each agent ai collects all the mes- sages received from its neighbor agents ∀ak ∈ Ni (including rcv = {M i,∀M k}. Using this message set, the M i) as Mi agent examines whether or not its own partition Πi is valid. If there exists any other partition Πk such that rk > ri, then the agent considers Πk more valid than Πi. This also happens if rk = ri and sk > si, which indicates the case where Πk and Πi have evolved over the same amount of times, but the former has a higher time stamp. Since Πk is considered as more valid, agent ai will need to re-examine if there is a more preferred coalition given Πk in the next iteration. Thus, the agent sets satisfied as f alse (Line 3 -- 10 in Algorithm 2). After completing this subroutine, depending on satisfied, each agent proceeds the decision-making process again (i.e., Line 4 -- 12 in Algorithm 1) and/or just broadcasts the existing locally-known partition to its neighbor agents (Line 13 in Algorithm 1). if (rk > ri) or (rk = ri & sk > si) then rcv) rcv do satisfied ← true for each message Mk ∈ Mi Algorithm 2 Distributed Mutex Subroutine 1: function D-MUTEX(Mi 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: end function ri ← rk si ← sk Πi ← Πk satisfied ← f alse end for return {ri, si, Πi}, satisfied end if In a nutshell, the distributed mutex algorithm makes sure that there is only one valid partition that dominates (or will finally dominate depending on the communication network) any other partitions. In other words, multiple partitions locally evolve, but one of them only eventually survive as long as a strongly-connected network is given. From each partition's viewpoint, it can be regarded as being evolved by a random sequence of the agents under the fully-connected network. Thus, the partition becomes Nash stable within the polynomial time as shown in Theorem 2. In an extreme case, we may encounter multiple Nash stable partitions at the very last. Nevertheless, thanks to the mutex algorithm, one of them can be distributedly selected by the agents. All the features imply that agents using Algorithm 1 can find a Nash stable partition in a decentralized manner and Theorems 1 and 2 still hold. IV. ANALYSIS A. Algorithmic Complexity (Scalability) Firstly, let us discuss about the running time for the pro- posed framework to find a Nash stable partition. This paper refers to a unit time required for each agent to proceed the main loop of Algorithm 1 (Line 4-15) as a time step. De- pending on the communication network considered, especially if it is not fully-connected, it may be possible that some of the given agents have to execute this loop to just propagate their locally-known partition information without affecting ri as Line 8. Because this process also spends a unit time step, we call it as dummy iteration to distinguish from a (normal) iteration, which increases ri. Notice that such dummy iterations happen sequentially at most dG times before a normal iteration occurs, where dG is the graph diameter of the communication network. Hence, thanks to Theorem 2, the total required time steps until finding a). For the fully-connected a Nash stable partition is O(dGn2 network case, it becomes O(n2 a) because of dG = 1. Note that this algorithmic complexity is less than that of the centralized algorithm, i.e., O(n2 a · nt), in [44]. Every agent at each iteration investigates nt + 1 of se- lectable task-coalition pairs including tφ given a locally-known valid partition (as shown in Line 5 in Algorithm 1). Therefore, the computational overhead for an agent is O(nt) per any iteration. With consideration of the total required time steps, the running time of the proposed approach for an agent can a). Note that the running time in be bounded by O(dGntn2 practice can be much less than the bound since Theorem 2 was conservatively analyzed, as described in the following remark. Remark 4 (The number of required iterations in practice). Algorithm 1 allows the entire agents in A to be involved in the decision-making process, whereas, in the proof for Theorem 2, a new agent can be involved after a Nash stable partition of existing agents is found. Since agents using Algorithm 1 do not need to find every Nash stable partition for each subset of the agents, unnecessary iterations can be reduced. Hence, the number of required iterations in practice may become less than that shown in Theorem 2, which is also supported by the experimental results in Section VI-B. Let us now discuss about the communication overhead for each agent per iteration. Given a network, agent ai should communicate with Ni of its neighbors, and the size of each message grows with regard to na. Hence, the communication overhead of the agent is O(Ni · na). It could be quadratic if Ni increases in proportional to na. However, this would rarely happen in practice due to spatial distribution of agents 7 and physical limits on communication such as range limitation. Instead, Ni would be most likely saturated in practice. Remark 5 (Communication overhead vs. Running time). To reduce the communication overhead, we may impose the maximum number of transactions per iteration, denoted by nc, on each agent. Even so, Theorems 1 and 2 are still valid as long as the union of underlying graphs of the communication networks over time intervals becomes connected. However, in return, the number of dummy iterations may increase, so does the framework's running time. In an extreme case where nc = 1 (i.e., unicast mode), dummy iterations may happen in a row at most na times. Thus, the total required time steps until finding a Nash stable partition could be O(n3 a), whereas the communication overhead is O(na). In short, the running time of the framework can be traded off against the communication overhead for each agent per iteration. B. Suboptimality This section investigates the suboptimality lower bound (or can be called approximation ratio) of the proposed framework in terms of the global utility, i.e., the objective function in Equation (1). Given a partition Π, the global utility value can be equivalently rewritten as J = ui(tΠ(i),SΠ(i)). (6) (cid:88) ∀ai∈A Note that we can simply derive {xij} for Equation (1) from Π for Equation (6), and vice versa. Let JGRAP E and JOP T represent the global utility of a Nash stable partition obtained by the proposed framework and the optimal value, respectively. This paper refers to the fraction of JGRAP E with respect to JOP T as the suboptimality of GRAPE, denoted by α, i.e., α := JGRAP E/JOP T . (7) The lower bound of the suboptimality can be determined by the following theorem. Theorem 3 (Suboptimality lower bound: general case). Given a Nash stable partition Π obtained by GRAPE, its suboptimal- ity in terms of the global utility is lower bounded as follows: α ≥ JGRAP E/(JGRAP E + λ), (8) (cid:8)p·(cid:2)ui(tj, p)−ui(tj,Sj∪{ai})(cid:3)(cid:9) (9) max ai∈A,p≤A where λ ≡ (cid:88) ∀Sj∈Π Proof. Let Π∗ denote the optimal partition for the objective function in Equation (6). Given a Nash stable partition Π, from Definition 3, it holds that, ∀ai ∈ A, ui(tΠ(i),SΠ(i)) ≥ ui(t∗ (10) j←i indicates task tj ∈ T to which agent ai should where t∗ have joined according to the optimal partition Π∗; and Sj ∈ Π is the coalition for task tj whose participants follow the Nash stable partition Π. j←i,Sj ∪ {ai}), 8 and (ii) all the individual utilities can derive SPAO preference relations, then a Nash stable partition Π obtained by GRAPE provides at least 50% of suboptimality in terms of the global utility. (cid:88) Proof. Firstly, we introduce some definitions and notations that facilitate to describe this proof. Given a partition Π of an instance (A,T ,P), the global utility is denoted by V (Π) := (15) We use operator ⊕ as follows. Given any two partitions ∀ai∈A ui(tΠ(i),SΠ(i)). ΠA = {SA 0 , ..., SA nt } and ΠB = {SB 0 ∪ SB j=0SB 1 ∪ SB 0 , SA j = A, 0 , ..., SB nt }, 1 , ..., SA nt ∪ SB }. there may exist nt ΠA ⊕ ΠB := {SA j = ∪nt j=0SA Since ∪nt the same agent ai even in two different coalitions in ΠA ⊕ ΠB. For instance, suppose that ΠA = {{a1},{a2},{a3}} and ΠB = {∅,{a1, a3},{a2}}. Then, ΠA ⊕ ΠB = {{a1},{a1, a2, a3},{a2, a3}}. We regard such an agent as two different agents in ΠA ⊕ ΠB. Accordingly, the operation may increase the number of total agents in the resultant partition. Using the definitions described above, condition (i) implies that V (ΠA) ≤ V (ΠA ⊕ ΠB). From now on, we will show that 1 (16) 2 V (Π∗) ≤ V ( Π), } is an optimal partition and where Π∗ = {S∗ Π = { S0, S1, ..., Snt} is a Nash stable partition. By doing so, this theorem can be proved. From the definition in Equation (15), it can be said that 1 , ..., S∗ 0 , S∗ nt (cid:88) ui(t Π(i), S Π(i) ∪ S∗ (cid:88) ∀ai∈A + ) Π(i) ui(tΠ∗(i), SΠ∗(i) ∪ S∗ Π∗(i)), (17) ∀ai∈A− where A− is the set of agents whose decisions follow not the Nash stable partition Π but only the optimal partition Π∗. Due to condition (ii), the first term of the right-hand side in Equation (17) is no more than ui(t Π(i), S Π(i)) ≡ V ( Π). Likewise, the second term is also at most ui(tΠ∗(i), SΠ∗(i) ∪ {ai}). By the definition of Nash stability (i.e., for every agent ai ∈ A, ui(t Π(i), S Π(i)) ≥ ui(tj, Sj ∪ {ai}), ∀ Sj ∈ Π), the above equation is at most (cid:88) (cid:88) which is also no more than, because of A− ⊆ A, ui(t Π(i), S Π(i)), ∀ai∈A− (20) ui(t Π(i), S Π(i)) ≡ V ( Π). (21) (cid:88) (cid:88) ∀ai∈A ∀ai∈A− (18) (19) ai∈A,p≤A Lij[p] ≡ λ, max (14) V ( Π ⊕ Π∗) = ui(tj,Sj ∪ {al}), ∀ai∈A The right-hand side of the inequality in Equation (10) can be rewritten as ui(t∗ (cid:8)ui(t∗ j←i,Sj ∪ {ai}) = ui(t∗ j←i,S∗ j ) − ui(t∗ j←i,S∗ j←i,Sj ∪ {ai})(cid:9), j )− where S∗ mizes the objective function. j ∈ Π∗ is the ideal coalition of task t∗ (11) j←i that maxi- By summing over all the agents, the inequality in Equation (10) can be said that (cid:88) ∀ai∈A ui(tΠ(i),SΠ(i)) ≥ (cid:88) − (cid:88) ∀ai∈A ∀ai∈A ui(t∗ j←i,S∗ j ) j←i,S∗ {ui(t∗ j ) − ui(t∗ j←i,Sj ∪ {ai})}. (12) The left-hand side of the inequality in Equation (12) represents the objective function value of the Nash stable partition Π, i.e., JGRAP E, and the first term of the right-hand side is the optimal value, i.e., JOP T . The second term in the right-hand side can be interpreted as the summation of the utility lost of each agent caused by the belated decision to its optimal task, provided that the other agents still follow the Nash stable partition. The upper bound of the second term is given by {ui(t∗ j←i,S∗ j )− ui(t∗ j←i,Sj ∪{ai})}. (13) S∗ j · max ai∈S∗ nt(cid:88) This is at most (cid:88) j=1 j ∀Sj∈Π where Lij[p] = p · (ui(tj, p) − ui(tj,Sj ∪ {ai})). Hence, the inequality in Eqn (12) can be rewritten as JGRAP E ≥ JOP T − λ. Dividing both sides by JGRAP E and rearranging them yield the suboptimality lower bound of the Nash stable partition, as given by Equation (8). Although Theorem 3 does not provide a fixed-value lower bound, it can be determined as long as a Nash stable partition and agents' individual utility functions are given. Nevertheless, as a special case, if the social utility for any coalition is non-decreasing (or monotonically increasing) in terms of the number of co-working agents, then we can obtain a fixed-value lower bound for the suboptimality of a Nash stable partition. Theorem 4 (Suboptimality lower bound: a special case). Given an instance (A,T ,P) of GRAPE, if (i) the social utility for any coalition is non-decreasing with regard to the number of participants, i.e., for any Sj ⊆ A and al ∈ A\ Sj, it holds that (cid:88) ui(tj,Sj) ≤ (cid:88) ∀ai∈Sj ∀ai∈Sj∪{al} Accordingly, the left-hand side of Equation (17) holds the following inequality: defined as those in Problem 1. Here, it is considered that, for ∀ai ∈ A and ∀tj ∈ T , 9 V ( Π ⊕ Π∗) ≤ 2V ( Π). Thanks to Equation (16), it follows that V (Π∗) ≤ V ( Π ⊕ Π∗). Therefore, V (Π∗) ≤ 2V ( Π), which completes this proof. (22) ui(tj, p) = 0 if p < Rj (27) because task tj cannot be completed in this case. Note that any task tj without such a requirement is regarded to have Rj = 0. For each task tj having Rj > 0, even if ui(tj, p) is monotonically decreasing at p ≥ Rj, the individual utility can not be simply transformed to a preference relation holding SPAO because of Equation (27). Thus, we need to modify the utility function to yield alternative values for the case when p < Rj. We refer to the modified utility as auxiliary individual utility ui, which is defined as ui(tj, p) = u0 i (tj, p) ui(tj, p) if p ≤ Rj otherwise, (28) (cid:40) where u0 to task tj when p ≤ Rj. i (tj, p) is the dummy utility of agent ai with regard The dummy utility is intentionally used also for the case when p = Rj in order to find an assignment that holds Equation (24). For this, the auxiliary individual utility should satisfy the following condition. Condition 1. For every agent ai ∈ A, its preference relation Pi holds that, for any two tasks tj, tk ∈ T , (tj, Rj) (cid:31)i (tk, Rk + 1). This condition enables every agent to prefer a task for which the number of co-working agents is less than its minimum requirement, over any other tasks whose requirements are already fulfilled. Under this condition, as long as the agent set A is such that A ≥(cid:80)∀tj∈T Rj and a Nash stable partition is found, the resultant assignment satisfies Equation (24). Proposition 1. Given an instance of Problem 2 where ui(tj, p) ∀i ∀j is a monotonically decreasing function with regard to ∀p ≥ Rj, if the dummy utilities u0 i (tj, p) ∀i ∀j in (28) are set to satisfy Condition 1 and SPAO for ∀p ≤ Rj, then all the resultant auxiliary individual utilities ui(tj, p) ∀i ∀j ∀p can be transformed to a na-tuple of preference relations P that hold Condition 1 as well as SPAO for ∀p ∈ {1, ..., na}. In the corresponding instance of GRAPE (A,T ,P), a Nash stable partition can be determined within polynomial times as shown in Theorems 1 and 2 because of SPAO, and the resultant partition can satisfy Equation (24) due to Condition 1. Let us give an example. Suppose that there exist 100 agents A, and 3 tasks T = {t1, t2, t3} where only t3 has its minimum requirement R3 = 5; for every agent ai ∈ A, individual utili- ties for t1 and t2, i.e., ui(t1, p) and ui(t2, p), are much higher than that for t3 in ∀p ∈ {1, ..., 100}. We can find a Nash stable partition for this example, as described in Proposition 1, by i (tj, p)= max∀tj{ui(tj, Rj + 1)} + β for ∀p ≤ Rj, setting u0 ∀ai ∈ A, where β > 0 is an arbitrary positive constant. After a Nash stable partition is found, in order to compute the objective function value in (23), the original individual C. Adaptability Our proposed framework is also adaptable to dynamic environments such as unexpected addition or loss of agents or tasks, owing to its fast convergence to a Nash stable partition. Thanks to Lemma 1, if a new agent additionally joins an ongoing mission in which an assignment was already determined, the number of iterations required for converging to a new Nash stable partition is at most the number of the total agents. Responding to any environmental change, the framework is able to establish a new agreed task assignment within polynomial time. D. Robustness in Asynchronous Environments it In the proposed framework, for every iteration, each agent does not need to wait until nor ensure that its locally-known information has been propagated to a certain neighbor group. Instead, as described in Remark 5, is enough for the agent to receive the local information from one of its neigh- bors, to make a decision, and to send the updated partition back to some of its neighbors. Temporary disconnection or non-operation of some agents may cause dummy iterations additionally. However, the existence of, the convergence toward, and the suboptimality of a Nash stable partition under the proposed framework, which is also supported by Section VI-E. it does not affect V. GRAPE WITH MINIMUM REQUIREMENTS This section addresses another task allocation problem where each task may require at least a certain number of agents for its completion. This problem can be defined as follows. Problem 2. Given a set of agents A and a set of tasks T , the objective is to find an assignment such that subject to (cid:88) (cid:88) ∀ai∈A ∀tj∈T ui(tj, p)xij, xij ≥ Rj, ∀tj ∈ T , xij ≤ 1, ∀ai ∈ A, max{xij} (cid:88) (cid:88) ∀ai∈A ∀tj∈T (23) (24) (25) xij ∈ {0, 1}, ∀(ai, tj) ∈ A × T , (26) where Rj ∈ N ∪ {0} is the number of minimum required agents for task tj, and all the other variables are identically utility function ui should be used instead of the auxiliary one ui. Proposition 2. Given a Nash stable partition Π obtained by implementing Proposition 1, its suboptimality bound α is such that α ≥ JGRAP E JGRAP E + λ · JGRAP E JGRAP E + δ . (29) Here, δ ≡ JGRAP E−JGRAP E, where JGRAP E (or JGRAP E) is the objective function value in (23) using ui (or using ui) given the Nash stable partition. Likewise, λ is the value in (9) using ui. In addition to this, if every ui satisfies the conditions for Theorem 4, then α ≥ 1 2 · JGRAP E JGRAP E + δ . (30) Proof. Since the Nash stable partition Π is obtained by using ui, it can be said from Equations (7) and (8) that JGRAP E JOP T ≥ JGRAP E . JGRAP E + λ (31) Due to the fact that ui(tj, p) ≥ ui(tj, p) for ∀i, j, p, it is clear that JGRAP E ≥ JGRAP E and JOP T ≥ JOP T . By letting that δ := JGRAP E − JGRAP E, the left term in (31) is at most (JGRAP E + δ)/JOP T . Besides, the right term in (31) is a monotonically-increasing function with regard to JGRAP E, and thus, it is lower bounded by JGRAP E/(JGRAP E + λ). From this, Equation (31) can be rewritten as Equation (29) by multiplying JGRAP E/(JGRAP E + δ). Likewise, for the case when every ui satisfies the conditions for Theorem 4, it can be said that JGRAP E ≥ 1/2 · JOP T , which can be transformed into Equation (30) as shown above. 10 (a) Social utility (b) Individual utility Fig. 1. Examples of utility functions used in the numerical experiment are shown, depending on the two different task types (i.e., peaked-reward and submodular-reward): (a) the social utility of a coalition; (b) an agent's individual utility. among the members. Therefore, the individual utility of agent ai executing task tj with coalition Sj is defined as ui(tj,Sj) = r(tj,Sj)/Sj − ci(tj), (32) where r(tj,Sj) is the reward from task tj when it is executed by Sj together, and ci(tj) is the cost that agent ai needs to pay for the task. Here, we simply set the cost as a function of the distance from agent ai to task tj. We set that if ui(tj,Sj) is not positive, agent ai prefers to join Sφ over Sj. This experiment considers two types of tasks. For the first type, a task's reward becomes higher as the number of participants gets close to a specific desired number. We refer to such a task as a peaked-reward task, and its reward can be defined as r(tj,Sj) = rmax j · e−Sj/nd j +1, (33) · Sj nd j Notice that if δ = 0 for the Nash stable partition in Propo- sition 2, then the suboptimality bounds become equivalent to those in Theorems 3 and 4. VI. SIMULATION AND RESULTS This section validates the performances of the proposed framework with respect to its scalability, suboptimality, adapt- ability against dynamic environments, and robustness in asyn- chronous environments. A. Mission Scenario and Settings 1) Utility functions: Firstly, we introduce the social and individual utilities used in this numerical experiment. We consider that if multiple robots execute a task together as a coalition, then they are given a certain level of reward for the task. The amount of the reward varies depending on the number of the co-working agents. The reward is shared with the agents, and each agent's individual utility is considered as the shared reward minus the cost required to personally spend on the task (e.g., fuel consumption for movement). In this experiment, the equal fair allocation rule [53], [54] is adopted. Under the rule, a task's reward is equally shared represents the desired number, and rmax is the where nd j peaked reward in case that nd j of agents are involved in. Consequently, the individual utility of agent ai with regard to task tj becomes the following equation: j ui(tj,Sj) = rmax j nd j · e−Sj/nd j +1 − ci(tj). (34) For the second type, a task's reward becomes higher as more agents are involved, but the corresponding marginal gain decreases. This type of tasks is said to be submodular-reward, and the reward can be defined as j j (35) · logj (Sj + j − 1), r(tj,Sj) = rmin where rmin indicates the reward obtained if there is only one agent is involved, and j > 1 is the design parameter regarding the diminishing marginal gain. The resultant individual utility becomes as follows: · logj (Sj + j − 1)/Sj − ci(tj). (36) ui(tj,Sj) = rmin Figure 1 illustrates examples of the social utilities and individual utilities for the task types introduced above. For simplification, agents' costs are ignored in the figure. We set and j to be 60, 15, 10, and 2, respectively. rmax j Notice that the individual utilities are monotonically decreas- ing in both cases, as depicted in Figure 1(b). Therefore, given j , rmin , nd j j 020406080100# Co-working Agents020406080Social UtilityPeaked-rewardSubmodular-reward020406080100# Co-working Agents0246810Individual UtilityPeaked-rewardSubmodular-rewarda mission that entails these task types, we can generate an instance (A,T ,P) of GRAPE that holds SPAO. 11 j j k /(cid:80)∀tk∈T ∗ rmax 2) Parameters generation: In the following sections, we will mainly utilize Monte Carlo simulations. At each run, nt tasks and na agents are uniform-randomly located in a 1000 m × 1000 m arena and a 250 m × 250 m arena within there, respectively. For a scenario including peaked-reward is randomly generated from a uniform distribution tasks, rmax over [1000, 2000] × na/nt, and nd j is set to be the rounded )×na. For a scenario including value of (rmax is uniform- submodular-reward tasks, j is set as 2, and rmin randomly generated over [1000, 2000]× 1/ logj (na/nt + 1). 3) Communication network: Given a set of agents, their communication network is strongly-connected in a way that only contains a bidirectional minimum spanning tree with consideration of the agents' positions. Furthermore, we also consider the fully-connected network in some experiments in order to examine the influence of the network. The communi- cation network is randomly generated at each instance, and is assumed to be sustained during a mission except the robustness test simulations in Section VI-E. j B. Scalability To investigate the effectiveness of nt and na upon the scalability of the proposed approach, we conduct Monte Carlo simulations with 100 runs for the scenarios introduced in Section VI-A with a fixed nt = 20 and various na ∈ {80, 160, 240, 320} and for those with na = 160 and nt ∈ {5, 10, 15, 20}. Figure 2 shows the statistical results using box- and-whisker plots, where the green boxes indicate the results from the scenarios with the peaked-reward tasks and the ma- genta boxes are those with the submodular-reward tasks. The blue and red lines connecting the boxes represent the average value for each test case (na, nt) under a strongly-connected network and the fully-connected network, respectively. The left subfigure in Figure 2(a) shows that the ratio of the number of required (normal) iterations to that of agents linearly increases as more agents are involved. This implies that the proposed framework has quadratic complexity with a), as stated in regard to the number of agents (i.e., C1n2 2, which is Theorem 2, but with C1 being much less than 1 the value from the theorem. C1 can become even lower (e.g., C1 = 5 × 10−4 in the experiments) under the fully-connected network. Such C1 being smaller than 1 2 may be explained by Remark 4: the algorithmic efficiency of Algorithm 1 can reduce unnecessary iterations that may be induced in the procedure of the proof for Theorem 2. On the other hand, the left subfigure in Figure 2(b) shows that the number of required iterations decreases with regard to the number of tasks. This trend may be caused by the fact that more selectable options provided to the fixed number of agents can reduce possible conflicts between the agents. Furthermore, in the two results, the trends regarding either na or nt have higher slopes under a strongly-connected network than those under the fully-connected network. This is because the former condition is more sensitive to conflicts iterations. For between agents, and thus causes additional (a) na ∈ {80, 160, 240, 320} with fixed nt = 20 (b) nt ∈ {5, 10, 15, 20} with fixed na = 160 Fig. 2. Convergence performance of the proposed framework is shown, depending on communication networks (i.e., Strongly-connected vs. Fully- connected) and utility function types (i.e., Peaked-reward vs. Submodular- reward) with different number of agents and tasks: (Left) the number of (normal) iterations happened relative to that of agents; (Right) the number of time steps happened (i.e., normal and dummy iterations) relative to that of iterations. example, agents at the middle nodes of the network may change their decisions (and thus increase the number of iterations) while the local partition information of the agent at one end node is being propagated to the other end nodes. Such unnecessary iterations in the middle might not have occurred if the agents at all the end nodes were directly connected to each other. The right subfigures in Figure 2(a) and (b) indicate that approximately 3 -- 4 times of dummy iterations, compared with the required number of normal iterations, are additionally needed under a strongly-connected network. Noting that the mean values of the graph diameter dG for the instances with na ∈ {80, 160, 240, 320} are 36, 58, 75 and 92, respectively, the results show that the amount of dummy iterations happened is much less than the bound value, which is dG as pointed out in Section IV-A. On the contrary, under the fully-connected network, there is no need of such a dummy iteration, and thus the required number of iterations and that of time steps are the same. C. Suboptimality This section examines the suboptimality of the proposed framework by using Monte Carlo simulations with 100 in- stances. In each instance, there are nt = 3 of tasks and na = 12 of agents who are strongly-connected. Figure 3 presents the true suboptimality of each instance, which is the ratio of the global utility obtained by the proposed framework to that by a brute-force search, i.e., JGRAP E/JOP T , and the lower bound given by Theorem 3. A blue circle and a red cross in the figure indicate the true suboptimality and the 80160240320# Agents (# Tasks = 20)11.522.533.54# Required Iterations / # AgentsPeaked-rewardSubmodular-rewardStrongly-connectedFully-connected80160240320# Agents (# Tasks = 20)123456# Required Time Steps / # IterationsPeaked-rewardSubmodular-rewardStrongly-connectedFully-connected5101520# Tasks (# Agents = 160)11.522.533.54# Required Iterations / # AgentsPeaked-rewardSubmodular-rewardStrongly-connectedFully-connected5101520# Tasks (# Agents = 160)1234567# Required Time Steps / # IterationsPeaked-rewardSubmodular-rewardStrongly-connectedFully-connected12 (a) Peaked-reward task (b) Submodular-reward task Fig. 3. True suboptimality of a Nash stable partition obtained by GRAPE for each run of the Monte Carlo simulation (denoted by a blue circle) and its lower bound provided by Theorem 3 (denoted by a red cross) under a strongly-connected communication network: (a) the scenarios with peaked- reward tasks; (b) the scenarios with submodular-reward tasks (a) (b) Fig. 4. The suboptimality lower bound, given by Theorem 3, of a Nash stable partition obtained by GRAPE, depending on communication networks (i.e., Strongly-connected vs. Fully-connected) and utility function types (i.e., Peaked-reward vs. Submodular-reward): (a) fixed nt = 20 with varying na ∈ {80, 160, 240, 320}; (b) fixed na = 160 with varying nt ∈ {5, 10, 15, 20} lower bound, respectively. The results show that the framework provides near-optimal solutions in almost all cases and the suboptimality of each Nash stable partition is enclosed by the corresponding lower bound. The suboptimality may be improved if the agents are allowed to investigate a larger search space, for example, possible coalitions caused by co-deviation of multiple agents. However, this strategy in return may increase communication transactions between the agents because they have to notice each other's willingness unless their individual utility func- tions are known to each other, which is in contradiction to Assumption 4. Besides, the computational overhead for each agent per iteration also becomes more expensive than O(nt), which is the complexity for unilateral searching, as shown in Section IV-A. Hence, the resultant algorithm's complexity may hinder its practical applicability to a large-scale multiple agent system. Figure 4 depicts the suboptimality lower bounds for the large-size problems that were previously addressed in Sec- tion VI-B. It is clearly shown that the agent communication network does not make any effect on the suboptimality lower bound of a Nash stable partition. Although there is no universal trend of the suboptimality with regard to na and nt in both utility types, it is suggested that the features of the lower bound given by Theorem 3 can be influenced by the utility functions considered. In the experiments, the suboptimality bound averagely remain above than 60 -- 70 %. D. Adaptability This section discusses the adaptability of our proposed framework in response to dynamic environments such as unexpected inclusion or loss of agents or tasks. Suppose that there are 10 tasks and 160 agents in a mission, and a Nash stable partition was already found as a baseline. During the mission, the number of agents (or tasks) changes; the range of the change is from losing 50% of the existing agents (or tasks) to additionally including new ones as much as 50% of them. For each dynamical environment, a Monte Carlo simulation with 100 instances is performed by randomly including or excluding a subset of the corresponding number of agents or tasks. Here, we consider a strongly-connected communication network. (a) Dynamic Agents (b) Dynamic Tasks Fig. 5. The plot shows the number of additional iterations required for re- converging toward a Nash stable partition relative to the number of agents in the case when some agents or tasks are partially lost or newly involved (Baseline: nt = 10, na = 160, and a Nash stable partition was already found). Negative values in the x-axis indicate that the corresponding number of existing agents or tasks are lost. Positive values indicate that the corresponding number of new agents or tasks are included in an ongoing mission. A strongly- connected communication network is used. Figure 5(a) illustrates that the more agents are involved additionally, the more iterations are required for re-converging to a new Nash stable partition. This is because the inclusion of a new agent may lead to additional iterations at most as much as the number of the total agents including the new agent (as shown in Lemma 1). On the contrary, the loss of existing agents does not seem to have any apparent relation with the number of iterations. A possible explanation is that the exclusion of an existing agent is favorable to the other agents due to SPAO preferences. This stimulates only a limited number of agents who are preferred to move to the coalition where the excluded agent was. This feature induces fewer 050100Monte Carlo Case ID 20% 40% 60% 80%100%Suboptimality (%)050100Monte Carlo Case ID 20% 40% 60% 80%100%Suboptimality (%)80160240320# Agents (# Tasks = 20)55%60%65%70%75%Suboptimality Lower BoundPeaked-rewardSubmodular-rewardStrongly-connectedFully-connected5101520# Tasks (# Agents = 160)50%55%60%65%70%Suboptimality Lower BoundPeaked-rewardSubmodular-rewardStrongly-connectedFully-connected-80-64-48-32-16ref+16+32+48+64+80# Lost or Newly-included Agents00.511.522.53# Addtional Required Iterations / # AgentsPeaked-rewardSubmodular-reward-5-4-3-2-1ref+1+2+3+4+5# Lost or Newly-included Tasks00.511.522.53# Addtional Required Iterations / # AgentsPeaked-rewardSubmodular-rewardadditional iterations to reach a new Nash stable partition, compared with the case of adding a new agent. Figure 5(b) shows that eliminating existing tasks causes more iterations than including new tasks. This can be ex- plained by the fact that removing any task releases the agents performing the task free and it results in extra iterations as much as the number of the freed agents. On the other hand, adding new tasks induces relatively fewer additional iterations because only some of the existing agents are attracted to these tasks. the number of additional In summary, as the ratio of the number of agents to that of tasks increases, iterations for convergence to a new Nash stable partition also increases. This result corresponds to the trend described in Section VI-B, i.e., the left subfigures in Figure 2(a) and (b). In all the cases of this experiment, the number of additionally induced iterations still remains at the same order of the number of the given agents, which implies that the proposed framework provides excellent adaptability. E. Robustness in Asynchronous Environments This section investigates the robustness of the proposed framework in asynchronous environments. This scenario as- sumes that a certain fraction of the given agents, which are randomly chosen at each time step, somehow can not execute Algorithm 1 and even can not communicate with other normally-working neighbor agents. We refer to such agents as non-operating agents. Given that nt = 5 and na = 40, the fractions of the non-operating agents are set as {0, 0.2, 0.4, 0.6, 0.8}. In each case, we conduct 100 instances of Monte Carlo experiments for which the submodular-reward tasks are used. Figure 6(a) presents that the number of (normal) iterations required for converging to a Nash stable partition remains at the same level regardless of the fraction of the non- operating agents. Despite that, the required time steps increase as more agents become non-operating, as shown in Figure 6(b). Note that time steps growth rate means the ratio of the total required time steps to those for the case when all the agents operate normally. These findings indicate that, due to communicational discontinuity caused by the non-operating agents, the framework may take more time to wait for these agents to operate again and then to disseminate locally-known partition information over the entire agents. As such, dummy iterations may increase in asynchronous environments, though the proposed framework is still able to find a Nash stable partition. Furthermore, the resultant Nash stable partition's suboptimality lower bound obtained by Theorem 3 is not affected, as presented in Figure 6(c). F. Visualization We have na = 320 agents and nt = 5 tasks. The initial locations of the given agents are randomly generated, and the overall formation shape is different in each test scenario such as being circle, skewed circle, and square (denoted by Scenario #1, #2, and #3, respectively). The tasks are also randomly located away from the agents. In this simulation, each agent 13 (b) (a) Fig. 6. Robustness test in asynchronous environments at scenarios with nt = 10, na = 160, and the submodular-reward tasks: the plot shows the effectiveness of the fraction of non-operating agents with regard to: (a) the number of iterations happened until convergence relative to that of agents; (b) the ratio of the time steps happened to those for the normal case; (c) the suboptimality lower bound by Theorem 3. (c) is able to communicate with its nearby agents within a radius of 50 m. Here, the submodular-reward tasks are used. Figure 7 shows the visualized task allocation results, where the circles and the squares indicate the positions of the agents and the tasks, respectively. The lines between the circles represent the communication networks of the agents. The colored agents are assigned to the same colored task, for example, yellow agents belong to the team for executing the yellow task. The size of a square indicates the reward of the corresponding task. The cost for an agent with regard to a task is considered as a function of the distance from the agent to the task. The allocation results seem to be reasonable with consideration of the task rewards and the costs. The number of iterations required to find a Nash stable partition is 1355, 1380, and 1295 for Scenario #1, #2, and #3, respectively. The number of dummy iterations happened is just 20 -- 30% of that of the iterations. This value is much fewer than the results in Figure 2 because the networks considered here are more connected than those in Section VI-B. VII. CONCLUSION This paper proposed a novel game-theoretical framework that addresses a task allocation problem for a robotic swarm consisting of self-interested agents. We showed that selfish agents whose individual interests are transformable to SPAO preferences can converge to a Nash stable partition by us- ing the proposed simple decentralized algorithm, which is executable even in asynchronous environments and under a strongly-connected communication network. We analytically and experimentally presented that the proposed framework provides scalability, a certain level of guaranteed suboptimal- ity, adaptability, robustness, and a potential to accommodate different interests of agents. As this framework can be considered as a new sub-branch of self-organized approaches, one of our ongoing works is to compare it with one of the existing methods. Defining a fair scenario for both methods is non-trivial and requires careful consideration; otherwise, a resultant unsuitable scenario may provide biased results. Secondly, another natural progression of this study is to relax anonymity of agents and thus to con- sider a combination of the agents' identities. Experimentally, 0%20%40%60%80%Fraction of Non-operating Agents11.522.533.54# Required Iterations / # Agents0%20%40%60%80%Fraction of Non-operating Agents010203040Time Steps Growth Rate0%20%40%60%80%Fraction of Non-operating Agents58%60%62%64%66%68%70%Suboptimality Lower Bound14 we have often observed that heterogeneous agents with social inhibition also can converge to a Nash stable partition. More research would be needed to analyze the quality of a Nash stable partition obtained by the proposed framework in terms of min max because our various experiments showed that the outcome provides individual utilities to agents in a balanced manner. ACKNOWLEDGMENT The authors gratefully acknowledge that this research was supported by International Joint Research Programme with Chungnam National University (No. EFA3004Z) REFERENCES [1] H.-S. Shin and P. Segui-Gasco, "UAV Swarms: Decision-Making John Wiley & Paradigms," in Encyclopedia of Aerospace Engineering. Sons, 2014, pp. 1 -- 13. [2] A. Khamis, A. Hussein, and A. Elmogy, "Multi-robot Task Allocation: A Review of the State-of-the-Art," in Cooperative Robots and Sensor Networks 2015. Cham: Springer International Publishing, 2015, vol. 604, pp. 31 -- 51. [3] A. Jevti´c, ´A. Gutierrez, D. Andina, and M. Jamshidi, "Distributed Bees Algorithm for Task Allocation in Swarm of Robots," IEEE Systems Journal, vol. 6, no. 2, pp. 296 -- 304, 2012. [4] E. Sahin, "Swarm Robotics: From Sources of Inspiration to Domains of Application," in Swarm Robotics. Berlin: Springer, 2005, pp. 10 -- 20. [5] M. Dorigo, M. Birattari, and M. Brambilla, "Swarm robotics," Scholar- pedia, vol. 9, no. 1, p. 1463, 2014. [6] K. Barton and D. Kingston, "Systematic Surveillance for UAVs: A Feedforward Iterative Learning Control Approach," in American Control Conference, Washington, DC, USA, 2013, pp. 5917 -- 5922. [7] I. Bekmezci, O. K. Sahingoz, and S. Temel, "Flying Ad-Hoc Networks (FANETs): A Survey," Ad Hoc Networks, vol. 11, no. 3, pp. 1254 -- 1270, 2013. [8] M. Erdelj, E. Natalizio, K. R. Chowdhury, and I. F. Akyildiz, "Help from the Sky: Leveraging UAVs for Disaster Management," IEEE Pervasive Computing, vol. 16, no. 1, pp. 24 -- 32, 2017. [9] I. Jang, J. Jeong, H.-S. Shin, S. Kim, A. Tsourdos, and J. Suk, "Cooperative Control for a Flight Array of UAVs and an Application in Radar Jamming," IFAC-PapersOnLine, vol. 50, no. 1, pp. 8011 -- 8018, 2017. [10] B. P. Gerkey and M. J. Matari´c, "A Formal Analysis and Taxonomy of Task Allocation in Multi-robot Systems," International Journal of Robotics Research, vol. 23, no. 9, pp. 939 -- 954, 2004. [11] G. A. Korsah, A. Stentz, and M. B. Dias, "A Comprehensive Taxonomy for Multi-robot Task Allocation," The International Journal of Robotics Research, vol. 32, no. 12, pp. 1495 -- 1512, 2013. [12] S. Bandyopadhyay, S.-J. Chung, and F. Y. Hadaegh, "Probabilistic and Distributed Control of a Large-Scale Swarm of Autonomous Agents," IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1103 -- 1123, 2017. [13] M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo, "Swarm Robotics: a Review from the Swarm Engineering Perspective," Swarm Intelligence, vol. 7, no. 1, pp. 1 -- 41, 2013. [14] L. Johnson, S. Ponda, H.-L. Choi, and J. How, "Asynchronous Decentralized Task Allocation for Dynamic Environments," in In- fotech@Aerospace, St.Louis, MO, USA, 2011. [15] C. M. Clark, R. Morton, and G. A. Bekey, "Altruistic relationships for optimizing task fulfillment in robot communities," Distributed Au- tonomous Robotic Systems 8, pp. 261 -- 270, 2009. [16] J. H. Dr`eze and J. Greenberg, "Hedonic Coalitions: Optimality and Stability," Econometrica, vol. 48, no. 4, pp. 987 -- 1003, 1980. [17] S. Banerjee, H. Konishi, and T. Sonmez, "Core in a Simple Coalition Formation Game," Social Choice and Welfare, vol. 18, no. 1, pp. 135 -- 153, 2001. [18] A. Bogomolnaia and M. O. Jackson, "The Stability of Hedonic Coalition Structures," Games and Economic Behavior, vol. 38, no. 2, pp. 201 -- 230, 2002. [19] A. Brutschy, G. Pini, C. Pinciroli, M. Birattari, and M. Dorigo, "Self- organized Task Allocation to Sequentially Interdependent Tasks in Swarm Robotics," Autonomous Agents and Multi-Agent Systems, vol. 28, no. 1, pp. 101 -- 125, 2014. (a) Scenario #1 (b) Scenario #2 (c) Scenario #3 Fig. 7. Visualized task allocation results are shown with different geographic scenarios (nt = 5, na = 320). Each square and its size represent each task's position and its reward (or demand), respectively. The circles and the lines between them indicate the positions of agents and their communication network, respectively. The color of each circle implies that the corresponding agent is assigned to the same colored task. 15 [43] A. Darmann, E. Elkind, S. Kurz, J. Lang, J. Schauer, and G. Woeginger, "Group Activity Selection Problem," in Proceedings of the 8th Interna- tional Conference on Internet and Network Economics, Liverpool, UK, 2012, pp. 156 -- 169. [44] A. Darmann, "Group Activity Selection from Ordinal Preferences," in Algorithmic Decision Theory. ADT 2015. Lecture Notes in Computer Science, ser. Lecture Notes in Computer Science, T. Walsh, Ed. Berlin, Heidelberg: Springer Cham, 2015, vol. 9346, pp. 35 -- 51. [45] M. Rubenstein, A. Cornejo, and R. Nagpal, "Programmable Self- assembly in a Thousand-robot Swarm," Science, vol. 345, no. 6198, pp. 795 -- 799, 2014. [46] S. C. Sung and D. Dimitrov, "On Myopic Stability Concepts for Hedonic Games," Theory and Decision, vol. 62, no. 1, pp. 31 -- 45, 2007. [47] M. Karakaya, "Hedonic Coalition Formation Games: A New Stability Notion," Mathematical Social Sciences, vol. 61, no. 3, pp. 157 -- 165, 2011. [48] H. Aziz and F. Brandl, "Existence of Stability in Hedonic Coalition Formation Games," in Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, Valencia, Spain, 2012, pp. 763 -- 770. [49] J. Guerrero and G. Oliver, "Multi-robot Coalition Formation in Real- time Scenarios," Robotics and Autonomous Systems, vol. 60, no. 10, pp. 1295 -- 1307, 2012. [50] O. Shehory and S. Kraus, "Feasible Formation of Coalitions Among Autonomous Agents in Nonsuperadditive Environments," Computational Intelligence, vol. 15, no. 3, pp. 218 -- 251, 1999. [51] C. Nam and D. A. Shell, "Assignment Algorithms for Modeling Re- source Contention in Multirobot Task Allocation," IEEE Transactions on Automation Science and Engineering, vol. 12, no. 3, pp. 889 -- 900, 2015. [52] L. B. Johnson, H.-L. Choi, and J. P. How, "The Role of Information As- sumptions in Decentralized Task Allocation: A Tutorial," IEEE Control Systems, vol. 36, no. 4, pp. 45 -- 58, 2016. [53] W. Saad, Z. Han, M. Debbah, and A. Hjørungnes, "A Distributed Coalition Formation Framework for Fair User Cooperation in Wireless Networks," IEEE Transactions on Wireless Communications, vol. 8, no. 9, pp. 4580 -- 4593, 2009. [54] W. Saad, Z. Han, T. Basar, M. Debbah, and A. Hjørungnes, "Hedonic Coalition Formation for Distributed Task Allocation among Wireless Agents," IEEE Transactions on Mobile Computing, vol. 10, no. 9, pp. 1327 -- 1344, 2011. IEEE Copyright Notice c(cid:13) 2018 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Accepted to be Published in: IEEE Transactions on Robotics [20] N. Kalra and A. Martinoli, "A Comparative Study of Market-Based and Threshold-Based Task Allocation," in Distributed Autonomous Robotic Systems 7, Tokyo, Ed. Springer Japan, 2006, pp. 91 -- 101. [21] Y. Zhang and L. E. Parker, "Considering Inter-task Resource Constraints in Task Allocation," Autonomous Agents and Multi-Agent Systems, vol. 26, no. 3, pp. 389 -- 419, 2013. [22] H. L. Choi, L. Brunet, and J. P. How, "Consensus-based Decentralized Auctions for Robust Task Allocation," IEEE Transactions on Robotics, vol. 25, no. 4, pp. 912 -- 926, 2009. [23] P. Segui-Gasco, H.-S. Shin, A. Tsourdos, and V. J. Segui, "Decentralised Submodular Multi-Robot Task Allocation," in IEEE/RSJ International Conference on Intelligent Robots and Systems, Hamburg, Germany, 2015, pp. 2829 -- 2834. [24] B. Acikmese and D. S. Bayard, "Markov Chain Approach to Proba- bilistic Guidance for Swarms of Autonomous Agents," Asian Journal of Control, vol. 17, no. 4, pp. 1105 -- 1124, 2015. [25] I. Chattopadhyay and A. Ray, "Supervised Self-Organization of Homo- geneous Swarms Using Ergodic Projections of Markov Chains," IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 39, no. 6, pp. 1505 -- 1515, 2009. [26] N. Demir and B. Acikmese, "Probabilistic Density Control for Swarm of Decentralized ON-OFF Agents with Safety Constraints," in American Control Conference, Chicago, IL, USA, 2015, pp. 5238 -- 5244. [27] S. Berman, A. Halasz, M. A. Hsieh, and V. Kumar, "Optimized Stochastic Policies for Task Allocation in Swarms of Robots," IEEE Transactions on Robotics, vol. 25, no. 4, pp. 927 -- 937, 2009. [28] A. Halasz, M. A. Hsieh, S. Berman, and V. Kumar, "Dynamic Redis- tribution of a Swarm of Robots among Multiple Sites," in IEEE/RSJ International Conference on Intelligent Robots and Systems, San Diego, CA, USA, 2007, pp. 2320 -- 2325. [29] M. A. Hsieh, A. Halasz, S. Berman, and V. Kumar, "Biologically Inspired Redistribution of a Swarm of Robots among Multiple Sites," Swarm Intelligence, vol. 2, no. 2-4, pp. 121 -- 141, 2008. [30] T. W. Mather and M. A. Hsieh, "Macroscopic Modeling of Stochastic Deployment Policies with Time Delays for Robot Ensembles," The International Journal of Robotics Research, vol. 30, no. 5, pp. 590 -- 600, 2011. [31] A. Prorok, M. A. Hsieh, and V. Kumar, "The Impact of Diversity on Optimal Control Policies for Heterogeneous Robot Swarms," IEEE Transactions on Robotics, vol. 33, no. 2, pp. 346 -- 358, 2017. [32] T. H. Labella, M. Dorigo, and J.-L. Deneubourg, "Division of Labor in a Group of Robots Inspired by Ants' Foraging Behavior," ACM Transactions on Autonomous and Adaptive Systems, vol. 1, no. 1, pp. 4 -- 25, 2006. [33] E. Castello, T. Yamamoto, Y. Nakamura, and H. Ishiguro, "Foraging Op- timization in Swarm Robotic Systems Based on an Adaptive Response Threshold Model," Advanced Robotics, vol. 28, no. 20, pp. 1343 -- 1356, 2014. [34] H. Kurdi, J. How, and G. Bautista, "Bio-Inspired Algorithm for Task Al- location in Multi-UAV Search and Rescue Missions," in AIAA Guidance, Navigation, and Control Conference, San Diego, CA, USA, 2016. [35] W. Liu, A. F. T. Winfield, J. Sa, J. Chen, and L. Dou, "Towards Energy Optimization: Emergent Task Allocation in a Swarm of Foraging Robots," Adaptive Behavior, vol. 15, no. 3, pp. 289 -- 305, 2007. [36] W. Liu and A. F. T. Winfield, "Modeling and Optimization of Adaptive Foraging in Swarm Robotic Systems," The International Journal of Robotics Research, vol. 29, no. 14, pp. 1743 -- 1760, 2010. [37] A. Martinoli, K. Easton, and W. Agassounon, "Modeling Swarm Robotic Systems: a Case Study in Collaborative Distributed Manipulation," The International Journal of Robotics Research, vol. 23, no. 4, pp. 415 -- 436, 2004. [38] K. Lerman, A. Martinoli, and A. Galstyan, "A Review of Probabilistic Macroscopic Models for Swarm Robotic Systems," in Swarm Robotics. Berlin: Springer, 2005, pp. 143 -- 152. [39] N. Correll and A. Martinoli, "System Identification of Self-Organizing Robotic Swarms," in Distributed Autonomous Robotic Systems 7. Tokyo: Springer Japan, 2006, pp. 31 -- 40. [40] A. Prorok, N. Correll, and A. Martinoli, "Multi-level Spatial Modeling for Stochastic Distributed Robotic Systems," The International Journal of Robotics Research, vol. 30, no. 5, pp. 574 -- 589, 2011. [41] A. Kanakia, J. Klingner, and N. Correll, "A Response Threshold Sigmoid Function Model for Swarm Robot Collaboration," Distributed Autonomous Robotic Systems, pp. 193 -- 206, 2016. [42] D. Dimitrov and S. C. Sung, "Top responstiveness and Nash stability in coalition formation games," Kybernetika, vol. 42, no. 4, pp. 453 -- 460, 2006.
1910.11027	2	1910	2019-11-15T15:02:18	Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support	[ "cs.MA", "cs.CY", "eess.SY", "eess.SY" ]	Demand for health care is constantly increasing due to the ongoing demographic change, while at the same time health service providers face difficulties in finding skilled personnel. This creates pressure on health care systems around the world, such that the efficient, nationwide provision of primary health care has become one of society's greatest challenges. Due to the complexity of health care systems, unforeseen future events, and a frequent lack of data, analyzing and optimizing the performance of health care systems means tackling a wicked problem. To support this task for primary care, this paper introduces the hybrid agent-based simulation model SiM-Care. SiM-Care models the interactions of patients and primary care physicians on an individual level. By tracking agent interactions, it enables modelers to assess multiple key indicators such as patient waiting times and physician utilization. Based on these indicators, primary care systems can be assessed and compared. Moreover, changes in the infrastructure, patient behavior, and service design can be directly evaluated. To showcase the opportunities offered by SiM-Care and aid model validation, we present a case study for a primary care system in Germany. Specifically, we investigate the effects of an aging population, a decrease in the number of physicians, as well as the combined effects.	cs.MA	cs	Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support Martin Comis · Catherine Cleophas · Christina Busing Received: date / Accepted: date Abstract Demand for health care is constantly in- creasing due to the ongoing demographic change, while at the same time health service providers face difficulties in finding skilled personnel. This cre- ates pressure on health care systems around the world, such that the efficient, nationwide provision of primary health care has become one of society's greatest challenges. Due to the complexity of health care systems, unforeseen future events, and a fre- quent lack of data, analyzing and optimizing the performance of health care systems means tackling a wicked problem. To support this task for primary care, this paper introduces the hybrid agent-based simulation model SiM-Care. SiM-Care models the interactions of patients and primary care physicians on an individual level. By tracking agent interac- tions, it enables modelers to assess multiple key in- dicators such as patient waiting times and physician utilization. Based on these indicators, primary care systems can be assessed and compared. Moreover, changes in the infrastructure, patient behavior, and service design can be directly evaluated. To show- case the opportunities offered by SiM-Care and aid model validation, we present a case study for a pri- mary care system in Germany. Specifically, we in- This work is supported by the Freigeist-Fellowship of the Volkswagen Stiftung. This work is supported by the German research council (DFG) Research Training Group 2236 UnRAVeL. Christina Busing · Martin Comis () Lehrstuhl II fur Mathematik, RWTH Aachen University, Pontdriesch 10 -- 12, 52062 Aachen, Germany E-mail: {buesing, comis}@math2.rwth-aachen.de Catherine Cleophas Working Group Service Analytics, Christian-Albrechts- Universitat zu Kiel, Westring 425, 24118 Kiel, Germany E-mail: [email protected] vestigate the effects of an aging population, a de- crease in the number of physicians, as well as the combined effects. Keywords Hybrid Simulation · ABM · DES · Primary Care · Decision Support 1 Introduction Health is one of the most important factors for the prosperity and well-being of a society. Therefore, all member states of the World Health Organiza- tion (WHO) made the commitment to ensure ev- eryone's access to health services [21]. The foun- dation of universally accessible health services is usually laid by a primary care system. Following the definition of the American Academy of Family Physicians [6], primary care systems "serve as the patient's first point of entry into the health care sys- tem and the continuing focal point for all needed health services". To that end, they feature a set of primary care physicians (PCPs) who provide "pri- mary care services to a defined population of pa- tients". Such primary care services include "health promotion, disease prevention, health maintenance, counseling, patient education, diagnosis and treat- ment of acute and chronic illnesses". Demographic change poses serious challenges to primary care systems: Medical and technological progress paired with improved living conditions and reduced birth rates lead to a rise in the share of elderly citizens. In the United States, the percent- age of individuals aged 65 and older is predicted to exceed 21% of the total population by 2030 [65]. 2 Martin Comis et al. In Germany, this demographic shift is even more severe with elderly citizens (aged 65 and above) being expected to account for more than 26% of the total population by 2030 [9]. As populations age, their demand for primary care services tends to increase. This is primarily due to the prevalence of chronic illnesses which disproportionately affect older adults [46,4]. On the supply side, primary care physicians are also aging, e.g., 34.1% of all primary care physicians in Germany were 60 years or older by the end of 2017 [2] and thus about to retire. Moreover, fewer medical students are will- ing to practice primary care [46], let alone open a private primary care practice [35]. This leads to re- duced treatment capacities and exacerbates the risk for supply disruptions. In the United States, the "confluence of a rising de- mand for primary care services and a decreasing supply of professionals providing these services" is considered a "crisis in primary care" [46]. In order to manage this crisis, existing primary care systems have to be fundamentally adjusted [53]. Various new concepts and policies to maintain the standard of health care provision are discussed by the statu- tory health insurances, the governments, and the Associations of Statutory Health Insurance Physi- cians [59,46]. However, all such concepts call for validation and evaluation prior to their potentially costly implementation [53]. Naturally, this leads to the pressing question: How can the quality of pri- mary care systems and the effect of changes to them be quantified? In German legislation, this question is answered by the 2012 GKV-Versorgungsstrukturgesetz [24] which defines adequate health care supply on the basis of profession-specific ratios. The law subdi- vides the country into zones and specifies the re- quired population-to-provider ratio for each med- ical specialization. For example, the predefined nominal ratio of primary care physicians is one PCP per 1,671 inhabitants [24, §11(4)]. This base indi- cator can be adjusted to account for a zone's in- dividual demographic and geographic characteris- tics [24, §2]. If the actual ratio of a zone is signif- icantly higher than the nominal ratio, closing prac- tices will not be replaced. If it is significantly lower, new practices are permitted to be opened. Beyond Germany, we can find similar ratio-based measures in other European countries like Bul- garia, Estonia, Italy, and Spain [43]. But also in the United States, the Health Resources and Services Administration (HRSA) defines adequate health care supply based on profession- and region- specific population-to-provider ratios. If the pre- defined population-to-provider ratio (for primary care 3,500 to 1; or 3,000 to 1 for unusually high needs [16]) of a geographic area is exceeded, HRSA designates it a health professional shortage area to which National Health Service Corps personnel is directed with priority. Obviously all ratio-based assessments have several shortcomings: Even after adjustment, population- to-provider ratios can only provide a very rough estimate for the actual demand. Furthermore, ad- justment rules are highly dependent on the defini- tion of the underlying zones or geographic areas. Factors such as the accessibility of practices and the PCP's individual workload are completely ne- glected. Finally, ratio-based assessments cannot ac- count for new health care concepts such as tele- medicine, mobile medical units, or centralized ap- pointment scheduling. To overcome these limitations, a new approach to model dynamic effects in primary care systems is required. However, analyzing and evaluating health care systems is a complex and complicated task due to the large number of involved individuals and uncertain nature of health care processes, e.g., fluctuating demand, arrival time of patients, emer- gency patients, and durations of treatments. To model such "wicked problems" [57], researchers have achieved promising results using agent-based modeling (ABM), which account for individual agents and their interactions on the micro-level. A general introduction to the concept of ABM is pro- vided in [26]. Existing studies implementing ABM have considered diverse social systems. Examples include matters such as tobacco control [55] and educational policy research [47]. But there are also numerous applications of ABM in field of health care [7], e.g., for accountable care organizations [5, 44], for medical workforce forecasting [45], and in epidemiology [52,50]. This paper introduces the hybrid agent-based sim- ulation tool SiM-Care (Simulation Model for pri- mary Care) to model the dynamics of primary care systems. SiM-Care models patients and PCPs on an individual level as illustrated by Figure 1: Patients and primary care physicians are modeled via a geo-social system, in which patients decide whether and where to request an appointment based on their state of health and PCPs handle appoint- ment requests, manage patient admission, and treat patients. By tracking the resulting interactions in Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 3 PCPs Patients Fig. 1 Geo-social system of patients and physicians.1 SiM-Care, planners can identify dependencies of different subproblems, evaluate new planning ap- proaches, and quantify the effects of interventions on the basis of multiple meaningful performance measurements. From empirical data, we develop re- alistic test scenarios including a controllable degree of uncertainty realized via stochastic simulation ex- periments. The main contribution of this paper can be sum- marized as follows: We introduce the simulation model SiM-Care, which provides decision makers with a versatile decision support tool for primary care planning. SiM-Care is very generic and can be easily modified and extended to meet each mod- eler's needs. Patients and physicians are modeled as individuals who follow their own objectives, learn, and adapt. To ensure computational tractability, the model incorporates a global event queue at its core. As such, SiM-Care can be considered an integrated hybrid simulation model that combines paradigms from agent-based modeling and discrete event sim- ulation. Based on empirical data for a German pri- mary care system, we illustrate how scenarios for the simulation model can be generated. Finally, we showcase the opportunities of SiM-Care through a case study. To the best of our knowledge, SiM-Care is the first model of its kind that captures entire pri- mary care system with all physicians and patients as individual agents and allows for the simultane- ous consideration of microsystem improvements as well as macrosystem reforms. 1 Map tiles by Stamen Design, under CC BY 3.0. Data by OpenStreetMap, under ODbL. The remainder of this paper is structured as fol- lows: Section 2 discusses related work on agent- based simulations with a focus on health care ap- plications. Section 3 introduces SiM-Care based on the ODD framework by Grimm et al. [27]. Section 4 presents a case study based on real-world data to aid model validation and showcase how SiM-Care can be applied to support health care planning. Sec- tion 5 discusses the potential applications and entry requirements of SiM-Care. Finally, Section 6 con- cludes and provides directions for future work. 2 Related Work Decision support for health care planning is an area with increasing importance [30]. To analyze health care systems, decision support tools have to deal with the detail complexity that is inherent to the health care sector [23]: Patients schedule appoint- ments based on their preferences and state of health, while PCPs offer appointments and treat patients. Thereby, micro-interactions affect macro-level in- dicators as agents observe, learn and adapt, decide and act, and - as a group - determine the system's behavior [26]. When a system's status depends on such micro-interactions, its behavior becomes diffi- cult to predict and a "wide range of possible out- comes may arise from any policy change" [23]. Simulation modeling can deal with this kind of complexity by "simulating the life histories of in- dividuals and then estimating the population effect from the sum of the individual effects" [23]. As such, simulation models represent a powerful tool to inform policy makers: They can provide valuable insights into the dependencies within health care systems and allow for the prediction of the outcome of a change in strategy ahead of a potentially costly and risky real-world intervention [23,30]. Given these potentials, the use of computer simula- tion to study health care delivery systems has sig- nificantly increased over the recent years [68]. The resulting body of literature on applications of simu- lation in health care is rich and can, from a method- ological point of view, be classified into four main groups [14]: System dynamics [32,11], discrete- event simulations [30,36], agent-based models [7, 64], and hybrid simulations that combine two or more of the former paradigms [13,12]. To catego- rize related work on the study of primary care sys- tems even further, we distinguish in the following models that focus on microsystem improvements and models that aim at macrosystem reforms. PCPsPatients4 Martin Comis et al. Simulation models aimed at studying microsystem improvements in primary care systems mostly in- clude a detailed model of a single (specific) outpa- tient practice and focus on a predefined subset of potential improvements. Zong et al. [68] present a discrete event simulation for a pediatric clinic at the University of Wisconsin Health. Their model includes a very detailed representation of the se- quential stages during a patient's visit. In a set of "what-if'" scenarios, the authors investigate how the overall performance of the clinic is impacted by different scheduling templates, a change in the medical assistant to physician ratio, and the pairing of resident doctors with clinicians. Shi et al. [62] developed a discrete event simulation model for a primary care clinic of the Department of Vet- eran Affairs. Within the model, the different patient flow routes for appointment patients, walk-ins, and nurse-only patients are distinguished. In a scenario analysis, the authors investigate how the clinic's performance is affected by six distinct factors that include walk-in and no-show rates as well as the double booking of appointments. Cayirli et al. [18] used empirical data collected at a primary care clinic in New York to devise a discrete event simu- lation of a generic single-server primary care prac- tice. The model distinguishes new and returning pa- tients and accounts for walk-ins, no-shows, patient punctuality, and service time variations. In a sim- ulation study, the authors evaluate 42 appointment systems that vary in the implemented sequencing- and appointment rules. A similar discrete event sim- ulation of a generic single-server primary care prac- tice is introduced by Schacht [61]. In his model, all arriving patients have a stochastic willingness to wait and always request an appointment. If the indirect waiting time for this appointment exceeds a patient's willingness to wait, they become walk- ins. The arrival rate of patients depends on the ses- sion, day, and month to model seasonality. In a case study, the author evaluates a class of appoint- ment systems that can account for seasonal varia- tions in demand through reconfigurations. Further simulation models aimed at the study of microsys- tem improvements in primary care practices can be found in [67,51,25]. In contrast to SiM-Care, all of the models above include only one single pri- mary care practice out of the many providers that make up a primary care system. Moreover, all of these models adopt a different approach to the rep- resentation of patients: While SiM-Care models a persistent patient population that is shared by all providers, the previous models perceive patients as non-persistent, i.e., patients are generated as they arrive at the practice and cease to exist as soon as they are discharged. As a result, the previous mod- els cannot account for the effects of individual mi- crosytem improvements on the entire primary care system itself. Simulation models aimed at investigating macrosys- tem reforms of primary care systems mostly include an entire primary care system, however they are usually much more high level. Matchar et al. [48] use the methodology of system dynamics to de- velop a simulation model to aid primary care plan- ning in Singapore. The model captures the causal relationships between the stakeholders' aims and the provision of services in an analytical frame- work. The authors evaluate three policy changes that constitute in reducing the service gap, reduc- ing the out-of-pocket costs, and increasing the num- ber of physicians. Through the use of system dy- namics, the model is much more high level than SiM-Care and does not model patients or physi- cians as individuals. Consequently, the model can- not account for the objectives and satisfaction of in- dividual stakeholders which limits the possibilities for evaluation. Homa et. al [31] present an agent- based model to investigate the paradox of primary care. Their model features patients, PCPs, and spe- cialists as individual agents. Every patient has a health status that changes over time: The contrac- tion of illnesses leads to a (temporary) decrease in the patients' health; the treatment of acute illnesses by PCPs and specialists as well as regular check- ups (performed exclusively by PCPs) lead to an in- crease in the patients' health. Tracking the evolu- tion of the patients' average health status over time, the authors investigate how public health is posi- tively affected by the interplay of different mecha- nisms in primary care. As such, the model of Homa et. al has a different objective than SiM-Care: While Homa et al. investigate the external effects of treat- ments in primary care on the entire health care sys- tem, SiM-Care focuses on the processes within pri- mary care systems. To that end, SiM-Care models the scheduling of appointments, the patients' actual practice visits that result in waiting times through the interaction of patients, and the physicians' treat- ments of patients with variable service times which are not part of the model by Homa et al. To the best of our knowledge, there is no previ- ous work on simulation models for the evaluation of primary care systems that allow for the simulta- neous consideration of microsystem improvements and macrosystem reforms as in SiM-Care. Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 5 3 Simulation Model SiM-Care models the interaction of patients and primary care physicians in a restricted geographical area over a given period of time; compare Figure 1. The model features two types of agents: a popu- lation of potential patients P and a population of primary care physicians G. Every patient ρ ∈ P re- sides at a specific location, belongs to a certain age group, has an individual health status, as well as in- dividual treatment preferences; compare Figure 2. Patients continuously develop acute illnesses that depend on their age and health status and require treatment until they subside. Additionally, patients may suffer from long term chronic illnesses which need to be monitored by a physician on a regular basis. To receive medical attention, patients either schedule an appointment or visit a PCP's practice without prior notice. Thereby, the patients' decision making process depends on their individual prefer- ences and health status which determine the choice of physician, the type of the visit (walk-in/ appoint- ment), as well as the time of the visit. Physicians φ ∈ G practice at a certain location and have weekly opening hours during which they admit patients for treatment; see Figure 2. Moreover, every physician φ ∈ G follows individual strategies that govern how they manage appointments, admit patients, and per- form treatments. As patients and physicians inter- act, they influence each other and adapt by adjust- ing their preferences and strategies. The simulation's purpose is to model trade-offs between the objectives pursued by three stake- holder groups: patients, primary care physicians, and policy makers. These objectives are assumed by the model as follows: While PCPs strive to effi- ciently utilize their time, patients strive for a quick response to their health concerns. Thereby, the model illustrates the trade-off between efficiency and patient-centered care. The objectives pursued by policy makers can range from minimizing the cost of health care to maximizing the degree of patient-centered care. Policy makers are not repre- sented by agents within SiM-Care. Instead, policy decisions set relevant model parameters such as the number of physicians in the area, treatment stan- dards, and the financial reimbursement system. To describe the simulation model in detail, we rely on guidance from the ODD framework de- scribed by Grimm et al. [27]. To concisely high- light the interaction of relevant components in that framework, we order and group the design ques- tions given by this framework as follows: First, Section 3.1 defines the temporal and geographical scales within SiM-Care. Second, Section 3.2 de- scribes the representation of the relevant entities and state variables, including patients and physi- cians with their sensing, predicting, adapting, in- teracting, and learning actions. Third, Section 3.3 provides process overviews and describes matters of scheduling. Fourth, Section 3.4 explains how and where the model captures the uncertain nature of health care systems through stochastic parame- ters. Fifth, Section 3.5 briefly reviews the indica- tors that result from running the simulation and ex- plains their emergent properties. Sixth, Section 3.6 discusses the initialization of a simulation experi- ment. Seventh, Section 3.7 documents the submod- els that implement, e.g., the PCP's strategies to han- dle appointments. Finally, Section 3.8 describes our structural validation as well as our approach to ver- ification taken when implementing the model. 3.1 Simulation Environment SiM-Care's environment entails the geographical and temporal structure as well as policy effects. Within the model, locations (cid:96) ∈ L := [−90,90] × [−180,180] are represented using the geographic coordinates latitude and longitude indicating the north-south and east-west position, respectively. The modeled time period is considered as a con- tinuum structured by points in time and durations. For any time object (point in time or duration) t = (δ,η) ∈ T := N× [0,1), δ ∈ N indicates the day and η ∈ [0,1) =: H specifies the time as an increment of day known as decimal time. That is, we use the same encoding for points in time and durations as context uniquely defines which of the former a time object refers to. For example, (38,0.55) ∈ T corre- sponds to day 38 and 24· 60· 0.55 = 792 minutes, i.e., 1:12 p.m. as a point in time or, analogously, to a duration of 38 days, 13 hours, and 12 minutes. To ease notation, we associate every point in time and duration (δ,η) ∈ T with the non-negative value δ+η∈ R≥0 which yields a bijection between T and R≥0. In addition to the continuous representation of time, we structure each day into a morning and an after- noon session as it is common practice in primary care [40]. Each session λ = (δ,γ) ∈ Λ := N×{0,1} is uniquely defined by a day δ ∈ N and a binary indicator γ ∈ {0,1}. Thereby, the binary indicator γ 6 Martin Comis et al. Patients • location • age class • health status • preferences (cid:8) develop illnesses → request appointments → visit with appointment → vist as walk-ins ↔ adjust preferences and strategies ← assign appointments ← manage admission ← perform treatments Physicians • location • opening hours • strategies Fig. 2 Concept of SiM-Care showing both types of agents with their main attributes as well as interactions between agents. defines whether it is the morning (γ = 0) or the after- noon (γ = 1) session. Sessions reoccur on a weekly basis which yields an equivalence relation ∼ on the set of sessions Λ via (δ1,γ1) ∼ (δ2,γ2) :⇔ δ1 ≡ δ2 mod7 ∧ γ1 = γ2. The resulting equivalence class for a session λ ∈ Λ defined as [λ] := {λ(cid:48) ∈ Λ : λ(cid:48) ∼ λ} contains all ses- sions sharing the same day of the week and time of the day, e.g., all Thursday afternoon sessions. Thus, we can associate the set of all equivalence classes Λ/∼ := {[λ] : λ ∈ Λ} with the 14 sessions of the week, i.e., Monday to Sunday with a respec- tive morning and afternoon session. 3.2 Entities and State Variables Modeled as interacting agents, patients ρ ∈ P and PCPs φ ∈ G are the active entities in the simula- tion. Their interaction is motivated by patients' suf- fering from illnesses and therefore seeking treat- ment with PCPs via appointments or walk-in vis- its. Both patients and PCPs are complex individu- als featuring characteristics that represent entities themselves. Going from simple to more elaborated, we begin by describing the self-containing entities of SiM-Care and end with the description of the agents representing patients and physicians. 3.2.1 Objectives When patients suffer from an acute illness, they want to be treated as soon as possible, ideally by their preferred physician. For the continuous treat- ment of chronic illnesses and the follow-up care of acute illnesses, patients prefer treatment by the same physician through appointments in regular in- tervals. Physicians, on the other hand, aim at effi- ciently utilizing their available time while minimiz- ing overtime. Thus, patients' and physicians' objec- tives are in conflict as it is ineffective for physicians to fully comply with patient demands: To ensure that all short-notice appointment requests can be accommodated, PCPs would have to withhold too much treatment time. Providing follow-up appoint- ments in strict intervals would prevent PCPs from reacting to demand fluctuations. Policy makers, while not explicitly modeled, fol- low a multitude of conflicting objectives. On the one hand, they need to ensure a certain minimum standard in health care quality to guarantee patients are treated when necessary. On the other hand, they cannot afford to subsidize an excessive number of physicians. Thus, policy makers necessarily aim at a trade-off: A purely patient-based system that dis- regards efficiency is likely to turn out to be unaf- fordable, a health system optimized only for effi- ciency might lead to unacceptable waiting and ac- cess times. SiM-Care represents policy decisions through their resulting parameter values, e.g., the number of physicians and their distribution. 3.2.2 Illnesses and Families of Illnesses Illnesses are health concerns that cause discomfort to patients and require treatment. They belong to a certain illness family (e.g. cold or heartburn), have a certain seriousness (e.g. mild or severe), persist over a certain period of time, and require an initial treatment within an acceptable time frame as well as subsequent follow-up visits in regular time inter- vals. In SiM-Care, we formalize illnesses as tuples 7 Unit [days] [days] [days] Type si ∈ [0,1] fi ∈ F di ∈ T ωi ∈ T νi ∈ T Table 1 Attributes of illnesses i ∈ I . Attribute seriousness illness family duration willingness to wait follow-up interval Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support i = (si, fi,di,ωi,νi) ∈ I with attributes as shown in Table 1. Thereby, si ∈ [0,1] defines the seriousness of the illness, fi ∈ F defines the illness family of the illness, and di ∈ T defines the duration of the illness. The parameter ωi ∈ T defines the illness' willingness to wait, which is the patient's maximum accepted waiting time for the initial treatment. The parameter νi ∈ T defines the illness' follow-up in- terval, which specifies the frequency of the required aftercare that follows the initial treatment. When we use this representation to model health concerns that are not strictly illnesses like the need for vac- cination, the characteristics duration and follow-up interval may not apply. In such cases, setting pa- rameter values di = /0 and νi = /0 indicates that the respective characteristic is not applicable for i ∈ I . Table 2 Attributes of families of illnesses f ∈ F . Type Attribute D f : [0,1] → T linear function for expected duration linear function for expected willingness Wf : [0,1] → T Nf : [0,1] → T linear function for follow-up interval κ f ∈ {0,1} chronic attribute Families of illnesses serve as the classification sys- tem of illnesses within SiM-Care. While emerg- ing illnesses vary in their manifestation, families of illnesses define the common constant traits of all illnesses belonging to the same illness family. In our model, the common constant traits of all ill- nesses i ∈ I with seriousness si ∈ [0,1] belong- ing to illness family fi ∈ F are the expected du- ration D fi(si) ∈ T , the expected willingness to wait Wfi(si) ∈ T , and the follow-up interval Nfi(si) ∈ T . The expected duration D fi(si) and expected willing- ness to wait Wfi(si) are exclusively used during the generation of new emerging illnesses and serve as the means for distributions from which we sample each illness' stochastic duration di and stochastic willingness to wait ωi. Thus for all emerged ill- nesses i ∈ I , it generally holds that di (cid:54)= D fi(si) and ωi (cid:54)= Wfi(si). Only the follow-up interval of emerged illnesses i ∈ I derives from the illness family in a deterministic way, i.e., νi = Nfi(si). In order to define the common traits of emerg- ing illnesses, families of illnesses f ∈ F are for- mally specified by three functions: A linear func- tion D f : [0,1] → T that defines the expected dura- tion D f (s) in days for all emerging illnesses with seriousness s ∈ [0,1] that derive from illness family f ∈ F . Moreover, linear functions Wf : [0,1] → T and Nf : [0,1] → T that analogously define the ex- pected willingness to wait in days and follow-up in- terval in days; see Table 2. As above, we indicate the inapplicability of the characteristics duration or follow-up interval to families of illnesses by setting D f = /0 and Nf = /0, respectively. To illustrate the concept of illnesses and families of illnesses, consider the illness family "common cold" with expected illness duration defined by D f (s) = 10s + 3, expected willingness to wait de- fined by Wf (s) = −3s + 3, and follow-up interval defined by Nf (s) = −2s + 7. When a patient de- velops a mild (si = 0.2) "common cold", the ill- ness family "common cold" defines the expected duration, expected willingness to wait, and follow- up interval of the mild cold as D f (si) = 5 days, Wf (si) = 2.4 days, and Nf (si) = 6.6 days. The ac- tual duration and willingness to wait of the devel- oped mild "common cold" are stochastic and vary around their expected counterparts, e.g., di = 5.5 days and ωi = 2.7 days. The illness' follow-up in- terval is deterministic and derives from the illness family via νi = Nfi(si) = 6.6 days. To model chronic health concerns such as diabetes that persist over an extended period of time, a chronic attribute κ f ∈ {0,1} identifies families of chronic illnesses. Thereby, κ f partitions F into the set of acute families of illnesses F act := { f ∈ F : κ f = 0} and the set of chronic families of illnesses F chro := { f ∈ F : κ f = 1}. This directly induces a partition of the set of illnesses I into the set of acute I act and the set of chronic illnesses I chro. Acute illnesses i ∈ I act develop and subside over time and patients can simultaneously suffer from an arbitrary number of acute illnesses I act ⊆ I act. Chronic illnesses ς ∈ I chro are conceived as static by the model -- they neither develop nor heal in the modeled time period. Instead, each patient ρ ∈ P either suffers from exactly one chronic illness ςρ ∈ I chro throughout the modeled time period, ρ = {ςρ} ⊆ I chro, or no chronic illness at i.e., I chro ρ = /0. To distinguish patients suffering all, i.e., I chro from a chronic illness from those who do not, we refer to the former as chronic patients. 8 Table 3 Attributes of age classes a ∈ A. Attribute lin. function exp. annual acute illnesses dev. from exp. illness duration dev. from exp. willingness to wait probability to cancel appointments Type Ia : [0,1] → R≥0 ∆d a > 0 a ≥ 0 ∆ω pa ∈ [0,1] 3.2.3 Appointments Appointments specify the point in time at which the treatment of a specific patient is scheduled to take place. To that end, appointments b ∈ B are defined by the time of the appointment tb ∈ T , the attend- ing primary care physician φb ∈ G, and the patient ρb ∈ P receiving treatment. At any point in time, non-chronic patients can have at most one sched- uled appointment bact ∈ B, called the acute appoint- ment. Acute appointments are intended for the ini- tial treatment of acute illnesses, the follow-up treat- ment of acute illnesses, or both. Chronic patients, may have a regular appointment breg ∈ B to treat their chronic illness in addition to the acute appoint- ment to treat their acute illnesses. While chronic illnesses are only treated during regular appoint- ments, acute illness are treated during any appoint- ment. Thus, all of a patients' acute illnesses I act are treated during every appointment. 3.2.4 Age Classes Age classes group the modeled set of patients and serve the purpose of defining the common charac- teristics of patients within the respective classes. For patients of age class a ∈ A, these characteristics are the deviation from the expected illness duration ∆d a > 0, the deviation from the expected willingness a ≥ 0, the probability to cancel an appoint- to wait ∆ω ment after full recovery pa ∈ [0,1], and the expected number of annual acute illnesses defined through the linear function Ia : [0,1] → R≥0; see Table 3. The deviation from the expected illness duration ∆d a is a multiplicative factor, that determines whether the expected illness duration D fi(si) ∈ T extends a > 1) or shortens (∆d (∆d a < 1) for patients of age class a ∈ A. Analogously, the deviation from the expected willingness to wait ∆ω a , determines how the expected willingness to wait Wfi(si) ∈ T of an illnesses changes for patients of age class a ∈ A. The linear function Ia : [0,1] → R≥0 defines the ex- pected number of annual acute illnesses Ia(c)∈ R≥0 for patients in age class a ∈ A which depends on Martin Comis et al. the patient's individual health condition c ∈ [0,1] which can range from perfectly healthy (c = 0) to extremely delicate (c = 1). 3.2.5 Age Class-Illness Distribution The age class-illness distribution πact : A × F act → [0,1] builds the connection between the set of age classes A and the set of acute families of illnesses F act. To that end, πact defines the expected distri- bution of acute illness families for each age class, i.e., among all developed acute illnesses by patients of age class a ∈ A, a fraction πact(a, fi) ∈ [0,1] is expected to belong to illness family fi ∈ F act. As a result, πact defines a discrete probability distribu- tion on the set of acute families of illnesses F act for fixed age class a ∈ A, i.e., ∑ fi∈F act πact(a, fi) = 1. 3.2.6 Patients Patients are the driving force of the simulation, as their health concerns trigger the events that underly most of the simulation's processes. All non-chronic patients ρ ∈ P are characterized by their geograph- ical location (cid:96) ∈ L, health condition c ∈ [0,1], acute illnesses I act ⊆ I act, age class a ∈ A, acute ap- pointment bact ∈ B, and preferences. While the lo- cation, health condition, and age class of each pa- tient remain constant throughout a simulation ex- periment, a patient's acute illnesses, acute appoint- ment and preferences are variable and change over time. Chronic patients possess all the characteristics of non-chronic patients, but are additionally identi- fied by a constant chronic illness ς ∈ I chro and a variable regular appointment breg ∈ B. Patients' preferences determine when, where and how they pursue treatment. Specifically, each pa- tient considers a set of PCPs G con ⊆ G and never seeks treatment with PCPs outside the considera- tion set. Since continuity in the treatment of chronic illnesses is particularly important, chronic patients select a distinguished family physician φfam ∈ G con with whom all regular appointments breg ∈ B are exclusively arranged. While every patients' consid- eration set G con remains constant throughout the modeled time period, patients reevaluate and vary their family physician. Naturally, patients have per- sonal schedules and cannot attend all weekly ses- sions. Thus, the model assumes that each patient has a constant set of weekly-reoccurring session availabilities given by α: Λ/∼ → {0,1}, where 0 Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 9 ρ encodes unavailability. Finally, patients maintain ρ (φ) ≥ 0 as well individual appointment ratings rapp (φ, [λ]) ≥ 0 as session-specific walk-in ratings rwalk for every weekly session [λ] ∈ Λ/∼ and every con- sidered physician φ ∈ G con. Ratings are the means by which patients express their satisfaction with a physician's services. When- ever a patient seeks consultation, the choice of physician is determined by the patient's current ratings. To that end, ratings incorporate patients' sense of geographic distance, matching of opening hours with availabilities, and previous positive and negative experiences. As patients adjust their rat- ings over time, they adjust their choice of PCP. If a physician is unable to meet an appointment re- quest, incurs excessive waiting time, or rejects pa- tients due to capacity overruns, patients reduce their rating. Positive experiences such as successful ap- pointment arrangements or short waiting times in- crease ratings. In other terms, through their sens- ing of the quality of treatment and the adaptation of their ratings, patients learn about the quality of PCPs throughout the simulation cycle. When patients begin to suffer from a new illness, they always seek treatment. To that end, patients first request an appointment from the set of con- sidered PCPs G con. Appointment requests are one of the ways in which patients and PCPs interact. Patients attempt up to two appointment requests in ρ (φ) ≥ 0 they order of the appointment rating rapp assign to the considered primary care physicians φ ∈ G con. If both requested PCPs fail to offer a fea- sible appointment within the patient's willingness to wait, patients resort to their second way of in- teracting with physicians: They forgo an appoint- ment and visit a PCP as a walk-in patient. The se- lection of the PCP for the walk-in visit is based on the corresponding walk-in rating rwalk (φ, [λ]) of the targeted session λ ∈ Λ. Upon arrival, patients may be rejected by physi- cians due to, e.g., capacity overloads. Following a rejection, patients update their rating of the reject- ing PCP and attempt a new visit as walk-in patient at the then-highest-rated PCP. Rejected patients are flagged as emergencies (ε = 1) for as long as they unsuccessfully continue to seek treatment. In our model, this emergency state does not enforce a par- ticular PCP behavior. Instead, PCPs may include the emergency state in their decision making. Until an illness i ∈ I act subsides, patients continu- ously try to arrange follow-up appointments to the initial treatment with the attending physician in the ρ Table 4 Attributes of (chronic) patients ρ ∈ P . Attribute location health condition age class acute illnesses emergency flag acute appointment considered PCPs availabilities appointment ratings walk-in ratings Domain (cid:96) ∈ L c ∈ [0,1] a ∈ A I act ⊆ I act ε ∈ {0,1} bact ∈ B G con ⊆ G α: Λ/∼ → {0,1} ρ (φ) ≥ 0, ∀φ∈G con rapp (φ, [λ]) ≥ 0, rwalk ρ ∀φ∈G con, ∀[λ]∈Λ/∼ I chro = {ς} ⊆ I chro breg ∈ B φfam ∈ G con chronic illness regular appointment family physician Type constant constant constant variable variable variable constant constant variable variable constant variable variable follow-up interval νi ∈ T . Analogously, chronic pa- tients continuously try to arrange regular appoint- ments with their family physician φfam ∈ G con in the follow-up interval νς ∈ T of their unique chronic illness ς ∈ I chro. Only if the arrangement of a follow-up or regular appointment fails and the after- care of the patients is endangered, do patients seek follow-up treatment as walk-in patients. As a result, a chronic patient's chronic illness ς ∈ I chro can be treated by a physician other than the family physi- cian φfam ∈ G con, but only through a walk-in visit triggered by the unavailability of a regular appoint- ment. In SiM-Care, patients do not directly interact with other patients. However, an indirect form interac- tion emerges as patients compete with each other for timely treatment by their preferred PCP. The attributes shared by all patients as well as the attributes specific to chronic patients are summa- rized in Table 4. 3.2.7 Primary Care Physicians Primary care physicians operate practices featur- ing an uncapacitated waiting room to offer medical services to patients in need. The model character- izes physicians φ ∈ G by their geographic location (cid:96) ∈ L, opening hours, as well as an individual set of strategies to schedule appointments, manage patient admission, and organize treatments. SiM-Care assumes that all physicians operate in clinical sessions. Opening hours for these sessions are weekly-reoccurring and therefore defined over the session of the week via o: Λ/∼ → H × H where H denotes the set of decimal times defined in 10 Martin Comis et al. Section 3.1. Opening hours specify for each session λ∈ Λ the time window o([λ]) := [o([λ]),o([λ])] dur- ing which a physician generally admits patients for treatment. The beginning of session λ = (δ,γ) ∈ Λ is defined as o(λ) := (δ,o([λ]) ∈ T , the session's end as o(λ) := (δ,o([λ]) ∈ T . To encode that a PCP is closed for a weekly session [λ] ∈ Λ/∼, we set o([λ]) = /0. Physicians utilize the first hour after the end of each session as time buffer to compensate for possible delays and walk-in patients. Buffers are considered anticipated working time so that only service time that extends beyond the buffer consti- tutes overtime. Figure 3 provides a schematic visu- alization of a PCP's working day. Primary care physicians implement a set of strate- gies to schedule appointments, decide on patient admissions, and organize the treatment of patients. These strategies govern the physicians' interactions with patients and incorporate all of their sensing, predicting, adapting, and learning. The PCP's appointment scheduling strategy S ∈ S app defines how consultation time is allocated to appointment slots and how the resulting slots are assigned to requesting patients. The feasible set of appointment scheduling strategies S app is defined via the interface shown in Figure 4. That is, ev- ery appointment scheduling strategy S ∈ S app has to provide the functionality to answer appointment re- quests with an appointment suggestion (that can be empty). Thereby, every appointment request speci- fies the requesting patient, earliest possible appoint- ment time, willingness to wait, whether the request is for a regular appointment, and whether patient's availabilities have to be respected. Furthermore, ev- ery appointment scheduling strategy S ∈ S app has to provide the functionality to schedule previously of- fered appointments as well as the functionality to cancel previously scheduled appointments. Finally, every appointment scheduling strategy S ∈ S app has to be able to compute the number of upcoming ap- pointments within a session that are scheduled to take place after specified point in time. The PCP's treatment strategy S ∈ S tmt defines the order of treatment among patients from the wait- ing room. Physicians sense their patients' waiting times as input for their strategy. To account for the observation that physicians consciously or uncon- sciously adjust service times depending on demand [29], treatment policies define when and how physi- cians adjust their consultation speed and thereby service times. The feasible set of treatment strate- gies S tmt is defined via the interface shown in Fig- Table 5 Attributes of PCPs φ ∈ G. Attribute location opening hours appointment scheduling strategy admission strategy treatment strategy Type (cid:96) ∈ L o: Λ/∼ → H × H S ∈ S app S ∈ S adm S ∈ S tmt ure 4. That is, every treatment strategy S ∈ S tmt has to keep track of admitted patients, count the number of waiting patients with and without appointment, and define how the treatment strategy is affected by the beginning of a session. Moreover, every treat- ment strategy S ∈ S tmt has to determine the next pa- tient to be treated (that might not exist) as well as the PCP's current consultation speed which is thor- oughly discussed in Section 3.7.4. The PCP's admission strategy S ∈ S adm determines whether a physician admits or rejects an arriving patient based on the current workload. Admitted patients await their treatment in the physician's waiting room. In SiM-Care, PCPs are required to treat all admitted patients. Thus, physicians under- estimating their workload due to faulty predictions might have to work overtime as they accept too many patients. On the other hand, physicians that overestimate their workload reject too many pa- tients and fail to fully utilize their available time. At the end of every session's buffer, physicians learn by reevaluating their predictions and adapt- ing their admission policy. The feasible set of ad- mission strategies S adm is defined via the interface shown in Figure 4. That is, every admission strategy S ∈ S adm has to be able to decide whether an arriv- ing patient is admitted or not given the PCP's treat- ment and appointment scheduling strategy. More- over, every admission strategy S ∈ S adm has to de- fine the adaptive traits that are performed at the end of every session's buffer and depend on the PCP's treatment strategy. Physicians do not directly interact with other physi- cians. However, an indirect form of interaction emerges as PCPs compete for the patients' favor while striving for optimal utilization. The attributes of PCP's are summarized in Table 5. 3.3 Process Overview and Scheduling Within SiM-Care, the progression of time is mod- eled via the discrete event paradigm. That is time Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 11 Buffer time Closed Service time Idle time Overtime o(λ0) o(λ0) o(λ1) o(λ1) Fig. 3 Schematic representation of a PCP's morning (λ0) and afternoon (λ1) session visualizing service-, idle- and overtime. Working Day I IAppointmentSchedulingStrategy Optional(cid:104)Appointment(cid:105) findAppointment( AppointmentRequest request) void scheduleAppointment(Appointment b) void cancelAppointment(Appointment b) int upcomingAppointmentsAfter(Time t) ITreatmentStrategy I void handleArrival(ArrivalEvent ae) int[] waitingPatients() void sessionStarted() Optional(cid:104)ArrivalEvent(cid:105) getNextPatient() float getConsultationSpeed() IAdmissionStrategy I boolean acceptPatient(IAppointmentSchedulingStrategy as, ITreatmentStrategy ts, ArrivalEvent ae) void adaptStrategy(Session session, ITreatmentStrategy ts) Fig. 4 Interfaces implemented by strategies. is a continuum which is traversed between discrete events at which the system state is updated. The model stores events of the form (t,e) in a sequential queue Q where t ∈ T is the point in time an event of type e ∈ E occurs. Events in Q happen chronologically, i.e., Q = {(t1,e1), . . . , (tn,en)} with ti ≤ ti+1 for 1 ≤ i ≤ n−1. As soon as an event (ti,ei) ∈ Q occurs, the simu- lation advances from time ti−1 to time ti, compare Figure 5. The simulation terminates at a specified point in time T ∈ T , i.e., when the first element (ti,ei) ∈ Q with ti ≥ T occurs. Any event (t,e) ∈ Q can generate new events or delete existing ones. To be introduced or affected by event (t,e) ∈ Q , events (t(cid:48),e(cid:48)) ∈ Q must hap- pen after time t, i.e., we require t(cid:48) > t, so that time progresses in a consistent fashion. By construction, event queue Q never runs empty. Every simulation run follows the structure depicted in Figure 6, chronologically processing the events in Q until time T ∈ T is reached. In this, the spe- cific process depends on the event type e ∈ E. We now describe the different event types. Arrival events are indicated by earv(φ,ρ). As illus- trated in Figure 7(a), they mark the event of patient ρ arriving at physician φ's practice, either for an ap- pointment or as a walk-in. The physician's decision to admit or reject arriving patients depends on φ's admission strategy. Every admitted patient is guar- anteed to receive treatment and enters the physi- cian's waiting room. If the physician is currently idle, this triggers the physician's treatment strategy and treatment commences. Follow-up events are indicated by efol(φ,ρ,i). Some families of illnesses fi ∈ F cannot be treated via a single visit. Instead, the related illnesses i ∈ I re- quire follow-up treatments in intervals defined by the parameter νi (cid:54)= /0. Ensuring continuous follow- up treatments, patients always try to arrange a follow-up appointment immediately after the treat- ment of illnesses requiring follow-up consultation. To account for the fact that no feasible follow-up appointment might be available, SiM-Care gener- ates a follow-up event efol(φ,ρ,i) at time ttreat + νi every time illness i ∈ I with νi (cid:54)= /0 suffered by pa- tient ρ ∈ P is treated by physician φ ∈ G at time ttreat ∈ T . Follow-up events serve as the patient's reminder to actively re-pursue follow-up consulta- tion for illness i after the duration of the follow-up interval. Triggered by a follow-up event efol(φ,ρ,i), patient ρ reattempts to arrange a follow-up appoint- ment with physician φ. Should φ once again be unable to provide a suitable appointment, ρ seeks follow-up consultation as a walk-in patient. Every follow-up treatment of an illness i∈ I invalidates all associated existing follow-up events, as the follow- up interval is reset. Therefore, SiM-Care deletes all existing follow-up events efol(φ,ρ,i) ∈ Q associ- ated with illnesses i ∈ I that were treated during a visit before the new follow-up events are gener- ated. As a result, follow-up events only trigger if an illness has not been treated for the duration of its follow-up interval νi ∈ T . Release events are indicated by erel(φ,ρ). As il- lustrated in Figure 7(b), release events mark the event of physician φ releasing patient ρ after a 12 Martin Comis et al. Q = {(t1,e1), (t2,e2), (t3,e3),(t4,e4), (t5,e5), (t6,e6), (t7,e7)} Past e1 e2 Now e4 e3 t4 −t3 e5 e6 Future e7 Fig. 5 Progression of time induced by the processing of event queue Q via the discrete event paradigm. initialize evaluators initialize agents no warm-up? yes perform warm-up poll first element (t,e) from Q no t < T yes process e poll next element (t,e) from Q evaluate simulation Fig. 6 Structure of simulation run with time horizon T . treatment is performed. Whenever a new treatment begins, the sampled service time determines the time of the subsequent release event erel(φ,ρ). All treated illnesses i ∈ I act without duration (di = /0) are cured through a one-time treatment and thus re- moved from I act. Subsequently, all existing follow- up events corresponding to treated illnesses are deleted and new follow-up events are generated in the previously described manner. The success- ful treatment revokes existing emergency flags, i.e., we set ε = 0. If the patient's chronic illness ς ∈ I was treated, the next recurrent regular appointment breg ∈ B with physician φfam is requested at time ttreat + νς. Then, patients request an acute appoint- ment bact ∈ B with physician φ for the follow- up treatment of the persisting acute illness i∗ = argmini∈I act:νi(cid:54)=/0 νi with smallest follow-up interval. The requested appointments ensure the follow-up treatment of all illnesses suffered by patient ρ and will preempt the previously generated follow-up events. Finally, physicians implement their treat- ment strategy to select the next patient from the waiting room if the latter is non-empty. Otherwise, physicians remains idle until the next arrival event triggers the treatment strategy. As a result of this behavior, physicians are never intentionally idle. Illness events are indicated by eill(ρ). As illustrated in Figure 7(c), they describe that patient ρ starts to suffer from a new acute illness. This means that the model generates a new acute illness i ∈ I act with stochastic qualities that depend on the patient's age and health condition and adds it to their set of ill- nesses I . To treat the emerged illness, patients re- quest an appointment from their preferred physi- cians or, in case this does not succeed, directly visit the preferred physician as a walk-in. As a result, each illness event generates a corresponding arrival event earv(φ,ρ) and adds it to the queue Q . Finally, each illness event generates a future illness event eill(ρ) for patient ρ and adds it to the queue Q to mark the next point in time patient ρ develops an acute illness. Recovery events are indicated by erec(ρ,i). They mark the event of patient ρ recovering from acute illness i ∈ I act. Whenever the model generates a new acute illness i ∈ I act with di (cid:54)= /0, it also gen- erates a corresponding recovery event erec(ρ,i) at time till + di, where till ∈ T is the point in time illness i is developed. Illnesses without duration (di = /0) do not require a recovery event as they are immediately cured through their initial treatment. A recovery event removes illness i from I and deletes any associated follow-up event efol(φ,ρ,i) ∈ Q . If patient ρ does not suffer from acute illnesses fol- lowing the removal of illness i, i.e., I act = /0, the model revokes existing emergency flags by set- ting ε = 0 and assumes that ρ may cancel sched- uled acute appointments. Such cancellations oc- Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 13 yes φ accepts ρ according to admission policy no admit ρ reject ρ φ currently idle? no yes initiate treatment of next patient ρ φ treats illnesses of ρ ρ arranges follow-up ∃ waiting patients? no yes initiate treatment of next patient ρ(cid:48) generate new illness i for ρ yes ∃ appointment satisfying patient ρ's requirements? no make appointment walk-in generate next illness event eill(ρ) (a) (b) (c) Fig. 7 Processing of (a) arrival events earv(φ,ρ), (b) release events erel(φ,ρ), and (c) illness events eill(ρ); ρ ∈ P and φ ∈ G. cur with the patient's age-class-specific probability pa ∈ [0,1] and consequently delete the associated arrival event earv(φ,ρ). As a result, some patients keep their existing acute appointment for a final debriefing. Should patient ρ be currently seeking walk-in treatment due to persisting chronic illness ς ∈ I , this effort is continued. Otherwise, current walk-in attempts are canceled and the associated ar- rival event earv(φ,ρ) is deleted. Open- and close events are indicated by eopn(φ) and eclo(φ), respectively. They mark the beginning and ending (including buffer) of a session λ ∈ Λ op- erated by physician φ. They ensure that treatment strategies become aware of a session's beginning, e.g., to allow for strategies that do not treat early- arriving patients before o(λ), and that overtime is incurred for all treatments performed beyond the anticipated buffer time of λ. 3.4 Modeling Variability SiM-Care relies on stochastic values to both ap- proximate real-world variability and control the frequency of events. This applies to aspects of illnesses as well as to patient arrivals, appoint- ment cancellations and service times. In conse- quence, every simulation experiment includes mul- tiple stochastic repetitions of the modeled time pe- riod, termed simulation runs. When examining sim- ulation output, we account for the resulting variabil- ity through confidence intervals. In the following, we highlight the aspects of the model that are probabilistic rather than determinis- tic and discuss how the distributions underlying the random values are parameterized. Frequency of Acute Illnesses. The occurrence of acute illnesses in SiM-Care is modeled via a Pois- sion process. Patients develop acute illnesses at a frequency that depends on their age and health con- dition. For patients ρ ∈ P of age class a ∈ A with health condition c ∈ [0,1], the expected number of acute illnesses per year is given by the parameter Ia(c). The intensity (or rate) of the Poission proc- cess is thus Ia(c)/364 per day. Moreover, the dura- tion between two consecutive illness events eill(ρ) for patient ρ can be sampled from an exponential distribution with rate Ia(c)/364; see [20, chapter 2]. Type of Acute Illnesses. Whenever an illness event eill(ρ) occurs and patient ρ ∈ P falls ill, the model generates an acute illness i ∈ I act according to the patients' age class a ∈ A and health condition c ∈ [0,1]. The model assumes a probabilistic link between illness family fi ∈ F act and the patient's age class a that is expressed via the age class-illness distribution πact; see Section 3.2.5. To that end, any emerging acute illness of patient ρ is randomly as- signed to an illness family according to the discrete probability distribution f (cid:55)→ πact(a, f ) for f ∈ F act. Qualities of Acute Illnesses. For any new illness i ∈ I act of family fi ∈ F generated through SiM- Care, its seriousness si ∈ [0,1] depends on a tri- angular distribution defined on the closed interval [0,1]. The distribution's mode is the health condi- tion c ∈ [0,1] of the patient ρ ∈ P developing illness Eω a ( fi,si) = 14 Eω a ( fi,si) = 10 Eω a ( fi,si) = 6 Eω a ( fi,si) = 2 14 y t i s n e D 3 0 . 2 0 . 1 0 . 0 0 . 0 10 5 20 Willingness to wait [days] 15 25 Fig. 8 Weibull distributions of ωi ∈ T for different values of patient's age adjusted expected willingness to wait Eω a ( fi,si). a( fi,si) := ∆d a( fi,si))− σ2/2. i. Thus, patients with a bad health condition tend to develop more serious illnesses. The duration di ∈ T of illness i depends on a log- normal distribution. Given i's family of illnesses fi ∈ F , seriousness si ∈ [0,1], and the patient's age class a ∈ A, we define the age-adjusted expected a · D fi(si). duration of illness i as Ed Therefore, SiM-Care samples the illness' duration di from a log-normal distribution with sdlog σ = 0.3 and meanlog µ = log(Ed Patient ρ's willingness to wait for the initial treat- ment of illness i as specified by ωi ∈ T depends on a Weibull distribution. Given i's family of ill- ness fi ∈ F , seriousness si ∈ [0,1], and the devel- oping patient's age class a ∈ A, the age-adjusted expected willingness to wait of illness i is defined a ·Wfi(si). Analogous to Wiesche as Eω et. al [67], we sample ωi from a Weibull distribu- tion with shape parameter p = 2 and derive the scale parameter from the age adjusted expected willing- ness to wait as q = Eω a ( fi,si)/Γ(1 + (1/p)) where Γ denotes the gamma function. Figure 8 visualizes the resulting density functions for various choices of the age-adjusted expected willingness to wait. a ( fi,si) := ∆ω Patient Punctuality. Patients do not always arrive on time for their scheduled appointments b ∈ B. In- stead, SiM-Care allows for patient arrivals to vary around the scheduled time tb ∈ T of the appoint- ment by including an arrival deviation. As sug- gested by Cayirli et al. [18], the arrival devia- tion from tb depends on a normal distribution. We choose a mean arrival deviation of µ = −5 minutes and standard deviation of σ = 6 minutes such that roughly 20% of all patients are expected to arrive late for their appointments which is consistent with the observations reported in [22]. Martin Comis et al. Empirical rates Beta(1.93,2.94) y t i s n e D 5 . 1 0 . 1 5 . 0 0 . 0 0.0 0.2 0.4 0.6 Relative arrival 0.8 1.0 Fig. 9 Histogram and beta distributed maximum-likelihood fit for empirical walk-in arrival rates from [66]. Walk-in Arrivals. Walk-in patients have no pre- specified time at which they are expected to arrive. Instead, SiM-Care defines for every walk-in patient an earliest arrival time a ∈ T as well as a latest arrival time b ∈ T which are both situational and thoroughly discussed in Section 3.7.3. The walk-in patients' actual arrival within the given feasible ar- rival interval [a,b] depends on a beta distribution. Specifically, we fit a beta distribution using maxi- mum likelihood estimation to the empirical arrival rates reported by Shan et al. [66]. As a result, we sample the arrival times of walk-in patients from the interval [a,b] of feasible arrival times according to a beta distribution with shape parameters p = 1.93 and q = 2.94; cf. Figure 9. Service Time. SiM-Care treats the service time per patient, i.e., the duration of treatments, as a random parameter. To sample service times, we collected a set of 21 service times in a local primary care practice. As suggested in literature [67,17], we di- vide the sample into patients with and without ap- pointment and apply a log-normal maximum likeli- hood fit. Histograms of our empirical samples and the resulting distributions for walk-ins and patients with appointment are depicted in Figures 10 and 11. Based on the fitted distributions, we sample the ser- vice times of patients with appointment from a log- normal distribution with meanlog µ = 1.82 and sd- log σ = 0.692 and the service times for walk-in pa- tients from a log-normal distribution with a mean- log µ = 1.254 and sdlog σ = 0.723. As our collected data set does not incorporate transition times, we prolong all sampled service times by one minute. Appointment Cancellations. Patients that recover from all their current acute illnesses, i.e., I act = /0, cancel their existing acute appointment bact ∈ B Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 15 y t i s n e D 5 2 0 . 0 2 0 . 5 1 0 . 0 1 0 . 5 0 0 . 0 0 0 . Empirical data Fitted log-normal meanlog sdlog mean sd median 1.82 0.692 7.843 6.144 6.174 0 5 10 15 Service times [min] 20 Fig. 10 Histogram and log-normal maximum-likelihood fit for empirical service times of patients with appointment. y t i s n e D 5 2 0 . 0 2 0 . 5 1 0 . 0 1 . 0 5 0 . 0 0 0 . 0 Empirical data Fitted log-normal meanlog sdlog mean sd median 1.254 0.723 4.551 3.768 3.505 0 5 10 15 Service times [min] 20 Fig. 11 Histogram and log-normal maximum-likelihood fit for empirical service times of walk-in patients. with the age-class-specific probability pa ∈ [0,1]; compare Section 3.2.4. As long as patients suffer from acute illnesses, they only cancel their acute ap- pointment if they require earlier treatment due to a newly emerged acute illness. All patients that have not canceled their appointment will arrive for it. As chronic illnesses are static within the model, regular appointments are never canceled. 3.5 Emergence and Observation SiM-Care tracks key performance indicators from the point of view of patients, primary care physi- cians, and policy makers. Thereby, it aims to illus- trate the trade-offs between the stakeholders' objec- tives. As these indicators emerge from agent inter- actions based on patients' evolving preferences and physicians' evolving strategies, they are difficult to predict in general. From the patients' point of view, key performance indicators include access time, travel distance, and waiting time. Access time measures the time a pa- tient has to wait for an appointment, i.e, given the earliest acceptable appointment time t ∈ T and the time of the arranged appointment tb ∈ T it is de- fined as ac-time := tb − t. The travel distance mea- sures the one-way distance patient ρ ∈ P has to travel when visiting physician φ ∈ G, i.e., tr-dist := dist((cid:96)ρ, (cid:96)φ), where dist((cid:96)1, (cid:96)2) denotes the driving distance between locations (cid:96)1 ∈ L and (cid:96)2 ∈ L in kilometers. The patient's waiting time measures the time spent on-site before the actual treatment com- mences. For walk-in patients, we define the waiting time for given walk-in arrival tarr ∈ T and treatment commencement ttreat ∈ T as wait-time := ttreat−tarr. For patients with appointment, we define the wait- ing time for given time of the appointment tb ∈ T , patient's arrival at the practice tarr ∈ T , and treatment commencement ttreat ∈ T as wait-time := max{ttreat − max{tb,tarr}, 0}. To evaluate patient's indicators, SiM-Care keeps track of the total access time of arranging acute and regular appointments, the total number of arranged acute and regular ap- pointments, the total number of attended appoint- ments, the total number of walk-in patients, the total distance traveled by patients to access physicians, and the total waiting time for both patients with ap- pointment and walk-ins. From the physicians' point of view, key perfor- mance indicators include the utilization, overtime, number of treatments, and number of rejected pa- tients with and without appointment. A physician's utilization describes the percentage of the available working time spent treating patients, i.e., for a ses- sion λ ∈ Λ with total treatment duration t ∈ T it is defined as util := t/(o(λ)− o(λ) + 1 24 ). Note that our definition of utilization clearly underestimates a physician's actual utilization as we do not account for additional tasks such as reporting, accounting, and answering phone calls that are not modeled in SiM-Care. Overtime describes the physician's working time beyond the anticipated buffer, i.e., if the last patient in session λ ∈ Λ is released at time trel ∈ T it is defined as over := max{trel − o(λ)− 24 , 0}; see Figure 3. To evaluate the physician's in- 1 dicators, SiM-Care collects on physician level the total service time spent treating patients, the total number of performed treatments, the total overtime, and the total number of rejected patients with and without appointment. The total available working time per PCP required to compute the utilization, can be derived from the opening hours o and the modeled time horizon T . 16 add φ to G con else if dist((cid:96)ρ, (cid:96)φ) < 15km then Algorithm 1 Determine patient's considered PCPs. Require: patient ρ ∈ P , set of primary care physicians G 1: set G con = /0 2: for φ ∈ G do 3: 4: 5: 6: 7: 8: end if 9: 10: end for 11: return G con if rand(20) < 1 then add φ to G con end if 3.6 Input, Initialization, and Warm-Up SiM-Care codes a large number of values as flexi- ble parameters. Setting up a simulation experiment requires an input scenario to specify these param- eter values. Each simulation scenario represents a particular setting, in which a specific set of patients interacts with a specific set of physicians under spe- cific circumstances. As part of every simulation scenario, the modeler specifies the families of illness F , the age classes A, the age class-illness distribution πact, and the set of physicians G with all their attributes. The set of patients P is only partially defined through the simulation scenario: Each scenario specifies the total number of chronic and non-chronic patients. Moreover, every patient's location (cid:96) ∈ L, health condition c ∈ [0,1], age class a ∈ A, availabili- ties α: Λ/∼ → {0,1}, and, for chronic patients, a chronic illness ς ∈ I chro are given. The remaining attributes of patients, e.g., ratings and illnesses, are initialized as described below. At initialization, patients do not suffer from acute illnesses, i.e., I act = /0 and are not considered emer- gencies, i.e., ε = 0. Furthermore, all patients are ini- tialized without scheduled appointments, i.e., bact = breg = /0. The consideration set of physicians G con ⊆ G per patient ρ ∈ P is determined according to Al- gorithm 1 where rand(x) for x > 0 denotes a uni- formly distributed float from the half-closed inter- val [0,x). As a result, each patient considers all physicians within a 15km driving radius. Physi- cians outside this radius are considered with a 5% chance as some patients may choose their physician according to criteria other than proximity to their home, e.g., for historical reasons or personal rec- ommendations. To initialize patients' appointment ratings rapp and walk-in ratings rwalk ρ (φ) (φ, [λ]) for every con- ρ Martin Comis et al. sidered physician φ ∈ G con and weekly session [λ] ∈ Λ/∼, we denote the number of matches be- tween the physician's opening hours and patient ρ's availabilities by m(ρ,φ) := {[λ] ∈ Λ/∼ : α([λ])∧ o([λ]) (cid:54)= /0} and the maximal shortest access dis- tance by distmax := maxρ∈P minφ∈G dist((cid:96)ρ, (cid:96)φ). The model then initializes appointment ratings as 3m(ρ,φ)− dist((cid:96)ρ, (cid:96)φ) + rand(2distmax) + 100 0 rapp ρ (φ) = if m(ρ,φ)>0 else. ρ Walk-in ratings rwalk (φ, [λ]) are session specific as immediate care requires physicians to be in service. Thus, sessions in which a physician is closed are not feasible for walk-in visits which is encoded by an empty rating. The model initializes walk-in ratings as rand(distmax) − dist((cid:96)ρ, (cid:96)φ) + 100 if o([λ]) (cid:54)= /0 /0 else. rwalk ρ (φ, [λ]) = From the initialized ratings, SiM-Care subsequently determines the family physician for chronic pa- tients as the physician from the consideration set that has the highest appointment rating, i.e., φfam = argmaxφ∈G con rapp ρ (φ) which completes the setup of all simulation entities. At this point in the initialization process, the global event queue Q is still empty and therefore running a simulation experiment would result in no agent actions. To make physicians take up their work, the model generates open- and close events eopn(φ) and eclo(φ) for every session operated by physician φ ∈ G and adds these to Q . To start the process of patients continuously developing acute illnesses, the model generates an initial illness event eill(ρ) for every patient ρ ∈ P and adds it to Q . Finally, to start the regular treatments of chronic illnesses ς ∈ I chro, an initial follow-up event efol(φfam,ρ,ς) is generated at a randomly chosen point in time within ς's follow-up interval νς ∈ T according to a uni- form distribution and subsequently added to Q . As discussed previously, several aspects of the sim- ulation model rely on emergent values that require a "warm-up" period before producing meaningful results for decision support. Therefore, we precede every simulation experiment with a warm-up dur- ing which patients develop acute illnesses, physi- cians fill their appointment books, and patients ad- just their ratings. The duration of the warm-up and the length of the modeled time horizon are both variable and specified by the modeler through the Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 17 input scenario. Note, that this does not correspond to solely analyzing a steady state, as the agents' emergent interactions can result in the development of meaningful trends in the data. 3.7 Submodels We consider different aspects of SiM-Care that rely on an internal logic as submodels. One of the most basic submodels describes the logic of distances and travel times. More complex exam- ples include the logic underlying patients' behav- ior when requesting appointments and visiting prac- tices as walk-ins, as well as the physician's strate- gies which are submodels by themselves. As SiM- Care allows for modular PCP strategies, we exem- plify each strategy through the specific approach that is used in the case study. Further submodels describe the consequences of rejecting patients, ser- vice time reductions, patients' rating adjustments, patients' choice of their family physician, and treat- ment effects. 3.7.1 Distances and Travel Times SiM-Care does not feature a road network to com- pute travel distances and travel times. Instead, it ap- proximates the driving distance dist: L × L → R between two locations in kilometers using the great circle distance computed through the haversine for- mula with a detour factor of 1.417 as determined by Boscoe et al. [10]. These authors also point out, that driving distances provide good approxima- tions for travel times in minutes, i.e., we compute travel times by assuming a constant driving speed of 60km/h. As a result we define the travel time τ: L × L → T as τ((cid:96)1, (cid:96)2) := dist((cid:96)1,(cid:96)2) 60·24 . 3.7.2 Patients Requesting Appointments Patient agents ρ ∈ P request an appointment with a physician φ ∈ G, by specifying the earliest ac- ceptable appointment time t ∈ T and their willing- ness to wait for this appointment ω ∈ T . As a re- sult, newly-arranged appointments are feasible, if and only if they are scheduled in the time interval [t,t + ω]. The earliest acceptable appointment time t ∈ T de- pends on the request. The initial treatment of acute illnesses i ∈ I act is urgent, so that patients seek to schedule an appointment as soon as possible. Thus, for these initial treatments, the earliest ac- ceptable appointment time is the time of the request treq ∈ T plus a 30 minute buffer (corresponding to 1 48 in decimal time) plus the direct travel time, i.e., t = treq + 1 48 + τ((cid:96)ρ, (cid:96)φ). Follow-up treatments are planned at regular intervals specified by the pa- rameter νi ∈ T . Patients request follow-up appoint- ments in two ways: First, at the very beginning of the follow-up interval as every patient requests a follow-up appointment directly after the treatment of illnesses that require aftercare. Second, at the very end of the follow-up interval (triggered by a follow-up event) in case no feasible appointment was available at the time of the previous treatment. In the latter case, the request is urgent and there- fore the earliest acceptable appointment time is de- fined as above, i.e., t = treq + 1 48 + τ((cid:96)ρ, (cid:96)φ). If the follow-up appointment is requested at the begin- ning of the follow-up interval, the next follow-up appointment for illness i ∈ I should be scheduled after the follow-up interval has passed, i.e., we set t = treq + νi. The willingness to wait ω ∈ T defines the maxi- mum acceptable waiting period between the earli- est appointment time and the actual time of the ap- pointment. As a result, it serves as an upper bound to the patient's access time defined in Section 3.5. Patients' willingness to wait for the initial treatment of acute illness i ∈ I act is illness specific and given by ω = ωi. Analogously, the maximum duration chronic patients are willing to wait for their reg- ular appointment depends on their chronic illness ς ∈ I chro, i.e., ω = ως. If patients request a follow- up appointment for acute illness i ∈ I act, the will- ingness to wait is proportional to the length of the follow-up interval νi ∈ T . To ensure that the follow- up interval is not exceeded by an excessive time span, the willingness to wait for follow-up appoint- ments regarding i∈ I act is ω = νi 5 +1. Finally, emer- gency patients who were denied treatment are ex- ceptionally impatient and their willingness to wait is ω = 0. Algorithm 2 describes how a patient requests an initial appointment for a newly emerged acute illness. First, patients check whether they have a pre-existing appointment within the acceptable time frame. From the patients' point of view, pre- existing appointments are particularly convenient as they require no further actions. Therefore, pa- tients accept pre-existing appointments as feasible, even if they exceed their willingness to wait by up 2 in decimal time); see lines 1− 2. to 12 hours (or 1 18 Martin Comis et al. Algorithm 2 Arranging appointment for acute illness. Require: patient ρ ∈ P , willingness to wait ω ∈ T , earliest 1: if ρ has acute or regular appointment before appointment time t ∈ T time t + ω + 1 return true 2 then cancel acute appointment delete associated arrival event earv(φ,ρ) 2: 3: else 4: 5: 6: end if 7: determine preferred physicians φ1,φ2 ∈ G con such that ρ (φ1) ≥ rapp rapp 8: for j = 1,2 do 9: 10: query φ j for an appointment if physician φi offers appointment within [t,t + ω] ∧ (satisfying ρ's availabilities α∨ ω ≤ 3) then ρ (φ) ∀φ ∈ G con \{φ1,φ2}. ρ (φ2) ≥ rapp accept appointment add earv(φ j,ρ) to Q return true refuse appointment continue else 11: 12: 13: 14: 15: 16: end if 17: 18: end for 19: return false # adapt rapp ρ (φ j) # adapt rapp ρ (φ j) If the patient's existing appointments are infeasible for the newly emerged illness, the existing acute ap- pointment is canceled to make room for a new, ear- lier, acute appointment (compare line 4 and 5). Patients ρ ∈ P request appointments from the two currently highest rated physicians φ1,φ2 ∈ G con in their consideration set (compare line 7). Physicians φ1 and φ2 are queried in order of their rating, i.e., patients first request an appointment with the higher rated PCP φ1 and only resort to φ2 if the request is unsuccessful. When a physician cannot offer a fit- ting slot, patients reduce their rating for the respec- tive PCP. When their willingness to wait is at least 3 days (compare line 10), patients only accept appoint- ments that fit their personal availability α: Λ/∼ → {0,1} (cf. Section 3.2). In case ω ≤ 3, the request is so urgent that patients are always available. If neither φ1 nor φ2 offer a feasible slot, the search for a feasible appointment is deemed unsuccessful and patients resort to a walk-in visit. When patients request follow-up appointments, they mostly follow the steps outlined in Algorithm 2. The main difference concerns the inquiry process (cf. line 7 and 8), as new follow-up appointments are exclusively arranged with the physician that per- formed the previous treatment. Only pre-existing appointments can be used for follow-up visits al- though they are not with the physician that per- formed the previous treatment; compare line 1. If the follow-up appointment request is made at the end of the follow-up interval triggered by a follow- up event, a failure initiates a walk-in attempt to ensure the patient's aftercare. If the follow-up ap- pointment is requested immediately after treatment at the beginning of the follow-up interval, a fail- ure does not lead to a walk-in attempt as the cor- responding follow-up event will eventually lead to a reattempt at arranging a follow-up appointment. Chronic patients' regular appointments are essen- tially follow-up appointments and thus arranged ac- cording to the same logic. The only difference con- cerns the evaluation of pre-existing appointments: As regular appointments are exclusively arranged with the patient's family physician φfam ∈ G con, pre- existing acute appointments are only perceived as feasible if they are with the family physician φfam (cf. line 1 and 2). Infeasible pre-existing acute ap- pointments are not canceled but instead an addi- tional regular appointment is arranged with the fam- ily physician φfam (cf. line 4 and 5). Only if the newly arranged regular appointment is before or at most 12 hours after an existing acute appointment, i.e., tbreg ≤ tbact + 1 2, the latter is canceled as all acute illnesses will be treated at the regular appointment. 3.7.3 Walk-in Decision Making Within SiM-Care, all walk-in visits are preceded by an unsuccessful appointment request. As walk-in visits are per se urgent, the earliest possible time t ∈ T for a walk-in visit of patient ρ ∈ P at physi- cian φ ∈ G con is, analogous to Section 3.7.2, de- fined as the current time tcurr ∈ T plus a 30 minute buffer plus the direct travel time, i.e., t = tcurr + 1 48 + τ((cid:96)ρ, (cid:96)φ). The patients' willing to wait for the walk-in visit is the willingness to wait ω ∈ T of the preceding appointment request. As a result, the pa- tient's walk-in visit takes place in the time interval [t,t + ω], unless this is impossible due to the physi- cians' opening hours. As part of the walk-in decision making, patients de- cide on a physician φ∗ ∈ G con and session λ∗ ∈ Λ for their walk-in visit. To that end, SiM-Care computes all physician-session combinations W ⊆ G con × Λ that fall into the interval [t,t + ω] and thus can be targeted for a walk-in visit. If W = /0, the model gradually increases the willingness to wait ω until W (cid:54)= /0. Patients select the physician-session combination (φ∗,λ∗) ∈ W targeted for their walk-in visit on the Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 19 ρ ρ via (φ, [λ]), basis of their walk-in ratings rwalk (φ∗,λ∗) = argmax(φ,λ)∈W 0.95wt (λ) · rwalk where wt (λ) := o(λ)−t denotes the time difference between the earliest possible walk-in time t ∈ T and the end of session λ ∈ Λ. This takes into ac- count that walk-in patients urgently want to visit a physician by discounting the ratings based on the approximate access time wt (λ) ≥ 0. Note that this discounting model yields undesired results if we al- low for negative ratings, motivating the models lim- itation to non-negative ratings. Given the targeted physician-session combination (φ∗,λ∗) ∈ W for the walk-in visit, the time interval during which the actual visit at φ∗ may take place is defined as follows: The earliest time for walk-in patients to arrive in session λ∗ ∈ Λ is 15 minutes before its beginning o(λ∗) ∈ T , but obviously not before the earliest possible arrival t ∈ T . The lat- est possible arrival in session λ∗ ∈ Λ is its ending o(λ∗) ∈ T , but not after the latest possible arrival t + ω. The resulting time interval for the patient's walk-in arrival is (cid:2)max(o(λ∗)− 1 96 , t), min(o(λ∗),t + ω)(cid:3) . The patient's actual arrival within the feasible time interval is stochastic and sampled according to the distribution specified in Section 3.4. As long as patients actively pursue walk-in treat- ment, they never arrange new appointments. That is if a walk-in patient develops a new acute illness or seeks an immediate follow-up appointment trig- gered by a follow-up event, their need for medical attention is met through the ongoing walk-in visit. 3.7.4 Service Time Reduction Physicians' treatment strategies let them reduce service times to prevent congestion and minimize overtime. Within the model, the service time re- duction operationalizes via a multiplicative factor ζ ∈ [0,1]. Thus, a treatment with an original service time of 10 minutes (sampled from the log-normal distribution described in Section 3.4) takes only 8 minutes when performed by a physician with cur- rent consultation speed ζ = 0.8. When there is no effort to reduce service times, i.e., ζ = 1, the ac- tual services time coincide with the sampled origi- nal service times. Table 6 Adaptation of patient ratings rapp ρ and rwalk ρ . Adjustment Positive Event +5 waiting time < 7min +4 successful arrangement of appointment +3ζ successful treatment as walk-in +2ζ successful treatment with appointment Adjustment Negative Event −10 waiting time > 30min −ω no appointment within willingness available −10 rejected as walk-in −20 rejected with appointment Parameter ω ∈ T describes patient's willingness to wait and ζ ∈ [0,1] the physician's consultation speed. 3.7.5 Consequences from Rejection of Patients ρ or rwalk Whenever a patient visits a physician either with an appointment or as a walk-in, the physician's admis- sion strategy determines whether the patient is ad- mitted or rejected. Following a rejection, patients reduce their personal ratings rapp depend- ρ ing on whether they arrived for an appointment or as a walk-in. As rejected patients have been denied treatment, they are subsequently flagged as emer- gencies, i.e, ε = 1. In order to be treated, rejected patients then start a walk-in attempt with reduced willingness to wait ω = 0, i.e., they visit their pre- ferred physician according to the updated walk-in preferences rwalk in the earliest possible session; compare Section 3.7.3. A patient's emergency flag is only revoked after the next successful treatment or if the patient fully recovers from all acute ill- nesses. ρ 3.7.6 Rating Adjustments Throughout the simulation, patients adjust their rat- ings of physicians according to their experiences via additive factors. To that end, patients increase ratings based on positive experiences and decrease ratings following negative experiences. Thereby, patients with appointment update their appoint- ment ratings rapp ρ while walk-in patients update their walk-in ratings rwalk . Table 6 lists all events that trigger a rating adjustment. ρ In SiM-Care, only the effect of a failed appointment request and the effect of a successful treatment are parameterized. All other event effects are hard- coded to represent the following intuition about patient perceptions: Unanticipated events cause a stronger adjustment, while anticipated events only 20 Martin Comis et al. Algorithm 3 Reevaluation of family physician. Require: chronic patient ρ ∈ P 1: let φ∗ = argmaxφ∈G con rapp ρ (φ) 2: if rapp 3: 4: end if ρ (φ∗) ≥ 1.2· rapp φfam = φ∗ ρ (φfam) then cause a slight adjustment. For example, visiting a physician with an appointment and not being ad- mitted is considered highly unlikely and therefore highly penalized. Furthermore, patients react more strongly to negative experiences, reflecting the so- called negativity bias [8]. When a physician fails to offer a fitting appoint- ment, the negative adjustment depends on the pa- tient's associated willingness to wait ω ∈ T . As ω ≥ 0, the adjustment −ω is always non-positive. When the willingness to wait is high, the expecta- tion of receiving a fitting slot is also high, so that the resulting disappointment leads to a stronger nega- tive adjustment. When physicians reduce their service time as part of their treatment strategy, patients feel rushed. There- fore, the model scales the positive adjustment fol- lowing a successful treatment as dependent on the physician's current consultation speed. For exam- ple, at a consultation speed of ζ = 0.5 a successful treatment with appointment increases rapp only by ρ a value of 0.5· 2 = 1. To ensure the desired behavior of discounting rat- ings as described in Section 3.7.3, we bound all rat- ρ (φ) ≥ ings from below by zero, i.e., we enforce rapp (φ, [λ]) ≥ 0 for all ρ ∈ P , φ ∈ G, [λ] ∈ 0 and rwalk Λ/∼. As a result, negative adjustments have no ef- fect on physicians with a rating of zero. ρ 3.7.7 Family Physician Adjustments Every time chronic patients adjust their appoint- ρ (φ) for any φ ∈ G con, they simul- ment ratings rapp taneously reevaluate their family physician φfam ac- cording to Algorithm 3. Thereby, chronic patients change their family physician when another physi- cian from the consideration set has a rating that is at least 20% higher than the current family physi- cian's rating. 3.7.8 Treatment Effects Physicians treat all of a patient's current acute ill- nesses i ∈ I act during the same appointment. As a result, all scheduled follow-up events efol(φ,ρ,i) for i ∈ I act are deleted. Moreover, all illnesses i ∈ I act that require only a single treatment, as indicated by di = /0, are cured and thus removed from I act. Fi- nally, new follow-up events are scheduled for all illnesses i ∈ I act that still require follow-up consul- tation as indicated by a positive follow-up interval νi > 0. Chronic illnesses are only treated during the recur- rent regular appointments or during walk-in visits triggered by the unavailability of a feasible regu- lar appointment. If ς ∈ I chro is treated, any existing follow-up event efol(φ,ρ,ς) ∈ Q is deleted and re- placed by a new, updated one. Finally, the successful treatment revokes any emer- gency flag the patient may have. 3.7.9 PCP Strategies PCP Strategies determine physicians' decision mak- ing through exchangeable submodels, that are de- fined as part of every scenario. For illustration, we describe the exemplary strategies implemented and evaluated in our case study. Appointment scheduling strategy: Individual-block/ Fixed-interval (IBFI) evenly spaces out appoint- ments throughout each session; see [17,40]. To that end, it divides the opening hours of each session in a 140 day rolling horizon into slots of 15 minutes length. Each slot can accommodate one appoint- ment and slots are offered to patients on a first- come-first-served (FCFS) basis. Thus, no appoint- ments are withheld and every patient is offered the earliest feasible appointment at the time of inquiry. Treatment strategy: Priority first come, first served (PFCFS) is popularly used in studies of health sys- tems [17]. In PFCFS, patients with appointment are prioritized over walk-ins and within their respective groups, patients are served in order of their arrival, i.e., FCFS; compare [56,19]. Patients that arrive be- fore the beginning o(λ) ∈ T of session λ ∈ Λ have to wait and the physician does not start treatments until the session has officially begun. The PCP's standard consultation speed in PFCFS is ζ = 1.0, which is adjusted to ζ = 0.8 whenever more than 3 patients await treatment; compare Section 3.7.4. Admission strategy: Priority threshold (PT) admits patients up to a certain utilization threshold; com- pare [39,54]. PT differentiates between appoint- ment, walk-in, and emergency patients: Emergency Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 21 patients are always admitted, i.e., they have an infi- nite admission threshold. Patients with an appoint- ment in session λ ∈ Λ are admitted as long as their time of arrival tarr ∈ T is before the end of the session's buffer, i.e., tarr ≤ o(λ) + 1 24. Appoint- ment patients that arrive after the session's antic- ipated buffer are rejected. For the admittance of walk-in patients, physicians predict their remain- ing workload by multiplying an expected service time with the number of currently waiting patients and upcoming scheduled appointments. If this esti- mated workload is lower than the remaining dura- tion of the current session including buffer, walk-in patients are admitted, otherwise rejected. The ex- pected service time is initialized to 7 minutes and adjusted at the end of each session as follows: On the one hand, the expected service time is increased by one minute if three or more patients are awaiting treatment at the end of the anticipated buffer. On the other hand, the expected service time is reduced by 20 seconds if the physician is idle at the end of the anticipated buffer although walk-ins were pre- viously rejected. 3.8 Structural Validation and Verification In SiM-Care, validation and verification were car- ried out according to the best practices docu- mented in the literature [41,60]. To ensure that our model implementation is correct (verification), we followed established good programming practices. That is, we used object oriented programming to write modular code. SiM-Care is implemented in Java 8. All random distributions are implemented using the Apache Common Math library [49]. Each module is individually verified through unit testing. Assertions ensure that variables remain within their specifications at runtime. As an additional mean to detect undesired model behavior, SiM-Care can trace the entire simulation process. Traces are spe- cialized logs that contain all information about the model's execution. In SiM-Care, traces are textual and comprehensible to modelers. They enable the tracking of agents through the overall model and contain all the information that would be required to animate the model. Analyzing traces and input output relationships, we performed dynamics tests for multiple simulation scenarios of various sizes with different system set-ups. To ensure that our conceptional model serves as an adequate representation of real primary care sys- tems (validation), we took several measures. With regard to face validity, we presented the concep- tual model to physicians and decision makers from health insurers as well as public authorities. Fur- thermore, SiM-Care builds on data from the liter- ature as well as empirical data collected on-site. Moreover, we visited a primary care practice and in- terviewed staff to capture and understand the daily processes and routines of PCPs. For the specific scenarios featured in the case study, we validated the simulation output with available empirical data. Details on this historical validation can be found in the baseline analysis of the following case study. 4 Case Study To demonstrate the potential of SiM-Care, we present a case study evaluating the effects of changes in the population of a primary care system. Specif- ically, we create a baseline scenario representing a real-world primary care system in the district of Aachen and investigate two possible changes in the primary care system's population from the status quo: On the one hand, a decline in the number of PCPs as a result of a decreasing interest in opening a primary care practice in rural areas; see [35]. On the other hand, an aging of the population causing a shift in the quality and intensity of illnesses and the resulting health care requirements. By considering both of these changes individually and in combi- nation, we create three "what-if" scenarios that we compare to the baseline scenario. Each scenario models a time period of one year pre- ceded by warm-up period. As SiM-Care relies on stochastic values, every simulation experiment in- cludes 20 independent runs. Section 4.1 details how the baseline scenario is derived from empirical data. Section 4.2 documents the analysis and validation of the baseline scenario. Sections 4.3, 4.4, and 4.5 describe how the considered changes in the three "what-if" scenarios are implemented in SiM-Care and subsequently benchmark these against the base- line scenario. 4.1 Baseline Scenario The real-world primary care system that serves as the template for our study comprises three predom- inantly rural municipalities (Roetgen, Simmerath, and Monschau) in western Germany with a total population of approximately 35,000 inhabitants and 22 20 primary care physicians. In order to capture a real-world primary care system in the form of a sim- ulation scenario, empirical data is required. Most of this data is specific to a primary care system or its country of origin such that data collection has to be carried out for each system individually. For the considered primary care system, empirical data concerning the physicians' distribution and open- ing hours was provided by the responsible depart- ment of public health or obtained from the responsi- ble association of statutory health insurance physi- cians [37]. The distribution of patients and their de- mographic composition is available from the na- tional census [33] and official population projec- tions by the federal state [34]. The distribution of illnesses and their characteristics can be estimated from publications of health insurances and federal government agencies [28,58]. All unavailable data was either empirically collected in a primary care practice or, where this was not possible, inferred. In the following, we discuss how the available em- pirical data translates into a simulation scenario. To that end, we detail the input parameter choices, i.e., the modeled physicians, patients, age classes, fami- lies of illnesses, and age class-illness distributions. 4.1.1 Primary Care Physicians The population of primary care physicians G in our baseline scenario aims to model the actual pri- mary care physicians in the considered primary care system. According to data provided by the Aachen department of public health in 2017, there are 20 primary care physicians with health insur- ance accreditation in the three municipalities under consideration. The physicians' exact locations are specified as part of the provided dataset (cf. Fig- ure 12) and the physicians opening hours were ob- tained from the Association of Statutory Health Insurance Physicians Nordrhein [37]. Concerning the employed strategies, all physicians φ ∈ G ap- ply the individual-block/fixed-interval appointment scheduling strategy (IBFI), priority first come, first served treatment strategy (PFCFS), and priority threshold admission strategy (PT); cf. Section 3.7.9. 4.1.2 Patients Martin Comis et al. PCPs Population 0− 20 20− 40 40− 60 60− 80 80− 100 Fig. 12 Locations of PCPs with health insurance accredita- tion and population cells reported by the 2011 census [33] publicly available high resolution population data for the considered region is the German Census conducted in 2011 [33]. At a resolution of 2,754 population cells measuring one hectare each, the 2011 Census reports a total population of 35,542 for the three municipalities Roetgen, Simmerath, and Monschau; compare Figure 12. This population includes children under the age of 16 who are ex- cluded from our considerations, as children mainly consult pediatricians who are not modeled in this study. Census data does not state the exact number of under 16-year-olds in each population cell. In- stead, the Census reports the total number of under 16-year-olds on municipality level: Roetgen 1,390, Simmerath 2,383, and Monschau 1,794. To exclude children under the age of 16, we proceed as fol- lows: First, we fix one adult per population cell as we assume that children under the age of 16 do not live on their own. Then, we sample the num- ber of under 16-year-olds from the remaining pop- ulation of each municipality according to a uniform distribution. Exemplifying this procedure for Roet- gen, the Census reports a total population of 8,288 distributed over 534 population cells. We fix one adult per population cell, and uniformly distribute the 1,390 under 16-year-olds among the 7,754 re- maining inhabitants. Performing this procedure for each municipality individually, we obtain the final patient population P consisting of 29,975 patient agents distributed over 2,754 population cells. The population of patient agents P in the base- line scenario aims to reflect the actual population in the considered primary care system. The latest The age-class independent attributes of each patient agent ρ ∈ P are determined as follows: The location (cid:96) ∈ L for each patient is sampled from the associ- Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 23 ated population cell according to a uniform distribu- tion. Patients' health conditions c ∈ [0,1] are sam- pled from a beta distribution with shape parameters p = q = 25 such that all patients have an expected health condition of E(c) = 0.5. 4.1.3 Age classes SiM-Care accounts for the age dependency of var- ious patient characteristics through the concept of age classes. The baseline scenario differentiates three patient age classes: young (16-24), middle- aged (25-65), and elderly (>65). The characteris- tics of the modeled age classes A are shown in Ta- ble 7. Young patients (16-24) are, on average, the healthiest among all patients. Thus, they are ex- pected to develop the fewest acute illnesses per year from which they recover relatively quickly. Their expected willingness to wait is prolonged and they are very unlikely to visit a PCP unless it is nec- essary. Middle aged patients (25-65) represent the working share of the population and we consider them to be our "nominal" patients. They thus do not deviate from the expected illness duration and the expected willingness to wait as specified by fami- lies of illnesses. On average, middle-aged patients (25-65) develop more acute illnesses per year than young patients (16-24) while keeping slightly more appointments after recovery. Elderly patients (>65) are expected to develop the most annual acute ill- nesses and it takes them more time to recover from these. Their expected willingness to wait is the low- est among all age classes and they are most likely to visit a PCP after all symptoms have subsided. Based on Census data [33], the age class a ∈ A of each patient depends on the discrete probabil- ity distribution shown in Table 8. The age-class de- pendent attributes of each patient agent ρ ∈ P are subsequently determined as follows: Each patient's session availabilities α are determined by perform- ing a Bernoulli trial for every session of the week [λ]∈ Λ/∼ based on the age-class dependent success probabilities from Table 8. To decide whether a pa- tient is chronically ill, we perform a Bernoulli trial using the age class dependent success probabilities from Table 8 that were estimated based on [58]. 4.1.4 Families of Illnesses The most important classification system for ill- nesses world-wide is the International Classifica- tion of Diseases and Related health Problems (ICD) Table 7 Age classes A. exp. illnesses dev. duration dev. willingness prob. cancel 16-24 Ia(c)=6c ∆d a=0.8 ∆ω a =1.2 pa=0.95 25-65 Ia(c)=7c+1 ∆d a=1.0 ∆ω a =1.0 pa=0.8 >65 Ia(c)=9c+1 ∆d a=1.2 ∆ω a =0.8 pa=0.7 Table 8 Age specific parameters for patient generation. age class distribution availability probability chronic illness probability 16-24 0.1196 0.85 0.12 25-65 0.6318 0.55 0.33 >65 0.2486 0.95 0.52 maintained by the World Health Organization. In its current revision, ICD-10 [3] distinguishes more than 14,000 codes. For the purpose of SiM-Care, such a granular illness distinction is generally not necessary. Thus, we can aggregate ICD-10 codes, e.g., using the 22 chapters of ICD-10, or consider- ing only a subset of all ICD-10 codes, e.g., the ones most frequently reported. In the baseline scenario, we consider a subset of the 100 ICD-10 codes most frequently reported to the Association of Statutory Health Insurance Physicians Nordrhein [38]. The attributes of families of illnesses can be estimated based on historical treatment data which is com- monly available to health insurers. Yet, such data is naturally protected by confidentiality and cannot be published. Thus, we choose a less elaborate ap- proach and only estimate all attributes which yields the families of illnesses F listed in Table 9. 4.1.5 Age Class-Illness Distributions Age class-illness distributions define the expected occurrence of acute families of illnesses fi ∈ F act per age class a ∈ A. For this distribution, the base- line scenario relies on the reported incidence rates of 8.2 million customers of a large German health insurer published in [28]. We aggregate this data by gender and age to obtain the age class-illness dis- tribution πact : A ×F act → [0,1] shown in Table 10. Analogously, we determine the expected distribu- tion of chronic families of illnesses F chro among the modeled age-classes A denoted by πchro : A × F chro → [0,1] shown in Table 10. The distribution πchro is not part of the baseline sce- nario itself. Instead, it is only required to gener- ate the unique chronic illness of chronically ill pa- tients. In the baseline scenario, we generate every chronic patient's chronic illness ς ∈ I chro analo- 24 Martin Comis et al. Table 9 Characteristics of considered families of illnesses f ∈ F . ICD I10 E11 I25 E78 M54 Z25 J06 Exp. willingness Wf Name Wf (s) = −10s + 20 high blood pressure Wf (s) = −4s + 14 diabetes ischemic heart disease Wf (s) = −4s + 10 Wf (s) = −5s + 8 high cholesterol level Wf (s) = −3s + 4 back pain Wf (s) = 40 vaccination Wf (s) = −2s + 2 cold Exp. duration D f not applicable not applicable not applicable D f (s) = 4s + 8 D f (s) = 9s + 5 not applicable D f (s) = 5s + 4 Table 10 Age class-illness distributions πact and πchro. high cholesterol level back pain vaccination cold high blood pressure diabetes ischemic heart disease 16-24 0.02 0.32 0.14 0.52 0.17 0.33 0.5 25-65 0.24 0.38 0.14 0.24 0.65 0.16 0.19 >65 0.36 0.28 0.27 0.09 0.61 0.2 0.19 e u l a V d e z i d r a d n a t S Treatment frequency Nf Nf (s) = −20s + 100 Nf (s) = −10s + 90 Nf (s) = −30s + 100 Nf (s) = −2s + 11 Nf (s) = −4s + 11 not applicable Nf (s) = −s + 6 Is chronic true true true false false false false Avg. access time Avg. daily overtime Avg. waiting time walk-in gously to the process of generating acute illnesses as described in Section 3.4: Given the patient's age class a ∈ A, the illness family fς ∈ F chro of ς de- pends on the discrete probability distribution f (cid:55)→ πchro(a, f ) for f ∈ F chro. The seriousness sς ∈ [0,1] of ς is sampled from a triangular distribution us- ing the patient's health condition c ∈ [0,1] as mode. In turn, the seriousness defines ς's treatment fre- quency via νς = Nfς(sς) and willingness to wait as ως = Wfς(sς). 4.1.6 Duration of Warm-up Every run of SiM-Care contains a warm-up pe- riod; compare Section 3.6. To determine an appro- priate length for the warm-up period, we simulate the baseline scenario for a time period of 70 years and track all performance indicator for each year individually. Figure 13 shows the resulting evolu- tion for the average access time of appointments, average daily overtime of physicians, and average waiting time of walk-in patients. As we can see, all performance indicators are evolving in the first 30 to 50 years before they stabilize. Similar behav- iors can be observed for all other measured perfor- mance indicators. Therefore, we set the duration of the warm-up period in each scenario to 60 years. 4.2 Baseline Analysis Table 11 reports the resulting expected key per- formance indicators as well as the associated ex- 0 10 20 30 40 50 60 70 Year Fig. 13 Evolution of performance indicators in the baseline scenario for every year in a 70 year time period. act 95%-confidence intervals for each tracked per- formance indicator; compare Section 3.5. The re- sults show that in the status quo, each physician in our primary care system performs, on average, 10,122.16 treatments per year. This amounts to an average number of 6.75 physician contacts per pa- tient which is slightly above the 6.6 annual PCP contacts reported back in 2006 [1]. Roughly 47% of patients visiting a physician in our baseline sce- nario are walk-in patients, which is consistent with the observed 48% share of walk-ins in our collected empirical dataset of service times; compare Sec- tion 3.4. Concerning overtimes, we were unable to obtain empirical data as most primary care physi- cians are self-employed and even the definition of overtime is unclear. However, the estimated average daily overtime per physician (according to our def- inition) seems to be too low at just 0.8 minutes per day. This can be explained by the fact that we incor- porate buffers at the end of each session and do not include additional mandatory physician's activities such as reporting and accounting into our simula- tion model. Patients in our baseline scenario are expected to travel almost 5km to visit a physician and have to wait an average number of 2.46 days for their ap- pointments. With regard to waiting times, we ob- tain an average expected waiting time of 2.09 min- Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 25 ] 8 7 . 2 3 2 , 0 1 , 4 5 . 1 1 2 , 0 1 [ 6 1 . 2 2 2 , 0 1 ] 5 9 . 9 1 0 , 5 1 , 1 6 . 2 9 9 , 4 1 [ 8 2 . 6 0 0 , 5 1 ] 3 7 . 4 2 4 , 2 1 , 7 8 . 9 9 3 , 2 1 [ t f i h S m r e t - t r o h S s t n e i t a P g n i g A t f i h S m r e t - m u i d e M s P C P n i e n i l c e D t f i h S m r e t - t r o h S s P C P n i e n i l c e D o i r a n e c S e n i l e s a B I - C % 5 9 n a e M I - C % 5 9 n a e M I - C % 5 9 n a e M I - C % 5 9 n a e M ] 2 0 . 2 4 8 , 4 ] 8 5 . 5 9 1 , 3 ] 6 1 . 8 9 1 , 2 , , , 9 7 . 0 2 8 , 4 [ 2 4 . 2 9 1 , 3 [ 5 3 . 5 9 1 , 2 [ 4 . 1 3 8 , 4 4 9 1 , 3 5 7 . 6 9 1 , 2 ] 5 6 . 2 7 , 4 5 . 2 7 [ ] 1 8 . 0 , 3 7 . 0 [ ] 8 1 . 6 1 , 5 9 . 3 1 [ ] 3 5 . 2 ] 3 5 . 1 , , 1 5 . 2 [ 8 4 . 1 [ ] 1 0 . 5 ] 2 1 . 2 , , 0 . 5 [ 1 . 2 [ ] 5 0 . 0 4 ] 1 0 . 1 6 , , 4 7 . 9 3 [ 6 8 . 0 6 [ 6 . 2 7 7 7 . 0 5 . 4 1 2 5 . 2 1 5 . 1 5 1 1 . 2 9 . 9 3 4 9 . 0 6 ] 1 1 . 5 7 3 , 9 ] 9 7 . 7 3 3 , 2 ] 5 2 . 9 1 3 , 3 , , , 7 7 . 7 4 3 , 9 [ 6 6 . 5 2 3 , 2 [ 8 9 . 6 0 3 , 3 [ ] 3 0 . 8 6 3 , 2 9 . 6 4 3 [ ] 9 5 . 8 8 , 4 4 . 8 8 [ ] 7 2 . 0 1 , 9 7 . 9 [ ] 2 1 . 4 ] 9 8 . 1 , , 5 0 . 4 [ 8 7 . 1 [ ] 2 5 . 7 , 5 . 7 [ ] 2 . 2 , 6 1 . 2 [ ] 6 9 . 5 6 ] 6 6 . 8 5 , , 6 5 . 5 6 [ 2 4 . 8 5 [ 1 5 . 8 8 3 0 . 0 1 1 . 7 5 3 9 0 . 4 4 8 . 1 1 5 . 7 8 1 . 2 6 7 . 5 6 4 5 . 8 5 4 4 . 1 6 3 , 9 2 7 . 1 3 3 , 2 2 1 . 3 1 3 , 3 ] 1 8 . 7 4 9 , 6 ] 7 7 . 0 7 7 , 2 ] 5 2 . 4 1 7 , 2 , , , 2 0 . 3 2 9 , 6 [ 8 7 . 2 6 7 , 2 [ 8 9 . 5 0 7 , 2 [ ] 9 7 . 0 8 , 5 6 . 0 8 [ ] 2 0 . 3 , 5 7 . 2 [ ] 2 9 . 2 7 , 1 2 . 7 6 [ ] 4 6 . 1 ] 6 6 . 6 ] 2 . 3 , , , 6 1 . 3 [ 6 5 . 1 [ 5 6 . 6 [ ] 3 2 . 2 , 2 . 2 [ ] 5 6 . 1 5 ] 3 0 . 9 5 , , 6 3 . 1 5 [ 6 8 . 8 5 [ 3 . 2 1 4 , 2 1 2 4 . 5 3 9 , 6 7 7 . 6 6 7 , 2 1 1 . 0 1 7 , 2 2 7 . 0 8 9 8 . 2 6 . 9 6 8 1 . 3 6 . 1 6 6 . 6 2 2 . 2 1 5 . 1 5 4 9 . 8 5 ] 2 2 . 2 3 1 , 0 1 ] 6 4 . 1 4 7 , 4 ] 5 0 . 7 1 2 , 3 , , , 1 . 2 1 1 , 0 1 [ 1 6 . 1 2 7 , 4 [ 4 1 . 4 1 2 , 3 [ ] 7 4 . 6 7 1 , 2 , 6 . 3 7 1 , 2 [ 6 1 . 2 2 1 , 0 1 3 5 . 1 3 7 , 4 9 5 . 5 1 2 , 3 3 0 . 5 7 1 , 2 ] 2 2 . 2 7 , 8 0 . 2 7 [ ] 6 8 . 0 , 4 7 . 0 [ ] 3 5 . 5 1 , 6 1 . 3 1 [ ] 7 4 . 2 ] 1 5 . 1 ] 6 9 . 4 ] 1 . 2 , , , , 5 4 . 2 [ 8 4 . 1 [ 4 9 . 4 [ 8 0 . 2 [ ] 5 8 . 9 3 ] 4 2 . 1 6 , , 4 6 . 9 3 [ 2 0 . 1 6 [ 5 1 . 2 7 8 . 0 5 8 . 3 1 6 4 . 2 9 4 . 1 5 9 . 4 9 0 . 2 5 7 . 9 3 3 1 . 1 6 ] 9 . 3 3 0 , 8 3 1 , 8 . 2 9 6 , 7 3 1 [ 5 3 . 3 6 8 , 7 3 1 ] 7 9 . 0 5 6 , 6 3 1 , 2 1 . 8 4 3 , 6 3 1 [ 5 5 . 9 9 4 , 6 3 1 ] 3 2 . 0 0 7 , 6 3 1 , 7 2 . 4 3 3 , 6 3 1 [ 5 2 . 7 1 5 , 6 3 1 ] 2 5 . 4 2 6 , 6 3 1 , 9 8 . 3 8 2 , 6 3 1 [ 2 . 4 5 4 , 6 3 1 -- -- 6 7 7 , 0 1 7 1 6 , 2 3 -- -- 2 6 6 , 0 1 9 3 1 , 2 2 -- -- 2 6 6 , 0 1 5 5 4 , 6 2 -- -- 2 6 6 , 0 1 7 1 6 , 2 3 t f i h S m r e t - m u i d e M s t c e f f E d e n i b m o C t f i h S m r e t - t r o h S s t c e f f E d e n i b m o C t f i h S m r e t - m u i d e M s t n e i t a P g n i g A I - C % 5 9 E I - C % 5 9 E I - C % 5 9 n a e M ] 8 7 . 5 6 2 , 2 ] 5 3 . 6 9 3 , 3 , , 5 8 . 8 4 2 , 2 [ 7 3 . 9 7 3 , 3 [ ] 9 4 . 5 3 6 , 9 , 2 . 3 1 6 , 9 [ 4 3 . 4 2 6 , 9 2 3 . 7 5 2 , 2 6 8 . 7 8 3 , 3 ] 4 0 . 3 7 0 , 7 ] 1 7 . 8 4 7 , 2 ] 1 7 . 7 3 7 , 2 , , , 4 1 . 5 4 0 , 7 [ 4 2 . 9 3 7 , 2 [ 5 2 . 8 2 7 , 2 [ 9 0 . 9 5 0 , 7 8 9 . 3 4 7 , 2 8 9 . 2 3 7 , 2 ] 6 6 . 0 2 9 , 4 ] 7 1 . 4 6 1 , 3 ] 6 5 . 0 3 2 , 2 , , , 2 6 . 7 9 8 , 4 [ 3 7 . 0 6 1 , 3 [ 9 9 . 6 2 2 , 2 [ ] 5 7 . 0 8 2 , 5 1 , 9 2 . 8 5 2 , 5 1 [ 2 5 . 9 6 2 , 5 1 ] 5 9 . 9 4 5 , 2 1 , 4 1 . 2 2 5 , 2 1 [ 5 0 . 6 3 5 , 2 1 ] 1 9 . 1 1 3 , 0 1 , 2 8 . 8 8 2 , 0 1 [ 7 3 . 0 0 3 , 0 1 ] 3 4 . 9 8 ] 7 5 . 1 1 , , 7 2 . 9 8 [ 8 0 . 1 1 [ ] 2 . 7 3 4 , 1 4 . 1 2 4 [ ] 1 0 . 2 ] 4 5 . 7 ] 6 1 . 2 , , , 8 8 . 1 [ 3 5 . 7 [ 4 1 . 2 [ ] 8 3 . 4 , 3 . 4 [ ] 9 3 . 7 6 , 1 0 . 7 6 [ ] 7 7 . 8 5 , 6 . 8 5 [ 5 3 . 9 8 2 3 . 1 1 9 . 8 2 4 4 3 . 4 4 9 . 1 4 5 . 7 5 1 . 2 2 . 7 6 8 6 . 8 5 ] 2 . 1 8 , 3 0 . 1 8 [ ] 8 0 . 3 , 1 8 . 2 [ ] 2 1 . 9 7 , 1 0 . 2 7 [ ] 2 3 . 3 ] 3 7 . 1 ] 5 7 . 6 ] 2 2 . 2 , , , , 7 2 . 3 [ 3 6 . 1 [ 3 7 . 6 [ 9 1 . 2 [ ] 3 2 . 2 5 ] 7 0 . 9 5 , , 1 9 . 1 5 [ 9 8 . 8 5 [ 1 1 . 1 8 4 9 . 2 5 1 . 5 7 8 6 . 1 4 7 . 6 1 2 . 2 3 . 3 7 0 . 2 5 8 9 . 8 5 ] 3 0 . 3 7 , 8 8 . 2 7 [ ] 6 8 . 0 , 4 7 . 0 [ ] 3 6 . 7 1 , 9 8 . 5 1 [ ] 8 5 . 2 ] 4 5 . 1 ] 5 0 . 5 , , , 7 5 . 2 [ 9 4 . 1 [ 4 0 . 5 [ ] 2 1 . 2 , 1 . 2 [ ] 4 2 . 0 4 ] 3 9 . 0 6 , , 7 9 . 9 3 [ 6 7 . 0 6 [ 4 1 . 9 0 9 , 4 5 4 . 2 6 1 , 3 8 7 . 8 2 2 , 2 5 9 . 2 7 4 . 6 1 8 5 . 2 1 5 . 1 4 0 . 5 1 1 . 2 8 . 0 1 1 . 0 4 5 8 . 0 6 ] 2 2 . 1 0 8 , 8 3 1 , 8 4 . 4 3 5 , 8 3 1 [ 5 8 . 7 6 6 , 8 3 1 ] 2 5 . 2 0 0 , 8 3 1 , 8 . 7 5 6 , 7 3 1 [ 5 1 . 0 3 8 , 7 3 1 ] 6 0 . 1 8 8 , 8 3 1 , 5 5 . 6 1 5 , 8 3 1 [ 8 . 8 9 6 , 8 3 1 -- -- 1 3 9 , 0 1 9 3 1 , 2 2 -- -- 6 7 7 , 0 1 5 5 4 , 6 2 -- -- 1 3 9 , 0 1 7 1 6 , 2 3 ] n i m [ n i - k l a w e m ] n i m [ . t p p a e m i t g n i t i a w i t g n i t i a w ] d [ r a l u g e r ] d [ e m e m i t i t s s e c c a s s e c c a ] m k [ e c n a t s i d s s e c c a s n i - k l a w d e t c e j e r # ] n i m [ . s t p p a d r a d n a t s # . s t p p a r a l u g e r # ] % [ n o i t a z i l i t u e m i t r e v o y l i a d s t n e m t a e r t # s n i - k l a w # . g v a . g v a . g v a . g v a . g v a . g v a . g v a . g v a . g v a . g v a . g v a . g v a ] h [ y t i c a p a c P C P l a t o t ] % [ . s t p p a e m i t - n o s t n e i t a p c i n o r h c s e s s e n l l i e t u c a # # ] n i m [ n i - k l a w e m ] n i m [ . t p p a e m i t i t g n i t i a w g n i t i a w ] d [ r a l u g e r ] d [ e m e m i t i t ] m k [ e c n a t s i d s s e c c a s s e c c a s s e c c a s n i - k l a w d e t c e j e r # ] n i m [ . s t p p a d r a d n a t s # . s t p p a r a l u g e r # ] % e m [ n o i t a z i l i t u i t r e v o y l i a d s t n e m t a e r t # s n i - k l a w # . . . . . . . . . . . . g v a g v a g v a g v a g v a g v a g v a g v a g v a g v a g v a g v a ] h [ y t i c a p a c P C P l a t o t ] % [ . s t p p a e m i t - n o s t n e i t a p c i n o r h c s e s s e n l l i e t u c a # # Table 11 Mean performance indicators and 95%-confidence intervals obtained by repeating each simulation experiment 20 times for each simulation scenario variant. 26 Martin Comis et al. utes for patients with appointment and 39.75 min- utes for walk-in patients. In comparison to the av- erage waiting times observed when recording our service time dataset (4 minutes with appointment, 15 minutes without appointment), our simulated waiting times are strikingly unfavorable for walk- in patients which suggests that physicians avoid ex- cessive waiting times for walk-in patients through more sophisticated treatment strategies, e.g., accu- mulating priority queues [63]. 4.3 Scenario One: Decline in PCPs Scenario one models a decline in the number of pri- mary care physicians for a short- and a medium- term shift in time. To that end, we exclude all pri- mary care physicians from our baseline PCP popu- lation G that reached the statutory retirement age of 65 by this point. Specifically, we consider the year 2023 by which 4 out 20 PCPs will have reached the statutory retirement age as well as the year 2027 by which 7 out 20 PCPs will have reached the statutory retirement age. Assuming that none of the excluded physicians are replaced by a successor, we obtain our decimated population of primary care physi- cians G s for the short-term and G m for the medium- term shift. By replacing the physician population G in our baseline scenario by G s and G m, respectively, we obtain two scenarios variants for scenario one. The patient and physician populations used in each scenario variant are summarized in Table 12. The simulation results for scenario one in Ta- ble 11 show a severe deterioration of all patient and physician indicators compared to the baseline scenario. The physicians' expected workload mea- sured through the average number of treatments in- creases by 23% for the short-term and 48% for the medium-term shift. Due to the increased scarcity of appointments, more and more patients are forced to visit physicians as walk-in patients (56% for short- term and 62% for medium-term shift). The average daily overtime for physicians (that neglects all the physicians' administrative and organizational tasks) increases by 2.09 minutes for the short-term and 9.23 minutes for the medium-term shift. On aver- age, patients wait 29% longer for their appoint- ments in the short-term and even 66% longer in the medium-term shift scenario variant. Similar in- creases can be observed for the patients' average ac- cess distance, which increases by 35% to 6.66km for the short-term and by 51% to 7.51km for the medium-term shift. The average waiting time for Table 12 Populations in each simulation scenario variant. Decl. PCPs Aging Patients Comb. Effects s P patients physicians G s m P m G m m P m G m P G m s P s G s P s G s s = short-term shift, m = medium-term shift. Table 13 Age class distributions for aged patient population. short-term shift medium-term shift 16-24 0.1051 0.1025 25-65 0.6283 0.6033 >65 0.2666 0.2942 patients with appointment is almost unaffected by the decline in the number of physicians, which can be explained by the strict prioritization in PFCFS. The average waiting time for walk-in patients in- creases by 30% for the short-term and 65% for the medium-term shift. 4.4 Scenario Two: Aging Patients Scenario two models the ongoing aging of the pa- tient population for a short- and medium-term shift in time. For this purpose, we adjust the discrete probability distribution determining the patients' age classes to generate two new patient populations. More precisely, we use current population projec- tions [34] for the years 2025 and 2030 to obtain the two adjusted discrete probability distributions for the patients' age classes shown in Table 13. Us- ing these distributions, we generate the aged patient population P s for the short-term and P m for the medium-term shift. By replacing the patient popu- lation P in our baseline scenario by P s and P m, re- spectively, we obtain two scenario variants for sce- nario two; compare Table 12. The simulation results for scenario two (Table 11) paint a similar picture as in scenario one, i.e., the majority of patient and physician indicators deteri- orate, albeit far less severe. As a result of the ag- ing of the patient population, the average number of treatments per physician increases by 1% and 2% for the short-term and medium-term shift, re- spectively. However, in contrast to scenario one, additional treatments distribute more evenly be- tween appointment and walk-in patients and thus the expected ratio of walk-in patients remains al- most unchanged (47% for short-term and 48% for Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 27 medium-term shift). Judging from the almost unaf- fected average overtime, physicians manage to ac- commodate the additional treatments mostly within their regular opening hours. As a result of the in- creased treatment demand, patients wait on average 2% longer for their appointments in the short-term and 5% longer in the medium-term shift scenario variant. Moreover, they are willing to accept 1% (2%) longer average access distances in the short- term (medium-term) shift scenario to receive more timely treatment or avoid longer waiting times. Pa- tient waiting times with appointment are unaffected by the increased patient demand. The average wait- ing times of walk-in patients in the short-term shift scenario variant remain almost unchanged, while they increase by 1% for the medium-term shift. 4.5 Scenario Three: Combined Effects Scenario three models a combined decline in the number of primary care physicians and aging of the patient population for a short- and medium-term shift in time. By replacing both, the patient and the physician population in our baseline scenario with the adjusted patient and physician populations from scenarios one and two, we obtain two scenario vari- ants for scenario three; compare Table 12. Analyzing our simulation results in Table 11 for scenario three, we can confirm that the combined effects of a decline in the number of PCPs and an aging population lead to the greatest deterio- ration of patient and physician indicators among all scenarios. However, the effect of the combined changes compared to the combination of the indi- vidual effects from scenarios one and two varies be- tween indicators: For the average number of treat- ments and the ratio of walk-in patients, the effects of the combined changes correspond to the sum of the effects for the individual changes, e.g., a 24% increase in the average number of treatments in short-term shift variant of scenario three versus a 23% and 1% increase in the respective variants of scenarios one and two. Concerning the physicians' average overtime, we can observe that a combined consideration of both changes has an amplifying ef- fect. For example, in the medium-term shift variants of scenarios one and two the average overtime in- creases by 9.23 and 0 minutes, respectively, while the combined changes in scenario three lead to an increase of 10.52 minutes. Similar amplifying ef- fects can be observed for the patients' average ac- cess time and walk-in waiting time. Considering the patients' average access distance, the combination of both changes leads to different effects in the two scenario variants. In the short-term shift variant, the effect of the combined changes corresponds to the sum of the effects for the individual changes. In the medium-term shift variant, we can observe a slight dampening effect resulting from a combined con- sideration of both changes, i.e., while the individual changes lead to a respective 52% and 2% increase of the expected average access distance, the com- bined effects lead to an increase of 52%. 5 Discussion The aim of SiM-Care is to provide decision mak- ers with a tool to analyze and optimize primary care systems. SiM-Care produces meaningful per- formance indicators that enable a far more detailed assessment of primary care systems compared to the current approaches based on patient-physician ratios. Next to more accurate evaluation of the sta- tus quo, SiM-Care can predict and quantify the in- fluence of policy decisions and changes in the sys- tems population, e.g., an aging of the population or a decline in the number of PCPs as illustrated in Section 4. Thereby, the model can particularly take several system changes into account at the same time which enables the analysis of combined ef- fects. As all components of a simulation scenario can be easily adjusted, this opens up a broad field of potential applications ranging from physicians' location planning to the evaluation of specific PCP strategies, e.g., in the field of appointment schedul- ing. Finally, the modular design of SiM-Care per- spectively allows for easy model extensions, e.g., to model prospective new supply concept such as mobile medical units or telemedicine. The greatest entry requirement to using SiM-Care is the complex and time consuming task of generat- ing and validating the input scenarios. As SiM-Care models each agent individually, it requires detailed empirical data which has to be obtained from var- ious parties or, even worse, could be unavailable. Moreover, some model components such as the ser- vice time distributions are tailored to the German system and thus might have to be adjusted when using SiM-Care to analyze, e.g., a primary care system in the United States. Each of these model changes, potentially change the model's behavior and thus require a new validation process to ensure that insights derived from SiM-Care are viable for the studied primary care system. 28 Martin Comis et al. To help overcoming this entry requirement, we ex- emplified the scenario generation and validation process for a real-world primary care system in Ger- many. Particularly, we detailed the generation pro- cess of all simulation entities and provided avail- able empirical data sources. Although data avail- ability may vary for other primary care systems, this may hint at where the required empirical data can be obtained. We validated our simulation scenario by comparing its output to available empirical data. To show internal validity, we performed 20 indepen- dent runs for each simulation experiment and cap- tured the resulting model variability through confi- dence intervals. However, we need to stress that ad- ditional validation should be performed before ac- tual policy decisions are derived from the presented case study. Such validation measures should partic- ularly include a sensitivity analysis as well as an expert validation which were out of scope for the purpose of this study. In summary, SiM-Care can serve as a versatile de- cision support tool in primary care planning when it is used with adequately validated simulation sce- narios. The process of generating and validating simulation scenarios is both challenging and time- consuming. However, once this process is com- plete, SiM-Care enables a detailed analysis and op- timization of primary care systems that is superior to basic ratio-based approaches. As a final motiva- tion to use SiM-Care, we present a potential real- world use-case from Germany: At the beginning of 2019, the German Bundestag passed the law for faster appointments and better care (TSVG) [15] that increased the minimum weekly opening hours for physicians with statutory health insurance ac- creditation from 20 to 25 hours [15, Art. 15]. The law is controversial, among other things, because there are doubts about the consequences of this pol- icy decision [42]. Using SiM-Care, decision mak- ers could have obtained insights into the effects of increased minimal opening hours from both the pa- tients' and the physicians' perspective before its im- plementation. 6 Conclusion and Further Steps In this paper, we presented the hybrid agent-based simulation model SiM-Care that serves as a deci- sion support tool in primary care. SiM-Care does not hard-code a specific primary care system, but requires an input scenario which encodes the speci- fics of the primary system the modeler wishes to evaluate. The generation of a simulation scenario is non-trivial and we exemplified this process for real- world primary care system in Germany. Based on this scenario, we performed a case study that serves as a proof of concept showcasing the capabilities of SiM-Care. Besides, it exemplifies the complex vali- dation and calibration process that has to be carried out for each simulation scenario individually. Future work will include further effort towards model validation and calibration as well as the implementation model extensions. Currently, ill- ness distributions are considered static by SiM- Care. By modeling dynamic illness distributions, we can make them dependent on seasonality or the patients' previous history of illnesses. In the cur- rent model, the duration of an illnesses is indepen- dent of the actual treatment. Interestingly, the re- sults are convincing even without this causal link. In the future, we want to compare whether imple- menting this link in the conceptual model signif- icantly affects our findings. A similar comparison shall investigate the influence of no show-patients, who introduce unexpected idle time into the physi- cians' schedules and require more sophisticated PCP strategies. Yet other possible model extensions include: Illness specific appointments such that not all acute illnesses are treated during every appoint- ment, intentional physicians' breaks, implementa- tion of additional patient attributes such as gender, and mobile patient agents that move between dif- ferent locations, e.g., their home and work. Finally, we plan to release our implementation of SiM-Care as open source software such that it can be easily accessed, studied, and adapted to the individual re- quirements of all modelers. Acknowledgements During the development of SiM-Care and the simulation scenarios in our case study, we received numerous support that we wish to acknowledge. We thank the city region Aachen and the Actimonda health insurance who provided both data and the decision makers' expertise. We are grateful to all physicians, who provided their perspec- tive and allowed us to collect empirical data in their prac- tices. We appreciate the help of Marcia Ruckbeil who ad- vised us on statistics as well as our (former) students Tabea Krabs, Stephan Marnach, and Mariia Anapolska who con- tributed towards the implementation of the simulation model. Finally, we want to thank Sebastian Rachuba for his helpful remarks on an early version of this manuscript. References 1. Bei Arztbesuchen sind die Deutschen Weltmeister. Arztliche Praxis p. 18 (2006) Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 29 2. Statistische Informationen aus dem Bundesarztregister. Kassenarztliche Bundesvereinigung (2017) 3. World Health Organization : ICD-10 : international sta- tistical classification of diseases and related health prob- lems, 10th revision, 2nd edn. World Health Organiza- tion, Geneva (2004) 4. Alemayehu, B., Warner, K.E.: The lifetime distribu- tion of health care costs. Health Services Research 39(3), 627 -- 642 (2004). DOI 10.1111/j.1475-6773. 2004.00248.x 5. Alibrahim, A., Wu, S.: An agent-based simulation model of patient choice of health care providers in ac- countable care organizations. Health Care Manage- ment Science 21(1), 131 -- 143 (2018). DOI 10.1007/ s10729-014-9279-x 6. American Academy of Family Physicians: Primary care. https://www.aafp.org/about/policies/all/primary- care.html (2019). Accessed: 2019-10-18 7. Barnes, S., Golden, B., Price, S.: Applications of Agent- Based Modeling and Simulation to Healthcare Opera- tions Management, pp. 45 -- 74. Springer New York, New York, NY (2013). DOI 10.1007/978-1-4614-5885-2 3 8. Baumeister, R., Bratslavsky, E., Finkenauer, C., Vohs, K.: Bad is stronger than good. Review of Gen- eral Psychology 5(4), 323 -- 370 (2001). DOI 10.1037/ /1089-2680.5.4.323 9. Bechtold, S., Sommer, B., Potzsch, O., Burg, F.: Bevolkerung im Wandel -- Annahmen und Ergebnisse der 14. koordinierten Bevolkerungsvorausberechnung. Statistisches Bundesamt, Wiesbaden (2019) 10. Boscoe, F.P., Henry, K.A., Zdeb, M.S.: A nationwide comparison of driving distance versus straight-line dis- tance to hospitals. The Professional Geographer 64(2), 188 -- 196 (2012). DOI 10.1080/00330124.2011.583586 11. Brailsford, S.C.: System dynamics: Whats in it for healthcare simulation modelers. In: Proceedings of the 2008 Winter Simulation Conference, pp. 1478 -- 1483 (2008). DOI 10.1109/WSC.2008.4736227 12. Brailsford, S.C., Desai, S.M., Viana, J.: Towards the holy grail: Combining system dynamics and discrete- event simulation in healthcare. In: Proceedings of the 2010 Winter Simulation Conference, pp. 2293 -- 2303 (2010). DOI 10.1109/WSC.2010.5678927 13. Brailsford, S.C., Eldabi, T., Kunc, M., Mustafee, N., Os- orio, A.F.: Hybrid simulation modelling in operational research: A state-of-the-art review. European Journal of Operational Research 278(3), 721 -- 737 (2019). DOI https://doi.org/10.1016/j.ejor.2018.10.025 14. Brailsford, S.C., Harper, P.R., Patel, B., Pitt, M.: An analysis of the academic literature on simulation and modelling in health care. Journal of Simulation 3(3), 130 -- 140 (2009). DOI 10.1057/jos.2009.10 15. Bundestag: Gesetz fur schnellere Termine und bessere Versorgung (Terminservice- und Versorgungsgesetz TSVG). Bundesgesetzblatt 1(18), 646 -- 691 (2019) 16. Bureau of Health Workforce, Health Resources and Ser- vices Administration (HRSA): Designated Health Pro- fessional Shortage Areas Statistics. U.S. Department of Health & Human Services (2019) 17. Cayirli, T., Veral, E.: Outpatient scheduling in health care: A review of literature. Production and Operations Management 12(4), 519 -- 549 (2003). DOI 10.1111/j. 1937-5956.2003.tb00218.x 18. Cayirli, T., Veral, E., Rosen, H.: Designing appointment scheduling systems for ambulatory care services. Health Care Management Science 9(1), 47 -- 58 (2006). DOI 10.1007/s10729-006-6279-5 19. Cox, T.F., Birchall, J.P., Wong, H.: Optimising the queu- ing system for an ear, nose and throat outpatient clinic. Journal of Applied Statistics 12(2), 113 -- 126 (1985). DOI 10.1080/02664768500000017 20. Daley, D., Vere-Jones, D.: An Introduction to the Theory of Point Processes: Volume I: Elementary Theory and Methods, 2 edn. Springer-Verlag New York, New York, NY (2003). DOI 10.1007/b97277 21. Dye, C., Reeder, J.C., Terry, R.F.: Research for univer- sal health coverage. Science Translational Medicine 5(199), 199ed13 (2013). DOI 10.1126/scitranslmed. 3006971 22. Fetter, R.B., Thompson, J.M.D.: Patients' waiting time and doctors' idle time in the outpatient setting. Health services research 1(1), 66 -- 90 (1966) 23. Fone, D., Hollinghurst, S., Temple, J., Round, A., Lester, N., Weightman, A., Roberts, K., Coyle, E., Be- van, G., Palmer, S.: Systematic review of the use and value of computer simulation modelling in population health and health care delivery. Journal of Public Health Medicine 25(4), 325 -- 335 (2003). DOI 10.1093/ pubmed/fdg075 24. Gemeinsamer Bundesausschuss: Bedarfsplanungs- richtlinie. Bundesanzeiger, Bekanntmachung BAnz AT 31.12.20102 B7 (2012) 25. Giachetti, R.E., Centeno, E.A., Centeno, M.A., Sun- daram, R.: Assessing the viability of an open access policy in an outpatient clinic: a discrete-event and con- tinuous simulation modeling approach. In: Proceedings of the 2005 Winter Simulation Conference, pp. 2246 -- 2255 (2005). DOI 10.1109/WSC.2005.1574513 26. Gilbert, N.: Agent-based models. No. 153 in Quantita- tive Applications in the Social Sciences. Sage Publica- tions, Thousand Oaks, CA (2008) 27. Grimm, V., Berger, U., DeAngelis, D.L., Polhill, J.G., Giske, J., Railsback, S.F.: The ODD protocol: A re- view and first update. Ecological Modelling 221(23), 2760 -- 2768 (2010). DOI http://dx.doi.org/10.1016/j. ecolmodel.2010.08.019 28. Grobe, T., Dorning, H., Schwartz, F.: BARMER GEK Arztreport 2011. Asgard-Verlag, St. Augustin (2011) 29. Gupta, D., Denton, B.: Appointment scheduling in health care: Challenges and opportunities. IIE Transactions 40(9), 800 -- 819 (2008). DOI 10.1080/ 07408170802165880 30. Hamrock, E., Paige, K., Parks, J., Scheulen, J., Levin, S.: Discrete event simulation for healthcare organiza- tions: a tool for decision making. Journal of Health- care Management 58(2), 110 -- 124 (2013). DOI 10.1097/ 00115514-201303000-00007 31. Homa, L., Rose, J., Hovmand, P.S., Cherng, S.T., Ri- olo, R.L., Kraus, A., Biswas, A., Burgess, K., Aungst, H., Stange, K.C., Brown, K., Brooks-Terry, M., Dec, E., Jackson, B., Gilliam, J., Kikano, G.E., Reichsman, A., Schaadt, D., Hilfer, J., Ticknor, C., Tyler, C.V., Van Der Meulen, A., Ways, H., Weinberger, R.F., Williams, C.: A participatory model of the paradox of primary care. An- nals of Family Medicine 13(5), 456 -- 465 (2015). DOI 10.1370/afm.1841 32. Homer, J.B., Hirsch, G.B.: System dynamics modeling for public health: background and opportunities. Amer- ican Journal of Public Health 96(3), 452 -- 458 (2006). DOI 10.2105/AJPH.2005.062059 33. Information und Technik Nordrhein-Westfalen: Zensus 2011: Vielfaltiges Deutschland. Statistische Amter des Bundes und der Lander (2016) 30 Martin Comis et al. 34. Information und Technik Nordrhein-Westfalen: Bevolkerungsentwicklung 2018 -- 2060 nach Alters- gruppen am 1. Januar . https://www.it.nrw/node/971/pdf (2019). Accessed: 2019-10-18 35. Jacob, R., Kopp, J., Schultz, S.: Berufsmonitoring Medi- zinstudenten 2014. Kassenarztliche Bundesvereinigung (2015) 36. Jacobson, S.H., Hall, S.N., Swisher, J.R.: Discrete-event simulation of health care systems. In: R.W. Hall (ed.) Patient Flow: Reducing Delay in Healthcare Delivery, International Series in Operations Research & Man- agement Science, vol. 91, pp. 211 -- 252. Springer US, Boston, MA (2006). DOI 10.1007/978-0-387-33636-7 8 37. Kassenartzliche Vereinigung Nordrhein: Suche nach Arzten und Psychotherapeuten in Nordrhein. https://www.kvno.de/20patienten/10arztsuche/. Ac- cessed: 2019-10-18 38. Kassenartzliche Vereinigung Nordrhein: Die 100 haufigsten ICD-10-Schlussel und Kurztexte (nach Fach- gruppen) - 4. Quartal 2018. pp. 1 -- 125 (2018) 39. Kim, S.H., Chan, C.W., Olivares, M., Escobar, G.: ICU Admission Control: An Empirical Study of Capacity Al- location and Its Implication for Patient Outcomes. Man- agement Science 61(1), 19 -- 38 (2015). DOI 10.1287/ mnsc.2014.2057 40. Klassen, K.J., Rohleder, T.R.: Scheduling outpatient ap- pointments in a dynamic environment. Journal of Op- erations Management 14(2), 83 -- 101 (1996). DOI https://doi.org/10.1016/0272-6963(95)00044-5 41. Kleijnen, J.P.: Verification and validation of simulation models. European Journal of Operational Research 82(1), 145 -- 162 (1995). DOI https://doi.org/10.1016/ 0377-2217(94)00016-6 42. Korzilius, H.: Terminservice- und Versorgungsgesetz: Deutsche Bundestag billigt umstrittenes Gesetz. Arzteblatt 116(12), A555 -- A557 (2019) 43. Kringos, D.S., Boerma, W.G., Hutchinson, A., Saltman, R.B.: Building primary care in a changing Europe: case studies. World Health Organization (2015) 44. Liu, P., Wu, S.: An agent-based simulation model to study accountable care organizations. Health Care Man- agement Science 19(1), 89 -- 101 (2016). DOI 10.1007/ s10729-014-9279-x 45. Lopes, M.A., Almeida, ´A.S., Almada-Lobo, B.: Fore- casting the medical workforce: a stochastic agent-based simulation approach. Health Care Management Science 21(1), 52 -- 75 (2018). DOI 10.1007/s10729-016-9379-x 46. Mann, E., Schuetz, B., Rubin-Johnston, E.: Remaking Primary Care. New England Healthcare Institute, Cam- bridge, MA (2010) 47. Maroulis, S., Guimer`a, R., Petry, H., Stringer, M.J., Gomez, L.M., Amaral, L.A.N., Wilensky, U.: Com- plex systems view of educational policy research. Sci- ence 330(6000), 38 -- 39 (2010). DOI 10.1126/science. 1195153 48. Matchar, D.B., Ansah, J.P., Hovmand, P., Bayer, S.: Simulation modeling for primary care planning in sin- gapore. In: Proceedings of the 2016 Winter Simulation Conference, pp. 2123 -- 2134 (2016). DOI 10.1109/WSC. 2016.7822255 49. Math Commons: The apache commons mathematics li- brary. Apache Commons (2016) 51. Oh, H.J.A., Muriel, A., Balasubramanian, H.: A user- friendly excel simulation for scheduling in primary care practices. In: Proceedings of the 2014 Winter Simula- tion Conference, pp. 1177 -- 1185 (2014). DOI 10.1109/ WSC.2014.7019975 52. Patlolla, P., Gunupudi, V., Mikler, A.R., Jacob, R.T.: Agent-based simulation tools in computational epidemi- In: T. Bohme, V.M. Larios Rosillo, H. Unger, ology. H. Unger (eds.) Innovative Internet Community Sys- tems, pp. 212 -- 223. Springer Berlin Heidelberg, Berlin, Heidelberg (2006). DOI 10.1007/11553762 21 53. Pfaff, H., Neugebauer, E., Glaeske, G., Schrappe, M., Rothmund, M., Schwartz, W.: Lehrbuch Versorgungs- forschung: Systematik - Methodik - Anwendung. Schat- tauer, Stuttgart (2017) 54. Qu, X., Peng, Y., Shi, J., LaGanga, L.: An MDP model for walk-in patient admission management in primary care clinics. International Journal of Production Eco- nomics 168, 303 -- 320 (2015). DOI https://doi.org/10. 1016/j.ijpe.2015.06.022 55. Rigotti, N., Wallace, R.: Using agent-based models to address "wicked problems" like tobacco use: A re- port from the institute of medicine. Annals of Inter- nal Medicine 163(6), 469 -- 471 (2015). DOI 10.7326/ M15-1567 56. Rising, E.J., Baron, R., Averill, B.: A systems analy- sis of a university-health-service outpatient clinic. Op- erations Research 21(5), 1030 -- 1047 (1973). DOI 10.1287/opre.21.5.1030 57. Rittel, H.W.J., Webber, M.M.: Dilemmas in a general Policy Sciences 4(2), 155 -- 169 theory of planning. (1973). DOI 10.1007/BF01405730 58. Robert Koch-Institut: Chronisches Kranksein. Fakten- blatt zu GEDA 2012: Ergebnisse der Studie Gesundheit in Deutschland aktuell 2012 pp. 1 -- 4 (2014) 59. Rosenbrock, R., Gerlinger, T.: Gesundheitspolitik. Eine systematische Einfuhrung. Verlag Hans Huber, Bern (2014) 60. Sargent, R.G.: Verification and validation of simulation models. Journal of Simulation 7(1), 12 -- 24 (2013). DOI 10.1057/jos.2012.20 61. Schacht, M.: Improving same-day access in primary care: Optimal reconfiguration of appointment system se- tups. Operations Research for Health Care 18, 119 -- 134 (2018). DOI https://doi.org/10.1016/j.orhc.2017.09.003 62. Shi, J., Peng, Y., Erdem, E.: Simulation analysis on patient visit efficiency of a typical VA primary care clinic with complex characteristics. Simulation Mod- elling Practice and Theory 47, 165 -- 181 (2014). DOI https://doi.org/10.1016/j.simpat.2014.06.003 63. Stanford, D.A., Taylor, P., Ziedins, I.: Waiting time dis- tributions in the accumulating priority queue. Queue- ing Systems 77(3), 297 -- 330 (2014). DOI 10.1007/ s11134-013-9382-6 in 64. Tracy, M., Cerd, M., Keyes, K.M.: Agent-based applica- Annual Review modeling tions of Public Health 39(1), 77 -- 94 (2018). 10.1146/annurev-publhealth-040617-014317 and future directions. health: Current public DOI States Census Na- 65. United tional Tables. https://census.gov/data/tables/2017/demo/popproj/2017- summary-tables.html (2017). Accessed: 2019-10-19 Bureau: Projections Population 2017 50. Meng, Y., Davies, R., Hardy, K., Hawkey, P.: An appli- cation of agent-based simulation to the management of hospital-acquired infection. Journal of Simulation 4(1), 60 -- 67 (2010). DOI 10.1057/jos.2009.17 66. Wang, S., Liu, N., Wan, G.: Managing appointment- based services in the presence of walk-in customers. Management Science, Forthcoming pp. 1 -- 41 (2018). DOI 10.1287/mnsc.2018.3239 Patients, Primary Care, and Policy: Simulation Modeling for Health Care Decision Support 31 67. Wiesche, L., Schacht, M., Werners, B.: Strategies for in- terday appointment scheduling in primary care. Health Care Management Science 20(3), 403 -- 418 (2017). DOI 10.1007/s10729-016-9361-7 68. Zhong, X., Williams, M., Li, J., A. Kraft, S., S. Sleeth, J.: Discrete-event simulation for primary care redesign: Review and a case study. In: H. Yang, E.K. Lee (eds.) Healthcare Analytics: From Data to Knowledge to Healthcare Improvement, pp. 361 -- 388. John Wiley & Sons, Inc., Hoboken, NJ (2016). DOI https://doi.org/ 10.1002/9781118919408.ch12
1901.03258	1	1901	2019-01-10T16:36:22	Sample Greedy Based Task Allocation for Multiple Robot Systems	[ "cs.MA", "cs.RO" ]	This paper addresses the task allocation problem for multi-robot systems. The main issue with the task allocation problem is inherent complexity that makes finding an optimal solution within a reasonable time almost impossible. To hand the issue, this paper develops a task allocation algorithm that can be decentralised by leveraging the submodularity concepts and sampling process. The theoretical analysis reveals that the proposed algorithm can provide approximation guarantee of $1/2$ for the monotone submodular case and $1/4$ for the non-monotone submodular case in average sense with polynomial time complexity. To examine the performance of the proposed algorithm and validate the theoretical analysis results, we design a task allocation problem and perform numerical simulations. The simulation results confirm that the proposed algorithm achieves solution quality, which is comparable to a state-of-the-art algorithm in the monotone case, and much better quality in the non-monotone case with significantly less computational complexity.	cs.MA	cs	JOURNAL OF LATEX CLASS FILES 1 Sample Greedy Based Task Allocation for Multiple Robot Systems Hyo-Sang Shin, Teng Li and Pau Segui-Gasco Abstract This paper addresses the task allocation problem for multi-robot systems. The main issue with the task allocation problem is inherent complexity that makes finding an optimal solution within a reasonable time almost impossible. To hand the issue, this paper develops a task allocation algorithm that can be decentralised by leveraging the submodularity concepts and sampling process. The theoretical analysis reveals that the proposed algorithm can provide approximation guarantee of 1/2 for the monotone submodular case and 1/4 for the non-monotone submodular case in average sense with polynomial time complexity. To examine the performance of the proposed algorithm and validate the theoretical analysis results, we design a task allocation problem and perform numerical simulations. The simulation results confirm that the proposed algorithm achieves solution quality, which is comparable to a state-of-the-art algorithm in the monotone case, and much better quality in the non-monotone case with significantly less computational complexity. Task allocation, Multi-robot system, Approximation guarantee, Submodularity, Sampling greedy Index Terms I. INTRODUCTION Multi-Robot Systems (MRS) have been gaining increasing attention thanks to their ability to coordinate simultaneous or co-operate to achieve common goals. MRS provides some fundamental strengths that could not be achieved with single agent systems, e.g., increased flexibility, enhanced reliability and resilience, simultaneous broad area coverage or capability to operate outside the communication range of base stations. Specific applications under consideration include surveillance and monitoring [1] -- [4], Hyo-Sang Shin, Teng Li and Pau Segui-Gasco are with the School of Aerospace, Transport and Manufacturing, Cranfield University, Cranfield MK43 0AL, UK (email: {h.shin,tengli}@cranfield.ac.uk) Corresponding author. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 2 border patrol [5] -- [7], police law enforcement [8], [9], forest-fire localisation [10], [11] and precision agriculture [12]. Efficient cooperation of MRS is a vital part for their successful operations and effective assignment, termed as task allocation, of the available resources is the key enabler of such cooperation. This is because the strength of MRS hinges on the distributed nature of the sensing resources available, making the successful assignment of these resources key to maximise its operational advantages. The main issue with the task allocation problem is that it has been proven to be NP-hard in general [13]. This means that task allocation problems require exponential time to be solved optimally, thus require a careful craft of approximation strategies. A key trade-off that must be performed concerns the optimality of the solution versus the computational complexity. There have been extensive studies to handle the NP-hardness issue in task allocation. The represen- tative task allocation algorithms can be generally classified into three categories: exact approach [14] -- [16], heuristic approach [17] -- [19] and approximation approach [20] -- [22]. Exact algorithms find optimal solutions, but they cannot normally guarantee polynomial time complexity. When the problem complexity becomes high or limited time is available for a solution, exact algorithms are often inapplicable. Heuristic algorithm can deliver good feasible solutions with certain convergence speed for the tested objective functions. As a consequence, most of the algorithms presented in this case are heuristic in their nature, see for example [23] for a classic review on market based algorithms, [24] and [13] for recent surveys. The issue with the heuristic approach is that they can guarantee neither optimality, nor the quality of the solution. Approximation algorithms find an approximated solution that balances optimality and the computational complexity. Moreover, they typically provide mathematical guarantee on the quality of solution and computational complexity if the problem satisfies certain conditions, e.g., submodularity condition. Note that some approaches exploit advantages of two or more different types of approaches and hence those are called hybrid approach [25]. Task allocation algorithms under consideration in this paper are the ones that exploit the approximation approach and also that can be leveraged for decentralised task allocation. Note that the centralised solution of the task allocation problem involves having to communicate all the agents and environment data to a centralised entity. This may not be possible in some realistic scenarios because relying on a central entity could remove resilience or the bandwidth to communicate all the information may not be available. Capability to be decentralised can relax such issues and thus enabling decentralisation provides an option that could be beneficial for extending MRS operations in practice. Most of the decentralised task algorithms have taken two qualitatively different approaches. The first approach targets a trivial instance by assuming linearity of the utility function, enabling optimal January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 3 algorithms [26] -- [30]. The second approach is the heuristic approach [31] -- [33]. A great breakthrough was the introduction of a new distinct approach: an algorithm that used non-trivial properties on the objective function to enable decentralised constant-factor approximation algorithms. In [28], Choi et al presented the Consensus-Based Bundle Auction (CBBA) algorithm, which was the first decentralised approximate algorithm that provided a solution guaranteed to be within a constant factor of the optimal. By assuming that the utility functions of the agents are monotone non-decreasing and submodular, whose definitions will be given in Section II, their algorithm is proven to provide a solution with guarantee of at least 50% of the value of the optimal solution. There are deep theoretical reasons behind the choice of submodularity, and they are intimately con- nected with the tractability of the task allocation problem. Note that submodularity means that the marginal value that an agent obtains by executing an extra task diminishes as the tasks that need to be carried out by the agent increases. The formal definition and its meaning will be discussed in detail in Section III. The task allocation problem can be viewed as optimisation of a set function subject to a matroid constraint. Hence, submodularity plays a key role defining the tractability boundaries of the task allocation problem. The matroid constrained optimisation problem is generally NP-Hard and inapproximable [34]. Although it remains NP-hard, the matroid constrained maximisation problem becomes approximable if submodularity is assumed. In the late 1970s, a seminal paper by Nemhauser et. al. [35] proved that the greedy algorithm achieves a 1/2 approximation guarantee with monotone submodular functions. This factor lies at the core of the reason why CBBA provides a constant factor approximation with monotone submodular functions as it is essentially a decentralised implementation of greedy algorithm. Immediate research questions would be whether 1/2 is the best constant factor approximation that can be achieved for the task allocation problem with monotone submodular functions and what about non-monotone submodular functions. Answering these two questions, Segui-Gasco and Shin developed a decentralised task allocation algorithm that can achieve (1 − 1/e − ǫ) approximation guarantee for monotone submodular functions and (1/e−ǫ) for non-monotone submodular functions [36]. The proposed algorithm leveraged a recent breakthrough in submodular optimisation, that is called measured continuous greedy algorithm. This algorithm, which is originally proposed by Feldman et al [37], obtains (1 − 1/e) approximation for matroid-constrained monotone submodular maximisation and a 1/e factor for the non- monotone case. The issue is that the measured greedy algorithm might be computationally too complex to be directly utilised for task allocation. Therefore, Segui-Gasco and Shin [36] developed a smoother version of measured continuous greedy and mitigated the computational complexity issue. Although the smoother version relaxes the computational complexity in orders of magnitude to the original measured continuous greedy, the complexity is still high compared with CBBA and it might be inefficient to be January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 4 used for large scale MRS. This paper aims to develop a decentralised task allocation algorithm that can be implemented in practice for large scale MRS based on submodular maximisation. The main objective is to provide mathematical guarantees on optimality not only for monotone submodular functions, but also for non- monotone ones with light computational complexity. To achieve the main objective, this paper proposes to leverage a sampling process and the main concepts of CBBA. The proposed algorithm is named as DSTA (Decentralised Sample-based Task Allocation). In DSTA, each agent first randomly selects tasks from the ground task set with a probability of p ∈ (0, 0.5]. Then, each agent selects the task that provides the maximum marginal gain only from the task samples, not from all the tasks. Next, all agents agree on one optimal task-agent pair through maximum consensus protocol. The greedy selection process repeats until it reaches certain stop criteria. The proposed DSTA algorithm unlocks three types of advantages. First, the algorithm achieves approximation guarantees in expectation for both monotone and non-monotone submodular functions. Second, the computational complexity can be relaxed as the algorithm considers only sampled tasks, not all tasks in the greedy selection phase. Third, DSTA enables the trade-off between the approximation guarantee and computational complexity by adjusting the sampling probability p. The advantages of the proposed algorithm are examined mainly by theoretical analysis. To demonstrate and confirm the analysis results, this paper defines a simple benchmark mission scenario and performs numerical simulations. The benchmark scenario selected is a multi-UAV surveillance mission and two different types of submodular functions to be maximised are defined in the scenario. From the simu- lation results, performance, i.e. optimality and computational complexity, of the proposed algorithm is thoroughly investigated and compared with that of CBBA. The rest of paper is organised as follows. Section II presents some preliminaries and backgrounds which will be essential for algorithm development and analysis. Section III develops a new task allocation algorithm, namely DSTA algorithm, and its analysis is detailed in Section IV. The performance and validity of the analysis results are investigated through numerical simulations in Section V. Section VI offers some conclusions and discusses future research directions. II. PRELIMINARIES This section provides some necessary preliminaries for the development and analysis of the proposed decentralised task allocation algorithm. Following general definition of the task allocation problem from [23], the particular instance of the task allocation problem under consideration in this paper is defined as: January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 5 Definition 1 (Task allocation). Given a set of tasks T , a set of agents A, and a non-negative submodular function for each agent a ∈ A specifying the utility of completing each subset of tasks fa : 2T → R+, find a non-overlapping allocation, S ∗ ∈ AT , that maximises a global objective function F : AT → R+ defined as F(S) =Pa∈A fa(Sa). The utility for each agent a is submodular of which definition given in the following. Definition 2 (Submodularity [38]). Let N be a finite set. A real-valued set function f : 2N → R is submodular if, for all X, Y ⊆ N , f (X) + f (Y ) ≥ f (X ∩ Y ) + f (X ∪ Y ). Equivalently, for all A ⊆ B ⊆ N and u ∈ N \B, f (A ∪ {u}) − f (A) ≥ f (B ∪ {u}) − f (B). (1) (2) Note that the submodular function considered in this paper is normalised (i.e. f (∅) = 0) and non- negative (i.e. f (S) ≥ 0 for all S ⊆ N ). This is because the submodular maximisation problem is inapproximable if the submodular function can take negative values [39]. Submodularity is quite an intuitive notion. It simply requires that the marginal value that an agent obtains by executing an extra task diminishes as the tasks that ought to be carried out by the agent increases. Many engineering problems such as control and estimation problems are not obviously convex; but scientists and engineers have devised useful models to "convexify" the problem, to enable practical algorithms that result in immensely useful applications. Like convexity, we believe that submodularity is a good modelling tool in designing utility functions for the task allocation problem. In our opinion, submodularity could be seen from the same perspective: a useful model that can be leveraged to induce the desired behaviour in MRS. In the machine learning community, there has been a great effort spearheading this idea: finding suitable submodular models to solve inherently discrete tasks, such as summarising documents, scene segmentation, or pattern discovery [40] -- [42]. Task allocation is an inherently discrete problem and thus "submodularising" task allocation problems should yield useful applications of efficient MRS cooperation. Eqn. (2) is known as the diminishing returns which is one of core properties of submodular functions. The diminishing return means that the marginal gain of a given element u will never increase as more elements have already been added into the set S. The range of applications holding this property could be wide [28], e.g. a surveillance mission in which the time that an agent will be able to monitor an additional January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 6 point decreases as a given agent is assigned more points to monitor. of The definition of marginal gain is provided in Definition 3. Definition 3 (Marginal gain [38]). For a set function f : 2N → R, S ⊆ N and u ∈ N , the marginal gain of f at S with respect to u is defined as ∆f (uS) := f (S ∪ {u}) − f (S). (3) As discussed in introduction, developing an algorithm that works with non-monotone submodular utilities could be of significant importance since non-monotonicity is a feature that arises naturally in many practical scenarios. For example, in a multi-robot surveillance mission, if a robot is assigned too many targets to track it is possible that it might end up spending its time travelling between targets and not gathering enough useful information at the targets' locations. Therefore, adding tasks to a robot's assignment could, indeed, reduce the utility obtained. Monotone submodular functions are structurally ill-suited to model such a scenario since, by definition, they do not contemplate a reduction in value due to an excessively large number of tasks. Therefore, this paper considers the submodular function that could be either monotone and non-monotone as described in Definition 1. Definition 4 (Monotonicity [38]). A set function f : 2N → R is monotone if, for every A ⊆ B ⊆ N , f (A) ≤ f (B). And f is non-monotone if it is not monotone. In Definition 1, the non-overlapping allocation implies that one task can be allocated at most to one agent, but one agent can take more than one tasks. Hence, the task allocation problem described in Definition 1 can be reduced to submodular maximisation subject to a partition matroid constraint. The definition of matroid is given in Definition 5. Definition 5 (Matroid [20]). A matroid is a pair M = (N , I) where N is a finite set and I ⊆ 2N is a collection of independent sets, satisfying: • ∅ ∈ I • A ⊆ B, B ∈ I ⇒ A ∈ I • A, B ∈ I, A < B ⇒ ∃ b ∈ B\A such that A ∪ {b} ∈ I Representative matroid constraints are uniform matroid and partition matroid constraints. In the uniform matroid constraint, which is also called cardinality constraint, any subset S ⊆ N satisfying S ≤ k is independent. The partition matroid constraint means that a subset S can contain at most one element from each partition. Therefore, in our task allocation problem, each partition is a set of agent-task pairs January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 7 for each task. The following is an important claim [21] that provides the mathematical foundation for the analysis of the proposed task allocation algorithm, especially for the non-monotone submodular case. Claim 1. Let h : 2N → R≥0 be non-negative and submodular, and let S be a random subset of N where each element appears with probability at most p (not necessarily independently). Then, E[h(S)] ≥ (1 − p)h(∅). III. ALGORITHM This section develops a decentralised task allocation algorithm and provides an equivalent centralised version that will be useful for the theoretical analysis. Algorithm 1 DSTA for Agent a Input: fa : 2T → R≥0, T , p Output: A set Ta ⊆ T 1: Na ← ∅, Ta ← ∅ 2: for j ∈ T do 3: 4: with probability p, Na ← Na ∪ {j} 5: end for 6: while ∃ j ∈ Na such that ∆fa(jTa) > 0 do fa(jTa) a∗ ← M axCons(a, j∗ a, ω∗ a) 7: 8: 9: j∗ a ← arg max j∈Na ω∗ a ← ∆fa(j∗ a∗, j∗ aTa) 10: 11: 12: 13: 14: 15: 16: 17: if a∗ == a then Ta ← Ta ∪ {j∗ a} Na ← Na\{j∗ a} else if j∗ a∗ ∈ Na then Na ← Na\{j∗ a∗ } end if end if 18: end while 19: return Ta The proposed DSTA algorithm is summarised in Algorithm 1. The main proposition in this paper is inclusion of a sampling process at the beginning the algorithm and integrate with the main concepts of CBBA. Therefore, in the proposed algorithm, each agent first randomly samples tasks from the ground task set with a probability p to form its own set of task samples, Na (lines 2 -- 5). Then, each agent selects a task that provides the maximum marginal gain from its own task samples Na using the auction process (line 7) and saves the corresponding marginal gain (line 8). Once each agent selects a task at each step, they communicate agent id, the selected task and the corresponding marginal gain with January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 8 their neighbour agents. Then, they negotiate and agree on one task-agent pair that provides the largest marginal gain through maximum consensus protocol. M axCons(a, j∗ a, ω∗ a) in the line 9 of Algorithm 1 is the max-consensus function, which represents the negotiation among all agents. For the max-consensus, each agent a sends its current best marginal value ω∗ a together with corresponding agent id a and task id j∗ a to neighbour agents. At the same time, each agent receives the same information from all other neighbour agents. After agreeing through max-consensus, the agent of the best agent-task pair adds the corresponding task to its task bundle and remove the task from the task samples (lines 10 -- 12). If any agent includes the task of the best agent-task pair in the task samples, they remove that task from their task samples (lines 14 -- 16). Each agent repeats these procedures until there is no task to be allocated. Remark 1. Thanks to sampling, each agent in DSTA is required to evaluate the function values only for a portion of entire tasks, i.e. only for sampled tasks. This should enable to accelerate task allocation by some degree, which will depend on the sampling probability p. Note that each agent in CBBA first constructs a bundle of tasks such that it optimises the score function, by continuously adding tasks until the algorithm is no longer able to add any other task. The score function considered in CBBA is a monotone submodular function that is a function of the task bundle and the path [28] . Then, CBBA applies a consensus strategy to enable convergence on the list of winding bids and agents. For the consensus strategy, a set of sophisticated decision rules was proposed and the convergence characteristics were analysed. The consensus strategy in [28] allows conflict resolution over all tasks and does not limit the network to a specific structure. Different from CBBA, our proposed algorithm utilises a simple strategy to build a bundle of tasks for each agent: each agent finds one best task and then applies a simple consensus strategy on the best tasks at each step. Following this strategy, all agents agree on one best task-agent pair and only the agent in the best task-agent pair adds the corresponding task to its bundle. By repeating the process until no agent can add anymore task to its bundle, each agent is able to construct individual bundles that are conflict-free. If necessary, the proposed algorithm can also use the same consensus strategy as CBBA. Remark 2. Once the bundles are constructed, CBBA initiates the consensus procedure and it might be able to reduce communication burden. However, the consensus strategy in CBBA could require a high number of function evaluations. Note that if a bundle of an agent includes any of tasks updated during consensus iterations in CBBA, the agent must release all of the tasks that were added to the bundle after the updated task and reconstruct the bundle at each iteration. Reconstructing the bundle requires function evaluations. On the other hand, our strategy and requires the same number of consensus procedures as the number of tasks. However, the strategy is relatively simple and doesn't require the agents to reconstruct January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 9 their task bundles. Consequently, the number of function evaluations in our proposed strategy could be reduced, compared with the consensus strategy in CBBA. Remark 3. Developing a max-consensus protocol is beyond the scope of this study. Hence, a simple max-consensus protocol is utilised for case studies in Section V. There are many max-consensus algo- rithms available in existing literature. Also, the convergence characteristics of such algorithms are well studied, considering various practical aspects of communication, e.g. dynamic network, asynchronous communication, time delay, etc. [43] -- [46]. For the application of the proposed DSTA algorithm, one can select an efficient consensus protocol suitable for the communication network placed. Algorithm 2 A Centralised Version of DSTA Input: fa : 2T → R≥0 ∀a ∈ A, T , p Output: Sets Ta ⊆ T ∀a ∈ A 1: for a ∈ A do 2: 3: 4: 5: 6: Na ← ∅, Ta ← ∅ for j ∈ T do with probability p, Na ← Na ∪ {j} end for 7: end for 8: while ∃ a ∈ A and j ∈ Na such that ∆fa(jTa) > 0 do for a ∈ A do if ∃ j ∈ Na s.t. ∆fa(jTa) > 0 then j∗ a ← arg max j∈Na ω∗ a ← ∆fa(j∗ aTa) fa(jTa) end if 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: end for a∗, j∗ a∗ ← arg max a ∈Na a∈A,j ∗ for a ∈ A do ω∗ a(a, j∗ a) if a∗ == a then Ta ← Ta ∪ {j∗ a} Na ← Na\{j∗ a} else if j∗ a∗ ∈ Na then Na ← Na\{j∗ a∗ } end if end if end for 26: end while 27: return Ta ∀a ∈ A Under the convergence of the max-consensus, the decentralised algorithm can be understood in an equivalent centralised version for the ease of analysis. Algorithm 2 summarises this equivalent centralised version of DSTA. Note that Algorithm 1 is for each agent, whereas Algorithm 2 is for all the agents since it is the centralised version. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 10 IV. ANALYSIS This section analyses the performance, especially optimality and computational complexity, of the DSTA algorithm. Assuming convergence of the max-consensus, which is well studied in many existing literature [43] -- [46], we perform analysis on the centralised version of DSTA. A. Algorithm Analysis Consider the ground set as a set of task-agent pairs (N := T × A) and each task-agent pair as an element of the ground set (uj,a := j × a ∀j ∈ T , a ∈ A). Then, the task-agent pairs can be considered as set elements and the task allocation problem can be understood as a submodular maximisation problem subject to a partition matroid constraint. Consequently, Algorithm 2 can be further simplified to Algorithm 3. The resultant algorithm given in in Algorithm 3 is equivalent to the sampling greedy algorithm for submodular maximisation subject to a partition matroid constraint. Algorithm 3 Sample Greedy Input: f : 2N → R≥0, N , I, p Output: A set S ⊆ I 1: Ns ← ∅, S ← ∅ 2: for u ∈ N do 3: 4: with probability p, Ns ← Ns ∪ {u} 5: end for 6: while ∃ u ∈ Ns such that S ∪ {u} ∈ I and ∆f (uS) > 0 do 7: 8: u∗ ← arg max u∈Ns\S S ← S ∪ {u∗} f (uS) 9: end while 10: return S This paper considers Algorithm 3 as the baseline algorithm for the performance analysis. Note that [47] provided a good analysis scheme for the sample greedy for k-extendable systems. Since the partition matroid constraint is a special case of k-extendable systems, the analysis scheme in [47] can be leveraged to prove the proposed DSTA algorithm. Therefore, we follow the analysis scheme in [47], but make necessary modifications to facilitate the partition matroid constraint. To ease the analysis, we can transform Algorithm 3 into an equivalent version, which is summarised in Algorithm 4. Algorithm 4 introduces four variables C, Sc, Q and Kc just for the convenience of analysis. It is clear that theses variables have no effect on the final output S and thus Algorithm 4 is equivalent to Algorithm 3. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 11 Algorithm 4 Equivalent Sample Greedy Input: f : 2N → R≥0, N , I, p Output: A set S ⊆ I 1: S ← ∅, R ← N , C ← ∅, Q ← OP T 2: while ∃ u ∈ R such that S ∪ {u} ∈ I and ∆f (uS) > 0 do c ← arg max f (uS) u∈R Sc ← S C ← C ∪ {c} with probability p do S ← S ∪ {c}, Q ← Q ∪ {c} 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: otherwise if c ∈ Q then Kc ← {c} else Kc ← ∅ end if Q ← Q\Kc R ← R\{c} Let Kc ⊆ Q\S be the smallest set such 17: end while that Q\Kc ∈ I 18: return S The meanings and roles of the four variables are given as follows. The variable C is the set that contains all task-agent pairs that have already been considered by Algorithm 4 no matter whether they are added to S or not. Sc is the set that contains the selected task-agent pairs at the beginning of the current iteration. At the end of the current iteration, Sc = S\{c} if c is added into S and Q, otherwise Sc = S. Q is the set that bridges the solution S and the optimal solution OP T . Q is initialised as OP T at the beginning of the algorithm and updates over time. Every task-agent pair that is added into S is also added into Q. At the same time, a set Kc is removed from Q at each iteration to keep the independence of Q if a task-agent pair c is added into Q. Notice that, if a task-agent pair c is already in Q and is considered but not added into S at the current iteration, then this task-agent pair c should be removed from Q. Kc is the set that is introduced to keep Q independent. According to the matroid properties, the Algorithm 4 is able to remove a set Kc ⊆ Q\S which contains at most one task-agent pair from Q if a task-agent pair is added into the currently independent set Q. B. Performance Analysis The main performance characteristics of the proposed DSTA algorithm are summarised in Theorem 1. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 12 Theorem 1. Suppose the max-consensus in Algorithm 1 assures the convergence. Then, the DSTA algorithm achieves an expected approximation guarantee of p p+Pmax for monotone submodular objective functions and of p(1−p) p+Pmax for non-monotone submodular utility functions with an expected total com- putational complexity of O(pnr) and individual complexity of O(pr2) for each agent, where p is the sampling probability, Pmax = max(p, 1 − p), r is the number of tasks, i.e. r = T , and n is the number of task-agent pairs i.e. n = T × A. The computational complexity of DSTA can be easily proven. As shown in Algorithm 2, there are at most r rounds of auctions. In each round, each agent requires function evaluations at most pr times. Since there are A number of agents, the total number of value oracle calls in each auction is O(pn). Therefore, the total time complexity is O(pnr). For each agent, the individual time complexity is equivalent to the total complexity divided by the number of agents i.e. O(pr2). Let us now investigate the approximation guarantee of DSTA through Algorithm 4. Since the task-agent pairs that have never been considered by Algorithm 4 have no contribution to the approximation guarantee, we only need to focus on the task-agent pairs that are contained in C. To prove the approximation guarantee, we will first show the bound of expectation of f (S), i.e. E [f (S)], with respect to E[f (S ∪ OP T )] in Lemma 3. Lemmas 1 and 2 will be required to prove the bound of E [f (S)] in Lemma 3. Lemma 1. E[Kc\S] ≤ Pmax where Pmax = max(p, 1 − p). Proof: For the proof, we have two cases to analyse, depending on whether the current task-agent pair c is already in Q at the beginning of the iteration in Algorithm 4. 1. If c ∈ Q at the beginning of the iteration, then Kc = ∅ if c is added to S and Kc = {c} if c is not added to S. As c is added to S with probability p, the law of total probability yields: E[Kc\S] = p · ∅ + (1 − p){c} = 1 − p (4) 2. If c /∈ Q at the beginning of the iteration, then Kc contains at most one element if c is added to S. This is because of the property of independent systems. If c is not added to S then Kc = ∅. Therefore, we have E[Kc\S] ≤ p · 1 + (1 − p)∅ = p (5) In summary, E[Kc\S] ≤ max(p, 1 − p). Lemma 2. E[f (S)] = p Pc∈C ∆f (cSc) . January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 13 Proof: According to the order by which the task-agent pairs are added into S, f (S) can be written as Note that the second equality is because f is assumed to be normalised, i.e. f (∅) = 0. Moreover, we have f (S) = f (∅) +Xc∈S ∆f (cSc) =Xc∈S ∆f (cSc) (6) the same number of Sc sets to the number of c elements by Algorithm 4. In each iteration, the task-agent pair c that is being considered is added to S with probability p. If this task-agent pair c is added to S, then its marginal value ∆f (cSc) will be added to the current function value f (Sc). Otherwise, the contribution of c is 0. Therefore, By the law of total probability, we have f (S) =Xc∈C [∆f (cSc)c∈S + 0 · ∆f (cSc)c∈C\S] E[f (S)] = p ·Xc∈C ∆f (cSc) + (1 − p) · 0 = p ·Xc∈C ∆f (cSc) (7) (8) Lemma 3. E[f (S)] ≥ p p+Pmax E[f (S ∪ OP T )]. Proof: From Algorithm 4 and definition of Q, it is clear that the set Q is independent, i.e. Q ∈ I and S is a subset of Q i.e. S ⊆ Q. Therefore, from Definition 5, we have S ∪ {q} ∈ I ∀q ∈ Q\S. By the termination condition, at the termination of Algorithm 4, ∆f (qS) ≤ 0 ∀q ∈ Q\S . Hence, Let Q\S = {q1, q2, · · · , qQ\S}, then ∆f (qS) ≤ 0 Xq∈Q\S (9) Q\S Q\S Xi=1 Xi=1 f (S) = f (Q) − ≥ f (Q) − ≥ f (Q) ∆f (qiS ∪ {q1, · · · , qi−1}) ∆f (qiS) (submodularity) If c is being considered, it implies that the marginal value of c is no less than any other element from Kc\S at that iteration, i.e. ∆f (cSc) ≥ ∆f (qSc), ∀q ∈ Kc\S (10) Additionally, any task-agent pair can be removed from Q at most once. In other words, the task-agent pair that is contained in Kc at one iteration is always different from other iterations when Kc is not January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 14 empty. Therefore, the sets {Kc}c∈C and sets {Kc\S}c∈C are disjoint. From the definition of Q, Q can be expressed as Q = (OP T \ ∪c∈C Kc) ∪ S = (S ∪ OP T )\ ∪c∈C (Kc\S) (11) Denote C as {c1, c2, · · · , cC}. Then, it holds that Sci ⊆ S ⊆ (S ∪OP T )\∪c∈C (Kc\S). From Eqn. (11), we obtain f (Q) = f ((S ∪ OP T )\ ∪c∈C (Kc\S)) = f (S ∪ OP T ) − ∆f (∪c∈C(Kc\S)(S ∪ OP T )\ ∪c∈C (Kc\S)) C = f (S ∪ OP T ) − ∆f ((Kci\S)(S ∪ OP T )\ ∪1≤j≤i (Kcj \S)) ∆f ((Kci\S)Sci) (submodularity) ∆f (qSci) (submodularity) C C ≥ f (S ∪ OP T ) − ≥ f (S ∪ OP T ) − Xi=1 Xi=1 Xi=1 Xq∈Kci \S = f (S ∪ OP T ) −Xc∈C Xq∈Kc\S ≥ f (S ∪ OP T ) −Xc∈C Xq∈Kc\S = f (S ∪ OP T ) −Xc∈C ∆f (qSc) ∆f (cSc) Kc\S∆f (cSc) By taking expectation over f (S), we have E[f (S)] ≥ E[f (Q)] (inequality (10)) (Lemma 1) Kc\S∆f (cSc)# ≥ E"f (S ∪ OP T ) −Xc∈C = E[f (S ∪ OP T )] − E[Kc\S] ·Xc∈C ≥ E[f (S ∪ OP T )] − Pmax ·Xc∈C ∆f (cSc) ∆f (cSc) = E[f (S ∪ OP T )] − Pmax · E[f (S)] (Lemma 2) 1 p The result is clear by rearranging the above inequation. We are now ready to complete the proof of Theorem 1. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 15 Proof of Theorem 1: In order to get the approximation guarantees for both monotone and non- monotone submodular utility functions, we need to analyze the relationship between f (S ∪ OP T ) and f (OP T ) respectively. If f is monotone, then From Lemma 3 and Eqn. (12), it is clear that f (S ∪ OP T ) ≥ f (OP T ) E[f (S)] ≥ ≥ p p + Pmax p p + Pmax · E[f (S ∪ OP T )] (Lemma 3) · E [f (OP T )] (12) (13) For the non-monotone case, define a new submodular and non-monotone function h : 2N → R≥0 as h(X) = f (X ∪ OP T ) ∀X ⊆ N . Since S contains every element with probability at most p, Claim 1 yields Hence, E[f (S ∪ OP T )] = E[h(S)] ≥ (1 − p)h(∅) = (1 − p)f (OP T ) E[f (S)] ≥ p(1 − p) p + Pmax · E [f (OP T )] (14) (15) Corollary 1. The trade-off between approximation guarantee and computational complexity can be controlled by adjusting the sampling probability p. When p = 1/2, the DSTA algorithm achieves the best approximation guarantee of 1 2 for the monotone case and of 1 4 for the non-monotone case. Proof: Recalling that Pmax = max(p, 1 − p), we have Therefore, for p ∈ (0, 0.5], the expected approximation ratios are obtained as: 1 − p for p ∈ (0, 0.5] p for p ∈ (0.5, 1] Pmax =  In case where p ∈ (0.5, 1], it is trivial that (16) (17) (18) E[f (S)] ≥  E[f (S)] ≥  p · f (OP T ) p(1 − p) · f (OP T ) if f is monotone if f is non-monotone 1 2 f (OP T ) 1−p 2 f (OP T ) if f is monotone if f is non-monotone As shown in Eqns. (17) and (18), there is no advantage in average sense to set the sampling probability bigger than 0.5. For p ∈ (0.5, 1], Eqn. (18) clearly shows that the expected approximation ratio becomes stagnated in the monotone case and is monotonically decreasing for the non-monotone case as the January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 16 sampling probability increases. Moreover, it is clear that the computational complexity increases as the sampling probability gets larger. On the other hand, for p ∈ (0, 0.5], the sampling probability provides trade-off capability between the approximation ratio and computational complexity. As the probability increases for p ∈ (0, 0.5], the expected approximation ratios improve for both monotone and non-monotone cases, but the computational complexity also increases. From Eqns. (17) and (18), the best expected approximation guarantees can be readily obtained as: E[f (S)] =  1 2 1 4 if f is monotone if f is non-monotone V. NUMERICAL SIMULATIONS (19) This section verifies the proposed DSTA algorithm and compare its performance with that of a benchmark task allocation algorithm through numerical simulations. The benchmark algorithm selected is CBBA since it is one of most well known and widely applied decentralised task allocation algorithms. The application scenario considered is a multi-UAV surveillance mission. To demonstrate the applica- bility of the submodular maximisation based task allocation algorithms such as CBBA and DSTA, this paper formulates the multi-UAV surveillance problem as submodular maximisation problems subject to the matroid constraint. Hence, the utility function of each UAV is formulated as a submodular function. To examine the performance of the proposed DSTA in comparison with CBBA also in the non-monotone case, a value function is formulated, considering various different factors. In the simulations, it is assumed that there are 300 tasks to be carried out by different numbers of UAVs from 10 to 50 denoted as Na. One task can be allocated to at most one agent, but one agent can carry out multiple tasks. The aim of the task allocation is to achieve the largest overall objective function value. The sampling probability used throughout the simulations is 0.5. A. Monotone Case This subsection investigates the performance of the proposed DSTA algorithm in the monotone case. For fair comparison, this paper adopts a simple monotone value function utilised in [28]. Suppose that, both agents and tasks are homogeneous. For each agent a ∈ A, the value function is given as [28] where Ta is the task set allocated to agent a, λj < 1 is the discounting factor for each task j ∈ T , τ (pa) fa(Ta) = Xj∈Ta λτ (pa) j bj (20) is the estimated distance of the path pa, and bj is the static score associated with performing task j. In the simulations, λ = 0.95 and bj = 1 January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 17 l e u a V y t i l i t U 200 175 150 125 100 75 50 25 0 10 15 20 25 30 Na 35 40 45 50 CBBA DSTA i ) c e s ( e m T l a n o i t a t u p m o C 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 CBBA DSTA 10 15 20 25 30 Na 35 40 45 50 (a) Utility function (b) Computational time Fig. 1: Performance comparison for the monotone submodular function (Nt = 200) We run 10 Monte Carlo simulations in which agents and tasks are randomly placed on a W × W 2-D space (W = 10 km). Depending on the number of tasks, there are two simulations cases: the number of tasks is 200 in the first case and 300 in the second case. The number of agents incrementally increases from 10 to 50 in both cases. The simulation results are depicted in Figs. 1 and 2. Note that the values in all figures in this subsection and the following subsection are the average values of the Monte Carlo simulation results. On the one hand, as shown in the figures, the proposed DSTA algorithm achieves comparable quality of solutions, i.e. comparable values of the utility function, to CBBA , in both simulation cases. In average, the quality of DSTA solution is around 95%, reaching to 97%, in the first simulation case and is around 94%, approaching to around 97%, in the second case. On the other hand, computational time of the DSTA algorithm is significantly less compared with that of CBBA. Note that computational time is measured as the total time taken to complete task allocation for all the agents. The simulation codes are written in Python 3.3.1 and the Monte Carlo simulations were run in late 2014 Mac Mini with an Intel i7 3.0 GHz processor and 16Gb of RAM, running OSX 10.13.5, running both algorithms sequentially in a single core without any parallelism. It is worth noting that as the size of the problem increases, the difference between the two algorithms on the function value becomes tighter, but the difference on computational time becomes more significant. B. Non-monotone Case For the non-monotone case, this paper adapts a surveillance mission in which a group of UAVs are deployed to explore an enemy area. The problem formulation and modelling are adapted in [48]. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 18 l e u a V y t i l i t U 300 250 200 150 100 50 0 10 15 20 25 30 Na 35 40 45 50 CBBA DSTA i ) c e s ( e m T l a n o i t a t u p m o C 8 6 4 2 0 CBBA DSTA 10 15 20 25 30 Na 35 40 45 50 (a) Utility function (b) Computational time Fig. 2: Performance comparison for the monotone submodular function (Nt = 300) The utility function adapted is non-monotone submodular and consists of various components. For the completeness, let us describe all the components of utility function and the resultant utility function. 1) Values: Each task could require different abilities or resources for its execution and might have different importance for the completion of mission. Considering these aspects, the value that an agent a can obtain by performing task j is formulated as waj =: σjmaj (21) where σj ∈ R+ is the importance factor of task j ∈ T , maj ∈ R≥0 is the task-agent fitness factor. For the simulations, this paper generates maj for each task-agent pair in a randomly or arbitrary manner. However, the fitness match maj can be defined arbitrarily as long as it is a positive real. By completing a set of tasks Ta ∈ T , the agent a can obtain a value of: v(Ta) = Xj∈Ta waj (22) 2) Probability of Mission Success: The probability of being detected generally increases as an UAV is exposed longer in the enemy environment, i.e. as it carries out more tasks. Hence, the probability of being detected at the (n + 1)th task, given that the UAV has executed previously n tasks, is modelled as [48]: PD,a(n + 1) = Pr(detection of a at the n + 1th taskn) , P0,a 1 − αanP0,a (23) where P0,a ∈ (0, 1) denotes the probability of UAV a being detected by the enemy when it executes a single task without any previous tasks being executed. αa ≥ 1 is a parameter that governs how January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 19 fast the detection probability increases. To guarantee PD(T ) < 1, α should be designed such that P0 1−α(T −1)P0 < 1. Then, for the agent a executing n number of tasks, the probability of being detected can be quantified as: PD,a(n) = PD,a(n − 1) + (1 − PD,a(n − 1)) P0,a 1 − αa(n − 1)P0,a (24) where P0,a = PD,a(0) by convention. For simplicity, it is assumed that if an UAV has been detected, the mission has failed and thus produced a zero value [48]. Then, we can define the probability of surviving, PS,a : 2T → (0, 1], during executing Ta task set as: PS,a(Ta) , 1 − PD,a(Ta). (25) 3) Inter-Task Effect: As discussed, the probability of an UAV being detected increases as the UAV is exposed more to the adversarial mission environment. Therefore, it is risky to allocate too many important tasks to one single agent. To remove this fragility, inter-task penalties are introduced. Define da ij as the penalty of UAV a when executing each pair i, j ∈ T of distinct tasks in its task set. Then, the total penalty function, ga : 2T → R+, for a given task set Ta ⊆ T is defined as: To discourage concentration of important tasks in a single UAV, we set da ij = eσiσj . g(Ta) = Xi,j∈Ta,i6=j da ij (26) 4) Utility Function: The utility function is obtained by combining all the terms described in Eqns. (22), (25) and (26). That is, the value of a task set Ta ⊆ T in agent a ∈ A is defined as the expected value obtained minus by inter-task penalties: fa(Ta) = ES[va(Ta)] − λaga(Ta) (27) where λa is a scaling factor ensuring that the value function is non-negative and ES[·] denotes the expected value that accounts the probability of the agent being detected and thus the corresponding value becoming zero. Plugging in all the definitions into Eqn. (27), the final value function is given as: fa(Ta) = PS,a(Ta)Xj∈Ta waj − λa Xi,j∈Ta,i6=j da ij (28) 5) Simulation Results: Initial parameter setting is as follows: αa = 1, λa = 0.01 for all a ∈ A. To guarantee PD(T ) < 1, the initial detection probability is defined as P0,a = 1 1+α(T ) for all a ∈ A. Assuming that there are certain tasks that are more important than other tasks, the importance factors of tasks are randomly drawn from two uniform distributions. The importance factors of Na number of tasks are drawn from a continuous uniform distribution between 5 and 7, i.e. U (5, 7). It is assumed that January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 20 200 175 150 125 100 75 50 25 l e u a V y t i l i t U CBBA DSTA i ) c e s ( e m T l a n o i t a t u p m o C 300 250 200 150 100 50 0 CBBA DSTA 10 15 20 25 30 Na 35 40 45 50 10 15 20 25 1.5 1.0 0.5 30 Na 20 40 35 40 45 50 (a) Utility function (b) Computational time Fig. 3: Performance comparison for the non-monotone submodular function (Nt = 200) each of these Na tasks is suitable for one agent and hence we set the corresponding task-agent fitness factor, waj , as 0.3 while setting other factors related to each of these tasks as 0.1. The importance factors of rest of tasks are randomly drawn from U (0.5, 1.5), implying that they are less important than the Na tasks. All task-agent fitness factors corresponding to these tasks are randomly selected from a continuous uniform distribution between 0.1 and 1. First, let us compare the performance, especially the quality of solution and computational time, of the proposed DSTA algorithm with CBBA. Like for the monotone submodular function, we have two simulations cases depending on the number of tasks. For each case, we run 10 Monte Carlo simulations. The number of agents incrementally increases from 10 to 50 in both cases. The sampling probability p is set to be equal to 1/2. The simulation results are depicted in Figs. 3 and 4. The results confirm that the proposed DSTA achieves much better function values than CBBA: the function values of CBBA achieves around 50% and 40% of DSTA for 200 and 300 tasks, respectively. The potential attribute the relatively poor performance of CBBA is that it makes the agents greedily select more important tasks at early iterations. If the agents select more tasks in CBBA, the total function values could start to decrease. Hence, CBBA seems to get trapped in local optima with a few tasks per agent. Note that CBBA only has constant factor guarantees for monotone submodular functions. By contrast, it may exhibit arbitrarily poor performance with non- monotone utility functions, which is confirmed by the simulation results. On the other hand, the sampling procedure in the proposed DSTA might enable abandoning important tasks with a certain probability and hence utilising most of the available tasks to find solutions without getting trapped in local optima. Also, the computational time of the proposed is less than that of CBBA. This could be attributed to the January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 21 300 250 200 150 100 50 l e u a V y t i l i t U l e u a V y t i l i t U 180 160 140 120 100 80 60 CBBA DSTA i ) c e s ( e m T l a n o i t a t u p m o C 1200 1000 800 600 400 200 0 CBBA DSTA 10 15 20 25 30 Na 35 40 45 50 10 15 20 25 3 2 1 20 40 35 40 45 50 30 Na (a) Utility function (b) Computational time Fig. 4: Performance comparison for the non-monotone submodular function (Nt = 300) i ) c e s ( e m T l a n o i t a t u p m o C 2.0 1.5 1.0 0.5 0.0 P=1/10 P=1/5 P=1/2 10 15 20 25 30 Na 35 40 45 50 P=1/10 P=1/5 P=1/2 10 15 20 25 30 Na 35 40 45 50 (a) Utility function (b) Computational time Fig. 5: Performance of the DSTA algorithm with respect to different sampling probabilities fact that CBBA requires reconstruction of task bundles and hence function evaluations whenever there exists a conflict. If the number of conflicts increases, the consensus strategy in CBBA might require more steps for conflict resolution, which results in increased computational complexity. To validate the analysis results in Section IV, we also run another set of simulations with respect to different sampling probabilities for the DSTA algorithm. The sampling probability is set to 1/10, 1/5 and 1/2, respectively, for 200 tasks. Fig. 5 shows the simulation results. As demonstrated in Fig. 5, the quality of solutions improves, but the computational time also increases as the sampling probability increases to 1/2. These results coincide with the analysis results in Section IV and hence confirm the analysis results. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 22 VI. CONCLUSIONS AND FUTURE WORK This paper presents an efficient decentralised task allocation algorithm for MRS. Considering the task allocation problem can be viewed as optimisation of a set function subject to a matroid constraint, we leveraged the submodular maximisation concepts for the theoretical tractability. The CBBA algorithm, which is one of the most applied and practical task allocation algorithms, also utilises the submodularity concept and hence provides an approximation guarantee. The issue is that it only provides an approxi- mation guarantee for monotone submodular utility functions. To overcome this issue, this paper utilises a sampling process, i.e. drawing task samples from the set of all tasks. Consequently, the proposed task allocation algorithm achieves an expected approximation guarantee not only for monotone submodular utility functions, but also for general non-monotone submodular utility functions. Moreover, the com- putational complexity of the proposed algorithm can be further relaxed and adjusted as introduction of the sampling process allows reduction of function evaluations: it requires to evaluate function values for only task samples, not for all the tasks. The performance of the proposed task allocation algorithm is investigated through theoretical analysis. The results of numerical simulations conform the validity of the theoretical analysis results. A future research direction would be improving the approximation ratio for both monotone and non- monotone submodular utility functions. There are some gaps between theoretically achievable approxi- mation guarantee and that of our algorithm. The key challenge will be how to improve the approxima- tion guarantee for both monotone and non-monotone cases while maintaining reasonable computational complexity. Another research direction would be further relaxing the computational complexity while guaranteeing the same or similar approximation guarantee. In our opinion, this could be achieved by introducing the lazy greedy concept to the proposed algorithm. REFERENCES [1] D. T. Cole, S. Sukkarieh, and A. H. Goktogan, "System Development and Demonstration of a UAV Control Architecture for Information Gathering Missions," Journal of Field Robotics, vol. 23, no. 6/7, pp. 417 -- 440, 2006. [2] R. R. Pitre, X. R. Li, and R. Delbalzo, "{UAV} Route Planning for Joint Search and Track Mission - An Informative-Value Approach," IEEE Transactions on Aerospace and Electronic Systems, vol. 48, no. 3, pp. 2551 -- 2565, 2012. [3] B. H. Lim, J. W. Kim, S. W. Ha, and Y. H. Moon, "Development of software platform for monitoring of multiple small UAVs," in AIAA/IEEE Digital Avionics Systems Conference - Proceedings, vol. 2016-Decem, 2016. [4] P. Li and H. Duan, "A potential game approach to multiple UAV cooperative search and surveillance," Aerospace Science and Technology, vol. 68, pp. 403 -- 415, 2017. [5] D. Kingston, R. W. Beard, and R. S. Holt, "Decentralized Perimeter Surveillance Using a Team of {UAVs}," IEEE Transactions on Robotics, vol. 24, no. 6, pp. 1394 -- 1404, 2008. January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 23 [6] D. Bein, W. Bein, A. Karki, and B. B. Madan, "Optimizing Border Patrol Operations Using Unmanned Aerial Vehicles," in Proceedings - 12th International Conference on Information Technology: New Generations, ITNG 2015, 2015, pp. 479 -- 484. [7] S. Minaeian, J. Liu, and Y.-J. Son, "Vision-Based Target Detection and Localization via a Team of Cooperative UAV and UGVs," IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 46, no. 7, pp. 1005 -- 1016, 2016. [8] A. Puri, "A survey of Unmanned Aerial Vehicles ({UAV}) for Traffic Surveillance," Department of Computer Science and Engineering, University of South Florida, Tech. Rep., 2004. [9] E. M. Carapezza and D. B. Law, "Sensors, C3I, Information, and Training Technologies for Law Enforcement," in Proc. SPIE, jan 1999. [10] L. Merino, F. Caballero, J. R. M.-d. Dios, J. Ferruz, and A. Ollero, "A Cooperative Perception System for Multiple UAVs: Application to Automatic Detection of Forest Fires," Journal of Field Robotics, vol. 23, no. 3/4, pp. 165 -- 184, 2006. [11] A. Belbachir and J.-A. Escareno, "Autonomous decisional high-level planning for UAVs-based forest-fire localization," in ICINCO 2016 - Proceedings of the 13th International Conference on Informatics in Control, Automation and Robotics, vol. 1, 2016, pp. 153 -- 159. [12] A. Barrientos, J. Colorado, J. del Cerro, A. Martinez, C. Rossi, D. Sanz, and J. Valente, "Aerial Remote Sensing in Agriculture: A Practical Approach to Area Coverage and Path Planning for Fleets of Mini Aerial Robots," Journal of Field Robotics, vol. 28, no. 5, pp. 667 -- 689, 2011. [13] H.-S. Shin and P. Segui-Gasco, "UAV Swarms: Decision-Making Paradigms," in Encyclopedia of Aerospace Engineering. John Wiley & Sons, Ltd, 2014. [14] E. A. Humo, "On a direct method of optimization," Mathematics and Computers in Simulation, vol. 12, no. 3, pp. 122 -- 129, 1970. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0378475470800106 [15] D. P. Bertsekas, Nonlinear programming. Athena scientific Belmont, 1999. [16] P. T. Boggs and J. W. Tolle, "Sequential quadratic programming for large-scale nonlinear optimization," Journal of computational and applied mathematics, vol. 124, no. 1-2, pp. 123 -- 137, 2000. [17] S. M. Elsayed, R. A. Sarker, and D. L. Essam, "A new genetic algorithm for solving optimization problems," Engineering Applications of Artificial Intelligence, vol. 27, pp. 57 -- 69, 2014. [18] S. Rathi, V. S. Rajput, and S. Gupta, "An improved evolution based optimization algorithm originated from the concept of sfla and simulated annealing," in Industrial and Information Systems (ICIIS), 2016 11th International Conference on. IEEE, 2016, pp. 193 -- 198. [19] K.-L. Du and M. Swamy, "Ant colony optimization," in Search and Optimization by Metaheuristics. Springer, 2016, pp. 191 -- 199. [20] A. Badanidiyuru and J. Vondr´ak, "Fast algorithms for maximizing submodular functions," in Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. SIAM, 2014, pp. 1497 -- 1514. [21] N. Buchbinder, M. Feldman, J. S. Naor, and R. Schwartz, "Submodular maximization with cardinality constraints," in Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. Society for Industrial and Applied Mathematics, 2014, pp. 1433 -- 1452. [22] M. Sviridenko, J. Vondr´ak, and J. Ward, "Optimal approximation for submodular and supermodular optimization with bounded curvature," Mathematics of Operations Research, vol. 42, no. 4, pp. 1197 -- 1218, 2017. [23] M. B. Dias, R. Zlot, N. Kalra, and a. Stentz, "Market-Based Multirobot Coordination: A Survey and Analysis," Proceedings of the IEEE, vol. 94, no. 7, pp. 1257 -- 1270, jul 2006. [24] Z. Yan, N. Jouandeau, and A. Ali, "A Survey and Analysis of Multi-Robot Coordination," January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 24 International Journal of Advanced Robotic Systems, vol. 10, p. 1, 2013. [Online]. Available: http://www.scopus.com/inward/record.url?eid=2-s2.0-84890512866{&}partnerID=tZOtx3y1 [25] X. Yuan, J. Zhao, Y. Yang, and Y. Wang, "Hybrid parallel chaos optimization algorithm with harmony search algorithm," Applied Soft Computing, vol. 17, pp. 12 -- 22, 2014. [26] D. P. Bertsekas and D. A. Castanon, "Parallel synchronous and asynchronous implementations of the auction algorithm," Parallel Computing, vol. 17, no. 6-7, pp. 707 -- 732, sep 1991. [Online]. Available: http://dl.acm.org/citation.cfm?id=1746086.1746168 [27] L. Liu and D. A. Shell, "An anytime assignment algorithm: From local task swapping to global optimality," Autonomous Robots, vol. 35, no. 4, pp. 271 -- 286, jul 2013. [Online]. Available: http://link.springer.com/10.1007/s10514-013-9351-2 [28] H.-l. Choi, L. Brunet, J. P. How, and S. Member, "Consensus-Based Decentralized Auctions for Robust Task Allocation," IEEE Transactions on Robotics, vol. 25, no. 4, pp. 912 -- 926, 2009. [29] S. Ismail and L. Sun, "Decentralized hungarian-based approach for fast and scalable task allocation," in Unmanned Aircraft Systems (ICUAS), 2017 International Conference on. IEEE, 2017, pp. 23 -- 28. [30] S. Moon, E. Oh, and D. H. Shim, "An Integral Framework of Task Assignment and Path Planning for Multiple Unmanned Aerial Vehicles in Dynamic Environments," Journal of Intelligent & Robotic Systems, vol. 70, no. 1-4, pp. 303 -- 313, sep 2012. [Online]. Available: http://www.springerlink.com/index/10.1007/s10846-012-9740-3 [31] L. E. Parker, "ALLIANCE: An architecture for fault tolerant multirobot cooperation," IEEE transactions on robotics and automation, vol. 14, no. 2, pp. 220 -- 240, 1998. [32] M. B. Dias, "Traderbots: A new paradigm for robust and efficient multirobot coordination in dynamic environments," Robotics Institute, p. 153, 2004. [33] B. P. Gerkey and M. J. Mataric, "Sold!: Auction methods for multirobot coordination," IEEE transactions on robotics and automation, vol. 18, no. 5, pp. 758 -- 768, 2002. [34] Z. Svitkina and L. Fleischer, "Submodular approximation: Sampling-based algorithms and lower bounds," SIAM Journal on Computing, vol. 40, no. 6, pp. 1715 -- 1737, 2011. [35] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, "An analysis of approximations for maximizing submodular set functions -- I," Mathematical Programming, vol. 14, no. 1, pp. 265 -- 294, dec 1978. [Online]. Available: http://link.springer.com/10.1007/BF01588971 [36] P. Segui-Gasco and H.-S. Shin, "Fast non-monotone submodular maximisation subject to a matroid constraint," arXiv preprint arXiv:1703.06053, 2017. [37] M. Feldman, J. Naor, and R. Schwartz, "A Unified Continuous Greedy Algorithm for Submodular Maximization," in 2011 IEEE 52nd Annual Symposium on Foundations of Computer Science. IEEE, oct 2011, pp. 570 -- 579. [Online]. Available: http://www.computer.org/csdl/proceedings/focs/2011/4571/00/4571a570-abs.html [38] A. Krause and D. Golovin, "Submodular function maximization," in Tractability: Practical Approaches to Hard Problems. Cambridge University Press, 2014, pp. 71 -- 104. [39] U. Feige, V. S. Mirrokni, and J. Vondrak, "Maximizing non-monotone submodular functions," SIAM Journal on Computing, vol. 40, no. 4, pp. 1133 -- 1153, 2011. [40] H. O. Song, Y. J. Lee, S. Jegelka, and T. Darrell, "Weakly-supervised discovery of visual pattern configurations," in Advances in Neural Information Processing Systems, 2014, pp. 1637 -- 1645. [41] B. Mirzasoleiman, A. Badanidiyuru, and A. Karbasi, "Fast constrained submodular maximization: Personalized data summarization," in ICLM'16: Proceedings of the 33rd International Conference on Machine Learning (ICML), 2016. [42] J. Bilmes and W. Bai, "Deep Submodular Functions," jan 2017. [Online]. Available: http://arxiv.org/abs/1701.08939 January 11, 2019 DRAFT JOURNAL OF LATEX CLASS FILES 25 [43] J. Cort´es, "Distributed algorithms for reaching consensus on general functions," Automatica, vol. 44, no. 3, pp. 726 -- 737, 2008. [44] S. Giannini, A. Petitti, D. Di Paola, and A. Rizzo, "Asynchronous max-consensus protocol with time delays: Convergence results and applications," IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 63, no. 2, pp. 256 -- 264, 2016. [45] F. Iutzeler, P. Ciblat, and J. Jakubowicz, "Analysis of max-consensus algorithms in wireless channels," IEEE Transactions on Signal Processing, vol. 60, no. 11, pp. 6103 -- 6107, 2012. [46] R. Olfati-Saber and R. M. Murray, "Consensus problems in networks of agents with switching topology and time-delays," IEEE Transactions on automatic control, vol. 49, no. 9, pp. 1520 -- 1533, 2004. [47] M. Feldman, C. Harshaw, and A. Karbasi, "Greed is good: Near-optimal submodular maximization via greedy optimization," arXiv preprint arXiv:1704.01652, 2017. [48] P. Segui-Gasco, "Decentralised multi-robot task allocation algorithms," Ph.D. dissertation, SATM, Cranfield University, 2018. January 11, 2019 DRAFT

Subsets and Splits

No saved queries yet

Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.